blob: c587858c36cb234f7921cbc53630379937b805f3 [file] [log] [blame]
Marat Dukhane903dff2021-07-16 19:43:41 -07001// 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
Frank Barchard447aa7b2021-12-28 14:11:40 -080016#include <xnnpack/allocator.h>
Marat Dukhane903dff2021-07-16 19:43:41 -070017#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
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026#if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY
27 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_eq_8) {
Frank Barchard914f57b2021-12-13 12:31:42 -080028 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080037 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -080038 }
39
Zhi An Ngc27f04b2022-01-11 09:34:07 -080040 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, strided_cn) {
Frank Barchard914f57b2021-12-13 12:31:42 -080041 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080051 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -080052 }
53
Zhi An Ngc27f04b2022-01-11 09:34:07 -080054 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_subtile) {
Frank Barchard914f57b2021-12-13 12:31:42 -080055 TEST_REQUIRES_ARM_NEON;
56 for (uint32_t m = 1; m <= 4; m++) {
57 for (uint32_t n = 1; n <= 8; n++) {
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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080067 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -080068 }
69 }
70 }
71
Zhi An Ngc27f04b2022-01-11 09:34:07 -080072 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_subtile_m) {
Frank Barchard914f57b2021-12-13 12:31:42 -080073 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080084 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -080085 }
86 }
87
Zhi An Ngc27f04b2022-01-11 09:34:07 -080088 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_subtile_n) {
Frank Barchard914f57b2021-12-13 12:31:42 -080089 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800100 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800101 }
102 }
103
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800104 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_lt_8) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800105 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800115 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800116 }
117 }
118
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800119 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_lt_8_subtile) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800120 TEST_REQUIRES_ARM_NEON;
121 for (size_t k = 1; k < 8; k++) {
122 for (uint32_t m = 1; m <= 4; m++) {
123 for (uint32_t n = 1; n <= 8; n++) {
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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800133 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800134 }
135 }
136 }
137 }
138
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800139 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_gt_8) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800140 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800150 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800151 }
152 }
153
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800154 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_gt_8_subtile) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800155 TEST_REQUIRES_ARM_NEON;
156 for (size_t k = 9; k < 16; k++) {
157 for (uint32_t m = 1; m <= 4; m++) {
158 for (uint32_t n = 1; n <= 8; n++) {
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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800168 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800169 }
170 }
171 }
172 }
173
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800174 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_div_8) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800175 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800185 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800186 }
187 }
188
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800189 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_div_8_subtile) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800190 TEST_REQUIRES_ARM_NEON;
191 for (size_t k = 16; k <= 80; k += 8) {
192 for (uint32_t m = 1; m <= 4; m++) {
193 for (uint32_t n = 1; n <= 8; n++) {
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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800203 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800204 }
205 }
206 }
207 }
208
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800209 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_gt_8) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800210 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)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800219 .n(n)
Frank Barchard914f57b2021-12-13 12:31:42 -0800220 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800221 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800222 }
223 }
224 }
225
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800226 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_gt_8_strided_cn) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800227 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)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800236 .n(n)
Frank Barchard914f57b2021-12-13 12:31:42 -0800237 .k(k)
238 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800239 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800240 }
241 }
242 }
243
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800244 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_gt_8_subtile) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800245 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800258 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800259 }
260 }
261 }
262 }
263
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800264 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_div_8) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800265 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)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800274 .n(n)
Frank Barchard914f57b2021-12-13 12:31:42 -0800275 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800276 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800277 }
278 }
279 }
280
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800281 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_div_8_strided_cn) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800282 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800294 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800295 }
296 }
297 }
298
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800299 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_div_8_subtile) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800300 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800313 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800314 }
315 }
316 }
317 }
318
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800319 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, small_kernel) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800320 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800331 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800332 }
333 }
334
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800335 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, small_kernel_subtile) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800336 TEST_REQUIRES_ARM_NEON;
337 for (size_t k = 1; k <= 40; k += 9) {
338 for (uint32_t m = 1; m <= 4; m++) {
339 for (uint32_t n = 1; n <= 8; n++) {
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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800350 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800351 }
352 }
353 }
354 }
355
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800356 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_gt_8_small_kernel) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800357 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)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800366 .n(n)
Frank Barchard914f57b2021-12-13 12:31:42 -0800367 .k(k)
368 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800369 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800370 }
371 }
372 }
373
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800374 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_div_8_small_kernel) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800375 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)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800384 .n(n)
Frank Barchard914f57b2021-12-13 12:31:42 -0800385 .k(k)
386 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800387 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800388 }
389 }
390 }
391
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800392 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, strided_cm_subtile) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800393 TEST_REQUIRES_ARM_NEON;
394 for (size_t k = 1; k <= 40; k += 9) {
395 for (uint32_t m = 1; m <= 4; m++) {
396 for (uint32_t n = 1; n <= 8; n++) {
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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800408 }
409 }
410 }
411 }
412
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800413 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, a_offset) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800414 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800426 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800427 }
428 }
429
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800430 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, zero) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800431 TEST_REQUIRES_ARM_NEON;
432 for (uint32_t mz = 0; mz < 4; mz++) {
433 for (size_t k = 1; k <= 40; k += 9) {
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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800445 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800446 }
447 }
448 }
449
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800450 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, qmin) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800451 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800461 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800462 }
463
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800464 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, qmax) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800465 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800475 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800476 }
477
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800478 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, strided_cm) {
Frank Barchard914f57b2021-12-13 12:31:42 -0800479 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)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800489 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard914f57b2021-12-13 12:31:42 -0800490 }
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800491#endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY
Frank Barchard914f57b2021-12-13 12:31:42 -0800492
493
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800494#if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY
495 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_eq_8) {
Frank Barchard15eec022021-11-17 13:26:20 -0800496 TEST_REQUIRES_ARM_NEON;
497 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800498 .mr(4)
Frank Barchard15eec022021-11-17 13:26:20 -0800499 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800500 .kr(1)
Frank Barchard15eec022021-11-17 13:26:20 -0800501 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800502 .m(4)
Frank Barchard15eec022021-11-17 13:26:20 -0800503 .n(8)
504 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800505 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -0800506 }
507
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800508 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, strided_cn) {
Frank Barchard15eec022021-11-17 13:26:20 -0800509 TEST_REQUIRES_ARM_NEON;
510 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800511 .mr(4)
Frank Barchard15eec022021-11-17 13:26:20 -0800512 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800513 .kr(1)
Frank Barchard15eec022021-11-17 13:26:20 -0800514 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800515 .m(4)
Frank Barchard15eec022021-11-17 13:26:20 -0800516 .n(8)
517 .k(8)
518 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800519 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -0800520 }
521
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800522 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_eq_8_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -0800523 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800524 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard15eec022021-11-17 13:26:20 -0800525 for (uint32_t n = 1; n <= 8; n++) {
526 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800527 .mr(4)
528 .nr(8)
529 .kr(1)
530 .sr(1)
531 .m(m)
532 .n(n)
533 .k(8)
534 .iterations(1)
535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
536 }
537 }
538 }
539
540 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_eq_8_subtile_m) {
541 TEST_REQUIRES_ARM_NEON;
542 for (uint32_t m = 1; m <= 4; m++) {
543 GemmMicrokernelTester()
544 .mr(4)
545 .nr(8)
546 .kr(1)
547 .sr(1)
548 .m(m)
549 .n(8)
550 .k(8)
551 .iterations(1)
552 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
553 }
554 }
555
556 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_eq_8_subtile_n) {
557 TEST_REQUIRES_ARM_NEON;
558 for (uint32_t n = 1; n <= 8; n++) {
559 GemmMicrokernelTester()
560 .mr(4)
561 .nr(8)
562 .kr(1)
563 .sr(1)
564 .m(4)
565 .n(n)
566 .k(8)
567 .iterations(1)
568 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
569 }
570 }
571
572 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_lt_8) {
573 TEST_REQUIRES_ARM_NEON;
574 for (size_t k = 1; k < 8; k++) {
575 GemmMicrokernelTester()
576 .mr(4)
577 .nr(8)
578 .kr(1)
579 .sr(1)
580 .m(4)
581 .n(8)
582 .k(k)
583 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
584 }
585 }
586
587 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_lt_8_subtile) {
588 TEST_REQUIRES_ARM_NEON;
589 for (size_t k = 1; k < 8; k++) {
590 for (uint32_t m = 1; m <= 4; m++) {
591 for (uint32_t n = 1; n <= 8; n++) {
592 GemmMicrokernelTester()
593 .mr(4)
594 .nr(8)
595 .kr(1)
596 .sr(1)
597 .m(m)
598 .n(n)
599 .k(k)
600 .iterations(1)
601 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
602 }
603 }
604 }
605 }
606
607 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_gt_8) {
608 TEST_REQUIRES_ARM_NEON;
609 for (size_t k = 9; k < 16; k++) {
610 GemmMicrokernelTester()
611 .mr(4)
612 .nr(8)
613 .kr(1)
614 .sr(1)
615 .m(4)
616 .n(8)
617 .k(k)
618 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
619 }
620 }
621
622 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_gt_8_subtile) {
623 TEST_REQUIRES_ARM_NEON;
624 for (size_t k = 9; k < 16; k++) {
625 for (uint32_t m = 1; m <= 4; m++) {
626 for (uint32_t n = 1; n <= 8; n++) {
627 GemmMicrokernelTester()
628 .mr(4)
629 .nr(8)
630 .kr(1)
631 .sr(1)
632 .m(m)
633 .n(n)
634 .k(k)
635 .iterations(1)
636 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
637 }
638 }
639 }
640 }
641
642 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_div_8) {
643 TEST_REQUIRES_ARM_NEON;
644 for (size_t k = 16; k <= 80; k += 8) {
645 GemmMicrokernelTester()
646 .mr(4)
647 .nr(8)
648 .kr(1)
649 .sr(1)
650 .m(4)
651 .n(8)
652 .k(k)
653 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
654 }
655 }
656
657 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, k_div_8_subtile) {
658 TEST_REQUIRES_ARM_NEON;
659 for (size_t k = 16; k <= 80; k += 8) {
660 for (uint32_t m = 1; m <= 4; m++) {
661 for (uint32_t n = 1; n <= 8; n++) {
662 GemmMicrokernelTester()
663 .mr(4)
664 .nr(8)
665 .kr(1)
666 .sr(1)
667 .m(m)
668 .n(n)
669 .k(k)
670 .iterations(1)
671 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
672 }
673 }
674 }
675 }
676
677 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, 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(4)
683 .nr(8)
684 .kr(1)
685 .sr(1)
686 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800687 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800688 .k(k)
689 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
690 }
691 }
692 }
693
694 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, 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(4)
700 .nr(8)
701 .kr(1)
702 .sr(1)
703 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800704 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800705 .k(k)
706 .cn_stride(11)
707 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
708 }
709 }
710 }
711
712 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, 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 <= 4; m++) {
717 GemmMicrokernelTester()
718 .mr(4)
719 .nr(8)
720 .kr(1)
721 .sr(1)
722 .m(m)
723 .n(n)
724 .k(k)
725 .iterations(1)
726 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
727 }
728 }
729 }
730 }
731
732 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, 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(4)
738 .nr(8)
739 .kr(1)
740 .sr(1)
741 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800742 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800743 .k(k)
744 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
745 }
746 }
747 }
748
749 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, 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(4)
755 .nr(8)
756 .kr(1)
757 .sr(1)
758 .m(4)
759 .n(n)
760 .k(k)
761 .cn_stride(11)
762 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
763 }
764 }
765 }
766
767 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, 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 <= 4; m++) {
772 GemmMicrokernelTester()
773 .mr(4)
774 .nr(8)
775 .kr(1)
776 .sr(1)
777 .m(m)
778 .n(n)
779 .k(k)
780 .iterations(1)
781 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
782 }
783 }
784 }
785 }
786
787 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, small_kernel) {
788 TEST_REQUIRES_ARM_NEON;
789 for (size_t k = 1; k <= 40; k += 9) {
790 GemmMicrokernelTester()
791 .mr(4)
792 .nr(8)
793 .kr(1)
794 .sr(1)
795 .m(4)
796 .n(8)
797 .k(k)
798 .ks(3)
799 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
800 }
801 }
802
803 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, small_kernel_subtile) {
804 TEST_REQUIRES_ARM_NEON;
805 for (size_t k = 1; k <= 40; k += 9) {
806 for (uint32_t m = 1; m <= 4; m++) {
807 for (uint32_t n = 1; n <= 8; n++) {
808 GemmMicrokernelTester()
809 .mr(4)
810 .nr(8)
811 .kr(1)
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_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
819 }
820 }
821 }
822 }
823
824 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, 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(4)
830 .nr(8)
831 .kr(1)
832 .sr(1)
833 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800834 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800835 .k(k)
836 .ks(3)
837 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
838 }
839 }
840 }
841
842 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, 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(4)
848 .nr(8)
849 .kr(1)
850 .sr(1)
851 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -0800852 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -0800853 .k(k)
854 .ks(3)
855 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
856 }
857 }
858 }
859
860 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, strided_cm_subtile) {
861 TEST_REQUIRES_ARM_NEON;
862 for (size_t k = 1; k <= 40; k += 9) {
863 for (uint32_t m = 1; m <= 4; m++) {
864 for (uint32_t n = 1; n <= 8; n++) {
865 GemmMicrokernelTester()
866 .mr(4)
867 .nr(8)
868 .kr(1)
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_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
876 }
877 }
878 }
879 }
880
881 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, a_offset) {
882 TEST_REQUIRES_ARM_NEON;
883 for (size_t k = 1; k <= 40; k += 9) {
884 GemmMicrokernelTester()
885 .mr(4)
886 .nr(8)
887 .kr(1)
888 .sr(1)
889 .m(4)
890 .n(8)
891 .k(k)
892 .ks(3)
893 .a_offset(163)
894 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
895 }
896 }
897
898 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, zero) {
899 TEST_REQUIRES_ARM_NEON;
900 for (uint32_t mz = 0; mz < 4; mz++) {
901 for (size_t k = 1; k <= 40; k += 9) {
902 GemmMicrokernelTester()
903 .mr(4)
904 .nr(8)
905 .kr(1)
906 .sr(1)
907 .m(4)
908 .n(8)
909 .k(k)
910 .ks(3)
911 .a_offset(163)
912 .zero_index(mz)
913 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
914 }
915 }
916 }
917
918 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, qmin) {
919 TEST_REQUIRES_ARM_NEON;
920 GemmMicrokernelTester()
921 .mr(4)
922 .nr(8)
923 .kr(1)
924 .sr(1)
925 .m(4)
926 .n(8)
927 .k(8)
928 .qmin(128)
929 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
930 }
931
932 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, qmax) {
933 TEST_REQUIRES_ARM_NEON;
934 GemmMicrokernelTester()
935 .mr(4)
936 .nr(8)
937 .kr(1)
938 .sr(1)
939 .m(4)
940 .n(8)
941 .k(8)
942 .qmax(128)
943 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
944 }
945
946 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_LD64, strided_cm) {
947 TEST_REQUIRES_ARM_NEON;
948 GemmMicrokernelTester()
949 .mr(4)
950 .nr(8)
951 .kr(1)
952 .sr(1)
953 .m(4)
954 .n(8)
955 .k(8)
956 .cm_stride(11)
957 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
958 }
959#endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY
960
961
962#if XNN_ARCH_ARM || XNN_ARCH_ARM64
963 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8) {
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(8)
973 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
974 }
975
976 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_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(8)
986 .cn_stride(11)
987 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
988 }
989
990 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_subtile) {
991 TEST_REQUIRES_ARM_NEON;
992 for (uint32_t m = 1; m <= 1; m++) {
993 for (uint32_t n = 1; n <= 8; n++) {
994 GemmMicrokernelTester()
995 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -0800996 .nr(8)
997 .kr(2)
998 .sr(1)
999 .m(m)
1000 .n(n)
1001 .k(8)
1002 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001004 }
1005 }
1006 }
1007
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001008 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_subtile_m) {
Frank Barchard15eec022021-11-17 13:26:20 -08001009 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001010 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard15eec022021-11-17 13:26:20 -08001011 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001012 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001013 .nr(8)
1014 .kr(2)
1015 .sr(1)
1016 .m(m)
1017 .n(8)
1018 .k(8)
1019 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001020 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001021 }
1022 }
1023
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001024 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_subtile_n) {
Frank Barchard15eec022021-11-17 13:26:20 -08001025 TEST_REQUIRES_ARM_NEON;
1026 for (uint32_t n = 1; n <= 8; n++) {
1027 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001028 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001029 .nr(8)
1030 .kr(2)
1031 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001032 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001033 .n(n)
1034 .k(8)
1035 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001036 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001037 }
1038 }
1039
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001040 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_lt_8) {
Frank Barchard15eec022021-11-17 13:26:20 -08001041 TEST_REQUIRES_ARM_NEON;
1042 for (size_t k = 1; k < 8; k++) {
1043 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001044 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001045 .nr(8)
1046 .kr(2)
1047 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001048 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001049 .n(8)
1050 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001051 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001052 }
1053 }
1054
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001055 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_lt_8_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08001056 TEST_REQUIRES_ARM_NEON;
1057 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001058 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard15eec022021-11-17 13:26:20 -08001059 for (uint32_t n = 1; n <= 8; n++) {
1060 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001061 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001062 .nr(8)
1063 .kr(2)
1064 .sr(1)
1065 .m(m)
1066 .n(n)
1067 .k(k)
1068 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001069 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001070 }
1071 }
1072 }
1073 }
1074
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001075 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_gt_8) {
Frank Barchard15eec022021-11-17 13:26:20 -08001076 TEST_REQUIRES_ARM_NEON;
1077 for (size_t k = 9; k < 16; k++) {
1078 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001079 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001080 .nr(8)
1081 .kr(2)
1082 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001083 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001084 .n(8)
1085 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001086 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001087 }
1088 }
1089
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001090 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_gt_8_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08001091 TEST_REQUIRES_ARM_NEON;
1092 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001093 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard15eec022021-11-17 13:26:20 -08001094 for (uint32_t n = 1; n <= 8; n++) {
1095 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001096 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001097 .nr(8)
1098 .kr(2)
1099 .sr(1)
1100 .m(m)
1101 .n(n)
1102 .k(k)
1103 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001104 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001105 }
1106 }
1107 }
1108 }
1109
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001110 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_div_8) {
Frank Barchard15eec022021-11-17 13:26:20 -08001111 TEST_REQUIRES_ARM_NEON;
1112 for (size_t k = 16; k <= 80; k += 8) {
1113 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001114 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001115 .nr(8)
1116 .kr(2)
1117 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001118 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001119 .n(8)
1120 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001121 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001122 }
1123 }
1124
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001125 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_div_8_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08001126 TEST_REQUIRES_ARM_NEON;
1127 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001128 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard15eec022021-11-17 13:26:20 -08001129 for (uint32_t n = 1; n <= 8; n++) {
1130 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001131 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001132 .nr(8)
1133 .kr(2)
1134 .sr(1)
1135 .m(m)
1136 .n(n)
1137 .k(k)
1138 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001139 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001140 }
1141 }
1142 }
1143 }
1144
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001145 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8) {
Frank Barchard15eec022021-11-17 13:26:20 -08001146 TEST_REQUIRES_ARM_NEON;
1147 for (uint32_t n = 9; n < 16; n++) {
1148 for (size_t k = 1; k <= 40; k += 9) {
1149 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001150 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001151 .nr(8)
1152 .kr(2)
1153 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001154 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001155 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08001156 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001157 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001158 }
1159 }
1160 }
1161
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001162 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8_strided_cn) {
Frank Barchard15eec022021-11-17 13:26:20 -08001163 TEST_REQUIRES_ARM_NEON;
1164 for (uint32_t n = 9; n < 16; n++) {
1165 for (size_t k = 1; k <= 40; k += 9) {
1166 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001167 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001168 .nr(8)
1169 .kr(2)
1170 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001171 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001172 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08001173 .k(k)
1174 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001175 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001176 }
1177 }
1178 }
1179
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001180 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08001181 TEST_REQUIRES_ARM_NEON;
1182 for (uint32_t n = 9; n < 16; n++) {
1183 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001184 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard15eec022021-11-17 13:26:20 -08001185 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001186 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001187 .nr(8)
1188 .kr(2)
1189 .sr(1)
1190 .m(m)
1191 .n(n)
1192 .k(k)
1193 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001194 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001195 }
1196 }
1197 }
1198 }
1199
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001200 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8) {
Frank Barchard15eec022021-11-17 13:26:20 -08001201 TEST_REQUIRES_ARM_NEON;
1202 for (uint32_t n = 16; n <= 24; n += 8) {
1203 for (size_t k = 1; k <= 40; k += 9) {
1204 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001205 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001206 .nr(8)
1207 .kr(2)
1208 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001209 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001210 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08001211 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001212 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001213 }
1214 }
1215 }
1216
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001217 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8_strided_cn) {
Frank Barchard15eec022021-11-17 13:26:20 -08001218 TEST_REQUIRES_ARM_NEON;
1219 for (uint32_t n = 16; n <= 24; n += 8) {
1220 for (size_t k = 1; k <= 40; k += 9) {
1221 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001222 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001223 .nr(8)
1224 .kr(2)
1225 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001226 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001227 .n(n)
1228 .k(k)
1229 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001230 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001231 }
1232 }
1233 }
1234
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001235 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08001236 TEST_REQUIRES_ARM_NEON;
1237 for (uint32_t n = 16; n <= 24; n += 8) {
1238 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001239 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard15eec022021-11-17 13:26:20 -08001240 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001241 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001242 .nr(8)
1243 .kr(2)
1244 .sr(1)
1245 .m(m)
1246 .n(n)
1247 .k(k)
1248 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001249 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001250 }
1251 }
1252 }
1253 }
1254
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001255 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, small_kernel) {
Frank Barchard15eec022021-11-17 13:26:20 -08001256 TEST_REQUIRES_ARM_NEON;
1257 for (size_t k = 1; k <= 40; k += 9) {
1258 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001259 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001260 .nr(8)
1261 .kr(2)
1262 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001263 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001264 .n(8)
1265 .k(k)
1266 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001267 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001268 }
1269 }
1270
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001271 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, small_kernel_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08001272 TEST_REQUIRES_ARM_NEON;
1273 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001274 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard15eec022021-11-17 13:26:20 -08001275 for (uint32_t n = 1; n <= 8; n++) {
1276 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001277 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001278 .nr(8)
1279 .kr(2)
1280 .sr(1)
1281 .m(m)
1282 .n(n)
1283 .k(k)
1284 .ks(3)
1285 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001286 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001287 }
1288 }
1289 }
1290 }
1291
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001292 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8_small_kernel) {
Frank Barchard15eec022021-11-17 13:26:20 -08001293 TEST_REQUIRES_ARM_NEON;
1294 for (uint32_t n = 9; n < 16; n++) {
1295 for (size_t k = 1; k <= 40; k += 9) {
1296 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001297 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001298 .nr(8)
1299 .kr(2)
1300 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001301 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001302 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08001303 .k(k)
1304 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001305 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001306 }
1307 }
1308 }
1309
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001310 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8_small_kernel) {
Frank Barchard15eec022021-11-17 13:26:20 -08001311 TEST_REQUIRES_ARM_NEON;
1312 for (uint32_t n = 16; n <= 24; n += 8) {
1313 for (size_t k = 1; k <= 40; k += 9) {
1314 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001315 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001316 .nr(8)
1317 .kr(2)
1318 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001319 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001320 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08001321 .k(k)
1322 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001323 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001324 }
1325 }
1326 }
1327
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001328 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, strided_cm_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08001329 TEST_REQUIRES_ARM_NEON;
1330 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001331 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard15eec022021-11-17 13:26:20 -08001332 for (uint32_t n = 1; n <= 8; n++) {
1333 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001334 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001335 .nr(8)
1336 .kr(2)
1337 .sr(1)
1338 .m(m)
1339 .n(n)
1340 .k(k)
1341 .cm_stride(11)
1342 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08001344 }
1345 }
1346 }
1347 }
1348
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001349 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, a_offset) {
Frank Barchard15eec022021-11-17 13:26:20 -08001350 TEST_REQUIRES_ARM_NEON;
1351 for (size_t k = 1; k <= 40; k += 9) {
1352 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001353 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08001354 .nr(8)
1355 .kr(2)
1356 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001357 .m(1)
1358 .n(8)
1359 .k(k)
1360 .ks(3)
1361 .a_offset(43)
1362 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1363 }
1364 }
1365
1366 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, zero) {
1367 TEST_REQUIRES_ARM_NEON;
1368 for (uint32_t mz = 0; mz < 1; mz++) {
1369 for (size_t k = 1; k <= 40; k += 9) {
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(43)
1380 .zero_index(mz)
1381 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_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_MULL_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(8)
1396 .qmin(128)
1397 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1398 }
1399
1400 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_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(8)
1410 .qmax(128)
1411 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1412 }
1413
1414 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_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(8)
1424 .cm_stride(11)
1425 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_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_4X16C2__NEON_MULL_LD1R, k_eq_8) {
1432 TEST_REQUIRES_ARM_NEON;
1433 GemmMicrokernelTester()
1434 .mr(4)
1435 .nr(16)
1436 .kr(2)
1437 .sr(1)
1438 .m(4)
1439 .n(16)
1440 .k(8)
1441 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1442 }
1443
1444 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, strided_cn) {
1445 TEST_REQUIRES_ARM_NEON;
1446 GemmMicrokernelTester()
1447 .mr(4)
1448 .nr(16)
1449 .kr(2)
1450 .sr(1)
1451 .m(4)
1452 .n(16)
1453 .k(8)
1454 .cn_stride(19)
1455 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1456 }
1457
1458 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, k_eq_8_subtile) {
1459 TEST_REQUIRES_ARM_NEON;
1460 for (uint32_t m = 1; m <= 4; m++) {
1461 for (uint32_t n = 1; n <= 16; n++) {
1462 GemmMicrokernelTester()
1463 .mr(4)
1464 .nr(16)
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_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1472 }
1473 }
1474 }
1475
1476 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, k_eq_8_subtile_m) {
1477 TEST_REQUIRES_ARM_NEON;
1478 for (uint32_t m = 1; m <= 4; m++) {
1479 GemmMicrokernelTester()
1480 .mr(4)
1481 .nr(16)
1482 .kr(2)
1483 .sr(1)
1484 .m(m)
1485 .n(16)
1486 .k(8)
1487 .iterations(1)
1488 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1489 }
1490 }
1491
1492 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, k_eq_8_subtile_n) {
1493 TEST_REQUIRES_ARM_NEON;
1494 for (uint32_t n = 1; n <= 16; n++) {
1495 GemmMicrokernelTester()
1496 .mr(4)
1497 .nr(16)
1498 .kr(2)
1499 .sr(1)
1500 .m(4)
1501 .n(n)
1502 .k(8)
1503 .iterations(1)
1504 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1505 }
1506 }
1507
1508 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, k_lt_8) {
1509 TEST_REQUIRES_ARM_NEON;
1510 for (size_t k = 1; k < 8; k++) {
1511 GemmMicrokernelTester()
1512 .mr(4)
1513 .nr(16)
1514 .kr(2)
1515 .sr(1)
1516 .m(4)
1517 .n(16)
1518 .k(k)
1519 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1520 }
1521 }
1522
1523 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, k_lt_8_subtile) {
1524 TEST_REQUIRES_ARM_NEON;
1525 for (size_t k = 1; k < 8; k++) {
1526 for (uint32_t m = 1; m <= 4; m++) {
1527 for (uint32_t n = 1; n <= 16; n++) {
1528 GemmMicrokernelTester()
1529 .mr(4)
1530 .nr(16)
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_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1538 }
1539 }
1540 }
1541 }
1542
1543 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, k_gt_8) {
1544 TEST_REQUIRES_ARM_NEON;
1545 for (size_t k = 9; k < 16; k++) {
1546 GemmMicrokernelTester()
1547 .mr(4)
1548 .nr(16)
1549 .kr(2)
1550 .sr(1)
1551 .m(4)
1552 .n(16)
1553 .k(k)
1554 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1555 }
1556 }
1557
1558 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, k_gt_8_subtile) {
1559 TEST_REQUIRES_ARM_NEON;
1560 for (size_t k = 9; k < 16; k++) {
1561 for (uint32_t m = 1; m <= 4; m++) {
1562 for (uint32_t n = 1; n <= 16; n++) {
1563 GemmMicrokernelTester()
1564 .mr(4)
1565 .nr(16)
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_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1573 }
1574 }
1575 }
1576 }
1577
1578 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, k_div_8) {
1579 TEST_REQUIRES_ARM_NEON;
1580 for (size_t k = 16; k <= 80; k += 8) {
1581 GemmMicrokernelTester()
1582 .mr(4)
1583 .nr(16)
1584 .kr(2)
1585 .sr(1)
1586 .m(4)
1587 .n(16)
1588 .k(k)
1589 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1590 }
1591 }
1592
1593 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, k_div_8_subtile) {
1594 TEST_REQUIRES_ARM_NEON;
1595 for (size_t k = 16; k <= 80; k += 8) {
1596 for (uint32_t m = 1; m <= 4; m++) {
1597 for (uint32_t n = 1; n <= 16; n++) {
1598 GemmMicrokernelTester()
1599 .mr(4)
1600 .nr(16)
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_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1608 }
1609 }
1610 }
1611 }
1612
1613 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, n_gt_16) {
1614 TEST_REQUIRES_ARM_NEON;
1615 for (uint32_t n = 17; n < 32; n++) {
1616 for (size_t k = 1; k <= 40; k += 9) {
1617 GemmMicrokernelTester()
1618 .mr(4)
1619 .nr(16)
1620 .kr(2)
1621 .sr(1)
1622 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001623 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001624 .k(k)
1625 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1626 }
1627 }
1628 }
1629
1630 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, n_gt_16_strided_cn) {
1631 TEST_REQUIRES_ARM_NEON;
1632 for (uint32_t n = 17; n < 32; n++) {
1633 for (size_t k = 1; k <= 40; k += 9) {
1634 GemmMicrokernelTester()
1635 .mr(4)
1636 .nr(16)
1637 .kr(2)
1638 .sr(1)
1639 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001640 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001641 .k(k)
1642 .cn_stride(19)
1643 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1644 }
1645 }
1646 }
1647
1648 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, n_gt_16_subtile) {
1649 TEST_REQUIRES_ARM_NEON;
1650 for (uint32_t n = 17; n < 32; n++) {
1651 for (size_t k = 1; k <= 40; k += 9) {
1652 for (uint32_t m = 1; m <= 4; m++) {
1653 GemmMicrokernelTester()
1654 .mr(4)
1655 .nr(16)
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_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1663 }
1664 }
1665 }
1666 }
1667
1668 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, n_div_16) {
1669 TEST_REQUIRES_ARM_NEON;
1670 for (uint32_t n = 32; n <= 48; n += 16) {
1671 for (size_t k = 1; k <= 40; k += 9) {
1672 GemmMicrokernelTester()
1673 .mr(4)
1674 .nr(16)
1675 .kr(2)
1676 .sr(1)
1677 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001678 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001679 .k(k)
1680 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1681 }
1682 }
1683 }
1684
1685 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, n_div_16_strided_cn) {
1686 TEST_REQUIRES_ARM_NEON;
1687 for (uint32_t n = 32; n <= 48; n += 16) {
1688 for (size_t k = 1; k <= 40; k += 9) {
1689 GemmMicrokernelTester()
1690 .mr(4)
1691 .nr(16)
1692 .kr(2)
1693 .sr(1)
1694 .m(4)
1695 .n(n)
1696 .k(k)
1697 .cn_stride(19)
1698 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1699 }
1700 }
1701 }
1702
1703 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, n_div_16_subtile) {
1704 TEST_REQUIRES_ARM_NEON;
1705 for (uint32_t n = 32; n <= 48; n += 16) {
1706 for (size_t k = 1; k <= 40; k += 9) {
1707 for (uint32_t m = 1; m <= 4; m++) {
1708 GemmMicrokernelTester()
1709 .mr(4)
1710 .nr(16)
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_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1718 }
1719 }
1720 }
1721 }
1722
1723 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, small_kernel) {
1724 TEST_REQUIRES_ARM_NEON;
1725 for (size_t k = 1; k <= 40; k += 9) {
1726 GemmMicrokernelTester()
1727 .mr(4)
1728 .nr(16)
1729 .kr(2)
1730 .sr(1)
1731 .m(4)
1732 .n(16)
1733 .k(k)
1734 .ks(3)
1735 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1736 }
1737 }
1738
1739 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, small_kernel_subtile) {
1740 TEST_REQUIRES_ARM_NEON;
1741 for (size_t k = 1; k <= 40; k += 9) {
1742 for (uint32_t m = 1; m <= 4; m++) {
1743 for (uint32_t n = 1; n <= 16; n++) {
1744 GemmMicrokernelTester()
1745 .mr(4)
1746 .nr(16)
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_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1755 }
1756 }
1757 }
1758 }
1759
1760 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, n_gt_16_small_kernel) {
1761 TEST_REQUIRES_ARM_NEON;
1762 for (uint32_t n = 17; n < 32; n++) {
1763 for (size_t k = 1; k <= 40; k += 9) {
1764 GemmMicrokernelTester()
1765 .mr(4)
1766 .nr(16)
1767 .kr(2)
1768 .sr(1)
1769 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001770 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001771 .k(k)
1772 .ks(3)
1773 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1774 }
1775 }
1776 }
1777
1778 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, n_div_16_small_kernel) {
1779 TEST_REQUIRES_ARM_NEON;
1780 for (uint32_t n = 32; n <= 48; n += 16) {
1781 for (size_t k = 1; k <= 40; k += 9) {
1782 GemmMicrokernelTester()
1783 .mr(4)
1784 .nr(16)
1785 .kr(2)
1786 .sr(1)
1787 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08001788 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08001789 .k(k)
1790 .ks(3)
1791 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1792 }
1793 }
1794 }
1795
1796 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, strided_cm_subtile) {
1797 TEST_REQUIRES_ARM_NEON;
1798 for (size_t k = 1; k <= 40; k += 9) {
1799 for (uint32_t m = 1; m <= 4; m++) {
1800 for (uint32_t n = 1; n <= 16; n++) {
1801 GemmMicrokernelTester()
1802 .mr(4)
1803 .nr(16)
1804 .kr(2)
1805 .sr(1)
1806 .m(m)
1807 .n(n)
1808 .k(k)
1809 .cm_stride(19)
1810 .iterations(1)
1811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1812 }
1813 }
1814 }
1815 }
1816
1817 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, a_offset) {
1818 TEST_REQUIRES_ARM_NEON;
1819 for (size_t k = 1; k <= 40; k += 9) {
1820 GemmMicrokernelTester()
1821 .mr(4)
1822 .nr(16)
1823 .kr(2)
1824 .sr(1)
1825 .m(4)
1826 .n(16)
1827 .k(k)
1828 .ks(3)
1829 .a_offset(163)
1830 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1831 }
1832 }
1833
1834 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, zero) {
1835 TEST_REQUIRES_ARM_NEON;
1836 for (uint32_t mz = 0; mz < 4; mz++) {
1837 for (size_t k = 1; k <= 40; k += 9) {
1838 GemmMicrokernelTester()
1839 .mr(4)
1840 .nr(16)
1841 .kr(2)
1842 .sr(1)
1843 .m(4)
1844 .n(16)
1845 .k(k)
1846 .ks(3)
1847 .a_offset(163)
1848 .zero_index(mz)
1849 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1850 }
1851 }
1852 }
1853
1854 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, qmin) {
1855 TEST_REQUIRES_ARM_NEON;
1856 GemmMicrokernelTester()
1857 .mr(4)
1858 .nr(16)
1859 .kr(2)
1860 .sr(1)
1861 .m(4)
1862 .n(16)
1863 .k(8)
1864 .qmin(128)
1865 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1866 }
1867
1868 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, qmax) {
1869 TEST_REQUIRES_ARM_NEON;
1870 GemmMicrokernelTester()
1871 .mr(4)
1872 .nr(16)
1873 .kr(2)
1874 .sr(1)
1875 .m(4)
1876 .n(16)
1877 .k(8)
1878 .qmax(128)
1879 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1880 }
1881
1882 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD1R, strided_cm) {
1883 TEST_REQUIRES_ARM_NEON;
1884 GemmMicrokernelTester()
1885 .mr(4)
1886 .nr(16)
1887 .kr(2)
1888 .sr(1)
1889 .m(4)
1890 .n(16)
1891 .k(8)
1892 .cm_stride(19)
1893 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r, 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_1X8C2__NEON_MLAL_LD1R, k_eq_16) {
1900 TEST_REQUIRES_ARM_NEON;
1901 GemmMicrokernelTester()
1902 .mr(1)
1903 .nr(8)
1904 .kr(2)
1905 .sr(1)
1906 .m(1)
1907 .n(8)
1908 .k(16)
1909 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1910 }
1911
1912 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, strided_cn) {
1913 TEST_REQUIRES_ARM_NEON;
1914 GemmMicrokernelTester()
1915 .mr(1)
1916 .nr(8)
1917 .kr(2)
1918 .sr(1)
1919 .m(1)
1920 .n(8)
1921 .k(16)
1922 .cn_stride(11)
1923 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1924 }
1925
1926 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_eq_16_subtile) {
1927 TEST_REQUIRES_ARM_NEON;
1928 for (uint32_t m = 1; m <= 1; m++) {
1929 for (uint32_t n = 1; n <= 8; n++) {
1930 GemmMicrokernelTester()
1931 .mr(1)
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_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1940 }
1941 }
1942 }
1943
1944 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_eq_16_subtile_m) {
1945 TEST_REQUIRES_ARM_NEON;
1946 for (uint32_t m = 1; m <= 1; m++) {
1947 GemmMicrokernelTester()
1948 .mr(1)
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_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1957 }
1958 }
1959
1960 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_eq_16_subtile_n) {
1961 TEST_REQUIRES_ARM_NEON;
1962 for (uint32_t n = 1; n <= 8; n++) {
1963 GemmMicrokernelTester()
1964 .mr(1)
1965 .nr(8)
1966 .kr(2)
1967 .sr(1)
1968 .m(1)
1969 .n(n)
1970 .k(16)
1971 .iterations(1)
1972 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1973 }
1974 }
1975
1976 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_lt_16) {
1977 TEST_REQUIRES_ARM_NEON;
1978 for (size_t k = 1; k < 16; k++) {
1979 GemmMicrokernelTester()
1980 .mr(1)
1981 .nr(8)
1982 .kr(2)
1983 .sr(1)
1984 .m(1)
1985 .n(8)
1986 .k(k)
1987 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1988 }
1989 }
1990
1991 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_lt_16_subtile) {
1992 TEST_REQUIRES_ARM_NEON;
1993 for (size_t k = 1; k < 16; k++) {
1994 for (uint32_t m = 1; m <= 1; m++) {
1995 for (uint32_t n = 1; n <= 8; n++) {
1996 GemmMicrokernelTester()
1997 .mr(1)
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_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2006 }
2007 }
2008 }
2009 }
2010
2011 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_gt_16) {
2012 TEST_REQUIRES_ARM_NEON;
2013 for (size_t k = 17; k < 32; k++) {
2014 GemmMicrokernelTester()
2015 .mr(1)
2016 .nr(8)
2017 .kr(2)
2018 .sr(1)
2019 .m(1)
2020 .n(8)
2021 .k(k)
2022 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2023 }
2024 }
2025
2026 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_gt_16_subtile) {
2027 TEST_REQUIRES_ARM_NEON;
2028 for (size_t k = 17; k < 32; k++) {
2029 for (uint32_t m = 1; m <= 1; m++) {
2030 for (uint32_t n = 1; n <= 8; n++) {
2031 GemmMicrokernelTester()
2032 .mr(1)
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_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2041 }
2042 }
2043 }
2044 }
2045
2046 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_div_16) {
2047 TEST_REQUIRES_ARM_NEON;
2048 for (size_t k = 32; k <= 160; k += 16) {
2049 GemmMicrokernelTester()
2050 .mr(1)
2051 .nr(8)
2052 .kr(2)
2053 .sr(1)
2054 .m(1)
2055 .n(8)
2056 .k(k)
2057 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2058 }
2059 }
2060
2061 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_div_16_subtile) {
2062 TEST_REQUIRES_ARM_NEON;
2063 for (size_t k = 32; k <= 160; k += 16) {
2064 for (uint32_t m = 1; m <= 1; m++) {
2065 for (uint32_t n = 1; n <= 8; n++) {
2066 GemmMicrokernelTester()
2067 .mr(1)
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_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2076 }
2077 }
2078 }
2079 }
2080
2081 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, 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(1)
2087 .nr(8)
2088 .kr(2)
2089 .sr(1)
2090 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002091 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002092 .k(k)
2093 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2094 }
2095 }
2096 }
2097
2098 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, 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(1)
2104 .nr(8)
2105 .kr(2)
2106 .sr(1)
2107 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002108 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002109 .k(k)
2110 .cn_stride(11)
2111 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2112 }
2113 }
2114 }
2115
2116 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, 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 <= 1; m++) {
2121 GemmMicrokernelTester()
2122 .mr(1)
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_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2131 }
2132 }
2133 }
2134 }
2135
2136 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, 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(1)
2142 .nr(8)
2143 .kr(2)
2144 .sr(1)
2145 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002146 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002147 .k(k)
2148 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2149 }
2150 }
2151 }
2152
2153 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, 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(1)
2159 .nr(8)
2160 .kr(2)
2161 .sr(1)
2162 .m(1)
2163 .n(n)
2164 .k(k)
2165 .cn_stride(11)
2166 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2167 }
2168 }
2169 }
2170
2171 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, 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 <= 1; m++) {
2176 GemmMicrokernelTester()
2177 .mr(1)
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_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2186 }
2187 }
2188 }
2189 }
2190
2191 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, small_kernel) {
2192 TEST_REQUIRES_ARM_NEON;
2193 for (size_t k = 1; k <= 80; k += 17) {
2194 GemmMicrokernelTester()
2195 .mr(1)
2196 .nr(8)
2197 .kr(2)
2198 .sr(1)
2199 .m(1)
2200 .n(8)
2201 .k(k)
2202 .ks(3)
2203 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2204 }
2205 }
2206
2207 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, small_kernel_subtile) {
2208 TEST_REQUIRES_ARM_NEON;
2209 for (size_t k = 1; k <= 80; k += 17) {
2210 for (uint32_t m = 1; m <= 1; m++) {
2211 for (uint32_t n = 1; n <= 8; n++) {
2212 GemmMicrokernelTester()
2213 .mr(1)
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_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2223 }
2224 }
2225 }
2226 }
2227
2228 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, 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(1)
2234 .nr(8)
2235 .kr(2)
2236 .sr(1)
2237 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002238 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002239 .k(k)
2240 .ks(3)
2241 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2242 }
2243 }
2244 }
2245
2246 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, 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(1)
2252 .nr(8)
2253 .kr(2)
2254 .sr(1)
2255 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002256 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002257 .k(k)
2258 .ks(3)
2259 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2260 }
2261 }
2262 }
2263
2264 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, strided_cm_subtile) {
2265 TEST_REQUIRES_ARM_NEON;
2266 for (size_t k = 1; k <= 80; k += 17) {
2267 for (uint32_t m = 1; m <= 1; m++) {
2268 for (uint32_t n = 1; n <= 8; n++) {
2269 GemmMicrokernelTester()
2270 .mr(1)
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_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2280 }
2281 }
2282 }
2283 }
2284
2285 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, a_offset) {
2286 TEST_REQUIRES_ARM_NEON;
2287 for (size_t k = 1; k <= 80; k += 17) {
2288 GemmMicrokernelTester()
2289 .mr(1)
2290 .nr(8)
2291 .kr(2)
2292 .sr(1)
2293 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08002294 .n(8)
2295 .k(k)
2296 .ks(3)
2297 .a_offset(83)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002298 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002299 }
2300 }
2301
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002302 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, zero) {
Frank Barchard15eec022021-11-17 13:26:20 -08002303 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002304 for (uint32_t mz = 0; mz < 1; mz++) {
2305 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08002306 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002307 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08002308 .nr(8)
2309 .kr(2)
2310 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002311 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08002312 .n(8)
2313 .k(k)
2314 .ks(3)
2315 .a_offset(83)
2316 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002317 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002318 }
2319 }
2320 }
2321
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002322 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, qmin) {
Frank Barchard15eec022021-11-17 13:26:20 -08002323 TEST_REQUIRES_ARM_NEON;
2324 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002325 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08002326 .nr(8)
2327 .kr(2)
2328 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002329 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08002330 .n(8)
2331 .k(16)
2332 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002333 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002334 }
2335
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002336 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, qmax) {
Frank Barchard15eec022021-11-17 13:26:20 -08002337 TEST_REQUIRES_ARM_NEON;
2338 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002339 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08002340 .nr(8)
2341 .kr(2)
2342 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002343 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08002344 .n(8)
2345 .k(16)
2346 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002347 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002348 }
2349
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002350 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, strided_cm) {
Frank Barchard15eec022021-11-17 13:26:20 -08002351 TEST_REQUIRES_ARM_NEON;
2352 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002353 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08002354 .nr(8)
2355 .kr(2)
2356 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002357 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08002358 .n(8)
2359 .k(16)
2360 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002361 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002362 }
2363#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
2364
2365
2366#if XNN_ARCH_ARM || XNN_ARCH_ARM64
2367 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_eq_16) {
2368 TEST_REQUIRES_ARM_NEON;
2369 GemmMicrokernelTester()
2370 .mr(4)
2371 .nr(8)
2372 .kr(2)
2373 .sr(1)
2374 .m(4)
2375 .n(8)
2376 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -08002377 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002378 }
2379
2380 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, strided_cn) {
2381 TEST_REQUIRES_ARM_NEON;
2382 GemmMicrokernelTester()
2383 .mr(4)
2384 .nr(8)
2385 .kr(2)
2386 .sr(1)
2387 .m(4)
2388 .n(8)
2389 .k(16)
2390 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08002391 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002392 }
2393
2394 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_eq_16_subtile) {
2395 TEST_REQUIRES_ARM_NEON;
2396 for (uint32_t m = 1; m <= 4; m++) {
2397 for (uint32_t n = 1; n <= 8; n++) {
2398 GemmMicrokernelTester()
2399 .mr(4)
2400 .nr(8)
2401 .kr(2)
2402 .sr(1)
2403 .m(m)
2404 .n(n)
2405 .k(16)
2406 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08002407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002408 }
2409 }
2410 }
2411
2412 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_eq_16_subtile_m) {
2413 TEST_REQUIRES_ARM_NEON;
2414 for (uint32_t m = 1; m <= 4; m++) {
2415 GemmMicrokernelTester()
2416 .mr(4)
2417 .nr(8)
2418 .kr(2)
2419 .sr(1)
2420 .m(m)
2421 .n(8)
2422 .k(16)
2423 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08002424 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002425 }
2426 }
2427
2428 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_eq_16_subtile_n) {
2429 TEST_REQUIRES_ARM_NEON;
2430 for (uint32_t n = 1; n <= 8; n++) {
2431 GemmMicrokernelTester()
2432 .mr(4)
2433 .nr(8)
2434 .kr(2)
2435 .sr(1)
2436 .m(4)
2437 .n(n)
2438 .k(16)
2439 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08002440 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002441 }
2442 }
2443
2444 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_lt_16) {
2445 TEST_REQUIRES_ARM_NEON;
2446 for (size_t k = 1; k < 16; k++) {
2447 GemmMicrokernelTester()
2448 .mr(4)
2449 .nr(8)
2450 .kr(2)
2451 .sr(1)
2452 .m(4)
2453 .n(8)
2454 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08002455 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002456 }
2457 }
2458
2459 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_lt_16_subtile) {
2460 TEST_REQUIRES_ARM_NEON;
2461 for (size_t k = 1; k < 16; k++) {
2462 for (uint32_t m = 1; m <= 4; m++) {
2463 for (uint32_t n = 1; n <= 8; n++) {
2464 GemmMicrokernelTester()
2465 .mr(4)
2466 .nr(8)
2467 .kr(2)
2468 .sr(1)
2469 .m(m)
2470 .n(n)
2471 .k(k)
2472 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08002473 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002474 }
2475 }
2476 }
2477 }
2478
2479 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_gt_16) {
2480 TEST_REQUIRES_ARM_NEON;
2481 for (size_t k = 17; k < 32; k++) {
2482 GemmMicrokernelTester()
2483 .mr(4)
2484 .nr(8)
2485 .kr(2)
2486 .sr(1)
2487 .m(4)
2488 .n(8)
2489 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08002490 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002491 }
2492 }
2493
2494 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_gt_16_subtile) {
2495 TEST_REQUIRES_ARM_NEON;
2496 for (size_t k = 17; k < 32; k++) {
2497 for (uint32_t m = 1; m <= 4; m++) {
2498 for (uint32_t n = 1; n <= 8; n++) {
2499 GemmMicrokernelTester()
2500 .mr(4)
2501 .nr(8)
2502 .kr(2)
2503 .sr(1)
2504 .m(m)
2505 .n(n)
2506 .k(k)
2507 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08002508 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002509 }
2510 }
2511 }
2512 }
2513
2514 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_div_16) {
2515 TEST_REQUIRES_ARM_NEON;
2516 for (size_t k = 32; k <= 160; k += 16) {
2517 GemmMicrokernelTester()
2518 .mr(4)
2519 .nr(8)
2520 .kr(2)
2521 .sr(1)
2522 .m(4)
2523 .n(8)
2524 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08002525 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002526 }
2527 }
2528
2529 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, k_div_16_subtile) {
2530 TEST_REQUIRES_ARM_NEON;
2531 for (size_t k = 32; k <= 160; k += 16) {
2532 for (uint32_t m = 1; m <= 4; m++) {
2533 for (uint32_t n = 1; n <= 8; n++) {
2534 GemmMicrokernelTester()
2535 .mr(4)
2536 .nr(8)
2537 .kr(2)
2538 .sr(1)
2539 .m(m)
2540 .n(n)
2541 .k(k)
2542 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08002543 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002544 }
2545 }
2546 }
2547 }
2548
2549 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, n_gt_8) {
2550 TEST_REQUIRES_ARM_NEON;
2551 for (uint32_t n = 9; n < 16; n++) {
2552 for (size_t k = 1; k <= 80; k += 17) {
2553 GemmMicrokernelTester()
2554 .mr(4)
2555 .nr(8)
2556 .kr(2)
2557 .sr(1)
2558 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002559 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08002560 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08002561 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002562 }
2563 }
2564 }
2565
2566 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, 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 <= 80; k += 17) {
2570 GemmMicrokernelTester()
2571 .mr(4)
2572 .nr(8)
2573 .kr(2)
2574 .sr(1)
2575 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002576 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08002577 .k(k)
2578 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08002579 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002580 }
2581 }
2582 }
2583
2584 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, 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 <= 80; k += 17) {
2588 for (uint32_t m = 1; m <= 4; m++) {
2589 GemmMicrokernelTester()
2590 .mr(4)
2591 .nr(8)
2592 .kr(2)
2593 .sr(1)
2594 .m(m)
2595 .n(n)
2596 .k(k)
2597 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08002598 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002599 }
2600 }
2601 }
2602 }
2603
2604 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, 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 <= 80; k += 17) {
2608 GemmMicrokernelTester()
2609 .mr(4)
2610 .nr(8)
2611 .kr(2)
2612 .sr(1)
2613 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002614 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08002615 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08002616 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002617 }
2618 }
2619 }
2620
2621 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, 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 <= 80; k += 17) {
2625 GemmMicrokernelTester()
2626 .mr(4)
2627 .nr(8)
2628 .kr(2)
2629 .sr(1)
2630 .m(4)
2631 .n(n)
2632 .k(k)
2633 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08002634 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002635 }
2636 }
2637 }
2638
2639 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, 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 <= 80; k += 17) {
2643 for (uint32_t m = 1; m <= 4; m++) {
2644 GemmMicrokernelTester()
2645 .mr(4)
2646 .nr(8)
2647 .kr(2)
2648 .sr(1)
2649 .m(m)
2650 .n(n)
2651 .k(k)
2652 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08002653 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002654 }
2655 }
2656 }
2657 }
2658
2659 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, small_kernel) {
2660 TEST_REQUIRES_ARM_NEON;
2661 for (size_t k = 1; k <= 80; k += 17) {
2662 GemmMicrokernelTester()
2663 .mr(4)
2664 .nr(8)
2665 .kr(2)
2666 .sr(1)
2667 .m(4)
2668 .n(8)
2669 .k(k)
2670 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08002671 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002672 }
2673 }
2674
2675 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, small_kernel_subtile) {
2676 TEST_REQUIRES_ARM_NEON;
2677 for (size_t k = 1; k <= 80; k += 17) {
2678 for (uint32_t m = 1; m <= 4; m++) {
2679 for (uint32_t n = 1; n <= 8; n++) {
2680 GemmMicrokernelTester()
2681 .mr(4)
2682 .nr(8)
2683 .kr(2)
2684 .sr(1)
2685 .m(m)
2686 .n(n)
2687 .k(k)
2688 .ks(3)
2689 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08002690 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002691 }
2692 }
2693 }
2694 }
2695
2696 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, 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 <= 80; k += 17) {
2700 GemmMicrokernelTester()
2701 .mr(4)
2702 .nr(8)
2703 .kr(2)
2704 .sr(1)
2705 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002706 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08002707 .k(k)
2708 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08002709 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002710 }
2711 }
2712 }
2713
2714 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, 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 <= 80; k += 17) {
2718 GemmMicrokernelTester()
2719 .mr(4)
2720 .nr(8)
2721 .kr(2)
2722 .sr(1)
2723 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08002724 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08002725 .k(k)
2726 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08002727 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002728 }
2729 }
2730 }
2731
2732 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, strided_cm_subtile) {
2733 TEST_REQUIRES_ARM_NEON;
2734 for (size_t k = 1; k <= 80; k += 17) {
2735 for (uint32_t m = 1; m <= 4; m++) {
2736 for (uint32_t n = 1; n <= 8; n++) {
2737 GemmMicrokernelTester()
2738 .mr(4)
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)
Marat Dukhan50323b82022-01-11 00:12:01 -08002747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002748 }
2749 }
2750 }
2751 }
2752
2753 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, a_offset) {
2754 TEST_REQUIRES_ARM_NEON;
2755 for (size_t k = 1; k <= 80; k += 17) {
2756 GemmMicrokernelTester()
2757 .mr(4)
2758 .nr(8)
2759 .kr(2)
2760 .sr(1)
2761 .m(4)
2762 .n(8)
2763 .k(k)
2764 .ks(3)
2765 .a_offset(331)
Marat Dukhan50323b82022-01-11 00:12:01 -08002766 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002767 }
2768 }
2769
2770 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, zero) {
2771 TEST_REQUIRES_ARM_NEON;
2772 for (uint32_t mz = 0; mz < 4; mz++) {
2773 for (size_t k = 1; k <= 80; k += 17) {
2774 GemmMicrokernelTester()
2775 .mr(4)
2776 .nr(8)
2777 .kr(2)
2778 .sr(1)
2779 .m(4)
2780 .n(8)
2781 .k(k)
2782 .ks(3)
2783 .a_offset(331)
2784 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08002785 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002786 }
2787 }
2788 }
2789
2790 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, qmin) {
2791 TEST_REQUIRES_ARM_NEON;
2792 GemmMicrokernelTester()
2793 .mr(4)
2794 .nr(8)
2795 .kr(2)
2796 .sr(1)
2797 .m(4)
2798 .n(8)
2799 .k(16)
2800 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08002801 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002802 }
2803
2804 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, qmax) {
2805 TEST_REQUIRES_ARM_NEON;
2806 GemmMicrokernelTester()
2807 .mr(4)
2808 .nr(8)
2809 .kr(2)
2810 .sr(1)
2811 .m(4)
2812 .n(8)
2813 .k(16)
2814 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08002815 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002816 }
2817
2818 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD1R, strided_cm) {
2819 TEST_REQUIRES_ARM_NEON;
2820 GemmMicrokernelTester()
2821 .mr(4)
2822 .nr(8)
2823 .kr(2)
2824 .sr(1)
2825 .m(4)
2826 .n(8)
2827 .k(16)
2828 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08002829 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002830 }
2831#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
2832
2833
2834#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002835 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_eq_16) {
Frank Barchard15eec022021-11-17 13:26:20 -08002836 TEST_REQUIRES_ARM_NEON;
2837 GemmMicrokernelTester()
2838 .mr(1)
2839 .nr(16)
2840 .kr(2)
2841 .sr(1)
2842 .m(1)
2843 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002844 .k(16)
2845 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002846 }
2847
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002848 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, strided_cn) {
Frank Barchard15eec022021-11-17 13:26:20 -08002849 TEST_REQUIRES_ARM_NEON;
2850 GemmMicrokernelTester()
2851 .mr(1)
2852 .nr(16)
2853 .kr(2)
2854 .sr(1)
2855 .m(1)
2856 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002857 .k(16)
Frank Barchard15eec022021-11-17 13:26:20 -08002858 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002859 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002860 }
2861
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002862 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_eq_16_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08002863 TEST_REQUIRES_ARM_NEON;
2864 for (uint32_t m = 1; m <= 1; m++) {
2865 for (uint32_t n = 1; n <= 16; n++) {
2866 GemmMicrokernelTester()
2867 .mr(1)
2868 .nr(16)
2869 .kr(2)
2870 .sr(1)
2871 .m(m)
2872 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002873 .k(16)
Frank Barchard15eec022021-11-17 13:26:20 -08002874 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002876 }
2877 }
2878 }
2879
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002880 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_eq_16_subtile_m) {
Frank Barchard15eec022021-11-17 13:26:20 -08002881 TEST_REQUIRES_ARM_NEON;
2882 for (uint32_t m = 1; m <= 1; m++) {
2883 GemmMicrokernelTester()
2884 .mr(1)
2885 .nr(16)
2886 .kr(2)
2887 .sr(1)
2888 .m(m)
2889 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002890 .k(16)
Frank Barchard15eec022021-11-17 13:26:20 -08002891 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002892 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002893 }
2894 }
2895
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002896 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_eq_16_subtile_n) {
Frank Barchard15eec022021-11-17 13:26:20 -08002897 TEST_REQUIRES_ARM_NEON;
2898 for (uint32_t n = 1; n <= 16; n++) {
2899 GemmMicrokernelTester()
2900 .mr(1)
2901 .nr(16)
2902 .kr(2)
2903 .sr(1)
2904 .m(1)
2905 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002906 .k(16)
Frank Barchard15eec022021-11-17 13:26:20 -08002907 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002908 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002909 }
2910 }
2911
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002912 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_lt_16) {
Frank Barchard15eec022021-11-17 13:26:20 -08002913 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002914 for (size_t k = 1; k < 16; k++) {
Frank Barchard15eec022021-11-17 13:26:20 -08002915 GemmMicrokernelTester()
2916 .mr(1)
2917 .nr(16)
2918 .kr(2)
2919 .sr(1)
2920 .m(1)
2921 .n(16)
2922 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002923 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002924 }
2925 }
2926
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002927 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_lt_16_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08002928 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002929 for (size_t k = 1; k < 16; k++) {
Frank Barchard15eec022021-11-17 13:26:20 -08002930 for (uint32_t m = 1; m <= 1; m++) {
2931 for (uint32_t n = 1; n <= 16; n++) {
2932 GemmMicrokernelTester()
2933 .mr(1)
2934 .nr(16)
2935 .kr(2)
2936 .sr(1)
2937 .m(m)
2938 .n(n)
2939 .k(k)
2940 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002941 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002942 }
2943 }
2944 }
2945 }
2946
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002947 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_gt_16) {
Frank Barchard15eec022021-11-17 13:26:20 -08002948 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002949 for (size_t k = 17; k < 32; k++) {
Frank Barchard15eec022021-11-17 13:26:20 -08002950 GemmMicrokernelTester()
2951 .mr(1)
2952 .nr(16)
2953 .kr(2)
2954 .sr(1)
2955 .m(1)
2956 .n(16)
2957 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002958 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002959 }
2960 }
2961
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002962 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_gt_16_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08002963 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002964 for (size_t k = 17; k < 32; k++) {
Frank Barchard15eec022021-11-17 13:26:20 -08002965 for (uint32_t m = 1; m <= 1; m++) {
2966 for (uint32_t n = 1; n <= 16; n++) {
2967 GemmMicrokernelTester()
2968 .mr(1)
2969 .nr(16)
2970 .kr(2)
2971 .sr(1)
2972 .m(m)
2973 .n(n)
2974 .k(k)
2975 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002976 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002977 }
2978 }
2979 }
2980 }
2981
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002982 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_div_16) {
Frank Barchard15eec022021-11-17 13:26:20 -08002983 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002984 for (size_t k = 32; k <= 160; k += 16) {
Frank Barchard15eec022021-11-17 13:26:20 -08002985 GemmMicrokernelTester()
2986 .mr(1)
2987 .nr(16)
2988 .kr(2)
2989 .sr(1)
2990 .m(1)
2991 .n(16)
2992 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002993 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08002994 }
2995 }
2996
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002997 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, k_div_16_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08002998 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08002999 for (size_t k = 32; k <= 160; k += 16) {
Frank Barchard15eec022021-11-17 13:26:20 -08003000 for (uint32_t m = 1; m <= 1; m++) {
3001 for (uint32_t n = 1; n <= 16; n++) {
3002 GemmMicrokernelTester()
3003 .mr(1)
3004 .nr(16)
3005 .kr(2)
3006 .sr(1)
3007 .m(m)
3008 .n(n)
3009 .k(k)
3010 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003011 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003012 }
3013 }
3014 }
3015 }
3016
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003017 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, n_gt_16) {
Frank Barchard15eec022021-11-17 13:26:20 -08003018 TEST_REQUIRES_ARM_NEON;
3019 for (uint32_t n = 17; n < 32; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003020 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003021 GemmMicrokernelTester()
3022 .mr(1)
3023 .nr(16)
3024 .kr(2)
3025 .sr(1)
3026 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003027 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003028 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003029 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003030 }
3031 }
3032 }
3033
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003034 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, n_gt_16_strided_cn) {
Frank Barchard15eec022021-11-17 13:26:20 -08003035 TEST_REQUIRES_ARM_NEON;
3036 for (uint32_t n = 17; n < 32; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003037 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003038 GemmMicrokernelTester()
3039 .mr(1)
3040 .nr(16)
3041 .kr(2)
3042 .sr(1)
3043 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003044 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003045 .k(k)
3046 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003047 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003048 }
3049 }
3050 }
3051
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003052 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, n_gt_16_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08003053 TEST_REQUIRES_ARM_NEON;
3054 for (uint32_t n = 17; n < 32; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003055 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003056 for (uint32_t m = 1; m <= 1; m++) {
3057 GemmMicrokernelTester()
3058 .mr(1)
3059 .nr(16)
3060 .kr(2)
3061 .sr(1)
3062 .m(m)
3063 .n(n)
3064 .k(k)
3065 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003066 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003067 }
3068 }
3069 }
3070 }
3071
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003072 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, n_div_16) {
Frank Barchard15eec022021-11-17 13:26:20 -08003073 TEST_REQUIRES_ARM_NEON;
3074 for (uint32_t n = 32; n <= 48; n += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003075 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003076 GemmMicrokernelTester()
3077 .mr(1)
3078 .nr(16)
3079 .kr(2)
3080 .sr(1)
3081 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003082 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003083 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003084 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003085 }
3086 }
3087 }
3088
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003089 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, n_div_16_strided_cn) {
Frank Barchard15eec022021-11-17 13:26:20 -08003090 TEST_REQUIRES_ARM_NEON;
3091 for (uint32_t n = 32; n <= 48; n += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003092 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003093 GemmMicrokernelTester()
3094 .mr(1)
3095 .nr(16)
3096 .kr(2)
3097 .sr(1)
3098 .m(1)
3099 .n(n)
3100 .k(k)
3101 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003102 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003103 }
3104 }
3105 }
3106
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003107 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, n_div_16_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08003108 TEST_REQUIRES_ARM_NEON;
3109 for (uint32_t n = 32; n <= 48; n += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003110 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003111 for (uint32_t m = 1; m <= 1; m++) {
3112 GemmMicrokernelTester()
3113 .mr(1)
3114 .nr(16)
3115 .kr(2)
3116 .sr(1)
3117 .m(m)
3118 .n(n)
3119 .k(k)
3120 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003121 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003122 }
3123 }
3124 }
3125 }
3126
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003127 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, small_kernel) {
Frank Barchard15eec022021-11-17 13:26:20 -08003128 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003129 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003130 GemmMicrokernelTester()
3131 .mr(1)
3132 .nr(16)
3133 .kr(2)
3134 .sr(1)
3135 .m(1)
3136 .n(16)
3137 .k(k)
3138 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003139 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003140 }
3141 }
3142
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003143 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, small_kernel_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08003144 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003145 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003146 for (uint32_t m = 1; m <= 1; m++) {
3147 for (uint32_t n = 1; n <= 16; n++) {
3148 GemmMicrokernelTester()
3149 .mr(1)
3150 .nr(16)
3151 .kr(2)
3152 .sr(1)
3153 .m(m)
3154 .n(n)
3155 .k(k)
3156 .ks(3)
3157 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003158 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003159 }
3160 }
3161 }
3162 }
3163
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003164 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, n_gt_16_small_kernel) {
Frank Barchard15eec022021-11-17 13:26:20 -08003165 TEST_REQUIRES_ARM_NEON;
3166 for (uint32_t n = 17; n < 32; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003167 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003168 GemmMicrokernelTester()
3169 .mr(1)
3170 .nr(16)
3171 .kr(2)
3172 .sr(1)
3173 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003174 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003175 .k(k)
3176 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003177 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003178 }
3179 }
3180 }
3181
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003182 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, n_div_16_small_kernel) {
Frank Barchard15eec022021-11-17 13:26:20 -08003183 TEST_REQUIRES_ARM_NEON;
3184 for (uint32_t n = 32; n <= 48; n += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003185 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003186 GemmMicrokernelTester()
3187 .mr(1)
3188 .nr(16)
3189 .kr(2)
3190 .sr(1)
3191 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003192 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003193 .k(k)
3194 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003195 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003196 }
3197 }
3198 }
3199
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003200 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, strided_cm_subtile) {
Frank Barchard15eec022021-11-17 13:26:20 -08003201 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003202 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003203 for (uint32_t m = 1; m <= 1; m++) {
3204 for (uint32_t n = 1; n <= 16; n++) {
3205 GemmMicrokernelTester()
3206 .mr(1)
3207 .nr(16)
3208 .kr(2)
3209 .sr(1)
3210 .m(m)
3211 .n(n)
3212 .k(k)
3213 .cm_stride(19)
3214 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003216 }
3217 }
3218 }
3219 }
3220
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003221 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, a_offset) {
Frank Barchard15eec022021-11-17 13:26:20 -08003222 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003223 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003224 GemmMicrokernelTester()
3225 .mr(1)
3226 .nr(16)
3227 .kr(2)
3228 .sr(1)
3229 .m(1)
3230 .n(16)
3231 .k(k)
3232 .ks(3)
Frank Barchard15eec022021-11-17 13:26:20 -08003233 .a_offset(83)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003234 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003235 }
3236 }
3237
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003238 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, zero) {
Frank Barchard15eec022021-11-17 13:26:20 -08003239 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003240 for (uint32_t mz = 0; mz < 1; mz++) {
3241 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard15eec022021-11-17 13:26:20 -08003242 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003243 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08003244 .nr(16)
3245 .kr(2)
3246 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003247 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08003248 .n(16)
3249 .k(k)
3250 .ks(3)
3251 .a_offset(83)
3252 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003253 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003254 }
3255 }
3256 }
3257
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003258 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, qmin) {
Frank Barchard15eec022021-11-17 13:26:20 -08003259 TEST_REQUIRES_ARM_NEON;
3260 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003261 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08003262 .nr(16)
3263 .kr(2)
3264 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003265 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08003266 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003267 .k(16)
Frank Barchard15eec022021-11-17 13:26:20 -08003268 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003269 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003270 }
3271
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003272 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, qmax) {
Frank Barchard15eec022021-11-17 13:26:20 -08003273 TEST_REQUIRES_ARM_NEON;
3274 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003275 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08003276 .nr(16)
3277 .kr(2)
3278 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003279 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08003280 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003281 .k(16)
Frank Barchard15eec022021-11-17 13:26:20 -08003282 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003283 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003284 }
3285
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003286 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_LD1R, strided_cm) {
Frank Barchard15eec022021-11-17 13:26:20 -08003287 TEST_REQUIRES_ARM_NEON;
3288 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003289 .mr(1)
Frank Barchard15eec022021-11-17 13:26:20 -08003290 .nr(16)
3291 .kr(2)
3292 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003293 .m(1)
Frank Barchard15eec022021-11-17 13:26:20 -08003294 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003295 .k(16)
Frank Barchard15eec022021-11-17 13:26:20 -08003296 .cm_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08003297 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003298 }
3299#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
3300
3301
3302#if XNN_ARCH_ARM || XNN_ARCH_ARM64
3303 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_eq_8) {
3304 TEST_REQUIRES_ARM_NEON;
3305 GemmMicrokernelTester()
3306 .mr(3)
3307 .nr(16)
3308 .kr(2)
3309 .sr(1)
3310 .m(3)
3311 .n(16)
3312 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -08003313 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003314 }
3315
3316 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, strided_cn) {
3317 TEST_REQUIRES_ARM_NEON;
3318 GemmMicrokernelTester()
3319 .mr(3)
3320 .nr(16)
3321 .kr(2)
3322 .sr(1)
3323 .m(3)
3324 .n(16)
3325 .k(8)
3326 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08003327 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003328 }
3329
3330 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_eq_8_subtile) {
3331 TEST_REQUIRES_ARM_NEON;
3332 for (uint32_t m = 1; m <= 3; m++) {
3333 for (uint32_t n = 1; n <= 16; n++) {
3334 GemmMicrokernelTester()
3335 .mr(3)
3336 .nr(16)
3337 .kr(2)
3338 .sr(1)
3339 .m(m)
3340 .n(n)
3341 .k(8)
3342 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003344 }
3345 }
3346 }
3347
3348 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_eq_8_subtile_m) {
3349 TEST_REQUIRES_ARM_NEON;
3350 for (uint32_t m = 1; m <= 3; m++) {
3351 GemmMicrokernelTester()
3352 .mr(3)
3353 .nr(16)
3354 .kr(2)
3355 .sr(1)
3356 .m(m)
3357 .n(16)
3358 .k(8)
3359 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003360 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003361 }
3362 }
3363
3364 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_eq_8_subtile_n) {
3365 TEST_REQUIRES_ARM_NEON;
3366 for (uint32_t n = 1; n <= 16; n++) {
3367 GemmMicrokernelTester()
3368 .mr(3)
3369 .nr(16)
3370 .kr(2)
3371 .sr(1)
3372 .m(3)
3373 .n(n)
3374 .k(8)
3375 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003376 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003377 }
3378 }
3379
3380 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_lt_8) {
3381 TEST_REQUIRES_ARM_NEON;
3382 for (size_t k = 1; k < 8; k++) {
3383 GemmMicrokernelTester()
3384 .mr(3)
3385 .nr(16)
3386 .kr(2)
3387 .sr(1)
3388 .m(3)
3389 .n(16)
3390 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08003391 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003392 }
3393 }
3394
3395 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_lt_8_subtile) {
3396 TEST_REQUIRES_ARM_NEON;
3397 for (size_t k = 1; k < 8; k++) {
3398 for (uint32_t m = 1; m <= 3; m++) {
3399 for (uint32_t n = 1; n <= 16; n++) {
3400 GemmMicrokernelTester()
3401 .mr(3)
3402 .nr(16)
3403 .kr(2)
3404 .sr(1)
3405 .m(m)
3406 .n(n)
3407 .k(k)
3408 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003409 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003410 }
3411 }
3412 }
3413 }
3414
3415 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_gt_8) {
3416 TEST_REQUIRES_ARM_NEON;
3417 for (size_t k = 9; k < 16; k++) {
3418 GemmMicrokernelTester()
3419 .mr(3)
3420 .nr(16)
3421 .kr(2)
3422 .sr(1)
3423 .m(3)
3424 .n(16)
3425 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08003426 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003427 }
3428 }
3429
3430 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_gt_8_subtile) {
3431 TEST_REQUIRES_ARM_NEON;
3432 for (size_t k = 9; k < 16; k++) {
3433 for (uint32_t m = 1; m <= 3; m++) {
3434 for (uint32_t n = 1; n <= 16; n++) {
3435 GemmMicrokernelTester()
3436 .mr(3)
3437 .nr(16)
3438 .kr(2)
3439 .sr(1)
3440 .m(m)
3441 .n(n)
3442 .k(k)
3443 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003444 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003445 }
3446 }
3447 }
3448 }
3449
3450 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_div_8) {
3451 TEST_REQUIRES_ARM_NEON;
3452 for (size_t k = 16; k <= 80; k += 8) {
3453 GemmMicrokernelTester()
3454 .mr(3)
3455 .nr(16)
3456 .kr(2)
3457 .sr(1)
3458 .m(3)
3459 .n(16)
3460 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08003461 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003462 }
3463 }
3464
3465 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_div_8_subtile) {
3466 TEST_REQUIRES_ARM_NEON;
3467 for (size_t k = 16; k <= 80; k += 8) {
3468 for (uint32_t m = 1; m <= 3; m++) {
3469 for (uint32_t n = 1; n <= 16; n++) {
3470 GemmMicrokernelTester()
3471 .mr(3)
3472 .nr(16)
3473 .kr(2)
3474 .sr(1)
3475 .m(m)
3476 .n(n)
3477 .k(k)
3478 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003479 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003480 }
3481 }
3482 }
3483 }
3484
3485 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_gt_16) {
3486 TEST_REQUIRES_ARM_NEON;
3487 for (uint32_t n = 17; n < 32; n++) {
3488 for (size_t k = 1; k <= 40; k += 9) {
3489 GemmMicrokernelTester()
3490 .mr(3)
3491 .nr(16)
3492 .kr(2)
3493 .sr(1)
3494 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003495 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003496 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08003497 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003498 }
3499 }
3500 }
3501
3502 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_gt_16_strided_cn) {
3503 TEST_REQUIRES_ARM_NEON;
3504 for (uint32_t n = 17; n < 32; n++) {
3505 for (size_t k = 1; k <= 40; k += 9) {
3506 GemmMicrokernelTester()
3507 .mr(3)
3508 .nr(16)
3509 .kr(2)
3510 .sr(1)
3511 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003512 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003513 .k(k)
3514 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08003515 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003516 }
3517 }
3518 }
3519
3520 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_gt_16_subtile) {
3521 TEST_REQUIRES_ARM_NEON;
3522 for (uint32_t n = 17; n < 32; n++) {
3523 for (size_t k = 1; k <= 40; k += 9) {
3524 for (uint32_t m = 1; m <= 3; m++) {
3525 GemmMicrokernelTester()
3526 .mr(3)
3527 .nr(16)
3528 .kr(2)
3529 .sr(1)
3530 .m(m)
3531 .n(n)
3532 .k(k)
3533 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003534 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003535 }
3536 }
3537 }
3538 }
3539
3540 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_div_16) {
3541 TEST_REQUIRES_ARM_NEON;
3542 for (uint32_t n = 32; n <= 48; n += 16) {
3543 for (size_t k = 1; k <= 40; k += 9) {
3544 GemmMicrokernelTester()
3545 .mr(3)
3546 .nr(16)
3547 .kr(2)
3548 .sr(1)
3549 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003550 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003551 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08003552 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003553 }
3554 }
3555 }
3556
3557 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_div_16_strided_cn) {
3558 TEST_REQUIRES_ARM_NEON;
3559 for (uint32_t n = 32; n <= 48; n += 16) {
3560 for (size_t k = 1; k <= 40; k += 9) {
3561 GemmMicrokernelTester()
3562 .mr(3)
3563 .nr(16)
3564 .kr(2)
3565 .sr(1)
3566 .m(3)
3567 .n(n)
3568 .k(k)
3569 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08003570 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003571 }
3572 }
3573 }
3574
3575 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_div_16_subtile) {
3576 TEST_REQUIRES_ARM_NEON;
3577 for (uint32_t n = 32; n <= 48; n += 16) {
3578 for (size_t k = 1; k <= 40; k += 9) {
3579 for (uint32_t m = 1; m <= 3; m++) {
3580 GemmMicrokernelTester()
3581 .mr(3)
3582 .nr(16)
3583 .kr(2)
3584 .sr(1)
3585 .m(m)
3586 .n(n)
3587 .k(k)
3588 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003589 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003590 }
3591 }
3592 }
3593 }
3594
3595 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, small_kernel) {
3596 TEST_REQUIRES_ARM_NEON;
3597 for (size_t k = 1; k <= 40; k += 9) {
3598 GemmMicrokernelTester()
3599 .mr(3)
3600 .nr(16)
3601 .kr(2)
3602 .sr(1)
3603 .m(3)
3604 .n(16)
3605 .k(k)
3606 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08003607 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003608 }
3609 }
3610
3611 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, small_kernel_subtile) {
3612 TEST_REQUIRES_ARM_NEON;
3613 for (size_t k = 1; k <= 40; k += 9) {
3614 for (uint32_t m = 1; m <= 3; m++) {
3615 for (uint32_t n = 1; n <= 16; n++) {
3616 GemmMicrokernelTester()
3617 .mr(3)
3618 .nr(16)
3619 .kr(2)
3620 .sr(1)
3621 .m(m)
3622 .n(n)
3623 .k(k)
3624 .ks(3)
3625 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003626 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003627 }
3628 }
3629 }
3630 }
3631
3632 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_gt_16_small_kernel) {
3633 TEST_REQUIRES_ARM_NEON;
3634 for (uint32_t n = 17; n < 32; n++) {
3635 for (size_t k = 1; k <= 40; k += 9) {
3636 GemmMicrokernelTester()
3637 .mr(3)
3638 .nr(16)
3639 .kr(2)
3640 .sr(1)
3641 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003642 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003643 .k(k)
3644 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08003645 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003646 }
3647 }
3648 }
3649
3650 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_div_16_small_kernel) {
3651 TEST_REQUIRES_ARM_NEON;
3652 for (uint32_t n = 32; n <= 48; n += 16) {
3653 for (size_t k = 1; k <= 40; k += 9) {
3654 GemmMicrokernelTester()
3655 .mr(3)
3656 .nr(16)
3657 .kr(2)
3658 .sr(1)
3659 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003660 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003661 .k(k)
3662 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08003663 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003664 }
3665 }
3666 }
3667
3668 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, strided_cm_subtile) {
3669 TEST_REQUIRES_ARM_NEON;
3670 for (size_t k = 1; k <= 40; k += 9) {
3671 for (uint32_t m = 1; m <= 3; m++) {
3672 for (uint32_t n = 1; n <= 16; n++) {
3673 GemmMicrokernelTester()
3674 .mr(3)
3675 .nr(16)
3676 .kr(2)
3677 .sr(1)
3678 .m(m)
3679 .n(n)
3680 .k(k)
3681 .cm_stride(19)
3682 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003683 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003684 }
3685 }
3686 }
3687 }
3688
3689 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, a_offset) {
3690 TEST_REQUIRES_ARM_NEON;
3691 for (size_t k = 1; k <= 40; k += 9) {
3692 GemmMicrokernelTester()
3693 .mr(3)
3694 .nr(16)
3695 .kr(2)
3696 .sr(1)
3697 .m(3)
3698 .n(16)
3699 .k(k)
3700 .ks(3)
3701 .a_offset(127)
Marat Dukhan50323b82022-01-11 00:12:01 -08003702 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003703 }
3704 }
3705
3706 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, zero) {
3707 TEST_REQUIRES_ARM_NEON;
3708 for (uint32_t mz = 0; mz < 3; mz++) {
3709 for (size_t k = 1; k <= 40; k += 9) {
3710 GemmMicrokernelTester()
3711 .mr(3)
3712 .nr(16)
3713 .kr(2)
3714 .sr(1)
3715 .m(3)
3716 .n(16)
3717 .k(k)
3718 .ks(3)
3719 .a_offset(127)
3720 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08003721 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003722 }
3723 }
3724 }
3725
3726 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, qmin) {
3727 TEST_REQUIRES_ARM_NEON;
3728 GemmMicrokernelTester()
3729 .mr(3)
3730 .nr(16)
3731 .kr(2)
3732 .sr(1)
3733 .m(3)
3734 .n(16)
3735 .k(8)
3736 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08003737 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003738 }
3739
3740 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, qmax) {
3741 TEST_REQUIRES_ARM_NEON;
3742 GemmMicrokernelTester()
3743 .mr(3)
3744 .nr(16)
3745 .kr(2)
3746 .sr(1)
3747 .m(3)
3748 .n(16)
3749 .k(8)
3750 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08003751 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003752 }
3753
3754 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, strided_cm) {
3755 TEST_REQUIRES_ARM_NEON;
3756 GemmMicrokernelTester()
3757 .mr(3)
3758 .nr(16)
3759 .kr(2)
3760 .sr(1)
3761 .m(3)
3762 .n(16)
3763 .k(8)
3764 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08003765 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003766 }
3767#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
3768
3769
3770#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barchard15eec022021-11-17 13:26:20 -08003771 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_eq_16) {
3772 TEST_REQUIRES_ARM_NEON;
3773 GemmMicrokernelTester()
3774 .mr(2)
3775 .nr(16)
3776 .kr(2)
3777 .sr(1)
3778 .m(2)
3779 .n(16)
3780 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -08003781 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003782 }
3783
3784 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, strided_cn) {
3785 TEST_REQUIRES_ARM_NEON;
3786 GemmMicrokernelTester()
3787 .mr(2)
3788 .nr(16)
3789 .kr(2)
3790 .sr(1)
3791 .m(2)
3792 .n(16)
3793 .k(16)
3794 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08003795 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003796 }
3797
3798 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_eq_16_subtile) {
3799 TEST_REQUIRES_ARM_NEON;
3800 for (uint32_t m = 1; m <= 2; m++) {
3801 for (uint32_t n = 1; n <= 16; n++) {
3802 GemmMicrokernelTester()
3803 .mr(2)
3804 .nr(16)
3805 .kr(2)
3806 .sr(1)
3807 .m(m)
3808 .n(n)
3809 .k(16)
3810 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003812 }
3813 }
3814 }
3815
3816 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_eq_16_subtile_m) {
3817 TEST_REQUIRES_ARM_NEON;
3818 for (uint32_t m = 1; m <= 2; m++) {
3819 GemmMicrokernelTester()
3820 .mr(2)
3821 .nr(16)
3822 .kr(2)
3823 .sr(1)
3824 .m(m)
3825 .n(16)
3826 .k(16)
3827 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003828 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003829 }
3830 }
3831
3832 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_eq_16_subtile_n) {
3833 TEST_REQUIRES_ARM_NEON;
3834 for (uint32_t n = 1; n <= 16; n++) {
3835 GemmMicrokernelTester()
3836 .mr(2)
3837 .nr(16)
3838 .kr(2)
3839 .sr(1)
3840 .m(2)
3841 .n(n)
3842 .k(16)
3843 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003844 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003845 }
3846 }
3847
3848 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_lt_16) {
3849 TEST_REQUIRES_ARM_NEON;
3850 for (size_t k = 1; k < 16; k++) {
3851 GemmMicrokernelTester()
3852 .mr(2)
3853 .nr(16)
3854 .kr(2)
3855 .sr(1)
3856 .m(2)
3857 .n(16)
3858 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08003859 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003860 }
3861 }
3862
3863 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_lt_16_subtile) {
3864 TEST_REQUIRES_ARM_NEON;
3865 for (size_t k = 1; k < 16; k++) {
3866 for (uint32_t m = 1; m <= 2; m++) {
3867 for (uint32_t n = 1; n <= 16; n++) {
3868 GemmMicrokernelTester()
3869 .mr(2)
3870 .nr(16)
3871 .kr(2)
3872 .sr(1)
3873 .m(m)
3874 .n(n)
3875 .k(k)
3876 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003877 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003878 }
3879 }
3880 }
3881 }
3882
3883 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_gt_16) {
3884 TEST_REQUIRES_ARM_NEON;
3885 for (size_t k = 17; k < 32; k++) {
3886 GemmMicrokernelTester()
3887 .mr(2)
3888 .nr(16)
3889 .kr(2)
3890 .sr(1)
3891 .m(2)
3892 .n(16)
3893 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08003894 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003895 }
3896 }
3897
3898 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_gt_16_subtile) {
3899 TEST_REQUIRES_ARM_NEON;
3900 for (size_t k = 17; k < 32; k++) {
3901 for (uint32_t m = 1; m <= 2; m++) {
3902 for (uint32_t n = 1; n <= 16; n++) {
3903 GemmMicrokernelTester()
3904 .mr(2)
3905 .nr(16)
3906 .kr(2)
3907 .sr(1)
3908 .m(m)
3909 .n(n)
3910 .k(k)
3911 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003912 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003913 }
3914 }
3915 }
3916 }
3917
3918 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_div_16) {
3919 TEST_REQUIRES_ARM_NEON;
3920 for (size_t k = 32; k <= 160; k += 16) {
3921 GemmMicrokernelTester()
3922 .mr(2)
3923 .nr(16)
3924 .kr(2)
3925 .sr(1)
3926 .m(2)
3927 .n(16)
3928 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08003929 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003930 }
3931 }
3932
3933 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, k_div_16_subtile) {
3934 TEST_REQUIRES_ARM_NEON;
3935 for (size_t k = 32; k <= 160; k += 16) {
3936 for (uint32_t m = 1; m <= 2; m++) {
3937 for (uint32_t n = 1; n <= 16; n++) {
3938 GemmMicrokernelTester()
3939 .mr(2)
3940 .nr(16)
3941 .kr(2)
3942 .sr(1)
3943 .m(m)
3944 .n(n)
3945 .k(k)
3946 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08003947 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003948 }
3949 }
3950 }
3951 }
3952
3953 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, n_gt_16) {
3954 TEST_REQUIRES_ARM_NEON;
3955 for (uint32_t n = 17; n < 32; n++) {
3956 for (size_t k = 1; k <= 80; k += 17) {
3957 GemmMicrokernelTester()
3958 .mr(2)
3959 .nr(16)
3960 .kr(2)
3961 .sr(1)
3962 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003963 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003964 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08003965 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003966 }
3967 }
3968 }
3969
3970 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, n_gt_16_strided_cn) {
3971 TEST_REQUIRES_ARM_NEON;
3972 for (uint32_t n = 17; n < 32; n++) {
3973 for (size_t k = 1; k <= 80; k += 17) {
3974 GemmMicrokernelTester()
3975 .mr(2)
3976 .nr(16)
3977 .kr(2)
3978 .sr(1)
3979 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08003980 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08003981 .k(k)
3982 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08003983 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08003984 }
3985 }
3986 }
3987
3988 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, n_gt_16_subtile) {
3989 TEST_REQUIRES_ARM_NEON;
3990 for (uint32_t n = 17; n < 32; n++) {
3991 for (size_t k = 1; k <= 80; k += 17) {
3992 for (uint32_t m = 1; m <= 2; m++) {
3993 GemmMicrokernelTester()
3994 .mr(2)
3995 .nr(16)
3996 .kr(2)
3997 .sr(1)
3998 .m(m)
3999 .n(n)
4000 .k(k)
4001 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004002 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004003 }
4004 }
4005 }
4006 }
4007
4008 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, n_div_16) {
4009 TEST_REQUIRES_ARM_NEON;
4010 for (uint32_t n = 32; n <= 48; n += 16) {
4011 for (size_t k = 1; k <= 80; k += 17) {
4012 GemmMicrokernelTester()
4013 .mr(2)
4014 .nr(16)
4015 .kr(2)
4016 .sr(1)
4017 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004018 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08004019 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004020 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004021 }
4022 }
4023 }
4024
4025 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, n_div_16_strided_cn) {
4026 TEST_REQUIRES_ARM_NEON;
4027 for (uint32_t n = 32; n <= 48; n += 16) {
4028 for (size_t k = 1; k <= 80; k += 17) {
4029 GemmMicrokernelTester()
4030 .mr(2)
4031 .nr(16)
4032 .kr(2)
4033 .sr(1)
4034 .m(2)
4035 .n(n)
4036 .k(k)
4037 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08004038 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004039 }
4040 }
4041 }
4042
4043 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, n_div_16_subtile) {
4044 TEST_REQUIRES_ARM_NEON;
4045 for (uint32_t n = 32; n <= 48; n += 16) {
4046 for (size_t k = 1; k <= 80; k += 17) {
4047 for (uint32_t m = 1; m <= 2; m++) {
4048 GemmMicrokernelTester()
4049 .mr(2)
4050 .nr(16)
4051 .kr(2)
4052 .sr(1)
4053 .m(m)
4054 .n(n)
4055 .k(k)
4056 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004057 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004058 }
4059 }
4060 }
4061 }
4062
4063 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, small_kernel) {
4064 TEST_REQUIRES_ARM_NEON;
4065 for (size_t k = 1; k <= 80; k += 17) {
4066 GemmMicrokernelTester()
4067 .mr(2)
4068 .nr(16)
4069 .kr(2)
4070 .sr(1)
4071 .m(2)
4072 .n(16)
4073 .k(k)
4074 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08004075 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004076 }
4077 }
4078
4079 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, small_kernel_subtile) {
4080 TEST_REQUIRES_ARM_NEON;
4081 for (size_t k = 1; k <= 80; k += 17) {
4082 for (uint32_t m = 1; m <= 2; m++) {
4083 for (uint32_t n = 1; n <= 16; n++) {
4084 GemmMicrokernelTester()
4085 .mr(2)
4086 .nr(16)
4087 .kr(2)
4088 .sr(1)
4089 .m(m)
4090 .n(n)
4091 .k(k)
4092 .ks(3)
4093 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004094 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004095 }
4096 }
4097 }
4098 }
4099
4100 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, n_gt_16_small_kernel) {
4101 TEST_REQUIRES_ARM_NEON;
4102 for (uint32_t n = 17; n < 32; n++) {
4103 for (size_t k = 1; k <= 80; k += 17) {
4104 GemmMicrokernelTester()
4105 .mr(2)
4106 .nr(16)
4107 .kr(2)
4108 .sr(1)
4109 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004110 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08004111 .k(k)
4112 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08004113 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004114 }
4115 }
4116 }
4117
4118 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, n_div_16_small_kernel) {
4119 TEST_REQUIRES_ARM_NEON;
4120 for (uint32_t n = 32; n <= 48; n += 16) {
4121 for (size_t k = 1; k <= 80; k += 17) {
4122 GemmMicrokernelTester()
4123 .mr(2)
4124 .nr(16)
4125 .kr(2)
4126 .sr(1)
4127 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004128 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08004129 .k(k)
4130 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08004131 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004132 }
4133 }
4134 }
4135
4136 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, strided_cm_subtile) {
4137 TEST_REQUIRES_ARM_NEON;
4138 for (size_t k = 1; k <= 80; k += 17) {
4139 for (uint32_t m = 1; m <= 2; m++) {
4140 for (uint32_t n = 1; n <= 16; n++) {
4141 GemmMicrokernelTester()
4142 .mr(2)
4143 .nr(16)
4144 .kr(2)
4145 .sr(1)
4146 .m(m)
4147 .n(n)
4148 .k(k)
4149 .cm_stride(19)
4150 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004151 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004152 }
4153 }
4154 }
4155 }
4156
4157 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, a_offset) {
4158 TEST_REQUIRES_ARM_NEON;
4159 for (size_t k = 1; k <= 80; k += 17) {
4160 GemmMicrokernelTester()
4161 .mr(2)
4162 .nr(16)
4163 .kr(2)
4164 .sr(1)
4165 .m(2)
4166 .n(16)
4167 .k(k)
4168 .ks(3)
4169 .a_offset(163)
Marat Dukhan50323b82022-01-11 00:12:01 -08004170 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004171 }
4172 }
4173
4174 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, zero) {
4175 TEST_REQUIRES_ARM_NEON;
4176 for (uint32_t mz = 0; mz < 2; mz++) {
4177 for (size_t k = 1; k <= 80; k += 17) {
4178 GemmMicrokernelTester()
4179 .mr(2)
4180 .nr(16)
4181 .kr(2)
4182 .sr(1)
4183 .m(2)
4184 .n(16)
4185 .k(k)
4186 .ks(3)
4187 .a_offset(163)
4188 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08004189 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004190 }
4191 }
4192 }
4193
4194 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, qmin) {
4195 TEST_REQUIRES_ARM_NEON;
4196 GemmMicrokernelTester()
4197 .mr(2)
4198 .nr(16)
4199 .kr(2)
4200 .sr(1)
4201 .m(2)
4202 .n(16)
4203 .k(16)
4204 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08004205 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004206 }
4207
4208 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, qmax) {
4209 TEST_REQUIRES_ARM_NEON;
4210 GemmMicrokernelTester()
4211 .mr(2)
4212 .nr(16)
4213 .kr(2)
4214 .sr(1)
4215 .m(2)
4216 .n(16)
4217 .k(16)
4218 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08004219 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004220 }
4221
4222 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MLAL_LD2R, strided_cm) {
4223 TEST_REQUIRES_ARM_NEON;
4224 GemmMicrokernelTester()
4225 .mr(2)
4226 .nr(16)
4227 .kr(2)
4228 .sr(1)
4229 .m(2)
4230 .n(16)
4231 .k(16)
4232 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08004233 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004234 }
4235#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
4236
4237
4238#if XNN_ARCH_ARM || XNN_ARCH_ARM64
4239 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_eq_16) {
4240 TEST_REQUIRES_ARM_NEON;
4241 GemmMicrokernelTester()
4242 .mr(3)
4243 .nr(16)
4244 .kr(2)
4245 .sr(1)
4246 .m(3)
4247 .n(16)
4248 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -08004249 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004250 }
4251
4252 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, strided_cn) {
4253 TEST_REQUIRES_ARM_NEON;
4254 GemmMicrokernelTester()
4255 .mr(3)
4256 .nr(16)
4257 .kr(2)
4258 .sr(1)
4259 .m(3)
4260 .n(16)
4261 .k(16)
4262 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08004263 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004264 }
4265
4266 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_eq_16_subtile) {
4267 TEST_REQUIRES_ARM_NEON;
4268 for (uint32_t m = 1; m <= 3; m++) {
4269 for (uint32_t n = 1; n <= 16; n++) {
4270 GemmMicrokernelTester()
4271 .mr(3)
4272 .nr(16)
4273 .kr(2)
4274 .sr(1)
4275 .m(m)
4276 .n(n)
4277 .k(16)
4278 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004280 }
4281 }
4282 }
4283
4284 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_eq_16_subtile_m) {
4285 TEST_REQUIRES_ARM_NEON;
4286 for (uint32_t m = 1; m <= 3; m++) {
4287 GemmMicrokernelTester()
4288 .mr(3)
4289 .nr(16)
4290 .kr(2)
4291 .sr(1)
4292 .m(m)
4293 .n(16)
4294 .k(16)
4295 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004296 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004297 }
4298 }
4299
4300 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_eq_16_subtile_n) {
4301 TEST_REQUIRES_ARM_NEON;
4302 for (uint32_t n = 1; n <= 16; n++) {
4303 GemmMicrokernelTester()
4304 .mr(3)
4305 .nr(16)
4306 .kr(2)
4307 .sr(1)
4308 .m(3)
4309 .n(n)
4310 .k(16)
4311 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004312 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004313 }
4314 }
4315
4316 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_lt_16) {
4317 TEST_REQUIRES_ARM_NEON;
4318 for (size_t k = 1; k < 16; k++) {
4319 GemmMicrokernelTester()
4320 .mr(3)
4321 .nr(16)
4322 .kr(2)
4323 .sr(1)
4324 .m(3)
4325 .n(16)
4326 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004327 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004328 }
4329 }
4330
4331 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_lt_16_subtile) {
4332 TEST_REQUIRES_ARM_NEON;
4333 for (size_t k = 1; k < 16; k++) {
4334 for (uint32_t m = 1; m <= 3; m++) {
4335 for (uint32_t n = 1; n <= 16; n++) {
4336 GemmMicrokernelTester()
4337 .mr(3)
4338 .nr(16)
4339 .kr(2)
4340 .sr(1)
4341 .m(m)
4342 .n(n)
4343 .k(k)
4344 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004345 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004346 }
4347 }
4348 }
4349 }
4350
4351 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_gt_16) {
4352 TEST_REQUIRES_ARM_NEON;
4353 for (size_t k = 17; k < 32; k++) {
4354 GemmMicrokernelTester()
4355 .mr(3)
4356 .nr(16)
4357 .kr(2)
4358 .sr(1)
4359 .m(3)
4360 .n(16)
4361 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004362 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004363 }
4364 }
4365
4366 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_gt_16_subtile) {
4367 TEST_REQUIRES_ARM_NEON;
4368 for (size_t k = 17; k < 32; k++) {
4369 for (uint32_t m = 1; m <= 3; m++) {
4370 for (uint32_t n = 1; n <= 16; n++) {
4371 GemmMicrokernelTester()
4372 .mr(3)
4373 .nr(16)
4374 .kr(2)
4375 .sr(1)
4376 .m(m)
4377 .n(n)
4378 .k(k)
4379 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004380 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004381 }
4382 }
4383 }
4384 }
4385
4386 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_div_16) {
4387 TEST_REQUIRES_ARM_NEON;
4388 for (size_t k = 32; k <= 160; k += 16) {
4389 GemmMicrokernelTester()
4390 .mr(3)
4391 .nr(16)
4392 .kr(2)
4393 .sr(1)
4394 .m(3)
4395 .n(16)
4396 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004397 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004398 }
4399 }
4400
4401 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, k_div_16_subtile) {
4402 TEST_REQUIRES_ARM_NEON;
4403 for (size_t k = 32; k <= 160; k += 16) {
4404 for (uint32_t m = 1; m <= 3; m++) {
4405 for (uint32_t n = 1; n <= 16; n++) {
4406 GemmMicrokernelTester()
4407 .mr(3)
4408 .nr(16)
4409 .kr(2)
4410 .sr(1)
4411 .m(m)
4412 .n(n)
4413 .k(k)
4414 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004415 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004416 }
4417 }
4418 }
4419 }
4420
4421 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, n_gt_16) {
4422 TEST_REQUIRES_ARM_NEON;
4423 for (uint32_t n = 17; n < 32; n++) {
4424 for (size_t k = 1; k <= 80; k += 17) {
4425 GemmMicrokernelTester()
4426 .mr(3)
4427 .nr(16)
4428 .kr(2)
4429 .sr(1)
4430 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004431 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08004432 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004433 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004434 }
4435 }
4436 }
4437
4438 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, n_gt_16_strided_cn) {
4439 TEST_REQUIRES_ARM_NEON;
4440 for (uint32_t n = 17; n < 32; n++) {
4441 for (size_t k = 1; k <= 80; k += 17) {
4442 GemmMicrokernelTester()
4443 .mr(3)
4444 .nr(16)
4445 .kr(2)
4446 .sr(1)
4447 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004448 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08004449 .k(k)
4450 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08004451 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004452 }
4453 }
4454 }
4455
4456 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, n_gt_16_subtile) {
4457 TEST_REQUIRES_ARM_NEON;
4458 for (uint32_t n = 17; n < 32; n++) {
4459 for (size_t k = 1; k <= 80; k += 17) {
4460 for (uint32_t m = 1; m <= 3; m++) {
4461 GemmMicrokernelTester()
4462 .mr(3)
4463 .nr(16)
4464 .kr(2)
4465 .sr(1)
4466 .m(m)
4467 .n(n)
4468 .k(k)
4469 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004470 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004471 }
4472 }
4473 }
4474 }
4475
4476 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, n_div_16) {
4477 TEST_REQUIRES_ARM_NEON;
4478 for (uint32_t n = 32; n <= 48; n += 16) {
4479 for (size_t k = 1; k <= 80; k += 17) {
4480 GemmMicrokernelTester()
4481 .mr(3)
4482 .nr(16)
4483 .kr(2)
4484 .sr(1)
4485 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004486 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08004487 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004488 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004489 }
4490 }
4491 }
4492
4493 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, n_div_16_strided_cn) {
4494 TEST_REQUIRES_ARM_NEON;
4495 for (uint32_t n = 32; n <= 48; n += 16) {
4496 for (size_t k = 1; k <= 80; k += 17) {
4497 GemmMicrokernelTester()
4498 .mr(3)
4499 .nr(16)
4500 .kr(2)
4501 .sr(1)
4502 .m(3)
4503 .n(n)
4504 .k(k)
4505 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08004506 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004507 }
4508 }
4509 }
4510
4511 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, n_div_16_subtile) {
4512 TEST_REQUIRES_ARM_NEON;
4513 for (uint32_t n = 32; n <= 48; n += 16) {
4514 for (size_t k = 1; k <= 80; k += 17) {
4515 for (uint32_t m = 1; m <= 3; m++) {
4516 GemmMicrokernelTester()
4517 .mr(3)
4518 .nr(16)
4519 .kr(2)
4520 .sr(1)
4521 .m(m)
4522 .n(n)
4523 .k(k)
4524 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004525 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004526 }
4527 }
4528 }
4529 }
4530
4531 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, small_kernel) {
4532 TEST_REQUIRES_ARM_NEON;
4533 for (size_t k = 1; k <= 80; k += 17) {
4534 GemmMicrokernelTester()
4535 .mr(3)
4536 .nr(16)
4537 .kr(2)
4538 .sr(1)
4539 .m(3)
4540 .n(16)
4541 .k(k)
4542 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08004543 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004544 }
4545 }
4546
4547 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, small_kernel_subtile) {
4548 TEST_REQUIRES_ARM_NEON;
4549 for (size_t k = 1; k <= 80; k += 17) {
4550 for (uint32_t m = 1; m <= 3; m++) {
4551 for (uint32_t n = 1; n <= 16; n++) {
4552 GemmMicrokernelTester()
4553 .mr(3)
4554 .nr(16)
4555 .kr(2)
4556 .sr(1)
4557 .m(m)
4558 .n(n)
4559 .k(k)
4560 .ks(3)
4561 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004562 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004563 }
4564 }
4565 }
4566 }
4567
4568 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, n_gt_16_small_kernel) {
4569 TEST_REQUIRES_ARM_NEON;
4570 for (uint32_t n = 17; n < 32; n++) {
4571 for (size_t k = 1; k <= 80; k += 17) {
4572 GemmMicrokernelTester()
4573 .mr(3)
4574 .nr(16)
4575 .kr(2)
4576 .sr(1)
4577 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004578 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08004579 .k(k)
4580 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08004581 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004582 }
4583 }
4584 }
4585
4586 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, n_div_16_small_kernel) {
4587 TEST_REQUIRES_ARM_NEON;
4588 for (uint32_t n = 32; n <= 48; n += 16) {
4589 for (size_t k = 1; k <= 80; k += 17) {
4590 GemmMicrokernelTester()
4591 .mr(3)
4592 .nr(16)
4593 .kr(2)
4594 .sr(1)
4595 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004596 .n(n)
Frank Barchard15eec022021-11-17 13:26:20 -08004597 .k(k)
4598 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08004599 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004600 }
4601 }
4602 }
4603
4604 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, strided_cm_subtile) {
4605 TEST_REQUIRES_ARM_NEON;
4606 for (size_t k = 1; k <= 80; k += 17) {
4607 for (uint32_t m = 1; m <= 3; m++) {
4608 for (uint32_t n = 1; n <= 16; n++) {
4609 GemmMicrokernelTester()
4610 .mr(3)
4611 .nr(16)
4612 .kr(2)
4613 .sr(1)
4614 .m(m)
4615 .n(n)
4616 .k(k)
4617 .cm_stride(19)
4618 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004619 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004620 }
4621 }
4622 }
4623 }
4624
4625 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, a_offset) {
4626 TEST_REQUIRES_ARM_NEON;
4627 for (size_t k = 1; k <= 80; k += 17) {
4628 GemmMicrokernelTester()
4629 .mr(3)
4630 .nr(16)
4631 .kr(2)
4632 .sr(1)
4633 .m(3)
4634 .n(16)
4635 .k(k)
4636 .ks(3)
4637 .a_offset(251)
Marat Dukhan50323b82022-01-11 00:12:01 -08004638 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004639 }
4640 }
4641
4642 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, zero) {
4643 TEST_REQUIRES_ARM_NEON;
4644 for (uint32_t mz = 0; mz < 3; mz++) {
4645 for (size_t k = 1; k <= 80; k += 17) {
4646 GemmMicrokernelTester()
4647 .mr(3)
4648 .nr(16)
4649 .kr(2)
4650 .sr(1)
4651 .m(3)
4652 .n(16)
4653 .k(k)
4654 .ks(3)
4655 .a_offset(251)
4656 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08004657 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004658 }
4659 }
4660 }
4661
4662 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, qmin) {
4663 TEST_REQUIRES_ARM_NEON;
4664 GemmMicrokernelTester()
4665 .mr(3)
4666 .nr(16)
4667 .kr(2)
4668 .sr(1)
4669 .m(3)
4670 .n(16)
4671 .k(16)
4672 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08004673 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004674 }
4675
4676 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, qmax) {
4677 TEST_REQUIRES_ARM_NEON;
4678 GemmMicrokernelTester()
4679 .mr(3)
4680 .nr(16)
4681 .kr(2)
4682 .sr(1)
4683 .m(3)
4684 .n(16)
4685 .k(16)
4686 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08004687 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004688 }
4689
4690 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MLAL_LD2R, strided_cm) {
4691 TEST_REQUIRES_ARM_NEON;
4692 GemmMicrokernelTester()
4693 .mr(3)
4694 .nr(16)
4695 .kr(2)
4696 .sr(1)
4697 .m(3)
4698 .n(16)
4699 .k(16)
4700 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08004701 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard15eec022021-11-17 13:26:20 -08004702 }
4703#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
4704
4705
4706#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barchard42f5c502021-11-16 10:04:21 -08004707 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_eq_8) {
4708 TEST_REQUIRES_ARM_NEON;
4709 GemmMicrokernelTester()
4710 .mr(1)
4711 .nr(8)
4712 .kr(2)
4713 .sr(1)
4714 .m(1)
4715 .n(8)
4716 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -08004717 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004718 }
4719
4720 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, strided_cn) {
4721 TEST_REQUIRES_ARM_NEON;
4722 GemmMicrokernelTester()
4723 .mr(1)
4724 .nr(8)
4725 .kr(2)
4726 .sr(1)
4727 .m(1)
4728 .n(8)
4729 .k(8)
4730 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08004731 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004732 }
4733
4734 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_eq_8_subtile) {
4735 TEST_REQUIRES_ARM_NEON;
4736 for (uint32_t m = 1; m <= 1; m++) {
4737 for (uint32_t n = 1; n <= 8; n++) {
4738 GemmMicrokernelTester()
4739 .mr(1)
4740 .nr(8)
4741 .kr(2)
4742 .sr(1)
4743 .m(m)
4744 .n(n)
4745 .k(8)
4746 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004748 }
4749 }
4750 }
4751
4752 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_eq_8_subtile_m) {
4753 TEST_REQUIRES_ARM_NEON;
4754 for (uint32_t m = 1; m <= 1; m++) {
4755 GemmMicrokernelTester()
4756 .mr(1)
4757 .nr(8)
4758 .kr(2)
4759 .sr(1)
4760 .m(m)
4761 .n(8)
4762 .k(8)
4763 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004764 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004765 }
4766 }
4767
4768 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_eq_8_subtile_n) {
4769 TEST_REQUIRES_ARM_NEON;
4770 for (uint32_t n = 1; n <= 8; n++) {
4771 GemmMicrokernelTester()
4772 .mr(1)
4773 .nr(8)
4774 .kr(2)
4775 .sr(1)
4776 .m(1)
4777 .n(n)
4778 .k(8)
4779 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004780 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004781 }
4782 }
4783
4784 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_lt_8) {
4785 TEST_REQUIRES_ARM_NEON;
4786 for (size_t k = 1; k < 8; k++) {
4787 GemmMicrokernelTester()
4788 .mr(1)
4789 .nr(8)
4790 .kr(2)
4791 .sr(1)
4792 .m(1)
4793 .n(8)
4794 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004795 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004796 }
4797 }
4798
4799 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_lt_8_subtile) {
4800 TEST_REQUIRES_ARM_NEON;
4801 for (size_t k = 1; k < 8; k++) {
4802 for (uint32_t m = 1; m <= 1; m++) {
4803 for (uint32_t n = 1; n <= 8; n++) {
4804 GemmMicrokernelTester()
4805 .mr(1)
4806 .nr(8)
4807 .kr(2)
4808 .sr(1)
4809 .m(m)
4810 .n(n)
4811 .k(k)
4812 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004813 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004814 }
4815 }
4816 }
4817 }
4818
4819 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_gt_8) {
4820 TEST_REQUIRES_ARM_NEON;
4821 for (size_t k = 9; k < 16; k++) {
4822 GemmMicrokernelTester()
4823 .mr(1)
4824 .nr(8)
4825 .kr(2)
4826 .sr(1)
4827 .m(1)
4828 .n(8)
4829 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004830 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004831 }
4832 }
4833
4834 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_gt_8_subtile) {
4835 TEST_REQUIRES_ARM_NEON;
4836 for (size_t k = 9; k < 16; k++) {
4837 for (uint32_t m = 1; m <= 1; m++) {
4838 for (uint32_t n = 1; n <= 8; n++) {
4839 GemmMicrokernelTester()
4840 .mr(1)
4841 .nr(8)
4842 .kr(2)
4843 .sr(1)
4844 .m(m)
4845 .n(n)
4846 .k(k)
4847 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004848 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004849 }
4850 }
4851 }
4852 }
4853
4854 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_div_8) {
4855 TEST_REQUIRES_ARM_NEON;
4856 for (size_t k = 16; k <= 80; k += 8) {
4857 GemmMicrokernelTester()
4858 .mr(1)
4859 .nr(8)
4860 .kr(2)
4861 .sr(1)
4862 .m(1)
4863 .n(8)
4864 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004865 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004866 }
4867 }
4868
4869 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, k_div_8_subtile) {
4870 TEST_REQUIRES_ARM_NEON;
4871 for (size_t k = 16; k <= 80; k += 8) {
4872 for (uint32_t m = 1; m <= 1; m++) {
4873 for (uint32_t n = 1; n <= 8; n++) {
4874 GemmMicrokernelTester()
4875 .mr(1)
4876 .nr(8)
4877 .kr(2)
4878 .sr(1)
4879 .m(m)
4880 .n(n)
4881 .k(k)
4882 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004883 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004884 }
4885 }
4886 }
4887 }
4888
4889 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, n_gt_8) {
4890 TEST_REQUIRES_ARM_NEON;
4891 for (uint32_t n = 9; n < 16; n++) {
4892 for (size_t k = 1; k <= 40; k += 9) {
4893 GemmMicrokernelTester()
4894 .mr(1)
4895 .nr(8)
4896 .kr(2)
4897 .sr(1)
4898 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004899 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08004900 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004901 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004902 }
4903 }
4904 }
4905
4906 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, n_gt_8_strided_cn) {
4907 TEST_REQUIRES_ARM_NEON;
4908 for (uint32_t n = 9; n < 16; n++) {
4909 for (size_t k = 1; k <= 40; k += 9) {
4910 GemmMicrokernelTester()
4911 .mr(1)
4912 .nr(8)
4913 .kr(2)
4914 .sr(1)
4915 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004916 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08004917 .k(k)
4918 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08004919 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004920 }
4921 }
4922 }
4923
4924 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, n_gt_8_subtile) {
4925 TEST_REQUIRES_ARM_NEON;
4926 for (uint32_t n = 9; n < 16; n++) {
4927 for (size_t k = 1; k <= 40; k += 9) {
4928 for (uint32_t m = 1; m <= 1; m++) {
4929 GemmMicrokernelTester()
4930 .mr(1)
4931 .nr(8)
4932 .kr(2)
4933 .sr(1)
4934 .m(m)
4935 .n(n)
4936 .k(k)
4937 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004938 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004939 }
4940 }
4941 }
4942 }
4943
4944 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, n_div_8) {
4945 TEST_REQUIRES_ARM_NEON;
4946 for (uint32_t n = 16; n <= 24; n += 8) {
4947 for (size_t k = 1; k <= 40; k += 9) {
4948 GemmMicrokernelTester()
4949 .mr(1)
4950 .nr(8)
4951 .kr(2)
4952 .sr(1)
4953 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08004954 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08004955 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08004956 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004957 }
4958 }
4959 }
4960
4961 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, n_div_8_strided_cn) {
4962 TEST_REQUIRES_ARM_NEON;
4963 for (uint32_t n = 16; n <= 24; n += 8) {
4964 for (size_t k = 1; k <= 40; k += 9) {
4965 GemmMicrokernelTester()
4966 .mr(1)
4967 .nr(8)
4968 .kr(2)
4969 .sr(1)
4970 .m(1)
4971 .n(n)
4972 .k(k)
4973 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08004974 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004975 }
4976 }
4977 }
4978
4979 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, n_div_8_subtile) {
4980 TEST_REQUIRES_ARM_NEON;
4981 for (uint32_t n = 16; n <= 24; n += 8) {
4982 for (size_t k = 1; k <= 40; k += 9) {
4983 for (uint32_t m = 1; m <= 1; m++) {
4984 GemmMicrokernelTester()
4985 .mr(1)
4986 .nr(8)
4987 .kr(2)
4988 .sr(1)
4989 .m(m)
4990 .n(n)
4991 .k(k)
4992 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08004993 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08004994 }
4995 }
4996 }
4997 }
4998
4999 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, small_kernel) {
5000 TEST_REQUIRES_ARM_NEON;
5001 for (size_t k = 1; k <= 40; k += 9) {
5002 GemmMicrokernelTester()
5003 .mr(1)
5004 .nr(8)
5005 .kr(2)
5006 .sr(1)
5007 .m(1)
5008 .n(8)
5009 .k(k)
5010 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08005011 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005012 }
5013 }
5014
5015 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, small_kernel_subtile) {
5016 TEST_REQUIRES_ARM_NEON;
5017 for (size_t k = 1; k <= 40; k += 9) {
5018 for (uint32_t m = 1; m <= 1; m++) {
5019 for (uint32_t n = 1; n <= 8; n++) {
5020 GemmMicrokernelTester()
5021 .mr(1)
5022 .nr(8)
5023 .kr(2)
5024 .sr(1)
5025 .m(m)
5026 .n(n)
5027 .k(k)
5028 .ks(3)
5029 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005030 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005031 }
5032 }
5033 }
5034 }
5035
5036 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, n_gt_8_small_kernel) {
5037 TEST_REQUIRES_ARM_NEON;
5038 for (uint32_t n = 9; n < 16; n++) {
5039 for (size_t k = 1; k <= 40; k += 9) {
5040 GemmMicrokernelTester()
5041 .mr(1)
5042 .nr(8)
5043 .kr(2)
5044 .sr(1)
5045 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005046 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08005047 .k(k)
5048 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08005049 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005050 }
5051 }
5052 }
5053
5054 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, n_div_8_small_kernel) {
5055 TEST_REQUIRES_ARM_NEON;
5056 for (uint32_t n = 16; n <= 24; n += 8) {
5057 for (size_t k = 1; k <= 40; k += 9) {
5058 GemmMicrokernelTester()
5059 .mr(1)
5060 .nr(8)
5061 .kr(2)
5062 .sr(1)
5063 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005064 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08005065 .k(k)
5066 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08005067 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005068 }
5069 }
5070 }
5071
5072 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, strided_cm_subtile) {
5073 TEST_REQUIRES_ARM_NEON;
5074 for (size_t k = 1; k <= 40; k += 9) {
5075 for (uint32_t m = 1; m <= 1; m++) {
5076 for (uint32_t n = 1; n <= 8; n++) {
5077 GemmMicrokernelTester()
5078 .mr(1)
5079 .nr(8)
5080 .kr(2)
5081 .sr(1)
5082 .m(m)
5083 .n(n)
5084 .k(k)
5085 .cm_stride(11)
5086 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005087 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005088 }
5089 }
5090 }
5091 }
5092
5093 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, a_offset) {
5094 TEST_REQUIRES_ARM_NEON;
5095 for (size_t k = 1; k <= 40; k += 9) {
5096 GemmMicrokernelTester()
5097 .mr(1)
5098 .nr(8)
5099 .kr(2)
5100 .sr(1)
5101 .m(1)
5102 .n(8)
5103 .k(k)
5104 .ks(3)
5105 .a_offset(43)
Marat Dukhan50323b82022-01-11 00:12:01 -08005106 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005107 }
5108 }
5109
5110 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, zero) {
5111 TEST_REQUIRES_ARM_NEON;
5112 for (uint32_t mz = 0; mz < 1; mz++) {
5113 for (size_t k = 1; k <= 40; k += 9) {
5114 GemmMicrokernelTester()
5115 .mr(1)
5116 .nr(8)
5117 .kr(2)
5118 .sr(1)
5119 .m(1)
5120 .n(8)
5121 .k(k)
5122 .ks(3)
5123 .a_offset(43)
5124 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08005125 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005126 }
5127 }
5128 }
5129
5130 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, qmin) {
5131 TEST_REQUIRES_ARM_NEON;
5132 GemmMicrokernelTester()
5133 .mr(1)
5134 .nr(8)
5135 .kr(2)
5136 .sr(1)
5137 .m(1)
5138 .n(8)
5139 .k(8)
5140 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08005141 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005142 }
5143
5144 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, qmax) {
5145 TEST_REQUIRES_ARM_NEON;
5146 GemmMicrokernelTester()
5147 .mr(1)
5148 .nr(8)
5149 .kr(2)
5150 .sr(1)
5151 .m(1)
5152 .n(8)
5153 .k(8)
5154 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08005155 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005156 }
5157
5158 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD4R, strided_cm) {
5159 TEST_REQUIRES_ARM_NEON;
5160 GemmMicrokernelTester()
5161 .mr(1)
5162 .nr(8)
5163 .kr(2)
5164 .sr(1)
5165 .m(1)
5166 .n(8)
5167 .k(8)
5168 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08005169 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005170 }
5171#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
5172
5173
5174#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -08005175 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_eq_8) {
5176 TEST_REQUIRES_ARM_NEON;
5177 GemmMicrokernelTester()
5178 .mr(2)
5179 .nr(8)
5180 .kr(2)
5181 .sr(1)
5182 .m(2)
5183 .n(8)
5184 .k(8)
5185 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5186 }
5187
5188 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, strided_cn) {
5189 TEST_REQUIRES_ARM_NEON;
5190 GemmMicrokernelTester()
5191 .mr(2)
5192 .nr(8)
5193 .kr(2)
5194 .sr(1)
5195 .m(2)
5196 .n(8)
5197 .k(8)
5198 .cn_stride(11)
5199 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5200 }
5201
5202 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_eq_8_subtile) {
5203 TEST_REQUIRES_ARM_NEON;
5204 for (uint32_t m = 1; m <= 2; m++) {
5205 for (uint32_t n = 1; n <= 8; n++) {
5206 GemmMicrokernelTester()
5207 .mr(2)
5208 .nr(8)
5209 .kr(2)
5210 .sr(1)
5211 .m(m)
5212 .n(n)
5213 .k(8)
5214 .iterations(1)
5215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5216 }
5217 }
5218 }
5219
5220 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_eq_8_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(2)
5227 .sr(1)
5228 .m(m)
5229 .n(8)
5230 .k(8)
5231 .iterations(1)
5232 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5233 }
5234 }
5235
5236 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_eq_8_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(2)
5243 .sr(1)
5244 .m(2)
5245 .n(n)
5246 .k(8)
5247 .iterations(1)
5248 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5249 }
5250 }
5251
5252 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_lt_8) {
5253 TEST_REQUIRES_ARM_NEON;
5254 for (size_t k = 1; k < 8; k++) {
5255 GemmMicrokernelTester()
5256 .mr(2)
5257 .nr(8)
5258 .kr(2)
5259 .sr(1)
5260 .m(2)
5261 .n(8)
5262 .k(k)
5263 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5264 }
5265 }
5266
5267 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_lt_8_subtile) {
5268 TEST_REQUIRES_ARM_NEON;
5269 for (size_t k = 1; k < 8; k++) {
5270 for (uint32_t m = 1; m <= 2; m++) {
5271 for (uint32_t n = 1; n <= 8; n++) {
5272 GemmMicrokernelTester()
5273 .mr(2)
5274 .nr(8)
5275 .kr(2)
5276 .sr(1)
5277 .m(m)
5278 .n(n)
5279 .k(k)
5280 .iterations(1)
5281 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5282 }
5283 }
5284 }
5285 }
5286
5287 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_gt_8) {
5288 TEST_REQUIRES_ARM_NEON;
5289 for (size_t k = 9; k < 16; k++) {
5290 GemmMicrokernelTester()
5291 .mr(2)
5292 .nr(8)
5293 .kr(2)
5294 .sr(1)
5295 .m(2)
5296 .n(8)
5297 .k(k)
5298 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5299 }
5300 }
5301
5302 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_gt_8_subtile) {
5303 TEST_REQUIRES_ARM_NEON;
5304 for (size_t k = 9; k < 16; k++) {
5305 for (uint32_t m = 1; m <= 2; m++) {
5306 for (uint32_t n = 1; n <= 8; n++) {
5307 GemmMicrokernelTester()
5308 .mr(2)
5309 .nr(8)
5310 .kr(2)
5311 .sr(1)
5312 .m(m)
5313 .n(n)
5314 .k(k)
5315 .iterations(1)
5316 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5317 }
5318 }
5319 }
5320 }
5321
5322 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_div_8) {
5323 TEST_REQUIRES_ARM_NEON;
5324 for (size_t k = 16; k <= 80; k += 8) {
5325 GemmMicrokernelTester()
5326 .mr(2)
5327 .nr(8)
5328 .kr(2)
5329 .sr(1)
5330 .m(2)
5331 .n(8)
5332 .k(k)
5333 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5334 }
5335 }
5336
5337 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, k_div_8_subtile) {
5338 TEST_REQUIRES_ARM_NEON;
5339 for (size_t k = 16; k <= 80; k += 8) {
5340 for (uint32_t m = 1; m <= 2; m++) {
5341 for (uint32_t n = 1; n <= 8; n++) {
5342 GemmMicrokernelTester()
5343 .mr(2)
5344 .nr(8)
5345 .kr(2)
5346 .sr(1)
5347 .m(m)
5348 .n(n)
5349 .k(k)
5350 .iterations(1)
5351 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5352 }
5353 }
5354 }
5355 }
5356
5357 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, n_gt_8) {
5358 TEST_REQUIRES_ARM_NEON;
5359 for (uint32_t n = 9; n < 16; n++) {
5360 for (size_t k = 1; k <= 40; k += 9) {
5361 GemmMicrokernelTester()
5362 .mr(2)
5363 .nr(8)
5364 .kr(2)
5365 .sr(1)
5366 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005367 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08005368 .k(k)
5369 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5370 }
5371 }
5372 }
5373
5374 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5378 GemmMicrokernelTester()
5379 .mr(2)
5380 .nr(8)
5381 .kr(2)
5382 .sr(1)
5383 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005384 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08005385 .k(k)
5386 .cn_stride(11)
5387 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5388 }
5389 }
5390 }
5391
5392 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5396 for (uint32_t m = 1; m <= 2; m++) {
5397 GemmMicrokernelTester()
5398 .mr(2)
5399 .nr(8)
5400 .kr(2)
5401 .sr(1)
5402 .m(m)
5403 .n(n)
5404 .k(k)
5405 .iterations(1)
5406 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5407 }
5408 }
5409 }
5410 }
5411
5412 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5416 GemmMicrokernelTester()
5417 .mr(2)
5418 .nr(8)
5419 .kr(2)
5420 .sr(1)
5421 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005422 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08005423 .k(k)
5424 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5425 }
5426 }
5427 }
5428
5429 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5433 GemmMicrokernelTester()
5434 .mr(2)
5435 .nr(8)
5436 .kr(2)
5437 .sr(1)
5438 .m(2)
5439 .n(n)
5440 .k(k)
5441 .cn_stride(11)
5442 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5443 }
5444 }
5445 }
5446
5447 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5451 for (uint32_t m = 1; m <= 2; m++) {
5452 GemmMicrokernelTester()
5453 .mr(2)
5454 .nr(8)
5455 .kr(2)
5456 .sr(1)
5457 .m(m)
5458 .n(n)
5459 .k(k)
5460 .iterations(1)
5461 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5462 }
5463 }
5464 }
5465 }
5466
5467 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, small_kernel) {
5468 TEST_REQUIRES_ARM_NEON;
5469 for (size_t k = 1; k <= 40; k += 9) {
5470 GemmMicrokernelTester()
5471 .mr(2)
5472 .nr(8)
5473 .kr(2)
5474 .sr(1)
5475 .m(2)
5476 .n(8)
5477 .k(k)
5478 .ks(3)
5479 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5480 }
5481 }
5482
5483 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, small_kernel_subtile) {
5484 TEST_REQUIRES_ARM_NEON;
5485 for (size_t k = 1; k <= 40; k += 9) {
5486 for (uint32_t m = 1; m <= 2; m++) {
5487 for (uint32_t n = 1; n <= 8; n++) {
5488 GemmMicrokernelTester()
5489 .mr(2)
5490 .nr(8)
5491 .kr(2)
5492 .sr(1)
5493 .m(m)
5494 .n(n)
5495 .k(k)
5496 .ks(3)
5497 .iterations(1)
5498 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5499 }
5500 }
5501 }
5502 }
5503
5504 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5508 GemmMicrokernelTester()
5509 .mr(2)
5510 .nr(8)
5511 .kr(2)
5512 .sr(1)
5513 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005514 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08005515 .k(k)
5516 .ks(3)
5517 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5518 }
5519 }
5520 }
5521
5522 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5526 GemmMicrokernelTester()
5527 .mr(2)
5528 .nr(8)
5529 .kr(2)
5530 .sr(1)
5531 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005532 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08005533 .k(k)
5534 .ks(3)
5535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5536 }
5537 }
5538 }
5539
5540 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, strided_cm_subtile) {
5541 TEST_REQUIRES_ARM_NEON;
5542 for (size_t k = 1; k <= 40; k += 9) {
5543 for (uint32_t m = 1; m <= 2; m++) {
5544 for (uint32_t n = 1; n <= 8; n++) {
5545 GemmMicrokernelTester()
5546 .mr(2)
5547 .nr(8)
5548 .kr(2)
5549 .sr(1)
5550 .m(m)
5551 .n(n)
5552 .k(k)
5553 .cm_stride(11)
5554 .iterations(1)
5555 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5556 }
5557 }
5558 }
5559 }
5560
5561 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, a_offset) {
5562 TEST_REQUIRES_ARM_NEON;
5563 for (size_t k = 1; k <= 40; k += 9) {
5564 GemmMicrokernelTester()
5565 .mr(2)
5566 .nr(8)
5567 .kr(2)
5568 .sr(1)
5569 .m(2)
5570 .n(8)
5571 .k(k)
5572 .ks(3)
5573 .a_offset(83)
5574 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5575 }
5576 }
5577
5578 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, zero) {
5579 TEST_REQUIRES_ARM_NEON;
5580 for (uint32_t mz = 0; mz < 2; mz++) {
5581 for (size_t k = 1; k <= 40; k += 9) {
5582 GemmMicrokernelTester()
5583 .mr(2)
5584 .nr(8)
5585 .kr(2)
5586 .sr(1)
5587 .m(2)
5588 .n(8)
5589 .k(k)
5590 .ks(3)
5591 .a_offset(83)
5592 .zero_index(mz)
5593 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5594 }
5595 }
5596 }
5597
5598 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, qmin) {
5599 TEST_REQUIRES_ARM_NEON;
5600 GemmMicrokernelTester()
5601 .mr(2)
5602 .nr(8)
5603 .kr(2)
5604 .sr(1)
5605 .m(2)
5606 .n(8)
5607 .k(8)
5608 .qmin(128)
5609 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5610 }
5611
5612 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, qmax) {
5613 TEST_REQUIRES_ARM_NEON;
5614 GemmMicrokernelTester()
5615 .mr(2)
5616 .nr(8)
5617 .kr(2)
5618 .sr(1)
5619 .m(2)
5620 .n(8)
5621 .k(8)
5622 .qmax(128)
5623 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5624 }
5625
5626 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD4R, strided_cm) {
5627 TEST_REQUIRES_ARM_NEON;
5628 GemmMicrokernelTester()
5629 .mr(2)
5630 .nr(8)
5631 .kr(2)
5632 .sr(1)
5633 .m(2)
5634 .n(8)
5635 .k(8)
5636 .cm_stride(11)
5637 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r, 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
Frank Barchard42f5c502021-11-16 10:04:21 -08005643 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_eq_8) {
5644 TEST_REQUIRES_ARM_NEON;
5645 GemmMicrokernelTester()
5646 .mr(1)
5647 .nr(16)
5648 .kr(2)
5649 .sr(1)
5650 .m(1)
5651 .n(16)
5652 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -08005653 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005654 }
5655
5656 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, strided_cn) {
5657 TEST_REQUIRES_ARM_NEON;
5658 GemmMicrokernelTester()
5659 .mr(1)
5660 .nr(16)
5661 .kr(2)
5662 .sr(1)
5663 .m(1)
5664 .n(16)
5665 .k(8)
5666 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08005667 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005668 }
5669
5670 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_eq_8_subtile) {
5671 TEST_REQUIRES_ARM_NEON;
5672 for (uint32_t m = 1; m <= 1; m++) {
5673 for (uint32_t n = 1; n <= 16; n++) {
5674 GemmMicrokernelTester()
5675 .mr(1)
5676 .nr(16)
5677 .kr(2)
5678 .sr(1)
5679 .m(m)
5680 .n(n)
5681 .k(8)
5682 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005683 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005684 }
5685 }
5686 }
5687
5688 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_eq_8_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(2)
5695 .sr(1)
5696 .m(m)
5697 .n(16)
5698 .k(8)
5699 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005700 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005701 }
5702 }
5703
5704 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_eq_8_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(2)
5711 .sr(1)
5712 .m(1)
5713 .n(n)
5714 .k(8)
5715 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005716 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005717 }
5718 }
5719
5720 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_lt_8) {
5721 TEST_REQUIRES_ARM_NEON;
5722 for (size_t k = 1; k < 8; k++) {
5723 GemmMicrokernelTester()
5724 .mr(1)
5725 .nr(16)
5726 .kr(2)
5727 .sr(1)
5728 .m(1)
5729 .n(16)
5730 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08005731 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005732 }
5733 }
5734
5735 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_lt_8_subtile) {
5736 TEST_REQUIRES_ARM_NEON;
5737 for (size_t k = 1; k < 8; k++) {
5738 for (uint32_t m = 1; m <= 1; m++) {
5739 for (uint32_t n = 1; n <= 16; n++) {
5740 GemmMicrokernelTester()
5741 .mr(1)
5742 .nr(16)
5743 .kr(2)
5744 .sr(1)
5745 .m(m)
5746 .n(n)
5747 .k(k)
5748 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005749 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005750 }
5751 }
5752 }
5753 }
5754
5755 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_gt_8) {
5756 TEST_REQUIRES_ARM_NEON;
5757 for (size_t k = 9; k < 16; k++) {
5758 GemmMicrokernelTester()
5759 .mr(1)
5760 .nr(16)
5761 .kr(2)
5762 .sr(1)
5763 .m(1)
5764 .n(16)
5765 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08005766 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005767 }
5768 }
5769
5770 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_gt_8_subtile) {
5771 TEST_REQUIRES_ARM_NEON;
5772 for (size_t k = 9; k < 16; k++) {
5773 for (uint32_t m = 1; m <= 1; m++) {
5774 for (uint32_t n = 1; n <= 16; n++) {
5775 GemmMicrokernelTester()
5776 .mr(1)
5777 .nr(16)
5778 .kr(2)
5779 .sr(1)
5780 .m(m)
5781 .n(n)
5782 .k(k)
5783 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005784 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005785 }
5786 }
5787 }
5788 }
5789
5790 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_div_8) {
5791 TEST_REQUIRES_ARM_NEON;
5792 for (size_t k = 16; k <= 80; k += 8) {
5793 GemmMicrokernelTester()
5794 .mr(1)
5795 .nr(16)
5796 .kr(2)
5797 .sr(1)
5798 .m(1)
5799 .n(16)
5800 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08005801 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005802 }
5803 }
5804
5805 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, k_div_8_subtile) {
5806 TEST_REQUIRES_ARM_NEON;
5807 for (size_t k = 16; k <= 80; k += 8) {
5808 for (uint32_t m = 1; m <= 1; m++) {
5809 for (uint32_t n = 1; n <= 16; n++) {
5810 GemmMicrokernelTester()
5811 .mr(1)
5812 .nr(16)
5813 .kr(2)
5814 .sr(1)
5815 .m(m)
5816 .n(n)
5817 .k(k)
5818 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005819 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005820 }
5821 }
5822 }
5823 }
5824
5825 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, n_gt_16) {
5826 TEST_REQUIRES_ARM_NEON;
5827 for (uint32_t n = 17; n < 32; n++) {
5828 for (size_t k = 1; k <= 40; k += 9) {
5829 GemmMicrokernelTester()
5830 .mr(1)
5831 .nr(16)
5832 .kr(2)
5833 .sr(1)
5834 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005835 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08005836 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08005837 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005838 }
5839 }
5840 }
5841
5842 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5846 GemmMicrokernelTester()
5847 .mr(1)
5848 .nr(16)
5849 .kr(2)
5850 .sr(1)
5851 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005852 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08005853 .k(k)
5854 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08005855 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005856 }
5857 }
5858 }
5859
5860 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5864 for (uint32_t m = 1; m <= 1; m++) {
5865 GemmMicrokernelTester()
5866 .mr(1)
5867 .nr(16)
5868 .kr(2)
5869 .sr(1)
5870 .m(m)
5871 .n(n)
5872 .k(k)
5873 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005874 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005875 }
5876 }
5877 }
5878 }
5879
5880 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5884 GemmMicrokernelTester()
5885 .mr(1)
5886 .nr(16)
5887 .kr(2)
5888 .sr(1)
5889 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005890 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08005891 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08005892 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005893 }
5894 }
5895 }
5896
5897 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5901 GemmMicrokernelTester()
5902 .mr(1)
5903 .nr(16)
5904 .kr(2)
5905 .sr(1)
5906 .m(1)
5907 .n(n)
5908 .k(k)
5909 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08005910 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005911 }
5912 }
5913 }
5914
5915 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5919 for (uint32_t m = 1; m <= 1; m++) {
5920 GemmMicrokernelTester()
5921 .mr(1)
5922 .nr(16)
5923 .kr(2)
5924 .sr(1)
5925 .m(m)
5926 .n(n)
5927 .k(k)
5928 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005929 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005930 }
5931 }
5932 }
5933 }
5934
5935 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, small_kernel) {
5936 TEST_REQUIRES_ARM_NEON;
5937 for (size_t k = 1; k <= 40; k += 9) {
5938 GemmMicrokernelTester()
5939 .mr(1)
5940 .nr(16)
5941 .kr(2)
5942 .sr(1)
5943 .m(1)
5944 .n(16)
5945 .k(k)
5946 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08005947 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005948 }
5949 }
5950
5951 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, small_kernel_subtile) {
5952 TEST_REQUIRES_ARM_NEON;
5953 for (size_t k = 1; k <= 40; k += 9) {
5954 for (uint32_t m = 1; m <= 1; m++) {
5955 for (uint32_t n = 1; n <= 16; n++) {
5956 GemmMicrokernelTester()
5957 .mr(1)
5958 .nr(16)
5959 .kr(2)
5960 .sr(1)
5961 .m(m)
5962 .n(n)
5963 .k(k)
5964 .ks(3)
5965 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08005966 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005967 }
5968 }
5969 }
5970 }
5971
5972 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5976 GemmMicrokernelTester()
5977 .mr(1)
5978 .nr(16)
5979 .kr(2)
5980 .sr(1)
5981 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08005982 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08005983 .k(k)
5984 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08005985 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08005986 }
5987 }
5988 }
5989
5990 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, 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 <= 40; k += 9) {
5994 GemmMicrokernelTester()
5995 .mr(1)
5996 .nr(16)
5997 .kr(2)
5998 .sr(1)
5999 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006000 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006001 .k(k)
6002 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08006003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006004 }
6005 }
6006 }
6007
6008 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, strided_cm_subtile) {
6009 TEST_REQUIRES_ARM_NEON;
6010 for (size_t k = 1; k <= 40; k += 9) {
6011 for (uint32_t m = 1; m <= 1; m++) {
6012 for (uint32_t n = 1; n <= 16; n++) {
6013 GemmMicrokernelTester()
6014 .mr(1)
6015 .nr(16)
6016 .kr(2)
6017 .sr(1)
6018 .m(m)
6019 .n(n)
6020 .k(k)
6021 .cm_stride(19)
6022 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006023 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006024 }
6025 }
6026 }
6027 }
6028
6029 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, a_offset) {
6030 TEST_REQUIRES_ARM_NEON;
6031 for (size_t k = 1; k <= 40; k += 9) {
6032 GemmMicrokernelTester()
6033 .mr(1)
6034 .nr(16)
6035 .kr(2)
6036 .sr(1)
6037 .m(1)
6038 .n(16)
6039 .k(k)
6040 .ks(3)
6041 .a_offset(43)
Marat Dukhan50323b82022-01-11 00:12:01 -08006042 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006043 }
6044 }
6045
6046 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, zero) {
6047 TEST_REQUIRES_ARM_NEON;
6048 for (uint32_t mz = 0; mz < 1; mz++) {
6049 for (size_t k = 1; k <= 40; k += 9) {
6050 GemmMicrokernelTester()
6051 .mr(1)
6052 .nr(16)
6053 .kr(2)
6054 .sr(1)
6055 .m(1)
6056 .n(16)
6057 .k(k)
6058 .ks(3)
6059 .a_offset(43)
6060 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08006061 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006062 }
6063 }
6064 }
6065
6066 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, qmin) {
6067 TEST_REQUIRES_ARM_NEON;
6068 GemmMicrokernelTester()
6069 .mr(1)
6070 .nr(16)
6071 .kr(2)
6072 .sr(1)
6073 .m(1)
6074 .n(16)
6075 .k(8)
6076 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08006077 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006078 }
6079
6080 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, qmax) {
6081 TEST_REQUIRES_ARM_NEON;
6082 GemmMicrokernelTester()
6083 .mr(1)
6084 .nr(16)
6085 .kr(2)
6086 .sr(1)
6087 .m(1)
6088 .n(16)
6089 .k(8)
6090 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08006091 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006092 }
6093
6094 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD4R, strided_cm) {
6095 TEST_REQUIRES_ARM_NEON;
6096 GemmMicrokernelTester()
6097 .mr(1)
6098 .nr(16)
6099 .kr(2)
6100 .sr(1)
6101 .m(1)
6102 .n(16)
6103 .k(8)
6104 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08006105 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006106 }
6107#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
6108
6109
6110#if XNN_ARCH_ARM || XNN_ARCH_ARM64
6111 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, k_eq_8) {
6112 TEST_REQUIRES_ARM_NEON;
6113 GemmMicrokernelTester()
6114 .mr(2)
6115 .nr(16)
6116 .kr(2)
6117 .sr(1)
6118 .m(2)
6119 .n(16)
6120 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -08006121 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006122 }
6123
6124 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, strided_cn) {
6125 TEST_REQUIRES_ARM_NEON;
6126 GemmMicrokernelTester()
6127 .mr(2)
6128 .nr(16)
6129 .kr(2)
6130 .sr(1)
6131 .m(2)
6132 .n(16)
6133 .k(8)
6134 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08006135 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006136 }
6137
6138 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, k_eq_8_subtile) {
6139 TEST_REQUIRES_ARM_NEON;
6140 for (uint32_t m = 1; m <= 2; m++) {
6141 for (uint32_t n = 1; n <= 16; n++) {
6142 GemmMicrokernelTester()
6143 .mr(2)
6144 .nr(16)
6145 .kr(2)
6146 .sr(1)
6147 .m(m)
6148 .n(n)
6149 .k(8)
6150 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006151 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006152 }
6153 }
6154 }
6155
6156 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, 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(16)
6162 .kr(2)
6163 .sr(1)
6164 .m(m)
6165 .n(16)
6166 .k(8)
6167 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006168 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006169 }
6170 }
6171
6172 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, k_eq_8_subtile_n) {
6173 TEST_REQUIRES_ARM_NEON;
6174 for (uint32_t n = 1; n <= 16; n++) {
6175 GemmMicrokernelTester()
6176 .mr(2)
6177 .nr(16)
6178 .kr(2)
6179 .sr(1)
6180 .m(2)
6181 .n(n)
6182 .k(8)
6183 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006184 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006185 }
6186 }
6187
6188 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, k_lt_8) {
6189 TEST_REQUIRES_ARM_NEON;
6190 for (size_t k = 1; k < 8; k++) {
6191 GemmMicrokernelTester()
6192 .mr(2)
6193 .nr(16)
6194 .kr(2)
6195 .sr(1)
6196 .m(2)
6197 .n(16)
6198 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006199 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006200 }
6201 }
6202
6203 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, k_lt_8_subtile) {
6204 TEST_REQUIRES_ARM_NEON;
6205 for (size_t k = 1; k < 8; k++) {
6206 for (uint32_t m = 1; m <= 2; m++) {
6207 for (uint32_t n = 1; n <= 16; n++) {
6208 GemmMicrokernelTester()
6209 .mr(2)
6210 .nr(16)
6211 .kr(2)
6212 .sr(1)
6213 .m(m)
6214 .n(n)
6215 .k(k)
6216 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006217 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006218 }
6219 }
6220 }
6221 }
6222
6223 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, k_gt_8) {
6224 TEST_REQUIRES_ARM_NEON;
6225 for (size_t k = 9; k < 16; k++) {
6226 GemmMicrokernelTester()
6227 .mr(2)
6228 .nr(16)
6229 .kr(2)
6230 .sr(1)
6231 .m(2)
6232 .n(16)
6233 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006234 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006235 }
6236 }
6237
6238 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, k_gt_8_subtile) {
6239 TEST_REQUIRES_ARM_NEON;
6240 for (size_t k = 9; k < 16; k++) {
6241 for (uint32_t m = 1; m <= 2; m++) {
6242 for (uint32_t n = 1; n <= 16; n++) {
6243 GemmMicrokernelTester()
6244 .mr(2)
6245 .nr(16)
6246 .kr(2)
6247 .sr(1)
6248 .m(m)
6249 .n(n)
6250 .k(k)
6251 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006252 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006253 }
6254 }
6255 }
6256 }
6257
6258 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, 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(16)
6264 .kr(2)
6265 .sr(1)
6266 .m(2)
6267 .n(16)
6268 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006269 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006270 }
6271 }
6272
6273 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, k_div_8_subtile) {
6274 TEST_REQUIRES_ARM_NEON;
6275 for (size_t k = 16; k <= 80; k += 8) {
6276 for (uint32_t m = 1; m <= 2; m++) {
6277 for (uint32_t n = 1; n <= 16; n++) {
6278 GemmMicrokernelTester()
6279 .mr(2)
6280 .nr(16)
6281 .kr(2)
6282 .sr(1)
6283 .m(m)
6284 .n(n)
6285 .k(k)
6286 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006287 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006288 }
6289 }
6290 }
6291 }
6292
6293 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, n_gt_16) {
6294 TEST_REQUIRES_ARM_NEON;
6295 for (uint32_t n = 17; n < 32; n++) {
6296 for (size_t k = 1; k <= 40; k += 9) {
6297 GemmMicrokernelTester()
6298 .mr(2)
6299 .nr(16)
6300 .kr(2)
6301 .sr(1)
6302 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006303 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006304 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006305 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006306 }
6307 }
6308 }
6309
6310 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, n_gt_16_strided_cn) {
6311 TEST_REQUIRES_ARM_NEON;
6312 for (uint32_t n = 17; n < 32; n++) {
6313 for (size_t k = 1; k <= 40; k += 9) {
6314 GemmMicrokernelTester()
6315 .mr(2)
6316 .nr(16)
6317 .kr(2)
6318 .sr(1)
6319 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006320 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006321 .k(k)
6322 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08006323 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006324 }
6325 }
6326 }
6327
6328 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, n_gt_16_subtile) {
6329 TEST_REQUIRES_ARM_NEON;
6330 for (uint32_t n = 17; n < 32; 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(16)
6336 .kr(2)
6337 .sr(1)
6338 .m(m)
6339 .n(n)
6340 .k(k)
6341 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006342 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006343 }
6344 }
6345 }
6346 }
6347
6348 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, n_div_16) {
6349 TEST_REQUIRES_ARM_NEON;
6350 for (uint32_t n = 32; n <= 48; n += 16) {
6351 for (size_t k = 1; k <= 40; k += 9) {
6352 GemmMicrokernelTester()
6353 .mr(2)
6354 .nr(16)
6355 .kr(2)
6356 .sr(1)
6357 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006358 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006359 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006360 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006361 }
6362 }
6363 }
6364
6365 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, n_div_16_strided_cn) {
6366 TEST_REQUIRES_ARM_NEON;
6367 for (uint32_t n = 32; n <= 48; n += 16) {
6368 for (size_t k = 1; k <= 40; k += 9) {
6369 GemmMicrokernelTester()
6370 .mr(2)
6371 .nr(16)
6372 .kr(2)
6373 .sr(1)
6374 .m(2)
6375 .n(n)
6376 .k(k)
6377 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08006378 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006379 }
6380 }
6381 }
6382
6383 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, n_div_16_subtile) {
6384 TEST_REQUIRES_ARM_NEON;
6385 for (uint32_t n = 32; n <= 48; n += 16) {
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(16)
6391 .kr(2)
6392 .sr(1)
6393 .m(m)
6394 .n(n)
6395 .k(k)
6396 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006397 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006398 }
6399 }
6400 }
6401 }
6402
6403 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, small_kernel) {
6404 TEST_REQUIRES_ARM_NEON;
6405 for (size_t k = 1; k <= 40; k += 9) {
6406 GemmMicrokernelTester()
6407 .mr(2)
6408 .nr(16)
6409 .kr(2)
6410 .sr(1)
6411 .m(2)
6412 .n(16)
6413 .k(k)
6414 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08006415 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006416 }
6417 }
6418
6419 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, small_kernel_subtile) {
6420 TEST_REQUIRES_ARM_NEON;
6421 for (size_t k = 1; k <= 40; k += 9) {
6422 for (uint32_t m = 1; m <= 2; m++) {
6423 for (uint32_t n = 1; n <= 16; n++) {
6424 GemmMicrokernelTester()
6425 .mr(2)
6426 .nr(16)
6427 .kr(2)
6428 .sr(1)
6429 .m(m)
6430 .n(n)
6431 .k(k)
6432 .ks(3)
6433 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006434 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006435 }
6436 }
6437 }
6438 }
6439
6440 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, n_gt_16_small_kernel) {
6441 TEST_REQUIRES_ARM_NEON;
6442 for (uint32_t n = 17; n < 32; n++) {
6443 for (size_t k = 1; k <= 40; k += 9) {
6444 GemmMicrokernelTester()
6445 .mr(2)
6446 .nr(16)
6447 .kr(2)
6448 .sr(1)
6449 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006450 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006451 .k(k)
6452 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08006453 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006454 }
6455 }
6456 }
6457
6458 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, n_div_16_small_kernel) {
6459 TEST_REQUIRES_ARM_NEON;
6460 for (uint32_t n = 32; n <= 48; n += 16) {
6461 for (size_t k = 1; k <= 40; k += 9) {
6462 GemmMicrokernelTester()
6463 .mr(2)
6464 .nr(16)
6465 .kr(2)
6466 .sr(1)
6467 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006468 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006469 .k(k)
6470 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08006471 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006472 }
6473 }
6474 }
6475
6476 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, strided_cm_subtile) {
6477 TEST_REQUIRES_ARM_NEON;
6478 for (size_t k = 1; k <= 40; k += 9) {
6479 for (uint32_t m = 1; m <= 2; m++) {
6480 for (uint32_t n = 1; n <= 16; n++) {
6481 GemmMicrokernelTester()
6482 .mr(2)
6483 .nr(16)
6484 .kr(2)
6485 .sr(1)
6486 .m(m)
6487 .n(n)
6488 .k(k)
6489 .cm_stride(19)
6490 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006491 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006492 }
6493 }
6494 }
6495 }
6496
6497 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, a_offset) {
6498 TEST_REQUIRES_ARM_NEON;
6499 for (size_t k = 1; k <= 40; k += 9) {
6500 GemmMicrokernelTester()
6501 .mr(2)
6502 .nr(16)
6503 .kr(2)
6504 .sr(1)
6505 .m(2)
6506 .n(16)
6507 .k(k)
6508 .ks(3)
6509 .a_offset(83)
Marat Dukhan50323b82022-01-11 00:12:01 -08006510 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006511 }
6512 }
6513
6514 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, zero) {
6515 TEST_REQUIRES_ARM_NEON;
6516 for (uint32_t mz = 0; mz < 2; mz++) {
6517 for (size_t k = 1; k <= 40; k += 9) {
6518 GemmMicrokernelTester()
6519 .mr(2)
6520 .nr(16)
6521 .kr(2)
6522 .sr(1)
6523 .m(2)
6524 .n(16)
6525 .k(k)
6526 .ks(3)
6527 .a_offset(83)
6528 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08006529 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006530 }
6531 }
6532 }
6533
6534 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, qmin) {
6535 TEST_REQUIRES_ARM_NEON;
6536 GemmMicrokernelTester()
6537 .mr(2)
6538 .nr(16)
6539 .kr(2)
6540 .sr(1)
6541 .m(2)
6542 .n(16)
6543 .k(8)
6544 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08006545 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006546 }
6547
6548 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, qmax) {
6549 TEST_REQUIRES_ARM_NEON;
6550 GemmMicrokernelTester()
6551 .mr(2)
6552 .nr(16)
6553 .kr(2)
6554 .sr(1)
6555 .m(2)
6556 .n(16)
6557 .k(8)
6558 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08006559 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006560 }
6561
6562 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_LD4R, strided_cm) {
6563 TEST_REQUIRES_ARM_NEON;
6564 GemmMicrokernelTester()
6565 .mr(2)
6566 .nr(16)
6567 .kr(2)
6568 .sr(1)
6569 .m(2)
6570 .n(16)
6571 .k(8)
6572 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08006573 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006574 }
6575#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
6576
6577
6578#if XNN_ARCH_ARM || XNN_ARCH_ARM64
6579 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_eq_8) {
6580 TEST_REQUIRES_ARM_NEON;
6581 GemmMicrokernelTester()
6582 .mr(3)
6583 .nr(16)
6584 .kr(2)
6585 .sr(1)
6586 .m(3)
6587 .n(16)
6588 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -08006589 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006590 }
6591
6592 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, strided_cn) {
6593 TEST_REQUIRES_ARM_NEON;
6594 GemmMicrokernelTester()
6595 .mr(3)
6596 .nr(16)
6597 .kr(2)
6598 .sr(1)
6599 .m(3)
6600 .n(16)
6601 .k(8)
6602 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08006603 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006604 }
6605
6606 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_eq_8_subtile) {
6607 TEST_REQUIRES_ARM_NEON;
6608 for (uint32_t m = 1; m <= 3; m++) {
6609 for (uint32_t n = 1; n <= 16; n++) {
6610 GemmMicrokernelTester()
6611 .mr(3)
6612 .nr(16)
6613 .kr(2)
6614 .sr(1)
6615 .m(m)
6616 .n(n)
6617 .k(8)
6618 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006619 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006620 }
6621 }
6622 }
6623
6624 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_eq_8_subtile_m) {
6625 TEST_REQUIRES_ARM_NEON;
6626 for (uint32_t m = 1; m <= 3; m++) {
6627 GemmMicrokernelTester()
6628 .mr(3)
6629 .nr(16)
6630 .kr(2)
6631 .sr(1)
6632 .m(m)
6633 .n(16)
6634 .k(8)
6635 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006636 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006637 }
6638 }
6639
6640 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_eq_8_subtile_n) {
6641 TEST_REQUIRES_ARM_NEON;
6642 for (uint32_t n = 1; n <= 16; n++) {
6643 GemmMicrokernelTester()
6644 .mr(3)
6645 .nr(16)
6646 .kr(2)
6647 .sr(1)
6648 .m(3)
6649 .n(n)
6650 .k(8)
6651 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006652 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006653 }
6654 }
6655
6656 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_lt_8) {
6657 TEST_REQUIRES_ARM_NEON;
6658 for (size_t k = 1; k < 8; k++) {
6659 GemmMicrokernelTester()
6660 .mr(3)
6661 .nr(16)
6662 .kr(2)
6663 .sr(1)
6664 .m(3)
6665 .n(16)
6666 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006667 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006668 }
6669 }
6670
6671 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_lt_8_subtile) {
6672 TEST_REQUIRES_ARM_NEON;
6673 for (size_t k = 1; k < 8; k++) {
6674 for (uint32_t m = 1; m <= 3; m++) {
6675 for (uint32_t n = 1; n <= 16; n++) {
6676 GemmMicrokernelTester()
6677 .mr(3)
6678 .nr(16)
6679 .kr(2)
6680 .sr(1)
6681 .m(m)
6682 .n(n)
6683 .k(k)
6684 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006685 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006686 }
6687 }
6688 }
6689 }
6690
6691 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_gt_8) {
6692 TEST_REQUIRES_ARM_NEON;
6693 for (size_t k = 9; k < 16; k++) {
6694 GemmMicrokernelTester()
6695 .mr(3)
6696 .nr(16)
6697 .kr(2)
6698 .sr(1)
6699 .m(3)
6700 .n(16)
6701 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006702 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006703 }
6704 }
6705
6706 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_gt_8_subtile) {
6707 TEST_REQUIRES_ARM_NEON;
6708 for (size_t k = 9; k < 16; k++) {
6709 for (uint32_t m = 1; m <= 3; m++) {
6710 for (uint32_t n = 1; n <= 16; n++) {
6711 GemmMicrokernelTester()
6712 .mr(3)
6713 .nr(16)
6714 .kr(2)
6715 .sr(1)
6716 .m(m)
6717 .n(n)
6718 .k(k)
6719 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006720 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006721 }
6722 }
6723 }
6724 }
6725
6726 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_div_8) {
6727 TEST_REQUIRES_ARM_NEON;
6728 for (size_t k = 16; k <= 80; k += 8) {
6729 GemmMicrokernelTester()
6730 .mr(3)
6731 .nr(16)
6732 .kr(2)
6733 .sr(1)
6734 .m(3)
6735 .n(16)
6736 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006737 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006738 }
6739 }
6740
6741 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, k_div_8_subtile) {
6742 TEST_REQUIRES_ARM_NEON;
6743 for (size_t k = 16; k <= 80; k += 8) {
6744 for (uint32_t m = 1; m <= 3; m++) {
6745 for (uint32_t n = 1; n <= 16; n++) {
6746 GemmMicrokernelTester()
6747 .mr(3)
6748 .nr(16)
6749 .kr(2)
6750 .sr(1)
6751 .m(m)
6752 .n(n)
6753 .k(k)
6754 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006755 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006756 }
6757 }
6758 }
6759 }
6760
6761 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, 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(3)
6767 .nr(16)
6768 .kr(2)
6769 .sr(1)
6770 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006771 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006772 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006773 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006774 }
6775 }
6776 }
6777
6778 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, 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(3)
6784 .nr(16)
6785 .kr(2)
6786 .sr(1)
6787 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006788 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006789 .k(k)
6790 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08006791 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006792 }
6793 }
6794 }
6795
6796 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, 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 <= 3; m++) {
6801 GemmMicrokernelTester()
6802 .mr(3)
6803 .nr(16)
6804 .kr(2)
6805 .sr(1)
6806 .m(m)
6807 .n(n)
6808 .k(k)
6809 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006810 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006811 }
6812 }
6813 }
6814 }
6815
6816 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, 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(3)
6822 .nr(16)
6823 .kr(2)
6824 .sr(1)
6825 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006826 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006827 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08006828 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006829 }
6830 }
6831 }
6832
6833 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, 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(3)
6839 .nr(16)
6840 .kr(2)
6841 .sr(1)
6842 .m(3)
6843 .n(n)
6844 .k(k)
6845 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08006846 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006847 }
6848 }
6849 }
6850
6851 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, 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 <= 3; m++) {
6856 GemmMicrokernelTester()
6857 .mr(3)
6858 .nr(16)
6859 .kr(2)
6860 .sr(1)
6861 .m(m)
6862 .n(n)
6863 .k(k)
6864 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006865 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006866 }
6867 }
6868 }
6869 }
6870
6871 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, small_kernel) {
6872 TEST_REQUIRES_ARM_NEON;
6873 for (size_t k = 1; k <= 40; k += 9) {
6874 GemmMicrokernelTester()
6875 .mr(3)
6876 .nr(16)
6877 .kr(2)
6878 .sr(1)
6879 .m(3)
6880 .n(16)
6881 .k(k)
6882 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08006883 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006884 }
6885 }
6886
6887 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, small_kernel_subtile) {
6888 TEST_REQUIRES_ARM_NEON;
6889 for (size_t k = 1; k <= 40; k += 9) {
6890 for (uint32_t m = 1; m <= 3; m++) {
6891 for (uint32_t n = 1; n <= 16; n++) {
6892 GemmMicrokernelTester()
6893 .mr(3)
6894 .nr(16)
6895 .kr(2)
6896 .sr(1)
6897 .m(m)
6898 .n(n)
6899 .k(k)
6900 .ks(3)
6901 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006902 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006903 }
6904 }
6905 }
6906 }
6907
6908 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, 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(3)
6914 .nr(16)
6915 .kr(2)
6916 .sr(1)
6917 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006918 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006919 .k(k)
6920 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08006921 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006922 }
6923 }
6924 }
6925
6926 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, 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(3)
6932 .nr(16)
6933 .kr(2)
6934 .sr(1)
6935 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08006936 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08006937 .k(k)
6938 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08006939 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006940 }
6941 }
6942 }
6943
6944 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, strided_cm_subtile) {
6945 TEST_REQUIRES_ARM_NEON;
6946 for (size_t k = 1; k <= 40; k += 9) {
6947 for (uint32_t m = 1; m <= 3; m++) {
6948 for (uint32_t n = 1; n <= 16; n++) {
6949 GemmMicrokernelTester()
6950 .mr(3)
6951 .nr(16)
6952 .kr(2)
6953 .sr(1)
6954 .m(m)
6955 .n(n)
6956 .k(k)
6957 .cm_stride(19)
6958 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08006959 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006960 }
6961 }
6962 }
6963 }
6964
6965 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, a_offset) {
6966 TEST_REQUIRES_ARM_NEON;
6967 for (size_t k = 1; k <= 40; k += 9) {
6968 GemmMicrokernelTester()
6969 .mr(3)
6970 .nr(16)
6971 .kr(2)
6972 .sr(1)
6973 .m(3)
6974 .n(16)
6975 .k(k)
6976 .ks(3)
6977 .a_offset(127)
Marat Dukhan50323b82022-01-11 00:12:01 -08006978 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006979 }
6980 }
6981
6982 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, zero) {
6983 TEST_REQUIRES_ARM_NEON;
6984 for (uint32_t mz = 0; mz < 3; mz++) {
6985 for (size_t k = 1; k <= 40; k += 9) {
6986 GemmMicrokernelTester()
6987 .mr(3)
6988 .nr(16)
6989 .kr(2)
6990 .sr(1)
6991 .m(3)
6992 .n(16)
6993 .k(k)
6994 .ks(3)
6995 .a_offset(127)
6996 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08006997 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08006998 }
6999 }
7000 }
7001
7002 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, qmin) {
7003 TEST_REQUIRES_ARM_NEON;
7004 GemmMicrokernelTester()
7005 .mr(3)
7006 .nr(16)
7007 .kr(2)
7008 .sr(1)
7009 .m(3)
7010 .n(16)
7011 .k(8)
7012 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08007013 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007014 }
7015
7016 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, qmax) {
7017 TEST_REQUIRES_ARM_NEON;
7018 GemmMicrokernelTester()
7019 .mr(3)
7020 .nr(16)
7021 .kr(2)
7022 .sr(1)
7023 .m(3)
7024 .n(16)
7025 .k(8)
7026 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08007027 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007028 }
7029
7030 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD4R, strided_cm) {
7031 TEST_REQUIRES_ARM_NEON;
7032 GemmMicrokernelTester()
7033 .mr(3)
7034 .nr(16)
7035 .kr(2)
7036 .sr(1)
7037 .m(3)
7038 .n(16)
7039 .k(8)
7040 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08007041 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007042 }
7043#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
7044
7045
7046#if XNN_ARCH_ARM || XNN_ARCH_ARM64
7047 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_eq_8) {
7048 TEST_REQUIRES_ARM_NEON;
7049 GemmMicrokernelTester()
7050 .mr(4)
7051 .nr(16)
7052 .kr(2)
7053 .sr(1)
7054 .m(4)
7055 .n(16)
7056 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -08007057 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007058 }
7059
7060 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, strided_cn) {
7061 TEST_REQUIRES_ARM_NEON;
7062 GemmMicrokernelTester()
7063 .mr(4)
7064 .nr(16)
7065 .kr(2)
7066 .sr(1)
7067 .m(4)
7068 .n(16)
7069 .k(8)
7070 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08007071 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007072 }
7073
7074 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_eq_8_subtile) {
7075 TEST_REQUIRES_ARM_NEON;
7076 for (uint32_t m = 1; m <= 4; m++) {
7077 for (uint32_t n = 1; n <= 16; n++) {
7078 GemmMicrokernelTester()
7079 .mr(4)
7080 .nr(16)
7081 .kr(2)
7082 .sr(1)
7083 .m(m)
7084 .n(n)
7085 .k(8)
7086 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007087 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007088 }
7089 }
7090 }
7091
7092 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_eq_8_subtile_m) {
7093 TEST_REQUIRES_ARM_NEON;
7094 for (uint32_t m = 1; m <= 4; m++) {
7095 GemmMicrokernelTester()
7096 .mr(4)
7097 .nr(16)
7098 .kr(2)
7099 .sr(1)
7100 .m(m)
7101 .n(16)
7102 .k(8)
7103 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007104 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007105 }
7106 }
7107
7108 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_eq_8_subtile_n) {
7109 TEST_REQUIRES_ARM_NEON;
7110 for (uint32_t n = 1; n <= 16; n++) {
7111 GemmMicrokernelTester()
7112 .mr(4)
7113 .nr(16)
7114 .kr(2)
7115 .sr(1)
7116 .m(4)
7117 .n(n)
7118 .k(8)
7119 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007120 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007121 }
7122 }
7123
7124 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_lt_8) {
7125 TEST_REQUIRES_ARM_NEON;
7126 for (size_t k = 1; k < 8; k++) {
7127 GemmMicrokernelTester()
7128 .mr(4)
7129 .nr(16)
7130 .kr(2)
7131 .sr(1)
7132 .m(4)
7133 .n(16)
7134 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007135 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007136 }
7137 }
7138
7139 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_lt_8_subtile) {
7140 TEST_REQUIRES_ARM_NEON;
7141 for (size_t k = 1; k < 8; k++) {
7142 for (uint32_t m = 1; m <= 4; m++) {
7143 for (uint32_t n = 1; n <= 16; n++) {
7144 GemmMicrokernelTester()
7145 .mr(4)
7146 .nr(16)
7147 .kr(2)
7148 .sr(1)
7149 .m(m)
7150 .n(n)
7151 .k(k)
7152 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007153 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007154 }
7155 }
7156 }
7157 }
7158
7159 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_gt_8) {
7160 TEST_REQUIRES_ARM_NEON;
7161 for (size_t k = 9; k < 16; k++) {
7162 GemmMicrokernelTester()
7163 .mr(4)
7164 .nr(16)
7165 .kr(2)
7166 .sr(1)
7167 .m(4)
7168 .n(16)
7169 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007170 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007171 }
7172 }
7173
7174 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_gt_8_subtile) {
7175 TEST_REQUIRES_ARM_NEON;
7176 for (size_t k = 9; k < 16; k++) {
7177 for (uint32_t m = 1; m <= 4; m++) {
7178 for (uint32_t n = 1; n <= 16; n++) {
7179 GemmMicrokernelTester()
7180 .mr(4)
7181 .nr(16)
7182 .kr(2)
7183 .sr(1)
7184 .m(m)
7185 .n(n)
7186 .k(k)
7187 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007188 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007189 }
7190 }
7191 }
7192 }
7193
7194 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_div_8) {
7195 TEST_REQUIRES_ARM_NEON;
7196 for (size_t k = 16; k <= 80; k += 8) {
7197 GemmMicrokernelTester()
7198 .mr(4)
7199 .nr(16)
7200 .kr(2)
7201 .sr(1)
7202 .m(4)
7203 .n(16)
7204 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007205 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007206 }
7207 }
7208
7209 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, k_div_8_subtile) {
7210 TEST_REQUIRES_ARM_NEON;
7211 for (size_t k = 16; k <= 80; k += 8) {
7212 for (uint32_t m = 1; m <= 4; m++) {
7213 for (uint32_t n = 1; n <= 16; n++) {
7214 GemmMicrokernelTester()
7215 .mr(4)
7216 .nr(16)
7217 .kr(2)
7218 .sr(1)
7219 .m(m)
7220 .n(n)
7221 .k(k)
7222 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007223 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007224 }
7225 }
7226 }
7227 }
7228
7229 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, n_gt_16) {
7230 TEST_REQUIRES_ARM_NEON;
7231 for (uint32_t n = 17; n < 32; n++) {
7232 for (size_t k = 1; k <= 40; k += 9) {
7233 GemmMicrokernelTester()
7234 .mr(4)
7235 .nr(16)
7236 .kr(2)
7237 .sr(1)
7238 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007239 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007240 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007241 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007242 }
7243 }
7244 }
7245
7246 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, n_gt_16_strided_cn) {
7247 TEST_REQUIRES_ARM_NEON;
7248 for (uint32_t n = 17; n < 32; n++) {
7249 for (size_t k = 1; k <= 40; k += 9) {
7250 GemmMicrokernelTester()
7251 .mr(4)
7252 .nr(16)
7253 .kr(2)
7254 .sr(1)
7255 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007256 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007257 .k(k)
7258 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08007259 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007260 }
7261 }
7262 }
7263
7264 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, n_gt_16_subtile) {
7265 TEST_REQUIRES_ARM_NEON;
7266 for (uint32_t n = 17; n < 32; n++) {
7267 for (size_t k = 1; k <= 40; k += 9) {
7268 for (uint32_t m = 1; m <= 4; m++) {
7269 GemmMicrokernelTester()
7270 .mr(4)
7271 .nr(16)
7272 .kr(2)
7273 .sr(1)
7274 .m(m)
7275 .n(n)
7276 .k(k)
7277 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007278 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007279 }
7280 }
7281 }
7282 }
7283
7284 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, n_div_16) {
7285 TEST_REQUIRES_ARM_NEON;
7286 for (uint32_t n = 32; n <= 48; n += 16) {
7287 for (size_t k = 1; k <= 40; k += 9) {
7288 GemmMicrokernelTester()
7289 .mr(4)
7290 .nr(16)
7291 .kr(2)
7292 .sr(1)
7293 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007294 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007295 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007296 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007297 }
7298 }
7299 }
7300
7301 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, n_div_16_strided_cn) {
7302 TEST_REQUIRES_ARM_NEON;
7303 for (uint32_t n = 32; n <= 48; n += 16) {
7304 for (size_t k = 1; k <= 40; k += 9) {
7305 GemmMicrokernelTester()
7306 .mr(4)
7307 .nr(16)
7308 .kr(2)
7309 .sr(1)
7310 .m(4)
7311 .n(n)
7312 .k(k)
7313 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08007314 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007315 }
7316 }
7317 }
7318
7319 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, n_div_16_subtile) {
7320 TEST_REQUIRES_ARM_NEON;
7321 for (uint32_t n = 32; n <= 48; n += 16) {
7322 for (size_t k = 1; k <= 40; k += 9) {
7323 for (uint32_t m = 1; m <= 4; m++) {
7324 GemmMicrokernelTester()
7325 .mr(4)
7326 .nr(16)
7327 .kr(2)
7328 .sr(1)
7329 .m(m)
7330 .n(n)
7331 .k(k)
7332 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007333 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007334 }
7335 }
7336 }
7337 }
7338
7339 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, small_kernel) {
7340 TEST_REQUIRES_ARM_NEON;
7341 for (size_t k = 1; k <= 40; k += 9) {
7342 GemmMicrokernelTester()
7343 .mr(4)
7344 .nr(16)
7345 .kr(2)
7346 .sr(1)
7347 .m(4)
7348 .n(16)
7349 .k(k)
7350 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08007351 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007352 }
7353 }
7354
7355 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, small_kernel_subtile) {
7356 TEST_REQUIRES_ARM_NEON;
7357 for (size_t k = 1; k <= 40; k += 9) {
7358 for (uint32_t m = 1; m <= 4; m++) {
7359 for (uint32_t n = 1; n <= 16; n++) {
7360 GemmMicrokernelTester()
7361 .mr(4)
7362 .nr(16)
7363 .kr(2)
7364 .sr(1)
7365 .m(m)
7366 .n(n)
7367 .k(k)
7368 .ks(3)
7369 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007370 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007371 }
7372 }
7373 }
7374 }
7375
7376 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, n_gt_16_small_kernel) {
7377 TEST_REQUIRES_ARM_NEON;
7378 for (uint32_t n = 17; n < 32; n++) {
7379 for (size_t k = 1; k <= 40; k += 9) {
7380 GemmMicrokernelTester()
7381 .mr(4)
7382 .nr(16)
7383 .kr(2)
7384 .sr(1)
7385 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007386 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007387 .k(k)
7388 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08007389 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007390 }
7391 }
7392 }
7393
7394 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, n_div_16_small_kernel) {
7395 TEST_REQUIRES_ARM_NEON;
7396 for (uint32_t n = 32; n <= 48; n += 16) {
7397 for (size_t k = 1; k <= 40; k += 9) {
7398 GemmMicrokernelTester()
7399 .mr(4)
7400 .nr(16)
7401 .kr(2)
7402 .sr(1)
7403 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007404 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007405 .k(k)
7406 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08007407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007408 }
7409 }
7410 }
7411
7412 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, strided_cm_subtile) {
7413 TEST_REQUIRES_ARM_NEON;
7414 for (size_t k = 1; k <= 40; k += 9) {
7415 for (uint32_t m = 1; m <= 4; m++) {
7416 for (uint32_t n = 1; n <= 16; n++) {
7417 GemmMicrokernelTester()
7418 .mr(4)
7419 .nr(16)
7420 .kr(2)
7421 .sr(1)
7422 .m(m)
7423 .n(n)
7424 .k(k)
7425 .cm_stride(19)
7426 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007427 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007428 }
7429 }
7430 }
7431 }
7432
7433 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, a_offset) {
7434 TEST_REQUIRES_ARM_NEON;
7435 for (size_t k = 1; k <= 40; k += 9) {
7436 GemmMicrokernelTester()
7437 .mr(4)
7438 .nr(16)
7439 .kr(2)
7440 .sr(1)
7441 .m(4)
7442 .n(16)
7443 .k(k)
7444 .ks(3)
7445 .a_offset(163)
Marat Dukhan50323b82022-01-11 00:12:01 -08007446 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007447 }
7448 }
7449
7450 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, zero) {
7451 TEST_REQUIRES_ARM_NEON;
7452 for (uint32_t mz = 0; mz < 4; mz++) {
7453 for (size_t k = 1; k <= 40; k += 9) {
7454 GemmMicrokernelTester()
7455 .mr(4)
7456 .nr(16)
7457 .kr(2)
7458 .sr(1)
7459 .m(4)
7460 .n(16)
7461 .k(k)
7462 .ks(3)
7463 .a_offset(163)
7464 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08007465 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007466 }
7467 }
7468 }
7469
7470 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, qmin) {
7471 TEST_REQUIRES_ARM_NEON;
7472 GemmMicrokernelTester()
7473 .mr(4)
7474 .nr(16)
7475 .kr(2)
7476 .sr(1)
7477 .m(4)
7478 .n(16)
7479 .k(8)
7480 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08007481 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007482 }
7483
7484 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, qmax) {
7485 TEST_REQUIRES_ARM_NEON;
7486 GemmMicrokernelTester()
7487 .mr(4)
7488 .nr(16)
7489 .kr(2)
7490 .sr(1)
7491 .m(4)
7492 .n(16)
7493 .k(8)
7494 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08007495 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007496 }
7497
7498 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_LD4R, strided_cm) {
7499 TEST_REQUIRES_ARM_NEON;
7500 GemmMicrokernelTester()
7501 .mr(4)
7502 .nr(16)
7503 .kr(2)
7504 .sr(1)
7505 .m(4)
7506 .n(16)
7507 .k(8)
7508 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08007509 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007510 }
7511#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
7512
7513
7514#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barchard42f5c502021-11-16 10:04:21 -08007515 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, k_eq_16) {
7516 TEST_REQUIRES_ARM_NEON;
7517 GemmMicrokernelTester()
7518 .mr(4)
7519 .nr(16)
7520 .kr(2)
7521 .sr(1)
7522 .m(4)
7523 .n(16)
7524 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -08007525 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007526 }
7527
7528 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, strided_cn) {
7529 TEST_REQUIRES_ARM_NEON;
7530 GemmMicrokernelTester()
7531 .mr(4)
7532 .nr(16)
7533 .kr(2)
7534 .sr(1)
7535 .m(4)
7536 .n(16)
7537 .k(16)
7538 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08007539 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007540 }
7541
7542 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, k_eq_16_subtile) {
7543 TEST_REQUIRES_ARM_NEON;
7544 for (uint32_t m = 1; m <= 4; m++) {
7545 for (uint32_t n = 1; n <= 16; n++) {
7546 GemmMicrokernelTester()
7547 .mr(4)
7548 .nr(16)
7549 .kr(2)
7550 .sr(1)
7551 .m(m)
7552 .n(n)
7553 .k(16)
7554 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007555 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007556 }
7557 }
7558 }
7559
7560 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7568 .m(m)
7569 .n(16)
7570 .k(16)
7571 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007572 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007573 }
7574 }
7575
7576 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7584 .m(4)
7585 .n(n)
7586 .k(16)
7587 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007588 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007589 }
7590 }
7591
7592 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7600 .m(4)
7601 .n(16)
7602 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007603 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007604 }
7605 }
7606
7607 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, k_lt_16_subtile) {
7608 TEST_REQUIRES_ARM_NEON;
7609 for (size_t k = 1; k < 16; k++) {
7610 for (uint32_t m = 1; m <= 4; m++) {
7611 for (uint32_t n = 1; n <= 16; n++) {
7612 GemmMicrokernelTester()
7613 .mr(4)
7614 .nr(16)
7615 .kr(2)
7616 .sr(1)
7617 .m(m)
7618 .n(n)
7619 .k(k)
7620 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007621 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007622 }
7623 }
7624 }
7625 }
7626
7627 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7635 .m(4)
7636 .n(16)
7637 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007638 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007639 }
7640 }
7641
7642 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, k_gt_16_subtile) {
7643 TEST_REQUIRES_ARM_NEON;
7644 for (size_t k = 17; k < 32; k++) {
7645 for (uint32_t m = 1; m <= 4; m++) {
7646 for (uint32_t n = 1; n <= 16; n++) {
7647 GemmMicrokernelTester()
7648 .mr(4)
7649 .nr(16)
7650 .kr(2)
7651 .sr(1)
7652 .m(m)
7653 .n(n)
7654 .k(k)
7655 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007656 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007657 }
7658 }
7659 }
7660 }
7661
7662 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7670 .m(4)
7671 .n(16)
7672 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007673 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007674 }
7675 }
7676
7677 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, k_div_16_subtile) {
7678 TEST_REQUIRES_ARM_NEON;
7679 for (size_t k = 32; k <= 160; k += 16) {
7680 for (uint32_t m = 1; m <= 4; m++) {
7681 for (uint32_t n = 1; n <= 16; n++) {
7682 GemmMicrokernelTester()
7683 .mr(4)
7684 .nr(16)
7685 .kr(2)
7686 .sr(1)
7687 .m(m)
7688 .n(n)
7689 .k(k)
7690 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007691 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007692 }
7693 }
7694 }
7695 }
7696
7697 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7706 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007707 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007708 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007709 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007710 }
7711 }
7712 }
7713
7714 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7723 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007724 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007725 .k(k)
7726 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08007727 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007728 }
7729 }
7730 }
7731
7732 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7742 .m(m)
7743 .n(n)
7744 .k(k)
7745 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007746 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007747 }
7748 }
7749 }
7750 }
7751
7752 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7761 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007762 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007763 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08007764 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007765 }
7766 }
7767 }
7768
7769 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7778 .m(4)
7779 .n(n)
7780 .k(k)
7781 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08007782 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007783 }
7784 }
7785 }
7786
7787 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7797 .m(m)
7798 .n(n)
7799 .k(k)
7800 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007801 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007802 }
7803 }
7804 }
7805 }
7806
7807 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7815 .m(4)
7816 .n(16)
7817 .k(k)
7818 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08007819 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007820 }
7821 }
7822
7823 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, small_kernel_subtile) {
7824 TEST_REQUIRES_ARM_NEON;
7825 for (size_t k = 1; k <= 80; k += 17) {
7826 for (uint32_t m = 1; m <= 4; m++) {
7827 for (uint32_t n = 1; n <= 16; n++) {
7828 GemmMicrokernelTester()
7829 .mr(4)
7830 .nr(16)
7831 .kr(2)
7832 .sr(1)
7833 .m(m)
7834 .n(n)
7835 .k(k)
7836 .ks(3)
7837 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007838 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007839 }
7840 }
7841 }
7842 }
7843
7844 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7853 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007854 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007855 .k(k)
7856 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08007857 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007858 }
7859 }
7860 }
7861
7862 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7871 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08007872 .n(n)
Frank Barchard42f5c502021-11-16 10:04:21 -08007873 .k(k)
7874 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08007875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007876 }
7877 }
7878 }
7879
7880 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, strided_cm_subtile) {
7881 TEST_REQUIRES_ARM_NEON;
7882 for (size_t k = 1; k <= 80; k += 17) {
7883 for (uint32_t m = 1; m <= 4; m++) {
7884 for (uint32_t n = 1; n <= 16; n++) {
7885 GemmMicrokernelTester()
7886 .mr(4)
7887 .nr(16)
7888 .kr(2)
7889 .sr(1)
7890 .m(m)
7891 .n(n)
7892 .k(k)
7893 .cm_stride(19)
7894 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08007895 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007896 }
7897 }
7898 }
7899 }
7900
7901 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, 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(1)
7909 .m(4)
7910 .n(16)
7911 .k(k)
7912 .ks(3)
7913 .a_offset(331)
Marat Dukhan50323b82022-01-11 00:12:01 -08007914 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007915 }
7916 }
7917
7918 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, zero) {
7919 TEST_REQUIRES_ARM_NEON;
7920 for (uint32_t mz = 0; mz < 4; mz++) {
7921 for (size_t k = 1; k <= 80; k += 17) {
7922 GemmMicrokernelTester()
7923 .mr(4)
7924 .nr(16)
7925 .kr(2)
7926 .sr(1)
7927 .m(4)
7928 .n(16)
7929 .k(k)
7930 .ks(3)
7931 .a_offset(331)
7932 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08007933 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007934 }
7935 }
7936 }
7937
7938 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, qmin) {
7939 TEST_REQUIRES_ARM_NEON;
7940 GemmMicrokernelTester()
7941 .mr(4)
7942 .nr(16)
7943 .kr(2)
7944 .sr(1)
7945 .m(4)
7946 .n(16)
7947 .k(16)
7948 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08007949 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007950 }
7951
7952 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, qmax) {
7953 TEST_REQUIRES_ARM_NEON;
7954 GemmMicrokernelTester()
7955 .mr(4)
7956 .nr(16)
7957 .kr(2)
7958 .sr(1)
7959 .m(4)
7960 .n(16)
7961 .k(16)
7962 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08007963 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007964 }
7965
7966 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD4R, strided_cm) {
7967 TEST_REQUIRES_ARM_NEON;
7968 GemmMicrokernelTester()
7969 .mr(4)
7970 .nr(16)
7971 .kr(2)
7972 .sr(1)
7973 .m(4)
7974 .n(16)
7975 .k(16)
7976 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08007977 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard42f5c502021-11-16 10:04:21 -08007978 }
7979#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
7980
7981
7982#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -08007983 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8) {
7984 TEST_REQUIRES_ARM_NEON;
7985 GemmMicrokernelTester()
7986 .mr(1)
7987 .nr(8)
7988 .kr(4)
7989 .sr(2)
7990 .m(1)
7991 .n(8)
7992 .k(8)
7993 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7994 }
7995
7996 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, strided_cn) {
7997 TEST_REQUIRES_ARM_NEON;
7998 GemmMicrokernelTester()
7999 .mr(1)
8000 .nr(8)
8001 .kr(4)
8002 .sr(2)
8003 .m(1)
8004 .n(8)
8005 .k(8)
8006 .cn_stride(11)
8007 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8008 }
8009
8010 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_subtile) {
8011 TEST_REQUIRES_ARM_NEON;
8012 for (uint32_t m = 1; m <= 1; m++) {
8013 for (uint32_t n = 1; n <= 8; n++) {
8014 GemmMicrokernelTester()
8015 .mr(1)
8016 .nr(8)
8017 .kr(4)
8018 .sr(2)
8019 .m(m)
8020 .n(n)
8021 .k(8)
8022 .iterations(1)
8023 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8024 }
8025 }
8026 }
8027
8028 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_subtile_m) {
8029 TEST_REQUIRES_ARM_NEON;
8030 for (uint32_t m = 1; m <= 1; m++) {
8031 GemmMicrokernelTester()
8032 .mr(1)
8033 .nr(8)
8034 .kr(4)
8035 .sr(2)
8036 .m(m)
8037 .n(8)
8038 .k(8)
8039 .iterations(1)
8040 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8041 }
8042 }
8043
8044 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_subtile_n) {
8045 TEST_REQUIRES_ARM_NEON;
8046 for (uint32_t n = 1; n <= 8; n++) {
8047 GemmMicrokernelTester()
8048 .mr(1)
8049 .nr(8)
8050 .kr(4)
8051 .sr(2)
8052 .m(1)
8053 .n(n)
8054 .k(8)
8055 .iterations(1)
8056 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8057 }
8058 }
8059
8060 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_lt_8) {
8061 TEST_REQUIRES_ARM_NEON;
8062 for (size_t k = 1; k < 8; k++) {
8063 GemmMicrokernelTester()
8064 .mr(1)
8065 .nr(8)
8066 .kr(4)
8067 .sr(2)
8068 .m(1)
8069 .n(8)
8070 .k(k)
8071 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8072 }
8073 }
8074
8075 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_lt_8_subtile) {
8076 TEST_REQUIRES_ARM_NEON;
8077 for (size_t k = 1; k < 8; k++) {
8078 for (uint32_t m = 1; m <= 1; m++) {
8079 for (uint32_t n = 1; n <= 8; n++) {
8080 GemmMicrokernelTester()
8081 .mr(1)
8082 .nr(8)
8083 .kr(4)
8084 .sr(2)
8085 .m(m)
8086 .n(n)
8087 .k(k)
8088 .iterations(1)
8089 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8090 }
8091 }
8092 }
8093 }
8094
8095 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_gt_8) {
8096 TEST_REQUIRES_ARM_NEON;
8097 for (size_t k = 9; k < 16; k++) {
8098 GemmMicrokernelTester()
8099 .mr(1)
8100 .nr(8)
8101 .kr(4)
8102 .sr(2)
8103 .m(1)
8104 .n(8)
8105 .k(k)
8106 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8107 }
8108 }
8109
8110 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_gt_8_subtile) {
8111 TEST_REQUIRES_ARM_NEON;
8112 for (size_t k = 9; k < 16; k++) {
8113 for (uint32_t m = 1; m <= 1; m++) {
8114 for (uint32_t n = 1; n <= 8; n++) {
8115 GemmMicrokernelTester()
8116 .mr(1)
8117 .nr(8)
8118 .kr(4)
8119 .sr(2)
8120 .m(m)
8121 .n(n)
8122 .k(k)
8123 .iterations(1)
8124 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8125 }
8126 }
8127 }
8128 }
8129
8130 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_div_8) {
8131 TEST_REQUIRES_ARM_NEON;
8132 for (size_t k = 16; k <= 80; k += 8) {
8133 GemmMicrokernelTester()
8134 .mr(1)
8135 .nr(8)
8136 .kr(4)
8137 .sr(2)
8138 .m(1)
8139 .n(8)
8140 .k(k)
8141 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8142 }
8143 }
8144
8145 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_div_8_subtile) {
8146 TEST_REQUIRES_ARM_NEON;
8147 for (size_t k = 16; k <= 80; k += 8) {
8148 for (uint32_t m = 1; m <= 1; m++) {
8149 for (uint32_t n = 1; n <= 8; n++) {
8150 GemmMicrokernelTester()
8151 .mr(1)
8152 .nr(8)
8153 .kr(4)
8154 .sr(2)
8155 .m(m)
8156 .n(n)
8157 .k(k)
8158 .iterations(1)
8159 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8160 }
8161 }
8162 }
8163 }
8164
8165 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, 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(1)
8171 .nr(8)
8172 .kr(4)
8173 .sr(2)
8174 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008175 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008176 .k(k)
8177 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8178 }
8179 }
8180 }
8181
8182 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, 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(1)
8188 .nr(8)
8189 .kr(4)
8190 .sr(2)
8191 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008192 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008193 .k(k)
8194 .cn_stride(11)
8195 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8196 }
8197 }
8198 }
8199
8200 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, 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 <= 1; m++) {
8205 GemmMicrokernelTester()
8206 .mr(1)
8207 .nr(8)
8208 .kr(4)
8209 .sr(2)
8210 .m(m)
8211 .n(n)
8212 .k(k)
8213 .iterations(1)
8214 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8215 }
8216 }
8217 }
8218 }
8219
8220 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, 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(1)
8226 .nr(8)
8227 .kr(4)
8228 .sr(2)
8229 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008230 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008231 .k(k)
8232 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8233 }
8234 }
8235 }
8236
8237 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, 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(1)
8243 .nr(8)
8244 .kr(4)
8245 .sr(2)
8246 .m(1)
8247 .n(n)
8248 .k(k)
8249 .cn_stride(11)
8250 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8251 }
8252 }
8253 }
8254
8255 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, 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 <= 1; m++) {
8260 GemmMicrokernelTester()
8261 .mr(1)
8262 .nr(8)
8263 .kr(4)
8264 .sr(2)
8265 .m(m)
8266 .n(n)
8267 .k(k)
8268 .iterations(1)
8269 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8270 }
8271 }
8272 }
8273 }
8274
8275 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, small_kernel) {
8276 TEST_REQUIRES_ARM_NEON;
8277 for (size_t k = 1; k <= 40; k += 9) {
8278 GemmMicrokernelTester()
8279 .mr(1)
8280 .nr(8)
8281 .kr(4)
8282 .sr(2)
8283 .m(1)
8284 .n(8)
8285 .k(k)
8286 .ks(3)
8287 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8288 }
8289 }
8290
8291 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, small_kernel_subtile) {
8292 TEST_REQUIRES_ARM_NEON;
8293 for (size_t k = 1; k <= 40; k += 9) {
8294 for (uint32_t m = 1; m <= 1; m++) {
8295 for (uint32_t n = 1; n <= 8; n++) {
8296 GemmMicrokernelTester()
8297 .mr(1)
8298 .nr(8)
8299 .kr(4)
8300 .sr(2)
8301 .m(m)
8302 .n(n)
8303 .k(k)
8304 .ks(3)
8305 .iterations(1)
8306 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8307 }
8308 }
8309 }
8310 }
8311
8312 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, 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(1)
8318 .nr(8)
8319 .kr(4)
8320 .sr(2)
8321 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008322 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008323 .k(k)
8324 .ks(3)
8325 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8326 }
8327 }
8328 }
8329
8330 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, 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(1)
8336 .nr(8)
8337 .kr(4)
8338 .sr(2)
8339 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008340 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008341 .k(k)
8342 .ks(3)
8343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8344 }
8345 }
8346 }
8347
8348 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, strided_cm_subtile) {
8349 TEST_REQUIRES_ARM_NEON;
8350 for (size_t k = 1; k <= 40; k += 9) {
8351 for (uint32_t m = 1; m <= 1; m++) {
8352 for (uint32_t n = 1; n <= 8; n++) {
8353 GemmMicrokernelTester()
8354 .mr(1)
8355 .nr(8)
8356 .kr(4)
8357 .sr(2)
8358 .m(m)
8359 .n(n)
8360 .k(k)
8361 .cm_stride(11)
8362 .iterations(1)
8363 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8364 }
8365 }
8366 }
8367 }
8368
8369 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, a_offset) {
8370 TEST_REQUIRES_ARM_NEON;
8371 for (size_t k = 1; k <= 40; k += 9) {
8372 GemmMicrokernelTester()
8373 .mr(1)
8374 .nr(8)
8375 .kr(4)
8376 .sr(2)
8377 .m(1)
8378 .n(8)
8379 .k(k)
8380 .ks(3)
8381 .a_offset(43)
8382 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8383 }
8384 }
8385
8386 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, zero) {
8387 TEST_REQUIRES_ARM_NEON;
8388 for (uint32_t mz = 0; mz < 1; mz++) {
8389 for (size_t k = 1; k <= 40; k += 9) {
8390 GemmMicrokernelTester()
8391 .mr(1)
8392 .nr(8)
8393 .kr(4)
8394 .sr(2)
8395 .m(1)
8396 .n(8)
8397 .k(k)
8398 .ks(3)
8399 .a_offset(43)
8400 .zero_index(mz)
8401 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8402 }
8403 }
8404 }
8405
8406 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, qmin) {
8407 TEST_REQUIRES_ARM_NEON;
8408 GemmMicrokernelTester()
8409 .mr(1)
8410 .nr(8)
8411 .kr(4)
8412 .sr(2)
8413 .m(1)
8414 .n(8)
8415 .k(8)
8416 .qmin(128)
8417 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8418 }
8419
8420 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, qmax) {
8421 TEST_REQUIRES_ARM_NEON;
8422 GemmMicrokernelTester()
8423 .mr(1)
8424 .nr(8)
8425 .kr(4)
8426 .sr(2)
8427 .m(1)
8428 .n(8)
8429 .k(8)
8430 .qmax(128)
8431 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8432 }
8433
8434 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, strided_cm) {
8435 TEST_REQUIRES_ARM_NEON;
8436 GemmMicrokernelTester()
8437 .mr(1)
8438 .nr(8)
8439 .kr(4)
8440 .sr(2)
8441 .m(1)
8442 .n(8)
8443 .k(8)
8444 .cm_stride(11)
8445 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, 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
Frank Barcharde22685a2021-11-12 11:36:58 -08008451 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_eq_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008452 TEST_REQUIRES_ARM_NEON;
8453 GemmMicrokernelTester()
8454 .mr(3)
8455 .nr(8)
8456 .kr(4)
8457 .sr(2)
8458 .m(3)
8459 .n(8)
8460 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -08008461 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008462 }
8463
Frank Barcharde22685a2021-11-12 11:36:58 -08008464 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008465 TEST_REQUIRES_ARM_NEON;
8466 GemmMicrokernelTester()
8467 .mr(3)
8468 .nr(8)
8469 .kr(4)
8470 .sr(2)
8471 .m(3)
8472 .n(8)
8473 .k(8)
8474 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08008475 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008476 }
8477
Frank Barcharde22685a2021-11-12 11:36:58 -08008478 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_eq_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008479 TEST_REQUIRES_ARM_NEON;
8480 for (uint32_t m = 1; m <= 3; m++) {
8481 for (uint32_t n = 1; n <= 8; n++) {
8482 GemmMicrokernelTester()
8483 .mr(3)
8484 .nr(8)
8485 .kr(4)
8486 .sr(2)
8487 .m(m)
8488 .n(n)
8489 .k(8)
8490 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008491 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008492 }
8493 }
8494 }
8495
Frank Barcharde22685a2021-11-12 11:36:58 -08008496 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_eq_8_subtile_m) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008497 TEST_REQUIRES_ARM_NEON;
8498 for (uint32_t m = 1; m <= 3; m++) {
8499 GemmMicrokernelTester()
8500 .mr(3)
8501 .nr(8)
8502 .kr(4)
8503 .sr(2)
8504 .m(m)
8505 .n(8)
8506 .k(8)
8507 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008508 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008509 }
8510 }
8511
Frank Barcharde22685a2021-11-12 11:36:58 -08008512 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_eq_8_subtile_n) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008513 TEST_REQUIRES_ARM_NEON;
8514 for (uint32_t n = 1; n <= 8; n++) {
8515 GemmMicrokernelTester()
8516 .mr(3)
8517 .nr(8)
8518 .kr(4)
8519 .sr(2)
8520 .m(3)
8521 .n(n)
8522 .k(8)
8523 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008524 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008525 }
8526 }
8527
Frank Barcharde22685a2021-11-12 11:36:58 -08008528 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_lt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008529 TEST_REQUIRES_ARM_NEON;
8530 for (size_t k = 1; k < 8; k++) {
8531 GemmMicrokernelTester()
8532 .mr(3)
8533 .nr(8)
8534 .kr(4)
8535 .sr(2)
8536 .m(3)
8537 .n(8)
8538 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08008539 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008540 }
8541 }
8542
Frank Barcharde22685a2021-11-12 11:36:58 -08008543 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_lt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008544 TEST_REQUIRES_ARM_NEON;
8545 for (size_t k = 1; k < 8; k++) {
8546 for (uint32_t m = 1; m <= 3; m++) {
8547 for (uint32_t n = 1; n <= 8; n++) {
8548 GemmMicrokernelTester()
8549 .mr(3)
8550 .nr(8)
8551 .kr(4)
8552 .sr(2)
8553 .m(m)
8554 .n(n)
8555 .k(k)
8556 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008557 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008558 }
8559 }
8560 }
8561 }
8562
Frank Barcharde22685a2021-11-12 11:36:58 -08008563 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_gt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008564 TEST_REQUIRES_ARM_NEON;
8565 for (size_t k = 9; k < 16; k++) {
8566 GemmMicrokernelTester()
8567 .mr(3)
8568 .nr(8)
8569 .kr(4)
8570 .sr(2)
8571 .m(3)
8572 .n(8)
8573 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08008574 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008575 }
8576 }
8577
Frank Barcharde22685a2021-11-12 11:36:58 -08008578 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_gt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008579 TEST_REQUIRES_ARM_NEON;
8580 for (size_t k = 9; k < 16; k++) {
8581 for (uint32_t m = 1; m <= 3; m++) {
8582 for (uint32_t n = 1; n <= 8; n++) {
8583 GemmMicrokernelTester()
8584 .mr(3)
8585 .nr(8)
8586 .kr(4)
8587 .sr(2)
8588 .m(m)
8589 .n(n)
8590 .k(k)
8591 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008592 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008593 }
8594 }
8595 }
8596 }
8597
Frank Barcharde22685a2021-11-12 11:36:58 -08008598 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_div_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008599 TEST_REQUIRES_ARM_NEON;
8600 for (size_t k = 16; k <= 80; k += 8) {
8601 GemmMicrokernelTester()
8602 .mr(3)
8603 .nr(8)
8604 .kr(4)
8605 .sr(2)
8606 .m(3)
8607 .n(8)
8608 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08008609 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008610 }
8611 }
8612
Frank Barcharde22685a2021-11-12 11:36:58 -08008613 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, k_div_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008614 TEST_REQUIRES_ARM_NEON;
8615 for (size_t k = 16; k <= 80; k += 8) {
8616 for (uint32_t m = 1; m <= 3; m++) {
8617 for (uint32_t n = 1; n <= 8; n++) {
8618 GemmMicrokernelTester()
8619 .mr(3)
8620 .nr(8)
8621 .kr(4)
8622 .sr(2)
8623 .m(m)
8624 .n(n)
8625 .k(k)
8626 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008627 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008628 }
8629 }
8630 }
8631 }
8632
Frank Barcharde22685a2021-11-12 11:36:58 -08008633 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, n_gt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008634 TEST_REQUIRES_ARM_NEON;
8635 for (uint32_t n = 9; n < 16; n++) {
8636 for (size_t k = 1; k <= 40; k += 9) {
8637 GemmMicrokernelTester()
8638 .mr(3)
8639 .nr(8)
8640 .kr(4)
8641 .sr(2)
8642 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008643 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08008644 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08008645 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008646 }
8647 }
8648 }
8649
Frank Barcharde22685a2021-11-12 11:36:58 -08008650 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, n_gt_8_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008651 TEST_REQUIRES_ARM_NEON;
8652 for (uint32_t n = 9; n < 16; n++) {
8653 for (size_t k = 1; k <= 40; k += 9) {
8654 GemmMicrokernelTester()
8655 .mr(3)
8656 .nr(8)
8657 .kr(4)
8658 .sr(2)
8659 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008660 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08008661 .k(k)
8662 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08008663 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008664 }
8665 }
8666 }
8667
Frank Barcharde22685a2021-11-12 11:36:58 -08008668 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, n_gt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008669 TEST_REQUIRES_ARM_NEON;
8670 for (uint32_t n = 9; n < 16; n++) {
8671 for (size_t k = 1; k <= 40; k += 9) {
8672 for (uint32_t m = 1; m <= 3; m++) {
8673 GemmMicrokernelTester()
8674 .mr(3)
8675 .nr(8)
8676 .kr(4)
8677 .sr(2)
8678 .m(m)
8679 .n(n)
8680 .k(k)
8681 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008682 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008683 }
8684 }
8685 }
8686 }
8687
Frank Barcharde22685a2021-11-12 11:36:58 -08008688 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, n_div_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008689 TEST_REQUIRES_ARM_NEON;
8690 for (uint32_t n = 16; n <= 24; n += 8) {
8691 for (size_t k = 1; k <= 40; k += 9) {
8692 GemmMicrokernelTester()
8693 .mr(3)
8694 .nr(8)
8695 .kr(4)
8696 .sr(2)
8697 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008698 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08008699 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08008700 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008701 }
8702 }
8703 }
8704
Frank Barcharde22685a2021-11-12 11:36:58 -08008705 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, n_div_8_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008706 TEST_REQUIRES_ARM_NEON;
8707 for (uint32_t n = 16; n <= 24; n += 8) {
8708 for (size_t k = 1; k <= 40; k += 9) {
8709 GemmMicrokernelTester()
8710 .mr(3)
8711 .nr(8)
8712 .kr(4)
8713 .sr(2)
8714 .m(3)
8715 .n(n)
8716 .k(k)
8717 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08008718 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008719 }
8720 }
8721 }
8722
Frank Barcharde22685a2021-11-12 11:36:58 -08008723 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, n_div_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008724 TEST_REQUIRES_ARM_NEON;
8725 for (uint32_t n = 16; n <= 24; n += 8) {
8726 for (size_t k = 1; k <= 40; k += 9) {
8727 for (uint32_t m = 1; m <= 3; m++) {
8728 GemmMicrokernelTester()
8729 .mr(3)
8730 .nr(8)
8731 .kr(4)
8732 .sr(2)
8733 .m(m)
8734 .n(n)
8735 .k(k)
8736 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008737 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008738 }
8739 }
8740 }
8741 }
8742
Frank Barcharde22685a2021-11-12 11:36:58 -08008743 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008744 TEST_REQUIRES_ARM_NEON;
8745 for (size_t k = 1; k <= 40; k += 9) {
8746 GemmMicrokernelTester()
8747 .mr(3)
8748 .nr(8)
8749 .kr(4)
8750 .sr(2)
8751 .m(3)
8752 .n(8)
8753 .k(k)
8754 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08008755 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008756 }
8757 }
8758
Frank Barcharde22685a2021-11-12 11:36:58 -08008759 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, small_kernel_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008760 TEST_REQUIRES_ARM_NEON;
8761 for (size_t k = 1; k <= 40; k += 9) {
8762 for (uint32_t m = 1; m <= 3; m++) {
8763 for (uint32_t n = 1; n <= 8; n++) {
8764 GemmMicrokernelTester()
8765 .mr(3)
8766 .nr(8)
8767 .kr(4)
8768 .sr(2)
8769 .m(m)
8770 .n(n)
8771 .k(k)
8772 .ks(3)
8773 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008774 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008775 }
8776 }
8777 }
8778 }
8779
Frank Barcharde22685a2021-11-12 11:36:58 -08008780 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, n_gt_8_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008781 TEST_REQUIRES_ARM_NEON;
8782 for (uint32_t n = 9; n < 16; n++) {
8783 for (size_t k = 1; k <= 40; k += 9) {
8784 GemmMicrokernelTester()
8785 .mr(3)
8786 .nr(8)
8787 .kr(4)
8788 .sr(2)
8789 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008790 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08008791 .k(k)
8792 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08008793 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008794 }
8795 }
8796 }
8797
Frank Barcharde22685a2021-11-12 11:36:58 -08008798 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, n_div_8_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008799 TEST_REQUIRES_ARM_NEON;
8800 for (uint32_t n = 16; n <= 24; n += 8) {
8801 for (size_t k = 1; k <= 40; k += 9) {
8802 GemmMicrokernelTester()
8803 .mr(3)
8804 .nr(8)
8805 .kr(4)
8806 .sr(2)
8807 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08008808 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08008809 .k(k)
8810 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08008811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008812 }
8813 }
8814 }
8815
Frank Barcharde22685a2021-11-12 11:36:58 -08008816 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, strided_cm_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008817 TEST_REQUIRES_ARM_NEON;
8818 for (size_t k = 1; k <= 40; k += 9) {
8819 for (uint32_t m = 1; m <= 3; m++) {
8820 for (uint32_t n = 1; n <= 8; n++) {
8821 GemmMicrokernelTester()
8822 .mr(3)
8823 .nr(8)
8824 .kr(4)
8825 .sr(2)
8826 .m(m)
8827 .n(n)
8828 .k(k)
8829 .cm_stride(11)
8830 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08008831 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008832 }
8833 }
8834 }
8835 }
8836
Frank Barcharde22685a2021-11-12 11:36:58 -08008837 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, a_offset) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008838 TEST_REQUIRES_ARM_NEON;
8839 for (size_t k = 1; k <= 40; k += 9) {
8840 GemmMicrokernelTester()
8841 .mr(3)
8842 .nr(8)
8843 .kr(4)
8844 .sr(2)
8845 .m(3)
8846 .n(8)
8847 .k(k)
8848 .ks(3)
8849 .a_offset(127)
Marat Dukhan50323b82022-01-11 00:12:01 -08008850 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008851 }
8852 }
8853
Frank Barcharde22685a2021-11-12 11:36:58 -08008854 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, zero) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008855 TEST_REQUIRES_ARM_NEON;
8856 for (uint32_t mz = 0; mz < 3; mz++) {
8857 for (size_t k = 1; k <= 40; k += 9) {
8858 GemmMicrokernelTester()
8859 .mr(3)
8860 .nr(8)
8861 .kr(4)
8862 .sr(2)
8863 .m(3)
8864 .n(8)
8865 .k(k)
8866 .ks(3)
8867 .a_offset(127)
8868 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08008869 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008870 }
8871 }
8872 }
8873
Frank Barcharde22685a2021-11-12 11:36:58 -08008874 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, qmin) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008875 TEST_REQUIRES_ARM_NEON;
8876 GemmMicrokernelTester()
8877 .mr(3)
8878 .nr(8)
8879 .kr(4)
8880 .sr(2)
8881 .m(3)
8882 .n(8)
8883 .k(8)
8884 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08008885 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008886 }
8887
Frank Barcharde22685a2021-11-12 11:36:58 -08008888 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, qmax) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008889 TEST_REQUIRES_ARM_NEON;
8890 GemmMicrokernelTester()
8891 .mr(3)
8892 .nr(8)
8893 .kr(4)
8894 .sr(2)
8895 .m(3)
8896 .n(8)
8897 .k(8)
8898 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08008899 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008900 }
8901
Frank Barcharde22685a2021-11-12 11:36:58 -08008902 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4S2__NEON_MULL, strided_cm) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008903 TEST_REQUIRES_ARM_NEON;
8904 GemmMicrokernelTester()
8905 .mr(3)
8906 .nr(8)
8907 .kr(4)
8908 .sr(2)
8909 .m(3)
8910 .n(8)
8911 .k(8)
8912 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -08008913 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008914 }
8915#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
8916
8917
8918#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008919 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_eq_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008920 TEST_REQUIRES_ARM_NEON;
8921 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008922 .mr(1)
8923 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08008924 .kr(4)
8925 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008926 .m(1)
8927 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08008928 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008929 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008930 }
8931
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008932 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008933 TEST_REQUIRES_ARM_NEON;
8934 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008935 .mr(1)
8936 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08008937 .kr(4)
8938 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008939 .m(1)
8940 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08008941 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008942 .cn_stride(19)
8943 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008944 }
8945
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008946 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_eq_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008947 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008948 for (uint32_t m = 1; m <= 1; m++) {
8949 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008950 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008951 .mr(1)
8952 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08008953 .kr(4)
8954 .sr(2)
8955 .m(m)
8956 .n(n)
8957 .k(8)
8958 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008959 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008960 }
8961 }
8962 }
8963
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008964 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_eq_8_subtile_m) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008965 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008966 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008967 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008968 .mr(1)
8969 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08008970 .kr(4)
8971 .sr(2)
8972 .m(m)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008973 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08008974 .k(8)
8975 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008976 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008977 }
8978 }
8979
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008980 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_eq_8_subtile_n) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008981 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008982 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008983 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008984 .mr(1)
8985 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08008986 .kr(4)
8987 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008988 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -08008989 .n(n)
8990 .k(8)
8991 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008992 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08008993 }
8994 }
8995
Zhi An Ngc27f04b2022-01-11 09:34:07 -08008996 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_lt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08008997 TEST_REQUIRES_ARM_NEON;
8998 for (size_t k = 1; k < 8; k++) {
8999 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009000 .mr(1)
9001 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009002 .kr(4)
9003 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009004 .m(1)
9005 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009006 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009007 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009008 }
9009 }
9010
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009011 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_lt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009012 TEST_REQUIRES_ARM_NEON;
9013 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009014 for (uint32_t m = 1; m <= 1; m++) {
9015 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009016 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009017 .mr(1)
9018 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009019 .kr(4)
9020 .sr(2)
9021 .m(m)
9022 .n(n)
9023 .k(k)
9024 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009025 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009026 }
9027 }
9028 }
9029 }
9030
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009031 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_gt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009032 TEST_REQUIRES_ARM_NEON;
9033 for (size_t k = 9; k < 16; k++) {
9034 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009035 .mr(1)
9036 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009037 .kr(4)
9038 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009039 .m(1)
9040 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009041 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009042 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009043 }
9044 }
9045
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009046 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_gt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009047 TEST_REQUIRES_ARM_NEON;
9048 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009049 for (uint32_t m = 1; m <= 1; m++) {
9050 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009051 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009052 .mr(1)
9053 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009054 .kr(4)
9055 .sr(2)
9056 .m(m)
9057 .n(n)
9058 .k(k)
9059 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009060 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009061 }
9062 }
9063 }
9064 }
9065
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009066 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_div_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009067 TEST_REQUIRES_ARM_NEON;
9068 for (size_t k = 16; k <= 80; k += 8) {
9069 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009070 .mr(1)
9071 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009072 .kr(4)
9073 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009074 .m(1)
9075 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009076 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009077 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009078 }
9079 }
9080
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009081 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, k_div_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009082 TEST_REQUIRES_ARM_NEON;
9083 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009084 for (uint32_t m = 1; m <= 1; m++) {
9085 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009086 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009087 .mr(1)
9088 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009089 .kr(4)
9090 .sr(2)
9091 .m(m)
9092 .n(n)
9093 .k(k)
9094 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009095 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009096 }
9097 }
9098 }
9099 }
9100
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009101 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, n_gt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009102 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009103 for (uint32_t n = 17; n < 32; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009104 for (size_t k = 1; k <= 40; k += 9) {
9105 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009106 .mr(1)
9107 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009108 .kr(4)
9109 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009110 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009111 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009112 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009113 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009114 }
9115 }
9116 }
9117
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009118 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, n_gt_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009119 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009120 for (uint32_t n = 17; n < 32; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009121 for (size_t k = 1; k <= 40; k += 9) {
9122 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009123 .mr(1)
9124 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009125 .kr(4)
9126 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009127 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009128 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009129 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009130 .cn_stride(19)
9131 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009132 }
9133 }
9134 }
9135
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009136 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, n_gt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009137 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009138 for (uint32_t n = 17; n < 32; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009139 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009140 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009141 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009142 .mr(1)
9143 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009144 .kr(4)
9145 .sr(2)
9146 .m(m)
9147 .n(n)
9148 .k(k)
9149 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009150 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009151 }
9152 }
9153 }
9154 }
9155
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009156 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, n_div_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009157 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009158 for (uint32_t n = 32; n <= 48; n += 16) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009159 for (size_t k = 1; k <= 40; k += 9) {
9160 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009161 .mr(1)
9162 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009163 .kr(4)
9164 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009165 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009166 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009167 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009168 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009169 }
9170 }
9171 }
9172
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009173 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, n_div_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009174 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009175 for (uint32_t n = 32; n <= 48; n += 16) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009176 for (size_t k = 1; k <= 40; k += 9) {
9177 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009178 .mr(1)
9179 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009180 .kr(4)
9181 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009182 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -08009183 .n(n)
9184 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009185 .cn_stride(19)
9186 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009187 }
9188 }
9189 }
9190
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009191 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, n_div_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009192 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009193 for (uint32_t n = 32; n <= 48; n += 16) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009194 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009195 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009196 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009197 .mr(1)
9198 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009199 .kr(4)
9200 .sr(2)
9201 .m(m)
9202 .n(n)
9203 .k(k)
9204 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009205 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009206 }
9207 }
9208 }
9209 }
9210
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009211 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009212 TEST_REQUIRES_ARM_NEON;
9213 for (size_t k = 1; k <= 40; k += 9) {
9214 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009215 .mr(1)
9216 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009217 .kr(4)
9218 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009219 .m(1)
9220 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009221 .k(k)
9222 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009223 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009224 }
9225 }
9226
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009227 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, small_kernel_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009228 TEST_REQUIRES_ARM_NEON;
9229 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009230 for (uint32_t m = 1; m <= 1; m++) {
9231 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009232 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009233 .mr(1)
9234 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009235 .kr(4)
9236 .sr(2)
9237 .m(m)
9238 .n(n)
9239 .k(k)
9240 .ks(3)
9241 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009242 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009243 }
9244 }
9245 }
9246 }
9247
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009248 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, n_gt_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009249 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009250 for (uint32_t n = 17; n < 32; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009251 for (size_t k = 1; k <= 40; k += 9) {
9252 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009253 .mr(1)
9254 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009255 .kr(4)
9256 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009257 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009258 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009259 .k(k)
9260 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009261 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009262 }
9263 }
9264 }
9265
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009266 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, n_div_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009267 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009268 for (uint32_t n = 32; n <= 48; n += 16) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009269 for (size_t k = 1; k <= 40; k += 9) {
9270 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009271 .mr(1)
9272 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009273 .kr(4)
9274 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009275 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009276 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009277 .k(k)
9278 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009280 }
9281 }
9282 }
9283
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009284 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, strided_cm_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009285 TEST_REQUIRES_ARM_NEON;
9286 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009287 for (uint32_t m = 1; m <= 1; m++) {
9288 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009289 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009290 .mr(1)
9291 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009292 .kr(4)
9293 .sr(2)
9294 .m(m)
9295 .n(n)
9296 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009297 .cm_stride(19)
Frank Barchardeb704f72021-11-12 01:26:50 -08009298 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009299 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009300 }
9301 }
9302 }
9303 }
9304
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009305 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, a_offset) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009306 TEST_REQUIRES_ARM_NEON;
9307 for (size_t k = 1; k <= 40; k += 9) {
9308 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009309 .mr(1)
9310 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009311 .kr(4)
9312 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009313 .m(1)
9314 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009315 .k(k)
9316 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009317 .a_offset(43)
9318 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009319 }
9320 }
9321
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009322 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, zero) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009323 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009324 for (uint32_t mz = 0; mz < 1; mz++) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009325 for (size_t k = 1; k <= 40; k += 9) {
9326 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009327 .mr(1)
9328 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009329 .kr(4)
9330 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009331 .m(1)
9332 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009333 .k(k)
9334 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009335 .a_offset(43)
Frank Barchardeb704f72021-11-12 01:26:50 -08009336 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009337 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009338 }
9339 }
9340 }
9341
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009342 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, qmin) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009343 TEST_REQUIRES_ARM_NEON;
9344 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009345 .mr(1)
9346 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009347 .kr(4)
9348 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009349 .m(1)
9350 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009351 .k(8)
9352 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009353 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009354 }
9355
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009356 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, qmax) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009357 TEST_REQUIRES_ARM_NEON;
9358 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009359 .mr(1)
9360 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009361 .kr(4)
9362 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009363 .m(1)
9364 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009365 .k(8)
9366 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009367 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009368 }
9369
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009370 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MULL, strided_cm) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009371 TEST_REQUIRES_ARM_NEON;
9372 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009373 .mr(1)
9374 .nr(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009375 .kr(4)
9376 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009377 .m(1)
9378 .n(16)
Frank Barchardeb704f72021-11-12 01:26:50 -08009379 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -08009380 .cm_stride(19)
9381 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009382 }
9383#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
9384
9385
9386#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -08009387 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_eq_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009388 TEST_REQUIRES_ARM_NEON;
9389 GemmMicrokernelTester()
9390 .mr(2)
9391 .nr(16)
9392 .kr(4)
9393 .sr(2)
9394 .m(2)
9395 .n(16)
9396 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -08009397 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009398 }
9399
Frank Barcharde22685a2021-11-12 11:36:58 -08009400 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009401 TEST_REQUIRES_ARM_NEON;
9402 GemmMicrokernelTester()
9403 .mr(2)
9404 .nr(16)
9405 .kr(4)
9406 .sr(2)
9407 .m(2)
9408 .n(16)
9409 .k(8)
9410 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08009411 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009412 }
9413
Frank Barcharde22685a2021-11-12 11:36:58 -08009414 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_eq_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009415 TEST_REQUIRES_ARM_NEON;
9416 for (uint32_t m = 1; m <= 2; m++) {
9417 for (uint32_t n = 1; n <= 16; n++) {
9418 GemmMicrokernelTester()
9419 .mr(2)
9420 .nr(16)
9421 .kr(4)
9422 .sr(2)
9423 .m(m)
9424 .n(n)
9425 .k(8)
9426 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009427 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009428 }
9429 }
9430 }
9431
Frank Barcharde22685a2021-11-12 11:36:58 -08009432 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_eq_8_subtile_m) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009433 TEST_REQUIRES_ARM_NEON;
9434 for (uint32_t m = 1; m <= 2; m++) {
9435 GemmMicrokernelTester()
9436 .mr(2)
9437 .nr(16)
9438 .kr(4)
9439 .sr(2)
9440 .m(m)
9441 .n(16)
9442 .k(8)
9443 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009444 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009445 }
9446 }
9447
Frank Barcharde22685a2021-11-12 11:36:58 -08009448 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_eq_8_subtile_n) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009449 TEST_REQUIRES_ARM_NEON;
9450 for (uint32_t n = 1; n <= 16; n++) {
9451 GemmMicrokernelTester()
9452 .mr(2)
9453 .nr(16)
9454 .kr(4)
9455 .sr(2)
9456 .m(2)
9457 .n(n)
9458 .k(8)
9459 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009460 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009461 }
9462 }
9463
Frank Barcharde22685a2021-11-12 11:36:58 -08009464 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_lt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009465 TEST_REQUIRES_ARM_NEON;
9466 for (size_t k = 1; k < 8; k++) {
9467 GemmMicrokernelTester()
9468 .mr(2)
9469 .nr(16)
9470 .kr(4)
9471 .sr(2)
9472 .m(2)
9473 .n(16)
9474 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08009475 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009476 }
9477 }
9478
Frank Barcharde22685a2021-11-12 11:36:58 -08009479 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_lt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009480 TEST_REQUIRES_ARM_NEON;
9481 for (size_t k = 1; k < 8; k++) {
9482 for (uint32_t m = 1; m <= 2; m++) {
9483 for (uint32_t n = 1; n <= 16; n++) {
9484 GemmMicrokernelTester()
9485 .mr(2)
9486 .nr(16)
9487 .kr(4)
9488 .sr(2)
9489 .m(m)
9490 .n(n)
9491 .k(k)
9492 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009493 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009494 }
9495 }
9496 }
9497 }
9498
Frank Barcharde22685a2021-11-12 11:36:58 -08009499 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_gt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009500 TEST_REQUIRES_ARM_NEON;
9501 for (size_t k = 9; k < 16; k++) {
9502 GemmMicrokernelTester()
9503 .mr(2)
9504 .nr(16)
9505 .kr(4)
9506 .sr(2)
9507 .m(2)
9508 .n(16)
9509 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08009510 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009511 }
9512 }
9513
Frank Barcharde22685a2021-11-12 11:36:58 -08009514 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_gt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009515 TEST_REQUIRES_ARM_NEON;
9516 for (size_t k = 9; k < 16; k++) {
9517 for (uint32_t m = 1; m <= 2; m++) {
9518 for (uint32_t n = 1; n <= 16; n++) {
9519 GemmMicrokernelTester()
9520 .mr(2)
9521 .nr(16)
9522 .kr(4)
9523 .sr(2)
9524 .m(m)
9525 .n(n)
9526 .k(k)
9527 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009528 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009529 }
9530 }
9531 }
9532 }
9533
Frank Barcharde22685a2021-11-12 11:36:58 -08009534 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_div_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009535 TEST_REQUIRES_ARM_NEON;
9536 for (size_t k = 16; k <= 80; k += 8) {
9537 GemmMicrokernelTester()
9538 .mr(2)
9539 .nr(16)
9540 .kr(4)
9541 .sr(2)
9542 .m(2)
9543 .n(16)
9544 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08009545 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009546 }
9547 }
9548
Frank Barcharde22685a2021-11-12 11:36:58 -08009549 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, k_div_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009550 TEST_REQUIRES_ARM_NEON;
9551 for (size_t k = 16; k <= 80; k += 8) {
9552 for (uint32_t m = 1; m <= 2; m++) {
9553 for (uint32_t n = 1; n <= 16; n++) {
9554 GemmMicrokernelTester()
9555 .mr(2)
9556 .nr(16)
9557 .kr(4)
9558 .sr(2)
9559 .m(m)
9560 .n(n)
9561 .k(k)
9562 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009563 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009564 }
9565 }
9566 }
9567 }
9568
Frank Barcharde22685a2021-11-12 11:36:58 -08009569 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, n_gt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009570 TEST_REQUIRES_ARM_NEON;
9571 for (uint32_t n = 17; n < 32; n++) {
9572 for (size_t k = 1; k <= 40; k += 9) {
9573 GemmMicrokernelTester()
9574 .mr(2)
9575 .nr(16)
9576 .kr(4)
9577 .sr(2)
9578 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009579 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009580 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08009581 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009582 }
9583 }
9584 }
9585
Frank Barcharde22685a2021-11-12 11:36:58 -08009586 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, n_gt_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009587 TEST_REQUIRES_ARM_NEON;
9588 for (uint32_t n = 17; n < 32; n++) {
9589 for (size_t k = 1; k <= 40; k += 9) {
9590 GemmMicrokernelTester()
9591 .mr(2)
9592 .nr(16)
9593 .kr(4)
9594 .sr(2)
9595 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009596 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009597 .k(k)
9598 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08009599 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009600 }
9601 }
9602 }
9603
Frank Barcharde22685a2021-11-12 11:36:58 -08009604 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, n_gt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009605 TEST_REQUIRES_ARM_NEON;
9606 for (uint32_t n = 17; n < 32; n++) {
9607 for (size_t k = 1; k <= 40; k += 9) {
9608 for (uint32_t m = 1; m <= 2; m++) {
9609 GemmMicrokernelTester()
9610 .mr(2)
9611 .nr(16)
9612 .kr(4)
9613 .sr(2)
9614 .m(m)
9615 .n(n)
9616 .k(k)
9617 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009618 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009619 }
9620 }
9621 }
9622 }
9623
Frank Barcharde22685a2021-11-12 11:36:58 -08009624 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, n_div_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009625 TEST_REQUIRES_ARM_NEON;
9626 for (uint32_t n = 32; n <= 48; n += 16) {
9627 for (size_t k = 1; k <= 40; k += 9) {
9628 GemmMicrokernelTester()
9629 .mr(2)
9630 .nr(16)
9631 .kr(4)
9632 .sr(2)
9633 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009634 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009635 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08009636 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009637 }
9638 }
9639 }
9640
Frank Barcharde22685a2021-11-12 11:36:58 -08009641 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, n_div_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009642 TEST_REQUIRES_ARM_NEON;
9643 for (uint32_t n = 32; n <= 48; n += 16) {
9644 for (size_t k = 1; k <= 40; k += 9) {
9645 GemmMicrokernelTester()
9646 .mr(2)
9647 .nr(16)
9648 .kr(4)
9649 .sr(2)
9650 .m(2)
9651 .n(n)
9652 .k(k)
9653 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08009654 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009655 }
9656 }
9657 }
9658
Frank Barcharde22685a2021-11-12 11:36:58 -08009659 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, n_div_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009660 TEST_REQUIRES_ARM_NEON;
9661 for (uint32_t n = 32; n <= 48; n += 16) {
9662 for (size_t k = 1; k <= 40; k += 9) {
9663 for (uint32_t m = 1; m <= 2; m++) {
9664 GemmMicrokernelTester()
9665 .mr(2)
9666 .nr(16)
9667 .kr(4)
9668 .sr(2)
9669 .m(m)
9670 .n(n)
9671 .k(k)
9672 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009673 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009674 }
9675 }
9676 }
9677 }
9678
Frank Barcharde22685a2021-11-12 11:36:58 -08009679 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009680 TEST_REQUIRES_ARM_NEON;
9681 for (size_t k = 1; k <= 40; k += 9) {
9682 GemmMicrokernelTester()
9683 .mr(2)
9684 .nr(16)
9685 .kr(4)
9686 .sr(2)
9687 .m(2)
9688 .n(16)
9689 .k(k)
9690 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08009691 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009692 }
9693 }
9694
Frank Barcharde22685a2021-11-12 11:36:58 -08009695 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, small_kernel_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009696 TEST_REQUIRES_ARM_NEON;
9697 for (size_t k = 1; k <= 40; k += 9) {
9698 for (uint32_t m = 1; m <= 2; m++) {
9699 for (uint32_t n = 1; n <= 16; n++) {
9700 GemmMicrokernelTester()
9701 .mr(2)
9702 .nr(16)
9703 .kr(4)
9704 .sr(2)
9705 .m(m)
9706 .n(n)
9707 .k(k)
9708 .ks(3)
9709 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009710 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009711 }
9712 }
9713 }
9714 }
9715
Frank Barcharde22685a2021-11-12 11:36:58 -08009716 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, n_gt_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009717 TEST_REQUIRES_ARM_NEON;
9718 for (uint32_t n = 17; n < 32; n++) {
9719 for (size_t k = 1; k <= 40; k += 9) {
9720 GemmMicrokernelTester()
9721 .mr(2)
9722 .nr(16)
9723 .kr(4)
9724 .sr(2)
9725 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009726 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009727 .k(k)
9728 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08009729 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009730 }
9731 }
9732 }
9733
Frank Barcharde22685a2021-11-12 11:36:58 -08009734 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, n_div_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009735 TEST_REQUIRES_ARM_NEON;
9736 for (uint32_t n = 32; n <= 48; n += 16) {
9737 for (size_t k = 1; k <= 40; k += 9) {
9738 GemmMicrokernelTester()
9739 .mr(2)
9740 .nr(16)
9741 .kr(4)
9742 .sr(2)
9743 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -08009744 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -08009745 .k(k)
9746 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -08009747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009748 }
9749 }
9750 }
9751
Frank Barcharde22685a2021-11-12 11:36:58 -08009752 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, strided_cm_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009753 TEST_REQUIRES_ARM_NEON;
9754 for (size_t k = 1; k <= 40; k += 9) {
9755 for (uint32_t m = 1; m <= 2; m++) {
9756 for (uint32_t n = 1; n <= 16; n++) {
9757 GemmMicrokernelTester()
9758 .mr(2)
9759 .nr(16)
9760 .kr(4)
9761 .sr(2)
9762 .m(m)
9763 .n(n)
9764 .k(k)
9765 .cm_stride(19)
9766 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009767 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009768 }
9769 }
9770 }
9771 }
9772
Frank Barcharde22685a2021-11-12 11:36:58 -08009773 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, a_offset) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009774 TEST_REQUIRES_ARM_NEON;
9775 for (size_t k = 1; k <= 40; k += 9) {
9776 GemmMicrokernelTester()
9777 .mr(2)
9778 .nr(16)
9779 .kr(4)
9780 .sr(2)
9781 .m(2)
9782 .n(16)
9783 .k(k)
9784 .ks(3)
9785 .a_offset(83)
Marat Dukhan50323b82022-01-11 00:12:01 -08009786 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009787 }
9788 }
9789
Frank Barcharde22685a2021-11-12 11:36:58 -08009790 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, zero) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009791 TEST_REQUIRES_ARM_NEON;
9792 for (uint32_t mz = 0; mz < 2; mz++) {
9793 for (size_t k = 1; k <= 40; k += 9) {
9794 GemmMicrokernelTester()
9795 .mr(2)
9796 .nr(16)
9797 .kr(4)
9798 .sr(2)
9799 .m(2)
9800 .n(16)
9801 .k(k)
9802 .ks(3)
9803 .a_offset(83)
9804 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -08009805 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009806 }
9807 }
9808 }
9809
Frank Barcharde22685a2021-11-12 11:36:58 -08009810 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, qmin) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009811 TEST_REQUIRES_ARM_NEON;
9812 GemmMicrokernelTester()
9813 .mr(2)
9814 .nr(16)
9815 .kr(4)
9816 .sr(2)
9817 .m(2)
9818 .n(16)
9819 .k(8)
9820 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08009821 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009822 }
9823
Frank Barcharde22685a2021-11-12 11:36:58 -08009824 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, qmax) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009825 TEST_REQUIRES_ARM_NEON;
9826 GemmMicrokernelTester()
9827 .mr(2)
9828 .nr(16)
9829 .kr(4)
9830 .sr(2)
9831 .m(2)
9832 .n(16)
9833 .k(8)
9834 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -08009835 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009836 }
9837
Frank Barcharde22685a2021-11-12 11:36:58 -08009838 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4S2__NEON_MULL, strided_cm) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009839 TEST_REQUIRES_ARM_NEON;
9840 GemmMicrokernelTester()
9841 .mr(2)
9842 .nr(16)
9843 .kr(4)
9844 .sr(2)
9845 .m(2)
9846 .n(16)
9847 .k(8)
9848 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08009849 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009850 }
9851#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
9852
9853
9854#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -08009855 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_eq_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009856 TEST_REQUIRES_ARM_NEON;
9857 GemmMicrokernelTester()
9858 .mr(3)
9859 .nr(16)
9860 .kr(4)
9861 .sr(2)
9862 .m(3)
9863 .n(16)
9864 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -08009865 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009866 }
9867
Frank Barcharde22685a2021-11-12 11:36:58 -08009868 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009869 TEST_REQUIRES_ARM_NEON;
9870 GemmMicrokernelTester()
9871 .mr(3)
9872 .nr(16)
9873 .kr(4)
9874 .sr(2)
9875 .m(3)
9876 .n(16)
9877 .k(8)
9878 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -08009879 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009880 }
9881
Frank Barcharde22685a2021-11-12 11:36:58 -08009882 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_eq_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009883 TEST_REQUIRES_ARM_NEON;
9884 for (uint32_t m = 1; m <= 3; m++) {
9885 for (uint32_t n = 1; n <= 16; n++) {
9886 GemmMicrokernelTester()
9887 .mr(3)
9888 .nr(16)
9889 .kr(4)
9890 .sr(2)
9891 .m(m)
9892 .n(n)
9893 .k(8)
9894 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009895 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009896 }
9897 }
9898 }
9899
Frank Barcharde22685a2021-11-12 11:36:58 -08009900 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_eq_8_subtile_m) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009901 TEST_REQUIRES_ARM_NEON;
9902 for (uint32_t m = 1; m <= 3; m++) {
9903 GemmMicrokernelTester()
9904 .mr(3)
9905 .nr(16)
9906 .kr(4)
9907 .sr(2)
9908 .m(m)
9909 .n(16)
9910 .k(8)
9911 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009912 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009913 }
9914 }
9915
Frank Barcharde22685a2021-11-12 11:36:58 -08009916 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_eq_8_subtile_n) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009917 TEST_REQUIRES_ARM_NEON;
9918 for (uint32_t n = 1; n <= 16; n++) {
9919 GemmMicrokernelTester()
9920 .mr(3)
9921 .nr(16)
9922 .kr(4)
9923 .sr(2)
9924 .m(3)
9925 .n(n)
9926 .k(8)
9927 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009928 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009929 }
9930 }
9931
Frank Barcharde22685a2021-11-12 11:36:58 -08009932 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_lt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009933 TEST_REQUIRES_ARM_NEON;
9934 for (size_t k = 1; k < 8; k++) {
9935 GemmMicrokernelTester()
9936 .mr(3)
9937 .nr(16)
9938 .kr(4)
9939 .sr(2)
9940 .m(3)
9941 .n(16)
9942 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08009943 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009944 }
9945 }
9946
Frank Barcharde22685a2021-11-12 11:36:58 -08009947 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_lt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009948 TEST_REQUIRES_ARM_NEON;
9949 for (size_t k = 1; k < 8; k++) {
9950 for (uint32_t m = 1; m <= 3; m++) {
9951 for (uint32_t n = 1; n <= 16; n++) {
9952 GemmMicrokernelTester()
9953 .mr(3)
9954 .nr(16)
9955 .kr(4)
9956 .sr(2)
9957 .m(m)
9958 .n(n)
9959 .k(k)
9960 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009961 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009962 }
9963 }
9964 }
9965 }
9966
Frank Barcharde22685a2021-11-12 11:36:58 -08009967 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_gt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009968 TEST_REQUIRES_ARM_NEON;
9969 for (size_t k = 9; k < 16; k++) {
9970 GemmMicrokernelTester()
9971 .mr(3)
9972 .nr(16)
9973 .kr(4)
9974 .sr(2)
9975 .m(3)
9976 .n(16)
9977 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -08009978 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009979 }
9980 }
9981
Frank Barcharde22685a2021-11-12 11:36:58 -08009982 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_gt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -08009983 TEST_REQUIRES_ARM_NEON;
9984 for (size_t k = 9; k < 16; k++) {
9985 for (uint32_t m = 1; m <= 3; m++) {
9986 for (uint32_t n = 1; n <= 16; n++) {
9987 GemmMicrokernelTester()
9988 .mr(3)
9989 .nr(16)
9990 .kr(4)
9991 .sr(2)
9992 .m(m)
9993 .n(n)
9994 .k(k)
9995 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -08009996 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -08009997 }
9998 }
9999 }
10000 }
10001
Frank Barcharde22685a2021-11-12 11:36:58 -080010002 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_div_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010003 TEST_REQUIRES_ARM_NEON;
10004 for (size_t k = 16; k <= 80; k += 8) {
10005 GemmMicrokernelTester()
10006 .mr(3)
10007 .nr(16)
10008 .kr(4)
10009 .sr(2)
10010 .m(3)
10011 .n(16)
10012 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080010013 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010014 }
10015 }
10016
Frank Barcharde22685a2021-11-12 11:36:58 -080010017 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, k_div_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010018 TEST_REQUIRES_ARM_NEON;
10019 for (size_t k = 16; k <= 80; k += 8) {
10020 for (uint32_t m = 1; m <= 3; m++) {
10021 for (uint32_t n = 1; n <= 16; n++) {
10022 GemmMicrokernelTester()
10023 .mr(3)
10024 .nr(16)
10025 .kr(4)
10026 .sr(2)
10027 .m(m)
10028 .n(n)
10029 .k(k)
10030 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010031 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010032 }
10033 }
10034 }
10035 }
10036
Frank Barcharde22685a2021-11-12 11:36:58 -080010037 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, n_gt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010038 TEST_REQUIRES_ARM_NEON;
10039 for (uint32_t n = 17; n < 32; n++) {
10040 for (size_t k = 1; k <= 40; k += 9) {
10041 GemmMicrokernelTester()
10042 .mr(3)
10043 .nr(16)
10044 .kr(4)
10045 .sr(2)
10046 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010047 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010048 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080010049 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010050 }
10051 }
10052 }
10053
Frank Barcharde22685a2021-11-12 11:36:58 -080010054 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, n_gt_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010055 TEST_REQUIRES_ARM_NEON;
10056 for (uint32_t n = 17; n < 32; n++) {
10057 for (size_t k = 1; k <= 40; k += 9) {
10058 GemmMicrokernelTester()
10059 .mr(3)
10060 .nr(16)
10061 .kr(4)
10062 .sr(2)
10063 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010064 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010065 .k(k)
10066 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080010067 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010068 }
10069 }
10070 }
10071
Frank Barcharde22685a2021-11-12 11:36:58 -080010072 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, n_gt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010073 TEST_REQUIRES_ARM_NEON;
10074 for (uint32_t n = 17; n < 32; n++) {
10075 for (size_t k = 1; k <= 40; k += 9) {
10076 for (uint32_t m = 1; m <= 3; m++) {
10077 GemmMicrokernelTester()
10078 .mr(3)
10079 .nr(16)
10080 .kr(4)
10081 .sr(2)
10082 .m(m)
10083 .n(n)
10084 .k(k)
10085 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010086 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010087 }
10088 }
10089 }
10090 }
10091
Frank Barcharde22685a2021-11-12 11:36:58 -080010092 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, n_div_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010093 TEST_REQUIRES_ARM_NEON;
10094 for (uint32_t n = 32; n <= 48; n += 16) {
10095 for (size_t k = 1; k <= 40; k += 9) {
10096 GemmMicrokernelTester()
10097 .mr(3)
10098 .nr(16)
10099 .kr(4)
10100 .sr(2)
10101 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010102 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010103 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080010104 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010105 }
10106 }
10107 }
10108
Frank Barcharde22685a2021-11-12 11:36:58 -080010109 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, n_div_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010110 TEST_REQUIRES_ARM_NEON;
10111 for (uint32_t n = 32; n <= 48; n += 16) {
10112 for (size_t k = 1; k <= 40; k += 9) {
10113 GemmMicrokernelTester()
10114 .mr(3)
10115 .nr(16)
10116 .kr(4)
10117 .sr(2)
10118 .m(3)
10119 .n(n)
10120 .k(k)
10121 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080010122 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010123 }
10124 }
10125 }
10126
Frank Barcharde22685a2021-11-12 11:36:58 -080010127 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, n_div_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010128 TEST_REQUIRES_ARM_NEON;
10129 for (uint32_t n = 32; n <= 48; n += 16) {
10130 for (size_t k = 1; k <= 40; k += 9) {
10131 for (uint32_t m = 1; m <= 3; m++) {
10132 GemmMicrokernelTester()
10133 .mr(3)
10134 .nr(16)
10135 .kr(4)
10136 .sr(2)
10137 .m(m)
10138 .n(n)
10139 .k(k)
10140 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010141 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010142 }
10143 }
10144 }
10145 }
10146
Frank Barcharde22685a2021-11-12 11:36:58 -080010147 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010148 TEST_REQUIRES_ARM_NEON;
10149 for (size_t k = 1; k <= 40; k += 9) {
10150 GemmMicrokernelTester()
10151 .mr(3)
10152 .nr(16)
10153 .kr(4)
10154 .sr(2)
10155 .m(3)
10156 .n(16)
10157 .k(k)
10158 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080010159 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010160 }
10161 }
10162
Frank Barcharde22685a2021-11-12 11:36:58 -080010163 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, small_kernel_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010164 TEST_REQUIRES_ARM_NEON;
10165 for (size_t k = 1; k <= 40; k += 9) {
10166 for (uint32_t m = 1; m <= 3; m++) {
10167 for (uint32_t n = 1; n <= 16; n++) {
10168 GemmMicrokernelTester()
10169 .mr(3)
10170 .nr(16)
10171 .kr(4)
10172 .sr(2)
10173 .m(m)
10174 .n(n)
10175 .k(k)
10176 .ks(3)
10177 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010178 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010179 }
10180 }
10181 }
10182 }
10183
Frank Barcharde22685a2021-11-12 11:36:58 -080010184 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, n_gt_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010185 TEST_REQUIRES_ARM_NEON;
10186 for (uint32_t n = 17; n < 32; n++) {
10187 for (size_t k = 1; k <= 40; k += 9) {
10188 GemmMicrokernelTester()
10189 .mr(3)
10190 .nr(16)
10191 .kr(4)
10192 .sr(2)
10193 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010194 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010195 .k(k)
10196 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080010197 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010198 }
10199 }
10200 }
10201
Frank Barcharde22685a2021-11-12 11:36:58 -080010202 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, n_div_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010203 TEST_REQUIRES_ARM_NEON;
10204 for (uint32_t n = 32; n <= 48; n += 16) {
10205 for (size_t k = 1; k <= 40; k += 9) {
10206 GemmMicrokernelTester()
10207 .mr(3)
10208 .nr(16)
10209 .kr(4)
10210 .sr(2)
10211 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010212 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010213 .k(k)
10214 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080010215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010216 }
10217 }
10218 }
10219
Frank Barcharde22685a2021-11-12 11:36:58 -080010220 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, strided_cm_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010221 TEST_REQUIRES_ARM_NEON;
10222 for (size_t k = 1; k <= 40; k += 9) {
10223 for (uint32_t m = 1; m <= 3; m++) {
10224 for (uint32_t n = 1; n <= 16; n++) {
10225 GemmMicrokernelTester()
10226 .mr(3)
10227 .nr(16)
10228 .kr(4)
10229 .sr(2)
10230 .m(m)
10231 .n(n)
10232 .k(k)
10233 .cm_stride(19)
10234 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010235 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010236 }
10237 }
10238 }
10239 }
10240
Frank Barcharde22685a2021-11-12 11:36:58 -080010241 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, a_offset) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010242 TEST_REQUIRES_ARM_NEON;
10243 for (size_t k = 1; k <= 40; k += 9) {
10244 GemmMicrokernelTester()
10245 .mr(3)
10246 .nr(16)
10247 .kr(4)
10248 .sr(2)
10249 .m(3)
10250 .n(16)
10251 .k(k)
10252 .ks(3)
10253 .a_offset(127)
Marat Dukhan50323b82022-01-11 00:12:01 -080010254 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010255 }
10256 }
10257
Frank Barcharde22685a2021-11-12 11:36:58 -080010258 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, zero) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010259 TEST_REQUIRES_ARM_NEON;
10260 for (uint32_t mz = 0; mz < 3; mz++) {
10261 for (size_t k = 1; k <= 40; k += 9) {
10262 GemmMicrokernelTester()
10263 .mr(3)
10264 .nr(16)
10265 .kr(4)
10266 .sr(2)
10267 .m(3)
10268 .n(16)
10269 .k(k)
10270 .ks(3)
10271 .a_offset(127)
10272 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080010273 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010274 }
10275 }
10276 }
10277
Frank Barcharde22685a2021-11-12 11:36:58 -080010278 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, qmin) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010279 TEST_REQUIRES_ARM_NEON;
10280 GemmMicrokernelTester()
10281 .mr(3)
10282 .nr(16)
10283 .kr(4)
10284 .sr(2)
10285 .m(3)
10286 .n(16)
10287 .k(8)
10288 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080010289 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010290 }
10291
Frank Barcharde22685a2021-11-12 11:36:58 -080010292 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, qmax) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010293 TEST_REQUIRES_ARM_NEON;
10294 GemmMicrokernelTester()
10295 .mr(3)
10296 .nr(16)
10297 .kr(4)
10298 .sr(2)
10299 .m(3)
10300 .n(16)
10301 .k(8)
10302 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080010303 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010304 }
10305
Frank Barcharde22685a2021-11-12 11:36:58 -080010306 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4S2__NEON_MULL, strided_cm) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010307 TEST_REQUIRES_ARM_NEON;
10308 GemmMicrokernelTester()
10309 .mr(3)
10310 .nr(16)
10311 .kr(4)
10312 .sr(2)
10313 .m(3)
10314 .n(16)
10315 .k(8)
10316 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080010317 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010318 }
10319#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
10320
10321
10322#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -080010323 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_eq_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010324 TEST_REQUIRES_ARM_NEON;
10325 GemmMicrokernelTester()
10326 .mr(4)
10327 .nr(8)
10328 .kr(4)
10329 .sr(2)
10330 .m(4)
10331 .n(8)
10332 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -080010333 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010334 }
10335
Frank Barcharde22685a2021-11-12 11:36:58 -080010336 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010337 TEST_REQUIRES_ARM_NEON;
10338 GemmMicrokernelTester()
10339 .mr(4)
10340 .nr(8)
10341 .kr(4)
10342 .sr(2)
10343 .m(4)
10344 .n(8)
10345 .k(16)
10346 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080010347 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010348 }
10349
Frank Barcharde22685a2021-11-12 11:36:58 -080010350 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_eq_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010351 TEST_REQUIRES_ARM_NEON;
10352 for (uint32_t m = 1; m <= 4; m++) {
10353 for (uint32_t n = 1; n <= 8; n++) {
10354 GemmMicrokernelTester()
10355 .mr(4)
10356 .nr(8)
10357 .kr(4)
10358 .sr(2)
10359 .m(m)
10360 .n(n)
10361 .k(16)
10362 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010363 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010364 }
10365 }
10366 }
10367
Frank Barcharde22685a2021-11-12 11:36:58 -080010368 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_eq_16_subtile_m) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010369 TEST_REQUIRES_ARM_NEON;
10370 for (uint32_t m = 1; m <= 4; m++) {
10371 GemmMicrokernelTester()
10372 .mr(4)
10373 .nr(8)
10374 .kr(4)
10375 .sr(2)
10376 .m(m)
10377 .n(8)
10378 .k(16)
10379 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010380 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010381 }
10382 }
10383
Frank Barcharde22685a2021-11-12 11:36:58 -080010384 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_eq_16_subtile_n) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010385 TEST_REQUIRES_ARM_NEON;
10386 for (uint32_t n = 1; n <= 8; n++) {
10387 GemmMicrokernelTester()
10388 .mr(4)
10389 .nr(8)
10390 .kr(4)
10391 .sr(2)
10392 .m(4)
10393 .n(n)
10394 .k(16)
10395 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010396 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010397 }
10398 }
10399
Frank Barcharde22685a2021-11-12 11:36:58 -080010400 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_lt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010401 TEST_REQUIRES_ARM_NEON;
10402 for (size_t k = 1; k < 16; k++) {
10403 GemmMicrokernelTester()
10404 .mr(4)
10405 .nr(8)
10406 .kr(4)
10407 .sr(2)
10408 .m(4)
10409 .n(8)
10410 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080010411 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010412 }
10413 }
10414
Frank Barcharde22685a2021-11-12 11:36:58 -080010415 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_lt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010416 TEST_REQUIRES_ARM_NEON;
10417 for (size_t k = 1; k < 16; k++) {
10418 for (uint32_t m = 1; m <= 4; m++) {
10419 for (uint32_t n = 1; n <= 8; n++) {
10420 GemmMicrokernelTester()
10421 .mr(4)
10422 .nr(8)
10423 .kr(4)
10424 .sr(2)
10425 .m(m)
10426 .n(n)
10427 .k(k)
10428 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010429 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010430 }
10431 }
10432 }
10433 }
10434
Frank Barcharde22685a2021-11-12 11:36:58 -080010435 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_gt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010436 TEST_REQUIRES_ARM_NEON;
10437 for (size_t k = 17; k < 32; k++) {
10438 GemmMicrokernelTester()
10439 .mr(4)
10440 .nr(8)
10441 .kr(4)
10442 .sr(2)
10443 .m(4)
10444 .n(8)
10445 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080010446 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010447 }
10448 }
10449
Frank Barcharde22685a2021-11-12 11:36:58 -080010450 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_gt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010451 TEST_REQUIRES_ARM_NEON;
10452 for (size_t k = 17; k < 32; k++) {
10453 for (uint32_t m = 1; m <= 4; m++) {
10454 for (uint32_t n = 1; n <= 8; n++) {
10455 GemmMicrokernelTester()
10456 .mr(4)
10457 .nr(8)
10458 .kr(4)
10459 .sr(2)
10460 .m(m)
10461 .n(n)
10462 .k(k)
10463 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010464 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010465 }
10466 }
10467 }
10468 }
10469
Frank Barcharde22685a2021-11-12 11:36:58 -080010470 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_div_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010471 TEST_REQUIRES_ARM_NEON;
10472 for (size_t k = 32; k <= 160; k += 16) {
10473 GemmMicrokernelTester()
10474 .mr(4)
10475 .nr(8)
10476 .kr(4)
10477 .sr(2)
10478 .m(4)
10479 .n(8)
10480 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080010481 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010482 }
10483 }
10484
Frank Barcharde22685a2021-11-12 11:36:58 -080010485 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, k_div_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010486 TEST_REQUIRES_ARM_NEON;
10487 for (size_t k = 32; k <= 160; k += 16) {
10488 for (uint32_t m = 1; m <= 4; m++) {
10489 for (uint32_t n = 1; n <= 8; n++) {
10490 GemmMicrokernelTester()
10491 .mr(4)
10492 .nr(8)
10493 .kr(4)
10494 .sr(2)
10495 .m(m)
10496 .n(n)
10497 .k(k)
10498 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010499 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010500 }
10501 }
10502 }
10503 }
10504
Frank Barcharde22685a2021-11-12 11:36:58 -080010505 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, n_gt_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010506 TEST_REQUIRES_ARM_NEON;
10507 for (uint32_t n = 9; n < 16; n++) {
10508 for (size_t k = 1; k <= 80; k += 17) {
10509 GemmMicrokernelTester()
10510 .mr(4)
10511 .nr(8)
10512 .kr(4)
10513 .sr(2)
10514 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010515 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010516 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080010517 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010518 }
10519 }
10520 }
10521
Frank Barcharde22685a2021-11-12 11:36:58 -080010522 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, n_gt_8_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010523 TEST_REQUIRES_ARM_NEON;
10524 for (uint32_t n = 9; n < 16; n++) {
10525 for (size_t k = 1; k <= 80; k += 17) {
10526 GemmMicrokernelTester()
10527 .mr(4)
10528 .nr(8)
10529 .kr(4)
10530 .sr(2)
10531 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010532 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010533 .k(k)
10534 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080010535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010536 }
10537 }
10538 }
10539
Frank Barcharde22685a2021-11-12 11:36:58 -080010540 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, n_gt_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010541 TEST_REQUIRES_ARM_NEON;
10542 for (uint32_t n = 9; n < 16; n++) {
10543 for (size_t k = 1; k <= 80; k += 17) {
10544 for (uint32_t m = 1; m <= 4; m++) {
10545 GemmMicrokernelTester()
10546 .mr(4)
10547 .nr(8)
10548 .kr(4)
10549 .sr(2)
10550 .m(m)
10551 .n(n)
10552 .k(k)
10553 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010554 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010555 }
10556 }
10557 }
10558 }
10559
Frank Barcharde22685a2021-11-12 11:36:58 -080010560 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, n_div_8) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010561 TEST_REQUIRES_ARM_NEON;
10562 for (uint32_t n = 16; n <= 24; n += 8) {
10563 for (size_t k = 1; k <= 80; k += 17) {
10564 GemmMicrokernelTester()
10565 .mr(4)
10566 .nr(8)
10567 .kr(4)
10568 .sr(2)
10569 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010570 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010571 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080010572 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010573 }
10574 }
10575 }
10576
Frank Barcharde22685a2021-11-12 11:36:58 -080010577 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, n_div_8_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010578 TEST_REQUIRES_ARM_NEON;
10579 for (uint32_t n = 16; n <= 24; n += 8) {
10580 for (size_t k = 1; k <= 80; k += 17) {
10581 GemmMicrokernelTester()
10582 .mr(4)
10583 .nr(8)
10584 .kr(4)
10585 .sr(2)
10586 .m(4)
10587 .n(n)
10588 .k(k)
10589 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080010590 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010591 }
10592 }
10593 }
10594
Frank Barcharde22685a2021-11-12 11:36:58 -080010595 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, n_div_8_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010596 TEST_REQUIRES_ARM_NEON;
10597 for (uint32_t n = 16; n <= 24; n += 8) {
10598 for (size_t k = 1; k <= 80; k += 17) {
10599 for (uint32_t m = 1; m <= 4; m++) {
10600 GemmMicrokernelTester()
10601 .mr(4)
10602 .nr(8)
10603 .kr(4)
10604 .sr(2)
10605 .m(m)
10606 .n(n)
10607 .k(k)
10608 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010609 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010610 }
10611 }
10612 }
10613 }
10614
Frank Barcharde22685a2021-11-12 11:36:58 -080010615 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010616 TEST_REQUIRES_ARM_NEON;
10617 for (size_t k = 1; k <= 80; k += 17) {
10618 GemmMicrokernelTester()
10619 .mr(4)
10620 .nr(8)
10621 .kr(4)
10622 .sr(2)
10623 .m(4)
10624 .n(8)
10625 .k(k)
10626 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080010627 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010628 }
10629 }
10630
Frank Barcharde22685a2021-11-12 11:36:58 -080010631 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, small_kernel_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010632 TEST_REQUIRES_ARM_NEON;
10633 for (size_t k = 1; k <= 80; k += 17) {
10634 for (uint32_t m = 1; m <= 4; m++) {
10635 for (uint32_t n = 1; n <= 8; n++) {
10636 GemmMicrokernelTester()
10637 .mr(4)
10638 .nr(8)
10639 .kr(4)
10640 .sr(2)
10641 .m(m)
10642 .n(n)
10643 .k(k)
10644 .ks(3)
10645 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010646 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010647 }
10648 }
10649 }
10650 }
10651
Frank Barcharde22685a2021-11-12 11:36:58 -080010652 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, n_gt_8_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010653 TEST_REQUIRES_ARM_NEON;
10654 for (uint32_t n = 9; n < 16; n++) {
10655 for (size_t k = 1; k <= 80; k += 17) {
10656 GemmMicrokernelTester()
10657 .mr(4)
10658 .nr(8)
10659 .kr(4)
10660 .sr(2)
10661 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010662 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010663 .k(k)
10664 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080010665 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010666 }
10667 }
10668 }
10669
Frank Barcharde22685a2021-11-12 11:36:58 -080010670 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, n_div_8_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010671 TEST_REQUIRES_ARM_NEON;
10672 for (uint32_t n = 16; n <= 24; n += 8) {
10673 for (size_t k = 1; k <= 80; k += 17) {
10674 GemmMicrokernelTester()
10675 .mr(4)
10676 .nr(8)
10677 .kr(4)
10678 .sr(2)
10679 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010680 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010681 .k(k)
10682 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080010683 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010684 }
10685 }
10686 }
10687
Frank Barcharde22685a2021-11-12 11:36:58 -080010688 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, strided_cm_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010689 TEST_REQUIRES_ARM_NEON;
10690 for (size_t k = 1; k <= 80; k += 17) {
10691 for (uint32_t m = 1; m <= 4; m++) {
10692 for (uint32_t n = 1; n <= 8; n++) {
10693 GemmMicrokernelTester()
10694 .mr(4)
10695 .nr(8)
10696 .kr(4)
10697 .sr(2)
10698 .m(m)
10699 .n(n)
10700 .k(k)
10701 .cm_stride(11)
10702 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080010703 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010704 }
10705 }
10706 }
10707 }
10708
Frank Barcharde22685a2021-11-12 11:36:58 -080010709 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, a_offset) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010710 TEST_REQUIRES_ARM_NEON;
10711 for (size_t k = 1; k <= 80; k += 17) {
10712 GemmMicrokernelTester()
10713 .mr(4)
10714 .nr(8)
10715 .kr(4)
10716 .sr(2)
10717 .m(4)
10718 .n(8)
10719 .k(k)
10720 .ks(3)
10721 .a_offset(331)
Marat Dukhan50323b82022-01-11 00:12:01 -080010722 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010723 }
10724 }
10725
Frank Barcharde22685a2021-11-12 11:36:58 -080010726 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, zero) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010727 TEST_REQUIRES_ARM_NEON;
10728 for (uint32_t mz = 0; mz < 4; mz++) {
10729 for (size_t k = 1; k <= 80; k += 17) {
10730 GemmMicrokernelTester()
10731 .mr(4)
10732 .nr(8)
10733 .kr(4)
10734 .sr(2)
10735 .m(4)
10736 .n(8)
10737 .k(k)
10738 .ks(3)
10739 .a_offset(331)
10740 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080010741 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010742 }
10743 }
10744 }
10745
Frank Barcharde22685a2021-11-12 11:36:58 -080010746 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, qmin) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010747 TEST_REQUIRES_ARM_NEON;
10748 GemmMicrokernelTester()
10749 .mr(4)
10750 .nr(8)
10751 .kr(4)
10752 .sr(2)
10753 .m(4)
10754 .n(8)
10755 .k(16)
10756 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080010757 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010758 }
10759
Frank Barcharde22685a2021-11-12 11:36:58 -080010760 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, qmax) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010761 TEST_REQUIRES_ARM_NEON;
10762 GemmMicrokernelTester()
10763 .mr(4)
10764 .nr(8)
10765 .kr(4)
10766 .sr(2)
10767 .m(4)
10768 .n(8)
10769 .k(16)
10770 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080010771 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010772 }
10773
Frank Barcharde22685a2021-11-12 11:36:58 -080010774 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4S2__NEON_MLAL, strided_cm) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010775 TEST_REQUIRES_ARM_NEON;
10776 GemmMicrokernelTester()
10777 .mr(4)
10778 .nr(8)
10779 .kr(4)
10780 .sr(2)
10781 .m(4)
10782 .n(8)
10783 .k(16)
10784 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080010785 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010786 }
10787#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
10788
10789
10790#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010791 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_eq_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010792 TEST_REQUIRES_ARM_NEON;
10793 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010794 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010795 .nr(16)
10796 .kr(4)
10797 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010798 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010799 .n(16)
10800 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010801 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010802 }
10803
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010804 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010805 TEST_REQUIRES_ARM_NEON;
10806 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010807 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010808 .nr(16)
10809 .kr(4)
10810 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010811 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010812 .n(16)
10813 .k(16)
10814 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010815 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010816 }
10817
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010818 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_eq_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010819 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010820 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010821 for (uint32_t n = 1; n <= 16; n++) {
10822 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010823 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010824 .nr(16)
10825 .kr(4)
10826 .sr(2)
10827 .m(m)
10828 .n(n)
10829 .k(16)
10830 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010831 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010832 }
10833 }
10834 }
10835
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010836 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_eq_16_subtile_m) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010837 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010838 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010839 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010840 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010841 .nr(16)
10842 .kr(4)
10843 .sr(2)
10844 .m(m)
10845 .n(16)
10846 .k(16)
10847 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010848 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010849 }
10850 }
10851
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010852 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_eq_16_subtile_n) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010853 TEST_REQUIRES_ARM_NEON;
10854 for (uint32_t n = 1; n <= 16; n++) {
10855 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010856 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010857 .nr(16)
10858 .kr(4)
10859 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010860 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010861 .n(n)
10862 .k(16)
10863 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010864 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010865 }
10866 }
10867
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010868 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_lt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010869 TEST_REQUIRES_ARM_NEON;
10870 for (size_t k = 1; k < 16; k++) {
10871 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010872 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010873 .nr(16)
10874 .kr(4)
10875 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010876 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010877 .n(16)
10878 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010879 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010880 }
10881 }
10882
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010883 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_lt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010884 TEST_REQUIRES_ARM_NEON;
10885 for (size_t k = 1; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010886 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010887 for (uint32_t n = 1; n <= 16; n++) {
10888 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010889 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010890 .nr(16)
10891 .kr(4)
10892 .sr(2)
10893 .m(m)
10894 .n(n)
10895 .k(k)
10896 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010897 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010898 }
10899 }
10900 }
10901 }
10902
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010903 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_gt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010904 TEST_REQUIRES_ARM_NEON;
10905 for (size_t k = 17; k < 32; k++) {
10906 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010907 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010908 .nr(16)
10909 .kr(4)
10910 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010911 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010912 .n(16)
10913 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010914 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010915 }
10916 }
10917
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010918 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_gt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010919 TEST_REQUIRES_ARM_NEON;
10920 for (size_t k = 17; k < 32; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010921 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010922 for (uint32_t n = 1; n <= 16; n++) {
10923 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010924 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010925 .nr(16)
10926 .kr(4)
10927 .sr(2)
10928 .m(m)
10929 .n(n)
10930 .k(k)
10931 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010932 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010933 }
10934 }
10935 }
10936 }
10937
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010938 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_div_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010939 TEST_REQUIRES_ARM_NEON;
10940 for (size_t k = 32; k <= 160; k += 16) {
10941 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010942 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010943 .nr(16)
10944 .kr(4)
10945 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010946 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010947 .n(16)
10948 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010949 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010950 }
10951 }
10952
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010953 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_div_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010954 TEST_REQUIRES_ARM_NEON;
10955 for (size_t k = 32; k <= 160; k += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010956 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010957 for (uint32_t n = 1; n <= 16; n++) {
10958 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010959 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010960 .nr(16)
10961 .kr(4)
10962 .sr(2)
10963 .m(m)
10964 .n(n)
10965 .k(k)
10966 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010967 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010968 }
10969 }
10970 }
10971 }
10972
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010973 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_gt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010974 TEST_REQUIRES_ARM_NEON;
10975 for (uint32_t n = 17; n < 32; n++) {
10976 for (size_t k = 1; k <= 80; k += 17) {
10977 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010978 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010979 .nr(16)
10980 .kr(4)
10981 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010982 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080010983 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080010984 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010985 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080010986 }
10987 }
10988 }
10989
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010990 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_gt_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080010991 TEST_REQUIRES_ARM_NEON;
10992 for (uint32_t n = 17; n < 32; n++) {
10993 for (size_t k = 1; k <= 80; k += 17) {
10994 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010995 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080010996 .nr(16)
10997 .kr(4)
10998 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080010999 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011000 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080011001 .k(k)
11002 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011004 }
11005 }
11006 }
11007
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011008 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_gt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011009 TEST_REQUIRES_ARM_NEON;
11010 for (uint32_t n = 17; n < 32; n++) {
11011 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011012 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011013 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011014 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011015 .nr(16)
11016 .kr(4)
11017 .sr(2)
11018 .m(m)
11019 .n(n)
11020 .k(k)
11021 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011022 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011023 }
11024 }
11025 }
11026 }
11027
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011028 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_div_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011029 TEST_REQUIRES_ARM_NEON;
11030 for (uint32_t n = 32; n <= 48; n += 16) {
11031 for (size_t k = 1; k <= 80; k += 17) {
11032 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011033 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011034 .nr(16)
11035 .kr(4)
11036 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011037 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011038 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080011039 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011040 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011041 }
11042 }
11043 }
11044
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011045 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_div_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011046 TEST_REQUIRES_ARM_NEON;
11047 for (uint32_t n = 32; n <= 48; n += 16) {
11048 for (size_t k = 1; k <= 80; k += 17) {
11049 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011050 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011051 .nr(16)
11052 .kr(4)
11053 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011054 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011055 .n(n)
11056 .k(k)
11057 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011058 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011059 }
11060 }
11061 }
11062
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011063 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_div_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011064 TEST_REQUIRES_ARM_NEON;
11065 for (uint32_t n = 32; n <= 48; n += 16) {
11066 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011067 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011068 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011069 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011070 .nr(16)
11071 .kr(4)
11072 .sr(2)
11073 .m(m)
11074 .n(n)
11075 .k(k)
11076 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011077 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011078 }
11079 }
11080 }
11081 }
11082
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011083 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011084 TEST_REQUIRES_ARM_NEON;
11085 for (size_t k = 1; k <= 80; k += 17) {
11086 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011087 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011088 .nr(16)
11089 .kr(4)
11090 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011091 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011092 .n(16)
11093 .k(k)
11094 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011095 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011096 }
11097 }
11098
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011099 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, small_kernel_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011100 TEST_REQUIRES_ARM_NEON;
11101 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011102 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011103 for (uint32_t n = 1; n <= 16; n++) {
11104 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011105 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011106 .nr(16)
11107 .kr(4)
11108 .sr(2)
11109 .m(m)
11110 .n(n)
11111 .k(k)
11112 .ks(3)
11113 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011114 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011115 }
11116 }
11117 }
11118 }
11119
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011120 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_gt_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011121 TEST_REQUIRES_ARM_NEON;
11122 for (uint32_t n = 17; n < 32; n++) {
11123 for (size_t k = 1; k <= 80; k += 17) {
11124 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011125 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011126 .nr(16)
11127 .kr(4)
11128 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011129 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011130 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080011131 .k(k)
11132 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011133 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011134 }
11135 }
11136 }
11137
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011138 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_div_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011139 TEST_REQUIRES_ARM_NEON;
11140 for (uint32_t n = 32; n <= 48; n += 16) {
11141 for (size_t k = 1; k <= 80; k += 17) {
11142 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011143 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011144 .nr(16)
11145 .kr(4)
11146 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011147 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011148 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080011149 .k(k)
11150 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011151 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011152 }
11153 }
11154 }
11155
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011156 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, strided_cm_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011157 TEST_REQUIRES_ARM_NEON;
11158 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011159 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011160 for (uint32_t n = 1; n <= 16; n++) {
11161 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011162 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011163 .nr(16)
11164 .kr(4)
11165 .sr(2)
11166 .m(m)
11167 .n(n)
11168 .k(k)
11169 .cm_stride(19)
11170 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011171 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011172 }
11173 }
11174 }
11175 }
11176
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011177 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, a_offset) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011178 TEST_REQUIRES_ARM_NEON;
11179 for (size_t k = 1; k <= 80; k += 17) {
11180 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011181 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011182 .nr(16)
11183 .kr(4)
11184 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011185 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011186 .n(16)
11187 .k(k)
11188 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011189 .a_offset(83)
11190 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011191 }
11192 }
11193
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011194 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, zero) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011195 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011196 for (uint32_t mz = 0; mz < 1; mz++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011197 for (size_t k = 1; k <= 80; k += 17) {
11198 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011199 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011200 .nr(16)
11201 .kr(4)
11202 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011203 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011204 .n(16)
11205 .k(k)
11206 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011207 .a_offset(83)
Frank Barchardeb704f72021-11-12 01:26:50 -080011208 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011209 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011210 }
11211 }
11212 }
11213
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011214 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, qmin) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011215 TEST_REQUIRES_ARM_NEON;
11216 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011217 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011218 .nr(16)
11219 .kr(4)
11220 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011221 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011222 .n(16)
11223 .k(16)
11224 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011225 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011226 }
11227
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011228 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, qmax) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011229 TEST_REQUIRES_ARM_NEON;
11230 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011231 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011232 .nr(16)
11233 .kr(4)
11234 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011235 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011236 .n(16)
11237 .k(16)
11238 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011239 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011240 }
11241
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011242 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, strided_cm) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011243 TEST_REQUIRES_ARM_NEON;
11244 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011245 .mr(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011246 .nr(16)
11247 .kr(4)
11248 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011249 .m(1)
Frank Barchardeb704f72021-11-12 01:26:50 -080011250 .n(16)
11251 .k(16)
11252 .cm_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011253 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011254 }
11255#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
11256
11257
11258#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011259 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_eq_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011260 TEST_REQUIRES_ARM_NEON;
11261 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011262 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011263 .nr(16)
11264 .kr(4)
11265 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011266 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011267 .n(16)
11268 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011269 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011270 }
11271
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011272 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011273 TEST_REQUIRES_ARM_NEON;
11274 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011275 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011276 .nr(16)
11277 .kr(4)
11278 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011279 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011280 .n(16)
11281 .k(16)
11282 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011283 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011284 }
11285
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011286 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_eq_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011287 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011288 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011289 for (uint32_t n = 1; n <= 16; n++) {
11290 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011291 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011292 .nr(16)
11293 .kr(4)
11294 .sr(2)
11295 .m(m)
11296 .n(n)
11297 .k(16)
11298 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011299 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011300 }
11301 }
11302 }
11303
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011304 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_eq_16_subtile_m) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011305 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011306 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011307 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011308 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011309 .nr(16)
11310 .kr(4)
11311 .sr(2)
11312 .m(m)
11313 .n(16)
11314 .k(16)
11315 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011316 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011317 }
11318 }
11319
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011320 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_eq_16_subtile_n) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011321 TEST_REQUIRES_ARM_NEON;
11322 for (uint32_t n = 1; n <= 16; n++) {
11323 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011324 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011325 .nr(16)
11326 .kr(4)
11327 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011328 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011329 .n(n)
11330 .k(16)
11331 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011332 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011333 }
11334 }
11335
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011336 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_lt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011337 TEST_REQUIRES_ARM_NEON;
11338 for (size_t k = 1; k < 16; k++) {
11339 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011340 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011341 .nr(16)
11342 .kr(4)
11343 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011344 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011345 .n(16)
11346 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011347 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011348 }
11349 }
11350
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011351 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_lt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011352 TEST_REQUIRES_ARM_NEON;
11353 for (size_t k = 1; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011354 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011355 for (uint32_t n = 1; n <= 16; n++) {
11356 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011357 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011358 .nr(16)
11359 .kr(4)
11360 .sr(2)
11361 .m(m)
11362 .n(n)
11363 .k(k)
11364 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011365 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011366 }
11367 }
11368 }
11369 }
11370
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011371 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_gt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011372 TEST_REQUIRES_ARM_NEON;
11373 for (size_t k = 17; k < 32; k++) {
11374 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011375 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011376 .nr(16)
11377 .kr(4)
11378 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011379 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011380 .n(16)
11381 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011382 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011383 }
11384 }
11385
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011386 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_gt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011387 TEST_REQUIRES_ARM_NEON;
11388 for (size_t k = 17; k < 32; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011389 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011390 for (uint32_t n = 1; n <= 16; n++) {
11391 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011392 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011393 .nr(16)
11394 .kr(4)
11395 .sr(2)
11396 .m(m)
11397 .n(n)
11398 .k(k)
11399 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011400 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011401 }
11402 }
11403 }
11404 }
11405
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011406 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_div_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011407 TEST_REQUIRES_ARM_NEON;
11408 for (size_t k = 32; k <= 160; k += 16) {
11409 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011410 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011411 .nr(16)
11412 .kr(4)
11413 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011414 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011415 .n(16)
11416 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011417 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011418 }
11419 }
11420
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011421 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_div_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011422 TEST_REQUIRES_ARM_NEON;
11423 for (size_t k = 32; k <= 160; k += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011424 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011425 for (uint32_t n = 1; n <= 16; n++) {
11426 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011427 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011428 .nr(16)
11429 .kr(4)
11430 .sr(2)
11431 .m(m)
11432 .n(n)
11433 .k(k)
11434 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011435 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011436 }
11437 }
11438 }
11439 }
11440
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011441 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_gt_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011442 TEST_REQUIRES_ARM_NEON;
11443 for (uint32_t n = 17; n < 32; n++) {
11444 for (size_t k = 1; k <= 80; k += 17) {
11445 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011446 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011447 .nr(16)
11448 .kr(4)
11449 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011450 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011451 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080011452 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011453 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011454 }
11455 }
11456 }
11457
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011458 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_gt_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011459 TEST_REQUIRES_ARM_NEON;
11460 for (uint32_t n = 17; n < 32; n++) {
11461 for (size_t k = 1; k <= 80; k += 17) {
11462 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011463 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011464 .nr(16)
11465 .kr(4)
11466 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011467 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011468 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080011469 .k(k)
11470 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011471 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011472 }
11473 }
11474 }
11475
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011476 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_gt_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011477 TEST_REQUIRES_ARM_NEON;
11478 for (uint32_t n = 17; n < 32; n++) {
11479 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011480 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011481 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011482 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011483 .nr(16)
11484 .kr(4)
11485 .sr(2)
11486 .m(m)
11487 .n(n)
11488 .k(k)
11489 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011490 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011491 }
11492 }
11493 }
11494 }
11495
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011496 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_div_16) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011497 TEST_REQUIRES_ARM_NEON;
11498 for (uint32_t n = 32; n <= 48; n += 16) {
11499 for (size_t k = 1; k <= 80; k += 17) {
11500 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011501 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011502 .nr(16)
11503 .kr(4)
11504 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011505 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011506 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080011507 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011508 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011509 }
11510 }
11511 }
11512
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011513 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_div_16_strided_cn) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011514 TEST_REQUIRES_ARM_NEON;
11515 for (uint32_t n = 32; n <= 48; n += 16) {
11516 for (size_t k = 1; k <= 80; k += 17) {
11517 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011518 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011519 .nr(16)
11520 .kr(4)
11521 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011522 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011523 .n(n)
11524 .k(k)
11525 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011526 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011527 }
11528 }
11529 }
11530
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011531 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_div_16_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011532 TEST_REQUIRES_ARM_NEON;
11533 for (uint32_t n = 32; n <= 48; n += 16) {
11534 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011535 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011536 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011537 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011538 .nr(16)
11539 .kr(4)
11540 .sr(2)
11541 .m(m)
11542 .n(n)
11543 .k(k)
11544 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011545 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011546 }
11547 }
11548 }
11549 }
11550
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011551 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011552 TEST_REQUIRES_ARM_NEON;
11553 for (size_t k = 1; k <= 80; k += 17) {
11554 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011555 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011556 .nr(16)
11557 .kr(4)
11558 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011559 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011560 .n(16)
11561 .k(k)
11562 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011563 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011564 }
11565 }
11566
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011567 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, small_kernel_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011568 TEST_REQUIRES_ARM_NEON;
11569 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011570 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011571 for (uint32_t n = 1; n <= 16; n++) {
11572 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011573 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011574 .nr(16)
11575 .kr(4)
11576 .sr(2)
11577 .m(m)
11578 .n(n)
11579 .k(k)
11580 .ks(3)
11581 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011582 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011583 }
11584 }
11585 }
11586 }
11587
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011588 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_gt_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011589 TEST_REQUIRES_ARM_NEON;
11590 for (uint32_t n = 17; n < 32; n++) {
11591 for (size_t k = 1; k <= 80; k += 17) {
11592 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011593 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011594 .nr(16)
11595 .kr(4)
11596 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011597 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011598 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080011599 .k(k)
11600 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011601 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011602 }
11603 }
11604 }
11605
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011606 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_div_16_small_kernel) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011607 TEST_REQUIRES_ARM_NEON;
11608 for (uint32_t n = 32; n <= 48; n += 16) {
11609 for (size_t k = 1; k <= 80; k += 17) {
11610 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011611 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011612 .nr(16)
11613 .kr(4)
11614 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011615 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011616 .n(n)
Frank Barchardeb704f72021-11-12 01:26:50 -080011617 .k(k)
11618 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011619 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011620 }
11621 }
11622 }
11623
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011624 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, strided_cm_subtile) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011625 TEST_REQUIRES_ARM_NEON;
11626 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011627 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011628 for (uint32_t n = 1; n <= 16; n++) {
11629 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011630 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011631 .nr(16)
11632 .kr(4)
11633 .sr(2)
11634 .m(m)
11635 .n(n)
11636 .k(k)
11637 .cm_stride(19)
11638 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011639 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011640 }
11641 }
11642 }
11643 }
11644
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011645 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, a_offset) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011646 TEST_REQUIRES_ARM_NEON;
11647 for (size_t k = 1; k <= 80; k += 17) {
11648 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011649 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011650 .nr(16)
11651 .kr(4)
11652 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011653 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011654 .n(16)
11655 .k(k)
11656 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011657 .a_offset(331)
11658 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011659 }
11660 }
11661
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011662 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, zero) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011663 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011664 for (uint32_t mz = 0; mz < 4; mz++) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011665 for (size_t k = 1; k <= 80; k += 17) {
11666 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011667 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011668 .nr(16)
11669 .kr(4)
11670 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011671 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011672 .n(16)
11673 .k(k)
11674 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011675 .a_offset(331)
Frank Barchardeb704f72021-11-12 01:26:50 -080011676 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011677 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011678 }
11679 }
11680 }
11681
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011682 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, qmin) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011683 TEST_REQUIRES_ARM_NEON;
11684 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011685 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011686 .nr(16)
11687 .kr(4)
11688 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011689 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011690 .n(16)
11691 .k(16)
11692 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011693 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011694 }
11695
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011696 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, qmax) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011697 TEST_REQUIRES_ARM_NEON;
11698 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011699 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011700 .nr(16)
11701 .kr(4)
11702 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011703 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011704 .n(16)
11705 .k(16)
11706 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011707 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080011708 }
11709
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011710 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, strided_cm) {
Frank Barchardeb704f72021-11-12 01:26:50 -080011711 TEST_REQUIRES_ARM_NEON;
11712 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011713 .mr(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011714 .nr(16)
11715 .kr(4)
11716 .sr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011717 .m(4)
Frank Barchardeb704f72021-11-12 01:26:50 -080011718 .n(16)
11719 .k(16)
11720 .cm_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011721 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, 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_1X8C2S4__NEON_MULL, k_eq_8) {
11728 TEST_REQUIRES_ARM_NEON;
11729 GemmMicrokernelTester()
11730 .mr(1)
11731 .nr(8)
11732 .kr(2)
11733 .sr(4)
11734 .m(1)
11735 .n(8)
11736 .k(8)
11737 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11738 }
11739
11740 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, strided_cn) {
11741 TEST_REQUIRES_ARM_NEON;
11742 GemmMicrokernelTester()
11743 .mr(1)
11744 .nr(8)
11745 .kr(2)
11746 .sr(4)
11747 .m(1)
11748 .n(8)
11749 .k(8)
11750 .cn_stride(11)
11751 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11752 }
11753
11754 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_eq_8_subtile) {
11755 TEST_REQUIRES_ARM_NEON;
11756 for (uint32_t m = 1; m <= 1; m++) {
11757 for (uint32_t n = 1; n <= 8; n++) {
11758 GemmMicrokernelTester()
11759 .mr(1)
11760 .nr(8)
11761 .kr(2)
11762 .sr(4)
11763 .m(m)
11764 .n(n)
11765 .k(8)
11766 .iterations(1)
11767 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11768 }
11769 }
11770 }
11771
11772 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_eq_8_subtile_m) {
11773 TEST_REQUIRES_ARM_NEON;
11774 for (uint32_t m = 1; m <= 1; m++) {
11775 GemmMicrokernelTester()
11776 .mr(1)
11777 .nr(8)
11778 .kr(2)
11779 .sr(4)
11780 .m(m)
11781 .n(8)
11782 .k(8)
11783 .iterations(1)
11784 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11785 }
11786 }
11787
11788 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_eq_8_subtile_n) {
11789 TEST_REQUIRES_ARM_NEON;
11790 for (uint32_t n = 1; n <= 8; n++) {
11791 GemmMicrokernelTester()
11792 .mr(1)
11793 .nr(8)
11794 .kr(2)
11795 .sr(4)
11796 .m(1)
11797 .n(n)
11798 .k(8)
11799 .iterations(1)
11800 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11801 }
11802 }
11803
11804 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_lt_8) {
11805 TEST_REQUIRES_ARM_NEON;
11806 for (size_t k = 1; k < 8; k++) {
11807 GemmMicrokernelTester()
11808 .mr(1)
11809 .nr(8)
11810 .kr(2)
11811 .sr(4)
11812 .m(1)
11813 .n(8)
11814 .k(k)
11815 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11816 }
11817 }
11818
11819 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_lt_8_subtile) {
11820 TEST_REQUIRES_ARM_NEON;
11821 for (size_t k = 1; k < 8; k++) {
11822 for (uint32_t m = 1; m <= 1; m++) {
11823 for (uint32_t n = 1; n <= 8; n++) {
11824 GemmMicrokernelTester()
11825 .mr(1)
11826 .nr(8)
11827 .kr(2)
11828 .sr(4)
11829 .m(m)
11830 .n(n)
11831 .k(k)
11832 .iterations(1)
11833 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11834 }
11835 }
11836 }
11837 }
11838
11839 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_gt_8) {
11840 TEST_REQUIRES_ARM_NEON;
11841 for (size_t k = 9; k < 16; k++) {
11842 GemmMicrokernelTester()
11843 .mr(1)
11844 .nr(8)
11845 .kr(2)
11846 .sr(4)
11847 .m(1)
11848 .n(8)
11849 .k(k)
11850 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11851 }
11852 }
11853
11854 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_gt_8_subtile) {
11855 TEST_REQUIRES_ARM_NEON;
11856 for (size_t k = 9; k < 16; k++) {
11857 for (uint32_t m = 1; m <= 1; m++) {
11858 for (uint32_t n = 1; n <= 8; n++) {
11859 GemmMicrokernelTester()
11860 .mr(1)
11861 .nr(8)
11862 .kr(2)
11863 .sr(4)
11864 .m(m)
11865 .n(n)
11866 .k(k)
11867 .iterations(1)
11868 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11869 }
11870 }
11871 }
11872 }
11873
11874 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_div_8) {
11875 TEST_REQUIRES_ARM_NEON;
11876 for (size_t k = 16; k <= 80; k += 8) {
11877 GemmMicrokernelTester()
11878 .mr(1)
11879 .nr(8)
11880 .kr(2)
11881 .sr(4)
11882 .m(1)
11883 .n(8)
11884 .k(k)
11885 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11886 }
11887 }
11888
11889 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_div_8_subtile) {
11890 TEST_REQUIRES_ARM_NEON;
11891 for (size_t k = 16; k <= 80; k += 8) {
11892 for (uint32_t m = 1; m <= 1; m++) {
11893 for (uint32_t n = 1; n <= 8; n++) {
11894 GemmMicrokernelTester()
11895 .mr(1)
11896 .nr(8)
11897 .kr(2)
11898 .sr(4)
11899 .m(m)
11900 .n(n)
11901 .k(k)
11902 .iterations(1)
11903 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11904 }
11905 }
11906 }
11907 }
11908
11909 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, n_gt_8) {
11910 TEST_REQUIRES_ARM_NEON;
11911 for (uint32_t n = 9; n < 16; n++) {
11912 for (size_t k = 1; k <= 40; k += 9) {
11913 GemmMicrokernelTester()
11914 .mr(1)
11915 .nr(8)
11916 .kr(2)
11917 .sr(4)
11918 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011919 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011920 .k(k)
11921 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11922 }
11923 }
11924 }
11925
11926 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, 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 <= 40; k += 9) {
11930 GemmMicrokernelTester()
11931 .mr(1)
11932 .nr(8)
11933 .kr(2)
11934 .sr(4)
11935 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011936 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011937 .k(k)
11938 .cn_stride(11)
11939 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11940 }
11941 }
11942 }
11943
11944 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, 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 <= 40; k += 9) {
11948 for (uint32_t m = 1; m <= 1; m++) {
11949 GemmMicrokernelTester()
11950 .mr(1)
11951 .nr(8)
11952 .kr(2)
11953 .sr(4)
11954 .m(m)
11955 .n(n)
11956 .k(k)
11957 .iterations(1)
11958 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11959 }
11960 }
11961 }
11962 }
11963
11964 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, 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 <= 40; k += 9) {
11968 GemmMicrokernelTester()
11969 .mr(1)
11970 .nr(8)
11971 .kr(2)
11972 .sr(4)
11973 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080011974 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080011975 .k(k)
11976 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11977 }
11978 }
11979 }
11980
11981 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, 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 <= 40; k += 9) {
11985 GemmMicrokernelTester()
11986 .mr(1)
11987 .nr(8)
11988 .kr(2)
11989 .sr(4)
11990 .m(1)
11991 .n(n)
11992 .k(k)
11993 .cn_stride(11)
11994 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11995 }
11996 }
11997 }
11998
11999 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, 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 <= 40; k += 9) {
12003 for (uint32_t m = 1; m <= 1; m++) {
12004 GemmMicrokernelTester()
12005 .mr(1)
12006 .nr(8)
12007 .kr(2)
12008 .sr(4)
12009 .m(m)
12010 .n(n)
12011 .k(k)
12012 .iterations(1)
12013 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12014 }
12015 }
12016 }
12017 }
12018
12019 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, small_kernel) {
12020 TEST_REQUIRES_ARM_NEON;
12021 for (size_t k = 1; k <= 40; k += 9) {
12022 GemmMicrokernelTester()
12023 .mr(1)
12024 .nr(8)
12025 .kr(2)
12026 .sr(4)
12027 .m(1)
12028 .n(8)
12029 .k(k)
12030 .ks(3)
12031 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12032 }
12033 }
12034
12035 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, small_kernel_subtile) {
12036 TEST_REQUIRES_ARM_NEON;
12037 for (size_t k = 1; k <= 40; k += 9) {
12038 for (uint32_t m = 1; m <= 1; m++) {
12039 for (uint32_t n = 1; n <= 8; n++) {
12040 GemmMicrokernelTester()
12041 .mr(1)
12042 .nr(8)
12043 .kr(2)
12044 .sr(4)
12045 .m(m)
12046 .n(n)
12047 .k(k)
12048 .ks(3)
12049 .iterations(1)
12050 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12051 }
12052 }
12053 }
12054 }
12055
12056 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, 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 <= 40; k += 9) {
12060 GemmMicrokernelTester()
12061 .mr(1)
12062 .nr(8)
12063 .kr(2)
12064 .sr(4)
12065 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012066 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080012067 .k(k)
12068 .ks(3)
12069 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12070 }
12071 }
12072 }
12073
12074 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, 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 <= 40; k += 9) {
12078 GemmMicrokernelTester()
12079 .mr(1)
12080 .nr(8)
12081 .kr(2)
12082 .sr(4)
12083 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012084 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080012085 .k(k)
12086 .ks(3)
12087 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12088 }
12089 }
12090 }
12091
12092 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, strided_cm_subtile) {
12093 TEST_REQUIRES_ARM_NEON;
12094 for (size_t k = 1; k <= 40; k += 9) {
12095 for (uint32_t m = 1; m <= 1; m++) {
12096 for (uint32_t n = 1; n <= 8; n++) {
12097 GemmMicrokernelTester()
12098 .mr(1)
12099 .nr(8)
12100 .kr(2)
12101 .sr(4)
12102 .m(m)
12103 .n(n)
12104 .k(k)
12105 .cm_stride(11)
12106 .iterations(1)
12107 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12108 }
12109 }
12110 }
12111 }
12112
12113 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, a_offset) {
12114 TEST_REQUIRES_ARM_NEON;
12115 for (size_t k = 1; k <= 40; k += 9) {
12116 GemmMicrokernelTester()
12117 .mr(1)
12118 .nr(8)
12119 .kr(2)
12120 .sr(4)
12121 .m(1)
12122 .n(8)
12123 .k(k)
12124 .ks(3)
12125 .a_offset(43)
12126 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12127 }
12128 }
12129
12130 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, zero) {
12131 TEST_REQUIRES_ARM_NEON;
12132 for (uint32_t mz = 0; mz < 1; mz++) {
12133 for (size_t k = 1; k <= 40; k += 9) {
12134 GemmMicrokernelTester()
12135 .mr(1)
12136 .nr(8)
12137 .kr(2)
12138 .sr(4)
12139 .m(1)
12140 .n(8)
12141 .k(k)
12142 .ks(3)
12143 .a_offset(43)
12144 .zero_index(mz)
12145 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12146 }
12147 }
12148 }
12149
12150 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, qmin) {
12151 TEST_REQUIRES_ARM_NEON;
12152 GemmMicrokernelTester()
12153 .mr(1)
12154 .nr(8)
12155 .kr(2)
12156 .sr(4)
12157 .m(1)
12158 .n(8)
12159 .k(8)
12160 .qmin(128)
12161 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12162 }
12163
12164 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, qmax) {
12165 TEST_REQUIRES_ARM_NEON;
12166 GemmMicrokernelTester()
12167 .mr(1)
12168 .nr(8)
12169 .kr(2)
12170 .sr(4)
12171 .m(1)
12172 .n(8)
12173 .k(8)
12174 .qmax(128)
12175 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12176 }
12177
12178 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, strided_cm) {
12179 TEST_REQUIRES_ARM_NEON;
12180 GemmMicrokernelTester()
12181 .mr(1)
12182 .nr(8)
12183 .kr(2)
12184 .sr(4)
12185 .m(1)
12186 .n(8)
12187 .k(8)
12188 .cm_stride(11)
12189 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardeb704f72021-11-12 01:26:50 -080012190 }
12191#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
12192
12193
12194#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -080012195 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_eq_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012196 TEST_REQUIRES_ARM_NEON;
12197 GemmMicrokernelTester()
12198 .mr(3)
12199 .nr(8)
12200 .kr(2)
12201 .sr(4)
12202 .m(3)
12203 .n(8)
12204 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080012205 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012206 }
12207
Frank Barcharde22685a2021-11-12 11:36:58 -080012208 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, strided_cn) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012209 TEST_REQUIRES_ARM_NEON;
12210 GemmMicrokernelTester()
12211 .mr(3)
12212 .nr(8)
12213 .kr(2)
12214 .sr(4)
12215 .m(3)
12216 .n(8)
12217 .k(8)
12218 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080012219 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012220 }
12221
Frank Barcharde22685a2021-11-12 11:36:58 -080012222 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_eq_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012223 TEST_REQUIRES_ARM_NEON;
12224 for (uint32_t m = 1; m <= 3; m++) {
12225 for (uint32_t n = 1; n <= 8; n++) {
12226 GemmMicrokernelTester()
12227 .mr(3)
12228 .nr(8)
12229 .kr(2)
12230 .sr(4)
12231 .m(m)
12232 .n(n)
12233 .k(8)
12234 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012235 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012236 }
12237 }
12238 }
12239
Frank Barcharde22685a2021-11-12 11:36:58 -080012240 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_eq_8_subtile_m) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012241 TEST_REQUIRES_ARM_NEON;
12242 for (uint32_t m = 1; m <= 3; m++) {
12243 GemmMicrokernelTester()
12244 .mr(3)
12245 .nr(8)
12246 .kr(2)
12247 .sr(4)
12248 .m(m)
12249 .n(8)
12250 .k(8)
12251 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012252 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012253 }
12254 }
12255
Frank Barcharde22685a2021-11-12 11:36:58 -080012256 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_eq_8_subtile_n) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012257 TEST_REQUIRES_ARM_NEON;
12258 for (uint32_t n = 1; n <= 8; n++) {
12259 GemmMicrokernelTester()
12260 .mr(3)
12261 .nr(8)
12262 .kr(2)
12263 .sr(4)
12264 .m(3)
12265 .n(n)
12266 .k(8)
12267 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012268 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012269 }
12270 }
12271
Frank Barcharde22685a2021-11-12 11:36:58 -080012272 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_lt_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012273 TEST_REQUIRES_ARM_NEON;
12274 for (size_t k = 1; k < 8; k++) {
12275 GemmMicrokernelTester()
12276 .mr(3)
12277 .nr(8)
12278 .kr(2)
12279 .sr(4)
12280 .m(3)
12281 .n(8)
12282 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012283 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012284 }
12285 }
12286
Frank Barcharde22685a2021-11-12 11:36:58 -080012287 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_lt_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012288 TEST_REQUIRES_ARM_NEON;
12289 for (size_t k = 1; k < 8; k++) {
12290 for (uint32_t m = 1; m <= 3; m++) {
12291 for (uint32_t n = 1; n <= 8; n++) {
12292 GemmMicrokernelTester()
12293 .mr(3)
12294 .nr(8)
12295 .kr(2)
12296 .sr(4)
12297 .m(m)
12298 .n(n)
12299 .k(k)
12300 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012301 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012302 }
12303 }
12304 }
12305 }
12306
Frank Barcharde22685a2021-11-12 11:36:58 -080012307 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_gt_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012308 TEST_REQUIRES_ARM_NEON;
12309 for (size_t k = 9; k < 16; k++) {
12310 GemmMicrokernelTester()
12311 .mr(3)
12312 .nr(8)
12313 .kr(2)
12314 .sr(4)
12315 .m(3)
12316 .n(8)
12317 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012318 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012319 }
12320 }
12321
Frank Barcharde22685a2021-11-12 11:36:58 -080012322 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_gt_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012323 TEST_REQUIRES_ARM_NEON;
12324 for (size_t k = 9; k < 16; k++) {
12325 for (uint32_t m = 1; m <= 3; m++) {
12326 for (uint32_t n = 1; n <= 8; n++) {
12327 GemmMicrokernelTester()
12328 .mr(3)
12329 .nr(8)
12330 .kr(2)
12331 .sr(4)
12332 .m(m)
12333 .n(n)
12334 .k(k)
12335 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012336 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012337 }
12338 }
12339 }
12340 }
12341
Frank Barcharde22685a2021-11-12 11:36:58 -080012342 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_div_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012343 TEST_REQUIRES_ARM_NEON;
12344 for (size_t k = 16; k <= 80; k += 8) {
12345 GemmMicrokernelTester()
12346 .mr(3)
12347 .nr(8)
12348 .kr(2)
12349 .sr(4)
12350 .m(3)
12351 .n(8)
12352 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012353 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012354 }
12355 }
12356
Frank Barcharde22685a2021-11-12 11:36:58 -080012357 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, k_div_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012358 TEST_REQUIRES_ARM_NEON;
12359 for (size_t k = 16; k <= 80; k += 8) {
12360 for (uint32_t m = 1; m <= 3; m++) {
12361 for (uint32_t n = 1; n <= 8; n++) {
12362 GemmMicrokernelTester()
12363 .mr(3)
12364 .nr(8)
12365 .kr(2)
12366 .sr(4)
12367 .m(m)
12368 .n(n)
12369 .k(k)
12370 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012371 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012372 }
12373 }
12374 }
12375 }
12376
Frank Barcharde22685a2021-11-12 11:36:58 -080012377 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, n_gt_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012378 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(3)
12383 .nr(8)
12384 .kr(2)
12385 .sr(4)
12386 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012387 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080012388 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012389 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012390 }
12391 }
12392 }
12393
Frank Barcharde22685a2021-11-12 11:36:58 -080012394 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, n_gt_8_strided_cn) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012395 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(3)
12400 .nr(8)
12401 .kr(2)
12402 .sr(4)
12403 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012404 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080012405 .k(k)
12406 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080012407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012408 }
12409 }
12410 }
12411
Frank Barcharde22685a2021-11-12 11:36:58 -080012412 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, n_gt_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012413 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 <= 3; m++) {
12417 GemmMicrokernelTester()
12418 .mr(3)
12419 .nr(8)
12420 .kr(2)
12421 .sr(4)
12422 .m(m)
12423 .n(n)
12424 .k(k)
12425 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012426 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012427 }
12428 }
12429 }
12430 }
12431
Frank Barcharde22685a2021-11-12 11:36:58 -080012432 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, n_div_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012433 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(3)
12438 .nr(8)
12439 .kr(2)
12440 .sr(4)
12441 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012442 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080012443 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012444 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012445 }
12446 }
12447 }
12448
Frank Barcharde22685a2021-11-12 11:36:58 -080012449 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, n_div_8_strided_cn) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012450 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(3)
12455 .nr(8)
12456 .kr(2)
12457 .sr(4)
12458 .m(3)
12459 .n(n)
12460 .k(k)
12461 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080012462 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012463 }
12464 }
12465 }
12466
Frank Barcharde22685a2021-11-12 11:36:58 -080012467 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, n_div_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012468 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 <= 3; m++) {
12472 GemmMicrokernelTester()
12473 .mr(3)
12474 .nr(8)
12475 .kr(2)
12476 .sr(4)
12477 .m(m)
12478 .n(n)
12479 .k(k)
12480 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012481 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012482 }
12483 }
12484 }
12485 }
12486
Frank Barcharde22685a2021-11-12 11:36:58 -080012487 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, small_kernel) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012488 TEST_REQUIRES_ARM_NEON;
12489 for (size_t k = 1; k <= 40; k += 9) {
12490 GemmMicrokernelTester()
12491 .mr(3)
12492 .nr(8)
12493 .kr(2)
12494 .sr(4)
12495 .m(3)
12496 .n(8)
12497 .k(k)
12498 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080012499 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012500 }
12501 }
12502
Frank Barcharde22685a2021-11-12 11:36:58 -080012503 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, small_kernel_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012504 TEST_REQUIRES_ARM_NEON;
12505 for (size_t k = 1; k <= 40; k += 9) {
12506 for (uint32_t m = 1; m <= 3; m++) {
12507 for (uint32_t n = 1; n <= 8; n++) {
12508 GemmMicrokernelTester()
12509 .mr(3)
12510 .nr(8)
12511 .kr(2)
12512 .sr(4)
12513 .m(m)
12514 .n(n)
12515 .k(k)
12516 .ks(3)
12517 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012518 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012519 }
12520 }
12521 }
12522 }
12523
Frank Barcharde22685a2021-11-12 11:36:58 -080012524 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, n_gt_8_small_kernel) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012525 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(3)
12530 .nr(8)
12531 .kr(2)
12532 .sr(4)
12533 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012534 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080012535 .k(k)
12536 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080012537 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012538 }
12539 }
12540 }
12541
Frank Barcharde22685a2021-11-12 11:36:58 -080012542 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, n_div_8_small_kernel) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012543 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(3)
12548 .nr(8)
12549 .kr(2)
12550 .sr(4)
12551 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012552 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080012553 .k(k)
12554 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080012555 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012556 }
12557 }
12558 }
12559
Frank Barcharde22685a2021-11-12 11:36:58 -080012560 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, strided_cm_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012561 TEST_REQUIRES_ARM_NEON;
12562 for (size_t k = 1; k <= 40; k += 9) {
12563 for (uint32_t m = 1; m <= 3; m++) {
12564 for (uint32_t n = 1; n <= 8; n++) {
12565 GemmMicrokernelTester()
12566 .mr(3)
12567 .nr(8)
12568 .kr(2)
12569 .sr(4)
12570 .m(m)
12571 .n(n)
12572 .k(k)
12573 .cm_stride(11)
12574 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012575 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012576 }
12577 }
12578 }
12579 }
12580
Frank Barcharde22685a2021-11-12 11:36:58 -080012581 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, a_offset) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012582 TEST_REQUIRES_ARM_NEON;
12583 for (size_t k = 1; k <= 40; k += 9) {
12584 GemmMicrokernelTester()
12585 .mr(3)
12586 .nr(8)
12587 .kr(2)
12588 .sr(4)
12589 .m(3)
12590 .n(8)
12591 .k(k)
12592 .ks(3)
12593 .a_offset(127)
Marat Dukhan50323b82022-01-11 00:12:01 -080012594 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012595 }
12596 }
12597
Frank Barcharde22685a2021-11-12 11:36:58 -080012598 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, zero) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012599 TEST_REQUIRES_ARM_NEON;
12600 for (uint32_t mz = 0; mz < 3; mz++) {
12601 for (size_t k = 1; k <= 40; k += 9) {
12602 GemmMicrokernelTester()
12603 .mr(3)
12604 .nr(8)
12605 .kr(2)
12606 .sr(4)
12607 .m(3)
12608 .n(8)
12609 .k(k)
12610 .ks(3)
12611 .a_offset(127)
12612 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080012613 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012614 }
12615 }
12616 }
12617
Frank Barcharde22685a2021-11-12 11:36:58 -080012618 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, qmin) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012619 TEST_REQUIRES_ARM_NEON;
12620 GemmMicrokernelTester()
12621 .mr(3)
12622 .nr(8)
12623 .kr(2)
12624 .sr(4)
12625 .m(3)
12626 .n(8)
12627 .k(8)
12628 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080012629 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012630 }
12631
Frank Barcharde22685a2021-11-12 11:36:58 -080012632 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, qmax) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012633 TEST_REQUIRES_ARM_NEON;
12634 GemmMicrokernelTester()
12635 .mr(3)
12636 .nr(8)
12637 .kr(2)
12638 .sr(4)
12639 .m(3)
12640 .n(8)
12641 .k(8)
12642 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080012643 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012644 }
12645
Frank Barcharde22685a2021-11-12 11:36:58 -080012646 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2S4__NEON_MULL, strided_cm) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012647 TEST_REQUIRES_ARM_NEON;
12648 GemmMicrokernelTester()
12649 .mr(3)
12650 .nr(8)
12651 .kr(2)
12652 .sr(4)
12653 .m(3)
12654 .n(8)
12655 .k(8)
12656 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080012657 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012658 }
12659#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
12660
12661
12662#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -080012663 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_eq_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012664 TEST_REQUIRES_ARM_NEON;
12665 GemmMicrokernelTester()
12666 .mr(4)
12667 .nr(8)
12668 .kr(2)
12669 .sr(4)
12670 .m(4)
12671 .n(8)
12672 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080012673 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012674 }
12675
Frank Barcharde22685a2021-11-12 11:36:58 -080012676 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, strided_cn) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012677 TEST_REQUIRES_ARM_NEON;
12678 GemmMicrokernelTester()
12679 .mr(4)
12680 .nr(8)
12681 .kr(2)
12682 .sr(4)
12683 .m(4)
12684 .n(8)
12685 .k(8)
12686 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080012687 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012688 }
12689
Frank Barcharde22685a2021-11-12 11:36:58 -080012690 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_eq_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012691 TEST_REQUIRES_ARM_NEON;
12692 for (uint32_t m = 1; m <= 4; m++) {
12693 for (uint32_t n = 1; n <= 8; n++) {
12694 GemmMicrokernelTester()
12695 .mr(4)
12696 .nr(8)
12697 .kr(2)
12698 .sr(4)
12699 .m(m)
12700 .n(n)
12701 .k(8)
12702 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012703 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012704 }
12705 }
12706 }
12707
Frank Barcharde22685a2021-11-12 11:36:58 -080012708 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_eq_8_subtile_m) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012709 TEST_REQUIRES_ARM_NEON;
12710 for (uint32_t m = 1; m <= 4; m++) {
12711 GemmMicrokernelTester()
12712 .mr(4)
12713 .nr(8)
12714 .kr(2)
12715 .sr(4)
12716 .m(m)
12717 .n(8)
12718 .k(8)
12719 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012720 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012721 }
12722 }
12723
Frank Barcharde22685a2021-11-12 11:36:58 -080012724 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_eq_8_subtile_n) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012725 TEST_REQUIRES_ARM_NEON;
12726 for (uint32_t n = 1; n <= 8; n++) {
12727 GemmMicrokernelTester()
12728 .mr(4)
12729 .nr(8)
12730 .kr(2)
12731 .sr(4)
12732 .m(4)
12733 .n(n)
12734 .k(8)
12735 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012736 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012737 }
12738 }
12739
Frank Barcharde22685a2021-11-12 11:36:58 -080012740 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_lt_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012741 TEST_REQUIRES_ARM_NEON;
12742 for (size_t k = 1; k < 8; k++) {
12743 GemmMicrokernelTester()
12744 .mr(4)
12745 .nr(8)
12746 .kr(2)
12747 .sr(4)
12748 .m(4)
12749 .n(8)
12750 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012751 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012752 }
12753 }
12754
Frank Barcharde22685a2021-11-12 11:36:58 -080012755 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_lt_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012756 TEST_REQUIRES_ARM_NEON;
12757 for (size_t k = 1; k < 8; k++) {
12758 for (uint32_t m = 1; m <= 4; m++) {
12759 for (uint32_t n = 1; n <= 8; n++) {
12760 GemmMicrokernelTester()
12761 .mr(4)
12762 .nr(8)
12763 .kr(2)
12764 .sr(4)
12765 .m(m)
12766 .n(n)
12767 .k(k)
12768 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012769 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012770 }
12771 }
12772 }
12773 }
12774
Frank Barcharde22685a2021-11-12 11:36:58 -080012775 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_gt_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012776 TEST_REQUIRES_ARM_NEON;
12777 for (size_t k = 9; k < 16; k++) {
12778 GemmMicrokernelTester()
12779 .mr(4)
12780 .nr(8)
12781 .kr(2)
12782 .sr(4)
12783 .m(4)
12784 .n(8)
12785 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012786 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012787 }
12788 }
12789
Frank Barcharde22685a2021-11-12 11:36:58 -080012790 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_gt_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012791 TEST_REQUIRES_ARM_NEON;
12792 for (size_t k = 9; k < 16; k++) {
12793 for (uint32_t m = 1; m <= 4; m++) {
12794 for (uint32_t n = 1; n <= 8; n++) {
12795 GemmMicrokernelTester()
12796 .mr(4)
12797 .nr(8)
12798 .kr(2)
12799 .sr(4)
12800 .m(m)
12801 .n(n)
12802 .k(k)
12803 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012804 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012805 }
12806 }
12807 }
12808 }
12809
Frank Barcharde22685a2021-11-12 11:36:58 -080012810 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_div_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012811 TEST_REQUIRES_ARM_NEON;
12812 for (size_t k = 16; k <= 80; k += 8) {
12813 GemmMicrokernelTester()
12814 .mr(4)
12815 .nr(8)
12816 .kr(2)
12817 .sr(4)
12818 .m(4)
12819 .n(8)
12820 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012821 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012822 }
12823 }
12824
Frank Barcharde22685a2021-11-12 11:36:58 -080012825 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_div_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012826 TEST_REQUIRES_ARM_NEON;
12827 for (size_t k = 16; k <= 80; k += 8) {
12828 for (uint32_t m = 1; m <= 4; m++) {
12829 for (uint32_t n = 1; n <= 8; n++) {
12830 GemmMicrokernelTester()
12831 .mr(4)
12832 .nr(8)
12833 .kr(2)
12834 .sr(4)
12835 .m(m)
12836 .n(n)
12837 .k(k)
12838 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012839 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012840 }
12841 }
12842 }
12843 }
12844
Frank Barcharde22685a2021-11-12 11:36:58 -080012845 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_gt_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012846 TEST_REQUIRES_ARM_NEON;
12847 for (uint32_t n = 9; n < 16; n++) {
12848 for (size_t k = 1; k <= 40; k += 9) {
12849 GemmMicrokernelTester()
12850 .mr(4)
12851 .nr(8)
12852 .kr(2)
12853 .sr(4)
12854 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012855 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080012856 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012857 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012858 }
12859 }
12860 }
12861
Frank Barcharde22685a2021-11-12 11:36:58 -080012862 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_gt_8_strided_cn) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012863 TEST_REQUIRES_ARM_NEON;
12864 for (uint32_t n = 9; n < 16; n++) {
12865 for (size_t k = 1; k <= 40; k += 9) {
12866 GemmMicrokernelTester()
12867 .mr(4)
12868 .nr(8)
12869 .kr(2)
12870 .sr(4)
12871 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012872 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080012873 .k(k)
12874 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080012875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012876 }
12877 }
12878 }
12879
Frank Barcharde22685a2021-11-12 11:36:58 -080012880 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_gt_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012881 TEST_REQUIRES_ARM_NEON;
12882 for (uint32_t n = 9; n < 16; n++) {
12883 for (size_t k = 1; k <= 40; k += 9) {
12884 for (uint32_t m = 1; m <= 4; m++) {
12885 GemmMicrokernelTester()
12886 .mr(4)
12887 .nr(8)
12888 .kr(2)
12889 .sr(4)
12890 .m(m)
12891 .n(n)
12892 .k(k)
12893 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012894 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012895 }
12896 }
12897 }
12898 }
12899
Frank Barcharde22685a2021-11-12 11:36:58 -080012900 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_div_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012901 TEST_REQUIRES_ARM_NEON;
12902 for (uint32_t n = 16; n <= 24; n += 8) {
12903 for (size_t k = 1; k <= 40; k += 9) {
12904 GemmMicrokernelTester()
12905 .mr(4)
12906 .nr(8)
12907 .kr(2)
12908 .sr(4)
12909 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080012910 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080012911 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080012912 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012913 }
12914 }
12915 }
12916
Frank Barcharde22685a2021-11-12 11:36:58 -080012917 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_div_8_strided_cn) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012918 TEST_REQUIRES_ARM_NEON;
12919 for (uint32_t n = 16; n <= 24; n += 8) {
12920 for (size_t k = 1; k <= 40; k += 9) {
12921 GemmMicrokernelTester()
12922 .mr(4)
12923 .nr(8)
12924 .kr(2)
12925 .sr(4)
12926 .m(4)
12927 .n(n)
12928 .k(k)
12929 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080012930 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012931 }
12932 }
12933 }
12934
Frank Barcharde22685a2021-11-12 11:36:58 -080012935 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_div_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012936 TEST_REQUIRES_ARM_NEON;
12937 for (uint32_t n = 16; n <= 24; n += 8) {
12938 for (size_t k = 1; k <= 40; k += 9) {
12939 for (uint32_t m = 1; m <= 4; m++) {
12940 GemmMicrokernelTester()
12941 .mr(4)
12942 .nr(8)
12943 .kr(2)
12944 .sr(4)
12945 .m(m)
12946 .n(n)
12947 .k(k)
12948 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012949 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012950 }
12951 }
12952 }
12953 }
12954
Frank Barcharde22685a2021-11-12 11:36:58 -080012955 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, small_kernel) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012956 TEST_REQUIRES_ARM_NEON;
12957 for (size_t k = 1; k <= 40; k += 9) {
12958 GemmMicrokernelTester()
12959 .mr(4)
12960 .nr(8)
12961 .kr(2)
12962 .sr(4)
12963 .m(4)
12964 .n(8)
12965 .k(k)
12966 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080012967 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012968 }
12969 }
12970
Frank Barcharde22685a2021-11-12 11:36:58 -080012971 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, small_kernel_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012972 TEST_REQUIRES_ARM_NEON;
12973 for (size_t k = 1; k <= 40; k += 9) {
12974 for (uint32_t m = 1; m <= 4; m++) {
12975 for (uint32_t n = 1; n <= 8; n++) {
12976 GemmMicrokernelTester()
12977 .mr(4)
12978 .nr(8)
12979 .kr(2)
12980 .sr(4)
12981 .m(m)
12982 .n(n)
12983 .k(k)
12984 .ks(3)
12985 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080012986 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080012987 }
12988 }
12989 }
12990 }
12991
Frank Barcharde22685a2021-11-12 11:36:58 -080012992 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_gt_8_small_kernel) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080012993 TEST_REQUIRES_ARM_NEON;
12994 for (uint32_t n = 9; n < 16; n++) {
12995 for (size_t k = 1; k <= 40; k += 9) {
12996 GemmMicrokernelTester()
12997 .mr(4)
12998 .nr(8)
12999 .kr(2)
13000 .sr(4)
13001 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013002 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080013003 .k(k)
13004 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080013005 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013006 }
13007 }
13008 }
13009
Frank Barcharde22685a2021-11-12 11:36:58 -080013010 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_div_8_small_kernel) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013011 TEST_REQUIRES_ARM_NEON;
13012 for (uint32_t n = 16; n <= 24; n += 8) {
13013 for (size_t k = 1; k <= 40; k += 9) {
13014 GemmMicrokernelTester()
13015 .mr(4)
13016 .nr(8)
13017 .kr(2)
13018 .sr(4)
13019 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013020 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080013021 .k(k)
13022 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080013023 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013024 }
13025 }
13026 }
13027
Frank Barcharde22685a2021-11-12 11:36:58 -080013028 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, strided_cm_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013029 TEST_REQUIRES_ARM_NEON;
13030 for (size_t k = 1; k <= 40; k += 9) {
13031 for (uint32_t m = 1; m <= 4; m++) {
13032 for (uint32_t n = 1; n <= 8; n++) {
13033 GemmMicrokernelTester()
13034 .mr(4)
13035 .nr(8)
13036 .kr(2)
13037 .sr(4)
13038 .m(m)
13039 .n(n)
13040 .k(k)
13041 .cm_stride(11)
13042 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013043 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013044 }
13045 }
13046 }
13047 }
13048
Frank Barcharde22685a2021-11-12 11:36:58 -080013049 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, a_offset) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013050 TEST_REQUIRES_ARM_NEON;
13051 for (size_t k = 1; k <= 40; k += 9) {
13052 GemmMicrokernelTester()
13053 .mr(4)
13054 .nr(8)
13055 .kr(2)
13056 .sr(4)
13057 .m(4)
13058 .n(8)
13059 .k(k)
13060 .ks(3)
13061 .a_offset(163)
Marat Dukhan50323b82022-01-11 00:12:01 -080013062 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013063 }
13064 }
13065
Frank Barcharde22685a2021-11-12 11:36:58 -080013066 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, zero) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013067 TEST_REQUIRES_ARM_NEON;
13068 for (uint32_t mz = 0; mz < 4; mz++) {
13069 for (size_t k = 1; k <= 40; k += 9) {
13070 GemmMicrokernelTester()
13071 .mr(4)
13072 .nr(8)
13073 .kr(2)
13074 .sr(4)
13075 .m(4)
13076 .n(8)
13077 .k(k)
13078 .ks(3)
13079 .a_offset(163)
13080 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080013081 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013082 }
13083 }
13084 }
13085
Frank Barcharde22685a2021-11-12 11:36:58 -080013086 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, qmin) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013087 TEST_REQUIRES_ARM_NEON;
13088 GemmMicrokernelTester()
13089 .mr(4)
13090 .nr(8)
13091 .kr(2)
13092 .sr(4)
13093 .m(4)
13094 .n(8)
13095 .k(8)
13096 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080013097 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013098 }
13099
Frank Barcharde22685a2021-11-12 11:36:58 -080013100 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, qmax) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013101 TEST_REQUIRES_ARM_NEON;
13102 GemmMicrokernelTester()
13103 .mr(4)
13104 .nr(8)
13105 .kr(2)
13106 .sr(4)
13107 .m(4)
13108 .n(8)
13109 .k(8)
13110 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080013111 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013112 }
13113
Frank Barcharde22685a2021-11-12 11:36:58 -080013114 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, strided_cm) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013115 TEST_REQUIRES_ARM_NEON;
13116 GemmMicrokernelTester()
13117 .mr(4)
13118 .nr(8)
13119 .kr(2)
13120 .sr(4)
13121 .m(4)
13122 .n(8)
13123 .k(8)
13124 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080013125 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013126 }
13127#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
13128
13129
13130#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -080013131 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_eq_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013132 TEST_REQUIRES_ARM_NEON;
13133 GemmMicrokernelTester()
13134 .mr(2)
13135 .nr(16)
13136 .kr(2)
13137 .sr(4)
13138 .m(2)
13139 .n(16)
13140 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080013141 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013142 }
13143
Frank Barcharde22685a2021-11-12 11:36:58 -080013144 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, strided_cn) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013145 TEST_REQUIRES_ARM_NEON;
13146 GemmMicrokernelTester()
13147 .mr(2)
13148 .nr(16)
13149 .kr(2)
13150 .sr(4)
13151 .m(2)
13152 .n(16)
13153 .k(8)
13154 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080013155 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013156 }
13157
Frank Barcharde22685a2021-11-12 11:36:58 -080013158 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_eq_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013159 TEST_REQUIRES_ARM_NEON;
13160 for (uint32_t m = 1; m <= 2; m++) {
13161 for (uint32_t n = 1; n <= 16; n++) {
13162 GemmMicrokernelTester()
13163 .mr(2)
13164 .nr(16)
13165 .kr(2)
13166 .sr(4)
13167 .m(m)
13168 .n(n)
13169 .k(8)
13170 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013171 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013172 }
13173 }
13174 }
13175
Frank Barcharde22685a2021-11-12 11:36:58 -080013176 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_eq_8_subtile_m) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013177 TEST_REQUIRES_ARM_NEON;
13178 for (uint32_t m = 1; m <= 2; m++) {
13179 GemmMicrokernelTester()
13180 .mr(2)
13181 .nr(16)
13182 .kr(2)
13183 .sr(4)
13184 .m(m)
13185 .n(16)
13186 .k(8)
13187 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013188 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013189 }
13190 }
13191
Frank Barcharde22685a2021-11-12 11:36:58 -080013192 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_eq_8_subtile_n) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013193 TEST_REQUIRES_ARM_NEON;
13194 for (uint32_t n = 1; n <= 16; n++) {
13195 GemmMicrokernelTester()
13196 .mr(2)
13197 .nr(16)
13198 .kr(2)
13199 .sr(4)
13200 .m(2)
13201 .n(n)
13202 .k(8)
13203 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013204 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013205 }
13206 }
13207
Frank Barcharde22685a2021-11-12 11:36:58 -080013208 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_lt_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013209 TEST_REQUIRES_ARM_NEON;
13210 for (size_t k = 1; k < 8; k++) {
13211 GemmMicrokernelTester()
13212 .mr(2)
13213 .nr(16)
13214 .kr(2)
13215 .sr(4)
13216 .m(2)
13217 .n(16)
13218 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080013219 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013220 }
13221 }
13222
Frank Barcharde22685a2021-11-12 11:36:58 -080013223 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_lt_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013224 TEST_REQUIRES_ARM_NEON;
13225 for (size_t k = 1; k < 8; k++) {
13226 for (uint32_t m = 1; m <= 2; m++) {
13227 for (uint32_t n = 1; n <= 16; n++) {
13228 GemmMicrokernelTester()
13229 .mr(2)
13230 .nr(16)
13231 .kr(2)
13232 .sr(4)
13233 .m(m)
13234 .n(n)
13235 .k(k)
13236 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013237 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013238 }
13239 }
13240 }
13241 }
13242
Frank Barcharde22685a2021-11-12 11:36:58 -080013243 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_gt_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013244 TEST_REQUIRES_ARM_NEON;
13245 for (size_t k = 9; k < 16; k++) {
13246 GemmMicrokernelTester()
13247 .mr(2)
13248 .nr(16)
13249 .kr(2)
13250 .sr(4)
13251 .m(2)
13252 .n(16)
13253 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080013254 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013255 }
13256 }
13257
Frank Barcharde22685a2021-11-12 11:36:58 -080013258 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_gt_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013259 TEST_REQUIRES_ARM_NEON;
13260 for (size_t k = 9; k < 16; k++) {
13261 for (uint32_t m = 1; m <= 2; m++) {
13262 for (uint32_t n = 1; n <= 16; n++) {
13263 GemmMicrokernelTester()
13264 .mr(2)
13265 .nr(16)
13266 .kr(2)
13267 .sr(4)
13268 .m(m)
13269 .n(n)
13270 .k(k)
13271 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013272 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013273 }
13274 }
13275 }
13276 }
13277
Frank Barcharde22685a2021-11-12 11:36:58 -080013278 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_div_8) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013279 TEST_REQUIRES_ARM_NEON;
13280 for (size_t k = 16; k <= 80; k += 8) {
13281 GemmMicrokernelTester()
13282 .mr(2)
13283 .nr(16)
13284 .kr(2)
13285 .sr(4)
13286 .m(2)
13287 .n(16)
13288 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080013289 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013290 }
13291 }
13292
Frank Barcharde22685a2021-11-12 11:36:58 -080013293 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, k_div_8_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013294 TEST_REQUIRES_ARM_NEON;
13295 for (size_t k = 16; k <= 80; k += 8) {
13296 for (uint32_t m = 1; m <= 2; m++) {
13297 for (uint32_t n = 1; n <= 16; n++) {
13298 GemmMicrokernelTester()
13299 .mr(2)
13300 .nr(16)
13301 .kr(2)
13302 .sr(4)
13303 .m(m)
13304 .n(n)
13305 .k(k)
13306 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013307 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013308 }
13309 }
13310 }
13311 }
13312
Frank Barcharde22685a2021-11-12 11:36:58 -080013313 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, n_gt_16) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013314 TEST_REQUIRES_ARM_NEON;
13315 for (uint32_t n = 17; n < 32; n++) {
13316 for (size_t k = 1; k <= 40; k += 9) {
13317 GemmMicrokernelTester()
13318 .mr(2)
13319 .nr(16)
13320 .kr(2)
13321 .sr(4)
13322 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013323 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080013324 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080013325 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013326 }
13327 }
13328 }
13329
Frank Barcharde22685a2021-11-12 11:36:58 -080013330 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, n_gt_16_strided_cn) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013331 TEST_REQUIRES_ARM_NEON;
13332 for (uint32_t n = 17; n < 32; n++) {
13333 for (size_t k = 1; k <= 40; k += 9) {
13334 GemmMicrokernelTester()
13335 .mr(2)
13336 .nr(16)
13337 .kr(2)
13338 .sr(4)
13339 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013340 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080013341 .k(k)
13342 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080013343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013344 }
13345 }
13346 }
13347
Frank Barcharde22685a2021-11-12 11:36:58 -080013348 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, n_gt_16_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013349 TEST_REQUIRES_ARM_NEON;
13350 for (uint32_t n = 17; n < 32; n++) {
13351 for (size_t k = 1; k <= 40; k += 9) {
13352 for (uint32_t m = 1; m <= 2; m++) {
13353 GemmMicrokernelTester()
13354 .mr(2)
13355 .nr(16)
13356 .kr(2)
13357 .sr(4)
13358 .m(m)
13359 .n(n)
13360 .k(k)
13361 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013362 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013363 }
13364 }
13365 }
13366 }
13367
Frank Barcharde22685a2021-11-12 11:36:58 -080013368 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, n_div_16) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013369 TEST_REQUIRES_ARM_NEON;
13370 for (uint32_t n = 32; n <= 48; n += 16) {
13371 for (size_t k = 1; k <= 40; k += 9) {
13372 GemmMicrokernelTester()
13373 .mr(2)
13374 .nr(16)
13375 .kr(2)
13376 .sr(4)
13377 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013378 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080013379 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080013380 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013381 }
13382 }
13383 }
13384
Frank Barcharde22685a2021-11-12 11:36:58 -080013385 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, n_div_16_strided_cn) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013386 TEST_REQUIRES_ARM_NEON;
13387 for (uint32_t n = 32; n <= 48; n += 16) {
13388 for (size_t k = 1; k <= 40; k += 9) {
13389 GemmMicrokernelTester()
13390 .mr(2)
13391 .nr(16)
13392 .kr(2)
13393 .sr(4)
13394 .m(2)
13395 .n(n)
13396 .k(k)
13397 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080013398 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013399 }
13400 }
13401 }
13402
Frank Barcharde22685a2021-11-12 11:36:58 -080013403 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, n_div_16_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013404 TEST_REQUIRES_ARM_NEON;
13405 for (uint32_t n = 32; n <= 48; n += 16) {
13406 for (size_t k = 1; k <= 40; k += 9) {
13407 for (uint32_t m = 1; m <= 2; m++) {
13408 GemmMicrokernelTester()
13409 .mr(2)
13410 .nr(16)
13411 .kr(2)
13412 .sr(4)
13413 .m(m)
13414 .n(n)
13415 .k(k)
13416 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013417 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013418 }
13419 }
13420 }
13421 }
13422
Frank Barcharde22685a2021-11-12 11:36:58 -080013423 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, small_kernel) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013424 TEST_REQUIRES_ARM_NEON;
13425 for (size_t k = 1; k <= 40; k += 9) {
13426 GemmMicrokernelTester()
13427 .mr(2)
13428 .nr(16)
13429 .kr(2)
13430 .sr(4)
13431 .m(2)
13432 .n(16)
13433 .k(k)
13434 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080013435 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013436 }
13437 }
13438
Frank Barcharde22685a2021-11-12 11:36:58 -080013439 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, small_kernel_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013440 TEST_REQUIRES_ARM_NEON;
13441 for (size_t k = 1; k <= 40; k += 9) {
13442 for (uint32_t m = 1; m <= 2; m++) {
13443 for (uint32_t n = 1; n <= 16; n++) {
13444 GemmMicrokernelTester()
13445 .mr(2)
13446 .nr(16)
13447 .kr(2)
13448 .sr(4)
13449 .m(m)
13450 .n(n)
13451 .k(k)
13452 .ks(3)
13453 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013454 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013455 }
13456 }
13457 }
13458 }
13459
Frank Barcharde22685a2021-11-12 11:36:58 -080013460 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, n_gt_16_small_kernel) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013461 TEST_REQUIRES_ARM_NEON;
13462 for (uint32_t n = 17; n < 32; n++) {
13463 for (size_t k = 1; k <= 40; k += 9) {
13464 GemmMicrokernelTester()
13465 .mr(2)
13466 .nr(16)
13467 .kr(2)
13468 .sr(4)
13469 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013470 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080013471 .k(k)
13472 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080013473 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013474 }
13475 }
13476 }
13477
Frank Barcharde22685a2021-11-12 11:36:58 -080013478 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, n_div_16_small_kernel) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013479 TEST_REQUIRES_ARM_NEON;
13480 for (uint32_t n = 32; n <= 48; n += 16) {
13481 for (size_t k = 1; k <= 40; k += 9) {
13482 GemmMicrokernelTester()
13483 .mr(2)
13484 .nr(16)
13485 .kr(2)
13486 .sr(4)
13487 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013488 .n(n)
Frank Barchardc7a032d2021-11-10 12:37:49 -080013489 .k(k)
13490 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080013491 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013492 }
13493 }
13494 }
13495
Frank Barcharde22685a2021-11-12 11:36:58 -080013496 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, strided_cm_subtile) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013497 TEST_REQUIRES_ARM_NEON;
13498 for (size_t k = 1; k <= 40; k += 9) {
13499 for (uint32_t m = 1; m <= 2; m++) {
13500 for (uint32_t n = 1; n <= 16; n++) {
13501 GemmMicrokernelTester()
13502 .mr(2)
13503 .nr(16)
13504 .kr(2)
13505 .sr(4)
13506 .m(m)
13507 .n(n)
13508 .k(k)
13509 .cm_stride(19)
13510 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080013511 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013512 }
13513 }
13514 }
13515 }
13516
Frank Barcharde22685a2021-11-12 11:36:58 -080013517 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, a_offset) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013518 TEST_REQUIRES_ARM_NEON;
13519 for (size_t k = 1; k <= 40; k += 9) {
13520 GemmMicrokernelTester()
13521 .mr(2)
13522 .nr(16)
13523 .kr(2)
13524 .sr(4)
13525 .m(2)
13526 .n(16)
13527 .k(k)
13528 .ks(3)
13529 .a_offset(83)
Marat Dukhan50323b82022-01-11 00:12:01 -080013530 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013531 }
13532 }
13533
Frank Barcharde22685a2021-11-12 11:36:58 -080013534 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, zero) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013535 TEST_REQUIRES_ARM_NEON;
13536 for (uint32_t mz = 0; mz < 2; mz++) {
13537 for (size_t k = 1; k <= 40; k += 9) {
13538 GemmMicrokernelTester()
13539 .mr(2)
13540 .nr(16)
13541 .kr(2)
13542 .sr(4)
13543 .m(2)
13544 .n(16)
13545 .k(k)
13546 .ks(3)
13547 .a_offset(83)
13548 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080013549 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013550 }
13551 }
13552 }
13553
Frank Barcharde22685a2021-11-12 11:36:58 -080013554 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, qmin) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013555 TEST_REQUIRES_ARM_NEON;
13556 GemmMicrokernelTester()
13557 .mr(2)
13558 .nr(16)
13559 .kr(2)
13560 .sr(4)
13561 .m(2)
13562 .n(16)
13563 .k(8)
13564 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080013565 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013566 }
13567
Frank Barcharde22685a2021-11-12 11:36:58 -080013568 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, qmax) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013569 TEST_REQUIRES_ARM_NEON;
13570 GemmMicrokernelTester()
13571 .mr(2)
13572 .nr(16)
13573 .kr(2)
13574 .sr(4)
13575 .m(2)
13576 .n(16)
13577 .k(8)
13578 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080013579 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013580 }
13581
Frank Barcharde22685a2021-11-12 11:36:58 -080013582 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2S4__NEON_MULL, strided_cm) {
Frank Barchardc7a032d2021-11-10 12:37:49 -080013583 TEST_REQUIRES_ARM_NEON;
13584 GemmMicrokernelTester()
13585 .mr(2)
13586 .nr(16)
13587 .kr(2)
13588 .sr(4)
13589 .m(2)
13590 .n(16)
13591 .k(8)
13592 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080013593 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchardc7a032d2021-11-10 12:37:49 -080013594 }
13595#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
13596
13597
13598#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013599 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_eq_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070013600 TEST_REQUIRES_ARM_NEON;
13601 GemmMicrokernelTester()
13602 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013603 .nr(16)
13604 .kr(2)
13605 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013606 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013607 .n(16)
13608 .k(16)
13609 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013610 }
13611
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013612 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070013613 TEST_REQUIRES_ARM_NEON;
13614 GemmMicrokernelTester()
13615 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013616 .nr(16)
13617 .kr(2)
13618 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013619 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013620 .n(16)
13621 .k(16)
13622 .cn_stride(19)
13623 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013624 }
13625
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013626 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_eq_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070013627 TEST_REQUIRES_ARM_NEON;
13628 for (uint32_t m = 1; m <= 1; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013629 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013630 GemmMicrokernelTester()
13631 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013632 .nr(16)
13633 .kr(2)
13634 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013635 .m(m)
13636 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013637 .k(16)
Frank Barchard287952a2021-11-03 15:26:45 -070013638 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013639 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013640 }
13641 }
13642 }
13643
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013644 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_eq_16_subtile_m) {
Frank Barchard287952a2021-11-03 15:26:45 -070013645 TEST_REQUIRES_ARM_NEON;
13646 for (uint32_t m = 1; m <= 1; m++) {
13647 GemmMicrokernelTester()
13648 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013649 .nr(16)
13650 .kr(2)
13651 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013652 .m(m)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013653 .n(16)
13654 .k(16)
Frank Barchard287952a2021-11-03 15:26:45 -070013655 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013656 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013657 }
13658 }
13659
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013660 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_eq_16_subtile_n) {
Frank Barchard287952a2021-11-03 15:26:45 -070013661 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013662 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013663 GemmMicrokernelTester()
13664 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013665 .nr(16)
13666 .kr(2)
13667 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013668 .m(1)
13669 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013670 .k(16)
Frank Barchard287952a2021-11-03 15:26:45 -070013671 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013672 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013673 }
13674 }
13675
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013676 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_lt_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070013677 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013678 for (size_t k = 1; k < 16; k++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013679 GemmMicrokernelTester()
13680 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013681 .nr(16)
13682 .kr(2)
13683 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013684 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013685 .n(16)
Frank Barchard287952a2021-11-03 15:26:45 -070013686 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013687 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013688 }
13689 }
13690
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013691 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_lt_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070013692 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013693 for (size_t k = 1; k < 16; k++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013694 for (uint32_t m = 1; m <= 1; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013695 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013696 GemmMicrokernelTester()
13697 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013698 .nr(16)
13699 .kr(2)
13700 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013701 .m(m)
13702 .n(n)
13703 .k(k)
13704 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013705 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013706 }
13707 }
13708 }
13709 }
13710
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013711 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_gt_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070013712 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013713 for (size_t k = 17; k < 32; k++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013714 GemmMicrokernelTester()
13715 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013716 .nr(16)
13717 .kr(2)
13718 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013719 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013720 .n(16)
Frank Barchard287952a2021-11-03 15:26:45 -070013721 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013722 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013723 }
13724 }
13725
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013726 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_gt_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070013727 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013728 for (size_t k = 17; k < 32; k++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013729 for (uint32_t m = 1; m <= 1; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013730 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013731 GemmMicrokernelTester()
13732 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013733 .nr(16)
13734 .kr(2)
13735 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013736 .m(m)
13737 .n(n)
13738 .k(k)
13739 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013740 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013741 }
13742 }
13743 }
13744 }
13745
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013746 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_div_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070013747 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013748 for (size_t k = 32; k <= 160; k += 16) {
Frank Barchard287952a2021-11-03 15:26:45 -070013749 GemmMicrokernelTester()
13750 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013751 .nr(16)
13752 .kr(2)
13753 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013754 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013755 .n(16)
Frank Barchard287952a2021-11-03 15:26:45 -070013756 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013757 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013758 }
13759 }
13760
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013761 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, k_div_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070013762 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013763 for (size_t k = 32; k <= 160; k += 16) {
Frank Barchard287952a2021-11-03 15:26:45 -070013764 for (uint32_t m = 1; m <= 1; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013765 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013766 GemmMicrokernelTester()
13767 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013768 .nr(16)
13769 .kr(2)
13770 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013771 .m(m)
13772 .n(n)
13773 .k(k)
13774 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013775 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013776 }
13777 }
13778 }
13779 }
13780
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013781 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, n_gt_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070013782 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013783 for (uint32_t n = 17; n < 32; n++) {
13784 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013785 GemmMicrokernelTester()
13786 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013787 .nr(16)
13788 .kr(2)
13789 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013790 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013791 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070013792 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013793 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013794 }
13795 }
13796 }
13797
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013798 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, n_gt_16_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070013799 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013800 for (uint32_t n = 17; n < 32; n++) {
13801 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013802 GemmMicrokernelTester()
13803 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013804 .nr(16)
13805 .kr(2)
13806 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013807 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013808 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070013809 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013810 .cn_stride(19)
13811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013812 }
13813 }
13814 }
13815
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013816 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, n_gt_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070013817 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013818 for (uint32_t n = 17; n < 32; n++) {
13819 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013820 for (uint32_t m = 1; m <= 1; m++) {
13821 GemmMicrokernelTester()
13822 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013823 .nr(16)
13824 .kr(2)
13825 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013826 .m(m)
13827 .n(n)
13828 .k(k)
13829 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013830 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013831 }
13832 }
13833 }
13834 }
13835
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013836 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, n_div_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070013837 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013838 for (uint32_t n = 32; n <= 48; n += 16) {
13839 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013840 GemmMicrokernelTester()
13841 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013842 .nr(16)
13843 .kr(2)
13844 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013845 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013846 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070013847 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013848 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013849 }
13850 }
13851 }
13852
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013853 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, n_div_16_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070013854 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013855 for (uint32_t n = 32; n <= 48; n += 16) {
13856 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013857 GemmMicrokernelTester()
13858 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013859 .nr(16)
13860 .kr(2)
13861 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013862 .m(1)
13863 .n(n)
13864 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013865 .cn_stride(19)
13866 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013867 }
13868 }
13869 }
13870
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013871 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, n_div_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070013872 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013873 for (uint32_t n = 32; n <= 48; n += 16) {
13874 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013875 for (uint32_t m = 1; m <= 1; m++) {
13876 GemmMicrokernelTester()
13877 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013878 .nr(16)
13879 .kr(2)
13880 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013881 .m(m)
13882 .n(n)
13883 .k(k)
13884 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013885 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013886 }
13887 }
13888 }
13889 }
13890
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013891 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070013892 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013893 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013894 GemmMicrokernelTester()
13895 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013896 .nr(16)
13897 .kr(2)
13898 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013899 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013900 .n(16)
Frank Barchard287952a2021-11-03 15:26:45 -070013901 .k(k)
13902 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013903 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013904 }
13905 }
13906
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013907 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, small_kernel_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070013908 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013909 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013910 for (uint32_t m = 1; m <= 1; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013911 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013912 GemmMicrokernelTester()
13913 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013914 .nr(16)
13915 .kr(2)
13916 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013917 .m(m)
13918 .n(n)
13919 .k(k)
13920 .ks(3)
13921 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013922 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013923 }
13924 }
13925 }
13926 }
13927
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013928 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, n_gt_16_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070013929 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013930 for (uint32_t n = 17; n < 32; n++) {
13931 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013932 GemmMicrokernelTester()
13933 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013934 .nr(16)
13935 .kr(2)
13936 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013937 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013938 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070013939 .k(k)
13940 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013941 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013942 }
13943 }
13944 }
13945
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013946 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, n_div_16_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070013947 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013948 for (uint32_t n = 32; n <= 48; n += 16) {
13949 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013950 GemmMicrokernelTester()
13951 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013952 .nr(16)
13953 .kr(2)
13954 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013955 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080013956 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070013957 .k(k)
13958 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013959 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013960 }
13961 }
13962 }
13963
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013964 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, strided_cm_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070013965 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013966 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013967 for (uint32_t m = 1; m <= 1; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013968 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard287952a2021-11-03 15:26:45 -070013969 GemmMicrokernelTester()
13970 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013971 .nr(16)
13972 .kr(2)
13973 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013974 .m(m)
13975 .n(n)
13976 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013977 .cm_stride(19)
Frank Barchard287952a2021-11-03 15:26:45 -070013978 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013979 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013980 }
13981 }
13982 }
13983 }
13984
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013985 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, a_offset) {
Frank Barchard287952a2021-11-03 15:26:45 -070013986 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013987 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070013988 GemmMicrokernelTester()
13989 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013990 .nr(16)
13991 .kr(2)
13992 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070013993 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013994 .n(16)
Frank Barchard287952a2021-11-03 15:26:45 -070013995 .k(k)
13996 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080013997 .a_offset(83)
13998 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070013999 }
14000 }
14001
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014002 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, zero) {
Frank Barchard287952a2021-11-03 15:26:45 -070014003 TEST_REQUIRES_ARM_NEON;
14004 for (uint32_t mz = 0; mz < 1; mz++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014005 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard287952a2021-11-03 15:26:45 -070014006 GemmMicrokernelTester()
14007 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014008 .nr(16)
14009 .kr(2)
14010 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070014011 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014012 .n(16)
Frank Barchard287952a2021-11-03 15:26:45 -070014013 .k(k)
14014 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014015 .a_offset(83)
Frank Barchard287952a2021-11-03 15:26:45 -070014016 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014017 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014018 }
14019 }
14020 }
14021
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014022 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, qmin) {
Frank Barchard287952a2021-11-03 15:26:45 -070014023 TEST_REQUIRES_ARM_NEON;
14024 GemmMicrokernelTester()
14025 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014026 .nr(16)
14027 .kr(2)
14028 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070014029 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014030 .n(16)
14031 .k(16)
Frank Barchard287952a2021-11-03 15:26:45 -070014032 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014033 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014034 }
14035
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014036 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, qmax) {
Frank Barchard287952a2021-11-03 15:26:45 -070014037 TEST_REQUIRES_ARM_NEON;
14038 GemmMicrokernelTester()
14039 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014040 .nr(16)
14041 .kr(2)
14042 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070014043 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014044 .n(16)
14045 .k(16)
Frank Barchard287952a2021-11-03 15:26:45 -070014046 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014047 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014048 }
14049
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014050 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MLAL, strided_cm) {
Frank Barchard287952a2021-11-03 15:26:45 -070014051 TEST_REQUIRES_ARM_NEON;
14052 GemmMicrokernelTester()
14053 .mr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014054 .nr(16)
14055 .kr(2)
14056 .sr(4)
Frank Barchard287952a2021-11-03 15:26:45 -070014057 .m(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080014058 .n(16)
14059 .k(16)
14060 .cm_stride(19)
14061 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014062 }
14063#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
14064
14065
14066#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -080014067 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_eq_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014068 TEST_REQUIRES_ARM_NEON;
14069 GemmMicrokernelTester()
14070 .mr(2)
14071 .nr(8)
14072 .kr(4)
14073 .sr(1)
14074 .m(2)
14075 .n(8)
14076 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080014077 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014078 }
14079
Frank Barcharde22685a2021-11-12 11:36:58 -080014080 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070014081 TEST_REQUIRES_ARM_NEON;
14082 GemmMicrokernelTester()
14083 .mr(2)
14084 .nr(8)
14085 .kr(4)
14086 .sr(1)
14087 .m(2)
14088 .n(8)
14089 .k(8)
14090 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080014091 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014092 }
14093
Frank Barcharde22685a2021-11-12 11:36:58 -080014094 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_eq_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014095 TEST_REQUIRES_ARM_NEON;
14096 for (uint32_t m = 1; m <= 2; m++) {
14097 for (uint32_t n = 1; n <= 8; n++) {
14098 GemmMicrokernelTester()
14099 .mr(2)
14100 .nr(8)
14101 .kr(4)
14102 .sr(1)
14103 .m(m)
14104 .n(n)
14105 .k(8)
14106 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014107 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014108 }
14109 }
14110 }
14111
Frank Barcharde22685a2021-11-12 11:36:58 -080014112 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_eq_8_subtile_m) {
Frank Barchard287952a2021-11-03 15:26:45 -070014113 TEST_REQUIRES_ARM_NEON;
14114 for (uint32_t m = 1; m <= 2; m++) {
14115 GemmMicrokernelTester()
14116 .mr(2)
14117 .nr(8)
14118 .kr(4)
14119 .sr(1)
14120 .m(m)
14121 .n(8)
14122 .k(8)
14123 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014124 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014125 }
14126 }
14127
Frank Barcharde22685a2021-11-12 11:36:58 -080014128 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_eq_8_subtile_n) {
Frank Barchard287952a2021-11-03 15:26:45 -070014129 TEST_REQUIRES_ARM_NEON;
14130 for (uint32_t n = 1; n <= 8; n++) {
14131 GemmMicrokernelTester()
14132 .mr(2)
14133 .nr(8)
14134 .kr(4)
14135 .sr(1)
14136 .m(2)
14137 .n(n)
14138 .k(8)
14139 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014140 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014141 }
14142 }
14143
Frank Barcharde22685a2021-11-12 11:36:58 -080014144 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_lt_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014145 TEST_REQUIRES_ARM_NEON;
14146 for (size_t k = 1; k < 8; k++) {
14147 GemmMicrokernelTester()
14148 .mr(2)
14149 .nr(8)
14150 .kr(4)
14151 .sr(1)
14152 .m(2)
14153 .n(8)
14154 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014155 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014156 }
14157 }
14158
Frank Barcharde22685a2021-11-12 11:36:58 -080014159 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_lt_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014160 TEST_REQUIRES_ARM_NEON;
14161 for (size_t k = 1; k < 8; k++) {
14162 for (uint32_t m = 1; m <= 2; m++) {
14163 for (uint32_t n = 1; n <= 8; n++) {
14164 GemmMicrokernelTester()
14165 .mr(2)
14166 .nr(8)
14167 .kr(4)
14168 .sr(1)
14169 .m(m)
14170 .n(n)
14171 .k(k)
14172 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014173 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014174 }
14175 }
14176 }
14177 }
14178
Frank Barcharde22685a2021-11-12 11:36:58 -080014179 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_gt_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014180 TEST_REQUIRES_ARM_NEON;
14181 for (size_t k = 9; k < 16; k++) {
14182 GemmMicrokernelTester()
14183 .mr(2)
14184 .nr(8)
14185 .kr(4)
14186 .sr(1)
14187 .m(2)
14188 .n(8)
14189 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014190 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014191 }
14192 }
14193
Frank Barcharde22685a2021-11-12 11:36:58 -080014194 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_gt_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014195 TEST_REQUIRES_ARM_NEON;
14196 for (size_t k = 9; k < 16; k++) {
14197 for (uint32_t m = 1; m <= 2; m++) {
14198 for (uint32_t n = 1; n <= 8; n++) {
14199 GemmMicrokernelTester()
14200 .mr(2)
14201 .nr(8)
14202 .kr(4)
14203 .sr(1)
14204 .m(m)
14205 .n(n)
14206 .k(k)
14207 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014208 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014209 }
14210 }
14211 }
14212 }
14213
Frank Barcharde22685a2021-11-12 11:36:58 -080014214 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_div_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014215 TEST_REQUIRES_ARM_NEON;
14216 for (size_t k = 16; k <= 80; k += 8) {
14217 GemmMicrokernelTester()
14218 .mr(2)
14219 .nr(8)
14220 .kr(4)
14221 .sr(1)
14222 .m(2)
14223 .n(8)
14224 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014225 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014226 }
14227 }
14228
Frank Barcharde22685a2021-11-12 11:36:58 -080014229 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, k_div_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014230 TEST_REQUIRES_ARM_NEON;
14231 for (size_t k = 16; k <= 80; k += 8) {
14232 for (uint32_t m = 1; m <= 2; m++) {
14233 for (uint32_t n = 1; n <= 8; n++) {
14234 GemmMicrokernelTester()
14235 .mr(2)
14236 .nr(8)
14237 .kr(4)
14238 .sr(1)
14239 .m(m)
14240 .n(n)
14241 .k(k)
14242 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014243 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014244 }
14245 }
14246 }
14247 }
14248
Frank Barcharde22685a2021-11-12 11:36:58 -080014249 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, n_gt_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014250 TEST_REQUIRES_ARM_NEON;
14251 for (uint32_t n = 9; n < 16; n++) {
14252 for (size_t k = 1; k <= 40; k += 9) {
14253 GemmMicrokernelTester()
14254 .mr(2)
14255 .nr(8)
14256 .kr(4)
14257 .sr(1)
14258 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014259 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014260 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014261 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014262 }
14263 }
14264 }
14265
Frank Barcharde22685a2021-11-12 11:36:58 -080014266 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, n_gt_8_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070014267 TEST_REQUIRES_ARM_NEON;
14268 for (uint32_t n = 9; n < 16; n++) {
14269 for (size_t k = 1; k <= 40; k += 9) {
14270 GemmMicrokernelTester()
14271 .mr(2)
14272 .nr(8)
14273 .kr(4)
14274 .sr(1)
14275 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014276 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014277 .k(k)
14278 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080014279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014280 }
14281 }
14282 }
14283
Frank Barcharde22685a2021-11-12 11:36:58 -080014284 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, n_gt_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014285 TEST_REQUIRES_ARM_NEON;
14286 for (uint32_t n = 9; n < 16; n++) {
14287 for (size_t k = 1; k <= 40; k += 9) {
14288 for (uint32_t m = 1; m <= 2; m++) {
14289 GemmMicrokernelTester()
14290 .mr(2)
14291 .nr(8)
14292 .kr(4)
14293 .sr(1)
14294 .m(m)
14295 .n(n)
14296 .k(k)
14297 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014298 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014299 }
14300 }
14301 }
14302 }
14303
Frank Barcharde22685a2021-11-12 11:36:58 -080014304 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, n_div_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014305 TEST_REQUIRES_ARM_NEON;
14306 for (uint32_t n = 16; n <= 24; n += 8) {
14307 for (size_t k = 1; k <= 40; k += 9) {
14308 GemmMicrokernelTester()
14309 .mr(2)
14310 .nr(8)
14311 .kr(4)
14312 .sr(1)
14313 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014314 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014315 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014316 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014317 }
14318 }
14319 }
14320
Frank Barcharde22685a2021-11-12 11:36:58 -080014321 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, n_div_8_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070014322 TEST_REQUIRES_ARM_NEON;
14323 for (uint32_t n = 16; n <= 24; n += 8) {
14324 for (size_t k = 1; k <= 40; k += 9) {
14325 GemmMicrokernelTester()
14326 .mr(2)
14327 .nr(8)
14328 .kr(4)
14329 .sr(1)
14330 .m(2)
14331 .n(n)
14332 .k(k)
14333 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080014334 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014335 }
14336 }
14337 }
14338
Frank Barcharde22685a2021-11-12 11:36:58 -080014339 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, n_div_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014340 TEST_REQUIRES_ARM_NEON;
14341 for (uint32_t n = 16; n <= 24; n += 8) {
14342 for (size_t k = 1; k <= 40; k += 9) {
14343 for (uint32_t m = 1; m <= 2; m++) {
14344 GemmMicrokernelTester()
14345 .mr(2)
14346 .nr(8)
14347 .kr(4)
14348 .sr(1)
14349 .m(m)
14350 .n(n)
14351 .k(k)
14352 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014353 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014354 }
14355 }
14356 }
14357 }
14358
Frank Barcharde22685a2021-11-12 11:36:58 -080014359 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070014360 TEST_REQUIRES_ARM_NEON;
14361 for (size_t k = 1; k <= 40; k += 9) {
14362 GemmMicrokernelTester()
14363 .mr(2)
14364 .nr(8)
14365 .kr(4)
14366 .sr(1)
14367 .m(2)
14368 .n(8)
14369 .k(k)
14370 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080014371 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014372 }
14373 }
14374
Frank Barcharde22685a2021-11-12 11:36:58 -080014375 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, small_kernel_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014376 TEST_REQUIRES_ARM_NEON;
14377 for (size_t k = 1; k <= 40; k += 9) {
14378 for (uint32_t m = 1; m <= 2; m++) {
14379 for (uint32_t n = 1; n <= 8; n++) {
14380 GemmMicrokernelTester()
14381 .mr(2)
14382 .nr(8)
14383 .kr(4)
14384 .sr(1)
14385 .m(m)
14386 .n(n)
14387 .k(k)
14388 .ks(3)
14389 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014390 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014391 }
14392 }
14393 }
14394 }
14395
Frank Barcharde22685a2021-11-12 11:36:58 -080014396 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, n_gt_8_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070014397 TEST_REQUIRES_ARM_NEON;
14398 for (uint32_t n = 9; n < 16; n++) {
14399 for (size_t k = 1; k <= 40; k += 9) {
14400 GemmMicrokernelTester()
14401 .mr(2)
14402 .nr(8)
14403 .kr(4)
14404 .sr(1)
14405 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014406 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014407 .k(k)
14408 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080014409 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014410 }
14411 }
14412 }
14413
Frank Barcharde22685a2021-11-12 11:36:58 -080014414 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, n_div_8_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070014415 TEST_REQUIRES_ARM_NEON;
14416 for (uint32_t n = 16; n <= 24; n += 8) {
14417 for (size_t k = 1; k <= 40; k += 9) {
14418 GemmMicrokernelTester()
14419 .mr(2)
14420 .nr(8)
14421 .kr(4)
14422 .sr(1)
14423 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014424 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014425 .k(k)
14426 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080014427 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014428 }
14429 }
14430 }
14431
Frank Barcharde22685a2021-11-12 11:36:58 -080014432 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, strided_cm_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014433 TEST_REQUIRES_ARM_NEON;
14434 for (size_t k = 1; k <= 40; k += 9) {
14435 for (uint32_t m = 1; m <= 2; m++) {
14436 for (uint32_t n = 1; n <= 8; n++) {
14437 GemmMicrokernelTester()
14438 .mr(2)
14439 .nr(8)
14440 .kr(4)
14441 .sr(1)
14442 .m(m)
14443 .n(n)
14444 .k(k)
14445 .cm_stride(11)
14446 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014447 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014448 }
14449 }
14450 }
14451 }
14452
Frank Barcharde22685a2021-11-12 11:36:58 -080014453 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, a_offset) {
Frank Barchard287952a2021-11-03 15:26:45 -070014454 TEST_REQUIRES_ARM_NEON;
14455 for (size_t k = 1; k <= 40; k += 9) {
14456 GemmMicrokernelTester()
14457 .mr(2)
14458 .nr(8)
14459 .kr(4)
14460 .sr(1)
14461 .m(2)
14462 .n(8)
14463 .k(k)
14464 .ks(3)
14465 .a_offset(83)
Marat Dukhan50323b82022-01-11 00:12:01 -080014466 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014467 }
14468 }
14469
Frank Barcharde22685a2021-11-12 11:36:58 -080014470 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, zero) {
Frank Barchard287952a2021-11-03 15:26:45 -070014471 TEST_REQUIRES_ARM_NEON;
14472 for (uint32_t mz = 0; mz < 2; mz++) {
14473 for (size_t k = 1; k <= 40; k += 9) {
14474 GemmMicrokernelTester()
14475 .mr(2)
14476 .nr(8)
14477 .kr(4)
14478 .sr(1)
14479 .m(2)
14480 .n(8)
14481 .k(k)
14482 .ks(3)
14483 .a_offset(83)
14484 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080014485 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014486 }
14487 }
14488 }
14489
Frank Barcharde22685a2021-11-12 11:36:58 -080014490 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, qmin) {
Frank Barchard287952a2021-11-03 15:26:45 -070014491 TEST_REQUIRES_ARM_NEON;
14492 GemmMicrokernelTester()
14493 .mr(2)
14494 .nr(8)
14495 .kr(4)
14496 .sr(1)
14497 .m(2)
14498 .n(8)
14499 .k(8)
14500 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080014501 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014502 }
14503
Frank Barcharde22685a2021-11-12 11:36:58 -080014504 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, qmax) {
Frank Barchard287952a2021-11-03 15:26:45 -070014505 TEST_REQUIRES_ARM_NEON;
14506 GemmMicrokernelTester()
14507 .mr(2)
14508 .nr(8)
14509 .kr(4)
14510 .sr(1)
14511 .m(2)
14512 .n(8)
14513 .k(8)
14514 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080014515 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014516 }
14517
Frank Barcharde22685a2021-11-12 11:36:58 -080014518 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_DUP, strided_cm) {
Frank Barchard287952a2021-11-03 15:26:45 -070014519 TEST_REQUIRES_ARM_NEON;
14520 GemmMicrokernelTester()
14521 .mr(2)
14522 .nr(8)
14523 .kr(4)
14524 .sr(1)
14525 .m(2)
14526 .n(8)
14527 .k(8)
14528 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080014529 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014530 }
14531#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
14532
14533
14534#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -080014535 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_eq_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014536 TEST_REQUIRES_ARM_NEON;
14537 GemmMicrokernelTester()
14538 .mr(3)
14539 .nr(8)
14540 .kr(4)
14541 .sr(1)
14542 .m(3)
14543 .n(8)
14544 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080014545 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014546 }
14547
Frank Barcharde22685a2021-11-12 11:36:58 -080014548 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070014549 TEST_REQUIRES_ARM_NEON;
14550 GemmMicrokernelTester()
14551 .mr(3)
14552 .nr(8)
14553 .kr(4)
14554 .sr(1)
14555 .m(3)
14556 .n(8)
14557 .k(8)
14558 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080014559 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014560 }
14561
Frank Barcharde22685a2021-11-12 11:36:58 -080014562 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_eq_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014563 TEST_REQUIRES_ARM_NEON;
14564 for (uint32_t m = 1; m <= 3; m++) {
14565 for (uint32_t n = 1; n <= 8; n++) {
14566 GemmMicrokernelTester()
14567 .mr(3)
14568 .nr(8)
14569 .kr(4)
14570 .sr(1)
14571 .m(m)
14572 .n(n)
14573 .k(8)
14574 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014575 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014576 }
14577 }
14578 }
14579
Frank Barcharde22685a2021-11-12 11:36:58 -080014580 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_eq_8_subtile_m) {
Frank Barchard287952a2021-11-03 15:26:45 -070014581 TEST_REQUIRES_ARM_NEON;
14582 for (uint32_t m = 1; m <= 3; m++) {
14583 GemmMicrokernelTester()
14584 .mr(3)
14585 .nr(8)
14586 .kr(4)
14587 .sr(1)
14588 .m(m)
14589 .n(8)
14590 .k(8)
14591 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014592 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014593 }
14594 }
14595
Frank Barcharde22685a2021-11-12 11:36:58 -080014596 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_eq_8_subtile_n) {
Frank Barchard287952a2021-11-03 15:26:45 -070014597 TEST_REQUIRES_ARM_NEON;
14598 for (uint32_t n = 1; n <= 8; n++) {
14599 GemmMicrokernelTester()
14600 .mr(3)
14601 .nr(8)
14602 .kr(4)
14603 .sr(1)
14604 .m(3)
14605 .n(n)
14606 .k(8)
14607 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014608 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014609 }
14610 }
14611
Frank Barcharde22685a2021-11-12 11:36:58 -080014612 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_lt_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014613 TEST_REQUIRES_ARM_NEON;
14614 for (size_t k = 1; k < 8; k++) {
14615 GemmMicrokernelTester()
14616 .mr(3)
14617 .nr(8)
14618 .kr(4)
14619 .sr(1)
14620 .m(3)
14621 .n(8)
14622 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014623 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014624 }
14625 }
14626
Frank Barcharde22685a2021-11-12 11:36:58 -080014627 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_lt_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014628 TEST_REQUIRES_ARM_NEON;
14629 for (size_t k = 1; k < 8; k++) {
14630 for (uint32_t m = 1; m <= 3; m++) {
14631 for (uint32_t n = 1; n <= 8; n++) {
14632 GemmMicrokernelTester()
14633 .mr(3)
14634 .nr(8)
14635 .kr(4)
14636 .sr(1)
14637 .m(m)
14638 .n(n)
14639 .k(k)
14640 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014641 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014642 }
14643 }
14644 }
14645 }
14646
Frank Barcharde22685a2021-11-12 11:36:58 -080014647 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_gt_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014648 TEST_REQUIRES_ARM_NEON;
14649 for (size_t k = 9; k < 16; k++) {
14650 GemmMicrokernelTester()
14651 .mr(3)
14652 .nr(8)
14653 .kr(4)
14654 .sr(1)
14655 .m(3)
14656 .n(8)
14657 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014658 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014659 }
14660 }
14661
Frank Barcharde22685a2021-11-12 11:36:58 -080014662 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_gt_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014663 TEST_REQUIRES_ARM_NEON;
14664 for (size_t k = 9; k < 16; k++) {
14665 for (uint32_t m = 1; m <= 3; m++) {
14666 for (uint32_t n = 1; n <= 8; n++) {
14667 GemmMicrokernelTester()
14668 .mr(3)
14669 .nr(8)
14670 .kr(4)
14671 .sr(1)
14672 .m(m)
14673 .n(n)
14674 .k(k)
14675 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014676 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014677 }
14678 }
14679 }
14680 }
14681
Frank Barcharde22685a2021-11-12 11:36:58 -080014682 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_div_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014683 TEST_REQUIRES_ARM_NEON;
14684 for (size_t k = 16; k <= 80; k += 8) {
14685 GemmMicrokernelTester()
14686 .mr(3)
14687 .nr(8)
14688 .kr(4)
14689 .sr(1)
14690 .m(3)
14691 .n(8)
14692 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014693 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014694 }
14695 }
14696
Frank Barcharde22685a2021-11-12 11:36:58 -080014697 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, k_div_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014698 TEST_REQUIRES_ARM_NEON;
14699 for (size_t k = 16; k <= 80; k += 8) {
14700 for (uint32_t m = 1; m <= 3; m++) {
14701 for (uint32_t n = 1; n <= 8; n++) {
14702 GemmMicrokernelTester()
14703 .mr(3)
14704 .nr(8)
14705 .kr(4)
14706 .sr(1)
14707 .m(m)
14708 .n(n)
14709 .k(k)
14710 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014711 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014712 }
14713 }
14714 }
14715 }
14716
Frank Barcharde22685a2021-11-12 11:36:58 -080014717 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, n_gt_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014718 TEST_REQUIRES_ARM_NEON;
14719 for (uint32_t n = 9; n < 16; n++) {
14720 for (size_t k = 1; k <= 40; k += 9) {
14721 GemmMicrokernelTester()
14722 .mr(3)
14723 .nr(8)
14724 .kr(4)
14725 .sr(1)
14726 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014727 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014728 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014729 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014730 }
14731 }
14732 }
14733
Frank Barcharde22685a2021-11-12 11:36:58 -080014734 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, n_gt_8_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070014735 TEST_REQUIRES_ARM_NEON;
14736 for (uint32_t n = 9; n < 16; n++) {
14737 for (size_t k = 1; k <= 40; k += 9) {
14738 GemmMicrokernelTester()
14739 .mr(3)
14740 .nr(8)
14741 .kr(4)
14742 .sr(1)
14743 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014744 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014745 .k(k)
14746 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080014747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014748 }
14749 }
14750 }
14751
Frank Barcharde22685a2021-11-12 11:36:58 -080014752 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, n_gt_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014753 TEST_REQUIRES_ARM_NEON;
14754 for (uint32_t n = 9; n < 16; n++) {
14755 for (size_t k = 1; k <= 40; k += 9) {
14756 for (uint32_t m = 1; m <= 3; m++) {
14757 GemmMicrokernelTester()
14758 .mr(3)
14759 .nr(8)
14760 .kr(4)
14761 .sr(1)
14762 .m(m)
14763 .n(n)
14764 .k(k)
14765 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014766 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014767 }
14768 }
14769 }
14770 }
14771
Frank Barcharde22685a2021-11-12 11:36:58 -080014772 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, n_div_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070014773 TEST_REQUIRES_ARM_NEON;
14774 for (uint32_t n = 16; n <= 24; n += 8) {
14775 for (size_t k = 1; k <= 40; k += 9) {
14776 GemmMicrokernelTester()
14777 .mr(3)
14778 .nr(8)
14779 .kr(4)
14780 .sr(1)
14781 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014782 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014783 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080014784 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014785 }
14786 }
14787 }
14788
Frank Barcharde22685a2021-11-12 11:36:58 -080014789 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, n_div_8_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070014790 TEST_REQUIRES_ARM_NEON;
14791 for (uint32_t n = 16; n <= 24; n += 8) {
14792 for (size_t k = 1; k <= 40; k += 9) {
14793 GemmMicrokernelTester()
14794 .mr(3)
14795 .nr(8)
14796 .kr(4)
14797 .sr(1)
14798 .m(3)
14799 .n(n)
14800 .k(k)
14801 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080014802 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014803 }
14804 }
14805 }
14806
Frank Barcharde22685a2021-11-12 11:36:58 -080014807 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, n_div_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014808 TEST_REQUIRES_ARM_NEON;
14809 for (uint32_t n = 16; n <= 24; n += 8) {
14810 for (size_t k = 1; k <= 40; k += 9) {
14811 for (uint32_t m = 1; m <= 3; m++) {
14812 GemmMicrokernelTester()
14813 .mr(3)
14814 .nr(8)
14815 .kr(4)
14816 .sr(1)
14817 .m(m)
14818 .n(n)
14819 .k(k)
14820 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014821 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014822 }
14823 }
14824 }
14825 }
14826
Frank Barcharde22685a2021-11-12 11:36:58 -080014827 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070014828 TEST_REQUIRES_ARM_NEON;
14829 for (size_t k = 1; k <= 40; k += 9) {
14830 GemmMicrokernelTester()
14831 .mr(3)
14832 .nr(8)
14833 .kr(4)
14834 .sr(1)
14835 .m(3)
14836 .n(8)
14837 .k(k)
14838 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080014839 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014840 }
14841 }
14842
Frank Barcharde22685a2021-11-12 11:36:58 -080014843 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, small_kernel_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014844 TEST_REQUIRES_ARM_NEON;
14845 for (size_t k = 1; k <= 40; k += 9) {
14846 for (uint32_t m = 1; m <= 3; m++) {
14847 for (uint32_t n = 1; n <= 8; n++) {
14848 GemmMicrokernelTester()
14849 .mr(3)
14850 .nr(8)
14851 .kr(4)
14852 .sr(1)
14853 .m(m)
14854 .n(n)
14855 .k(k)
14856 .ks(3)
14857 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014858 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014859 }
14860 }
14861 }
14862 }
14863
Frank Barcharde22685a2021-11-12 11:36:58 -080014864 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, n_gt_8_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070014865 TEST_REQUIRES_ARM_NEON;
14866 for (uint32_t n = 9; n < 16; n++) {
14867 for (size_t k = 1; k <= 40; k += 9) {
14868 GemmMicrokernelTester()
14869 .mr(3)
14870 .nr(8)
14871 .kr(4)
14872 .sr(1)
14873 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014874 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014875 .k(k)
14876 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080014877 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014878 }
14879 }
14880 }
14881
Frank Barcharde22685a2021-11-12 11:36:58 -080014882 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, n_div_8_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070014883 TEST_REQUIRES_ARM_NEON;
14884 for (uint32_t n = 16; n <= 24; n += 8) {
14885 for (size_t k = 1; k <= 40; k += 9) {
14886 GemmMicrokernelTester()
14887 .mr(3)
14888 .nr(8)
14889 .kr(4)
14890 .sr(1)
14891 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080014892 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070014893 .k(k)
14894 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080014895 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014896 }
14897 }
14898 }
14899
Frank Barcharde22685a2021-11-12 11:36:58 -080014900 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, strided_cm_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070014901 TEST_REQUIRES_ARM_NEON;
14902 for (size_t k = 1; k <= 40; k += 9) {
14903 for (uint32_t m = 1; m <= 3; m++) {
14904 for (uint32_t n = 1; n <= 8; n++) {
14905 GemmMicrokernelTester()
14906 .mr(3)
14907 .nr(8)
14908 .kr(4)
14909 .sr(1)
14910 .m(m)
14911 .n(n)
14912 .k(k)
14913 .cm_stride(11)
14914 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080014915 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014916 }
14917 }
14918 }
14919 }
14920
Frank Barcharde22685a2021-11-12 11:36:58 -080014921 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, a_offset) {
Frank Barchard287952a2021-11-03 15:26:45 -070014922 TEST_REQUIRES_ARM_NEON;
14923 for (size_t k = 1; k <= 40; k += 9) {
14924 GemmMicrokernelTester()
14925 .mr(3)
14926 .nr(8)
14927 .kr(4)
14928 .sr(1)
14929 .m(3)
14930 .n(8)
14931 .k(k)
14932 .ks(3)
14933 .a_offset(127)
Marat Dukhan50323b82022-01-11 00:12:01 -080014934 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014935 }
14936 }
14937
Frank Barcharde22685a2021-11-12 11:36:58 -080014938 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, zero) {
Frank Barchard287952a2021-11-03 15:26:45 -070014939 TEST_REQUIRES_ARM_NEON;
14940 for (uint32_t mz = 0; mz < 3; mz++) {
14941 for (size_t k = 1; k <= 40; k += 9) {
14942 GemmMicrokernelTester()
14943 .mr(3)
14944 .nr(8)
14945 .kr(4)
14946 .sr(1)
14947 .m(3)
14948 .n(8)
14949 .k(k)
14950 .ks(3)
14951 .a_offset(127)
14952 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080014953 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014954 }
14955 }
14956 }
14957
Frank Barcharde22685a2021-11-12 11:36:58 -080014958 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, qmin) {
Frank Barchard287952a2021-11-03 15:26:45 -070014959 TEST_REQUIRES_ARM_NEON;
14960 GemmMicrokernelTester()
14961 .mr(3)
14962 .nr(8)
14963 .kr(4)
14964 .sr(1)
14965 .m(3)
14966 .n(8)
14967 .k(8)
14968 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080014969 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014970 }
14971
Frank Barcharde22685a2021-11-12 11:36:58 -080014972 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, qmax) {
Frank Barchard287952a2021-11-03 15:26:45 -070014973 TEST_REQUIRES_ARM_NEON;
14974 GemmMicrokernelTester()
14975 .mr(3)
14976 .nr(8)
14977 .kr(4)
14978 .sr(1)
14979 .m(3)
14980 .n(8)
14981 .k(8)
14982 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080014983 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014984 }
14985
Frank Barcharde22685a2021-11-12 11:36:58 -080014986 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_DUP, strided_cm) {
Frank Barchard287952a2021-11-03 15:26:45 -070014987 TEST_REQUIRES_ARM_NEON;
14988 GemmMicrokernelTester()
14989 .mr(3)
14990 .nr(8)
14991 .kr(4)
14992 .sr(1)
14993 .m(3)
14994 .n(8)
14995 .k(8)
14996 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080014997 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070014998 }
14999#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
15000
15001
15002#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015003 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_eq_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070015004 TEST_REQUIRES_ARM_NEON;
15005 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015006 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015007 .nr(16)
15008 .kr(4)
15009 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015010 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015011 .n(16)
15012 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015013 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015014 }
15015
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015016 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070015017 TEST_REQUIRES_ARM_NEON;
15018 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015019 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015020 .nr(16)
15021 .kr(4)
15022 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015023 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015024 .n(16)
15025 .k(8)
15026 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015027 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015028 }
15029
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015030 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_eq_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070015031 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015032 for (uint32_t m = 1; m <= 3; m++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015033 for (uint32_t n = 1; n <= 16; n++) {
15034 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015035 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015036 .nr(16)
15037 .kr(4)
15038 .sr(1)
15039 .m(m)
15040 .n(n)
15041 .k(8)
15042 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015043 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015044 }
15045 }
15046 }
15047
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015048 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_eq_8_subtile_m) {
Frank Barchard287952a2021-11-03 15:26:45 -070015049 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015050 for (uint32_t m = 1; m <= 3; m++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015051 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015052 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015053 .nr(16)
15054 .kr(4)
15055 .sr(1)
15056 .m(m)
15057 .n(16)
15058 .k(8)
15059 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015060 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015061 }
15062 }
15063
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015064 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_eq_8_subtile_n) {
Frank Barchard287952a2021-11-03 15:26:45 -070015065 TEST_REQUIRES_ARM_NEON;
15066 for (uint32_t n = 1; n <= 16; n++) {
15067 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015068 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015069 .nr(16)
15070 .kr(4)
15071 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015072 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015073 .n(n)
15074 .k(8)
15075 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015076 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015077 }
15078 }
15079
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015080 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_lt_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070015081 TEST_REQUIRES_ARM_NEON;
15082 for (size_t k = 1; k < 8; k++) {
15083 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015084 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015085 .nr(16)
15086 .kr(4)
15087 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015088 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015089 .n(16)
15090 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015091 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015092 }
15093 }
15094
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015095 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_lt_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070015096 TEST_REQUIRES_ARM_NEON;
15097 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015098 for (uint32_t m = 1; m <= 3; m++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015099 for (uint32_t n = 1; n <= 16; n++) {
15100 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015101 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015102 .nr(16)
15103 .kr(4)
15104 .sr(1)
15105 .m(m)
15106 .n(n)
15107 .k(k)
15108 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015109 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015110 }
15111 }
15112 }
15113 }
15114
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015115 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_gt_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070015116 TEST_REQUIRES_ARM_NEON;
15117 for (size_t k = 9; k < 16; k++) {
15118 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015119 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015120 .nr(16)
15121 .kr(4)
15122 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015123 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015124 .n(16)
15125 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015126 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015127 }
15128 }
15129
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015130 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_gt_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070015131 TEST_REQUIRES_ARM_NEON;
15132 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015133 for (uint32_t m = 1; m <= 3; m++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015134 for (uint32_t n = 1; n <= 16; n++) {
15135 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015136 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015137 .nr(16)
15138 .kr(4)
15139 .sr(1)
15140 .m(m)
15141 .n(n)
15142 .k(k)
15143 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015144 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015145 }
15146 }
15147 }
15148 }
15149
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015150 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_div_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070015151 TEST_REQUIRES_ARM_NEON;
15152 for (size_t k = 16; k <= 80; k += 8) {
15153 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015154 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015155 .nr(16)
15156 .kr(4)
15157 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015158 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015159 .n(16)
15160 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015161 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015162 }
15163 }
15164
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015165 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, k_div_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070015166 TEST_REQUIRES_ARM_NEON;
15167 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015168 for (uint32_t m = 1; m <= 3; m++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015169 for (uint32_t n = 1; n <= 16; n++) {
15170 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015171 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015172 .nr(16)
15173 .kr(4)
15174 .sr(1)
15175 .m(m)
15176 .n(n)
15177 .k(k)
15178 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015179 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015180 }
15181 }
15182 }
15183 }
15184
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015185 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, n_gt_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070015186 TEST_REQUIRES_ARM_NEON;
15187 for (uint32_t n = 17; n < 32; n++) {
15188 for (size_t k = 1; k <= 40; k += 9) {
15189 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015190 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015191 .nr(16)
15192 .kr(4)
15193 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015194 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015195 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070015196 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015197 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015198 }
15199 }
15200 }
15201
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015202 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, n_gt_16_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070015203 TEST_REQUIRES_ARM_NEON;
15204 for (uint32_t n = 17; n < 32; n++) {
15205 for (size_t k = 1; k <= 40; k += 9) {
15206 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015207 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015208 .nr(16)
15209 .kr(4)
15210 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015211 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015212 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070015213 .k(k)
15214 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015216 }
15217 }
15218 }
15219
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015220 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, n_gt_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070015221 TEST_REQUIRES_ARM_NEON;
15222 for (uint32_t n = 17; n < 32; n++) {
15223 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015224 for (uint32_t m = 1; m <= 3; m++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015225 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015226 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015227 .nr(16)
15228 .kr(4)
15229 .sr(1)
15230 .m(m)
15231 .n(n)
15232 .k(k)
15233 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015234 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015235 }
15236 }
15237 }
15238 }
15239
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015240 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, n_div_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070015241 TEST_REQUIRES_ARM_NEON;
15242 for (uint32_t n = 32; n <= 48; n += 16) {
15243 for (size_t k = 1; k <= 40; k += 9) {
15244 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015245 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015246 .nr(16)
15247 .kr(4)
15248 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015249 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015250 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070015251 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015252 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015253 }
15254 }
15255 }
15256
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015257 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, n_div_16_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070015258 TEST_REQUIRES_ARM_NEON;
15259 for (uint32_t n = 32; n <= 48; n += 16) {
15260 for (size_t k = 1; k <= 40; k += 9) {
15261 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015262 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015263 .nr(16)
15264 .kr(4)
15265 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015266 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015267 .n(n)
15268 .k(k)
15269 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015270 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015271 }
15272 }
15273 }
15274
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015275 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, n_div_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070015276 TEST_REQUIRES_ARM_NEON;
15277 for (uint32_t n = 32; n <= 48; n += 16) {
15278 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015279 for (uint32_t m = 1; m <= 3; m++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015280 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015281 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015282 .nr(16)
15283 .kr(4)
15284 .sr(1)
15285 .m(m)
15286 .n(n)
15287 .k(k)
15288 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015289 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015290 }
15291 }
15292 }
15293 }
15294
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015295 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070015296 TEST_REQUIRES_ARM_NEON;
15297 for (size_t k = 1; k <= 40; k += 9) {
15298 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015299 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015300 .nr(16)
15301 .kr(4)
15302 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015303 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015304 .n(16)
15305 .k(k)
15306 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015307 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015308 }
15309 }
15310
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015311 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, small_kernel_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070015312 TEST_REQUIRES_ARM_NEON;
15313 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015314 for (uint32_t m = 1; m <= 3; m++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015315 for (uint32_t n = 1; n <= 16; n++) {
15316 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015317 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015318 .nr(16)
15319 .kr(4)
15320 .sr(1)
15321 .m(m)
15322 .n(n)
15323 .k(k)
15324 .ks(3)
15325 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015326 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015327 }
15328 }
15329 }
15330 }
15331
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015332 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, n_gt_16_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070015333 TEST_REQUIRES_ARM_NEON;
15334 for (uint32_t n = 17; n < 32; n++) {
15335 for (size_t k = 1; k <= 40; k += 9) {
15336 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015337 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015338 .nr(16)
15339 .kr(4)
15340 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015341 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015342 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070015343 .k(k)
15344 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015345 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015346 }
15347 }
15348 }
15349
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015350 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, n_div_16_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070015351 TEST_REQUIRES_ARM_NEON;
15352 for (uint32_t n = 32; n <= 48; n += 16) {
15353 for (size_t k = 1; k <= 40; k += 9) {
15354 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015355 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015356 .nr(16)
15357 .kr(4)
15358 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015359 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015360 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070015361 .k(k)
15362 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015363 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015364 }
15365 }
15366 }
15367
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015368 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, strided_cm_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070015369 TEST_REQUIRES_ARM_NEON;
15370 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015371 for (uint32_t m = 1; m <= 3; m++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015372 for (uint32_t n = 1; n <= 16; n++) {
15373 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015374 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015375 .nr(16)
15376 .kr(4)
15377 .sr(1)
15378 .m(m)
15379 .n(n)
15380 .k(k)
15381 .cm_stride(19)
15382 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015383 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015384 }
15385 }
15386 }
15387 }
15388
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015389 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, a_offset) {
Frank Barchard287952a2021-11-03 15:26:45 -070015390 TEST_REQUIRES_ARM_NEON;
15391 for (size_t k = 1; k <= 40; k += 9) {
15392 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015393 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015394 .nr(16)
15395 .kr(4)
15396 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015397 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015398 .n(16)
15399 .k(k)
15400 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015401 .a_offset(127)
15402 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015403 }
15404 }
15405
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015406 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, zero) {
Frank Barchard287952a2021-11-03 15:26:45 -070015407 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015408 for (uint32_t mz = 0; mz < 3; mz++) {
Frank Barchard287952a2021-11-03 15:26:45 -070015409 for (size_t k = 1; k <= 40; k += 9) {
15410 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015411 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015412 .nr(16)
15413 .kr(4)
15414 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015415 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015416 .n(16)
15417 .k(k)
15418 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015419 .a_offset(127)
Frank Barchard287952a2021-11-03 15:26:45 -070015420 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015421 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015422 }
15423 }
15424 }
15425
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015426 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, qmin) {
Frank Barchard287952a2021-11-03 15:26:45 -070015427 TEST_REQUIRES_ARM_NEON;
15428 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015429 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015430 .nr(16)
15431 .kr(4)
15432 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015433 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015434 .n(16)
15435 .k(8)
15436 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015437 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015438 }
15439
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015440 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, qmax) {
Frank Barchard287952a2021-11-03 15:26:45 -070015441 TEST_REQUIRES_ARM_NEON;
15442 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015443 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015444 .nr(16)
15445 .kr(4)
15446 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015447 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015448 .n(16)
15449 .k(8)
15450 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015451 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015452 }
15453
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015454 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_DUP, strided_cm) {
Frank Barchard287952a2021-11-03 15:26:45 -070015455 TEST_REQUIRES_ARM_NEON;
15456 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015457 .mr(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015458 .nr(16)
15459 .kr(4)
15460 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015461 .m(3)
Frank Barchard287952a2021-11-03 15:26:45 -070015462 .n(16)
15463 .k(8)
15464 .cm_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015465 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15466 }
15467#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
15468
15469
15470#if XNN_ARCH_ARM || XNN_ARCH_ARM64
15471 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_eq_16) {
15472 TEST_REQUIRES_ARM_NEON;
15473 GemmMicrokernelTester()
15474 .mr(2)
15475 .nr(8)
15476 .kr(4)
15477 .sr(1)
15478 .m(2)
15479 .n(8)
15480 .k(16)
15481 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15482 }
15483
15484 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, strided_cn) {
15485 TEST_REQUIRES_ARM_NEON;
15486 GemmMicrokernelTester()
15487 .mr(2)
15488 .nr(8)
15489 .kr(4)
15490 .sr(1)
15491 .m(2)
15492 .n(8)
15493 .k(16)
15494 .cn_stride(11)
15495 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15496 }
15497
15498 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_eq_16_subtile) {
15499 TEST_REQUIRES_ARM_NEON;
15500 for (uint32_t m = 1; m <= 2; m++) {
15501 for (uint32_t n = 1; n <= 8; n++) {
15502 GemmMicrokernelTester()
15503 .mr(2)
15504 .nr(8)
15505 .kr(4)
15506 .sr(1)
15507 .m(m)
15508 .n(n)
15509 .k(16)
15510 .iterations(1)
15511 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15512 }
15513 }
15514 }
15515
15516 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_eq_16_subtile_m) {
15517 TEST_REQUIRES_ARM_NEON;
15518 for (uint32_t m = 1; m <= 2; m++) {
15519 GemmMicrokernelTester()
15520 .mr(2)
15521 .nr(8)
15522 .kr(4)
15523 .sr(1)
15524 .m(m)
15525 .n(8)
15526 .k(16)
15527 .iterations(1)
15528 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15529 }
15530 }
15531
15532 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_eq_16_subtile_n) {
15533 TEST_REQUIRES_ARM_NEON;
15534 for (uint32_t n = 1; n <= 8; n++) {
15535 GemmMicrokernelTester()
15536 .mr(2)
15537 .nr(8)
15538 .kr(4)
15539 .sr(1)
15540 .m(2)
15541 .n(n)
15542 .k(16)
15543 .iterations(1)
15544 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15545 }
15546 }
15547
15548 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_lt_16) {
15549 TEST_REQUIRES_ARM_NEON;
15550 for (size_t k = 1; k < 16; k++) {
15551 GemmMicrokernelTester()
15552 .mr(2)
15553 .nr(8)
15554 .kr(4)
15555 .sr(1)
15556 .m(2)
15557 .n(8)
15558 .k(k)
15559 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15560 }
15561 }
15562
15563 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_lt_16_subtile) {
15564 TEST_REQUIRES_ARM_NEON;
15565 for (size_t k = 1; k < 16; k++) {
15566 for (uint32_t m = 1; m <= 2; m++) {
15567 for (uint32_t n = 1; n <= 8; n++) {
15568 GemmMicrokernelTester()
15569 .mr(2)
15570 .nr(8)
15571 .kr(4)
15572 .sr(1)
15573 .m(m)
15574 .n(n)
15575 .k(k)
15576 .iterations(1)
15577 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15578 }
15579 }
15580 }
15581 }
15582
15583 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_gt_16) {
15584 TEST_REQUIRES_ARM_NEON;
15585 for (size_t k = 17; k < 32; k++) {
15586 GemmMicrokernelTester()
15587 .mr(2)
15588 .nr(8)
15589 .kr(4)
15590 .sr(1)
15591 .m(2)
15592 .n(8)
15593 .k(k)
15594 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15595 }
15596 }
15597
15598 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_gt_16_subtile) {
15599 TEST_REQUIRES_ARM_NEON;
15600 for (size_t k = 17; k < 32; k++) {
15601 for (uint32_t m = 1; m <= 2; m++) {
15602 for (uint32_t n = 1; n <= 8; n++) {
15603 GemmMicrokernelTester()
15604 .mr(2)
15605 .nr(8)
15606 .kr(4)
15607 .sr(1)
15608 .m(m)
15609 .n(n)
15610 .k(k)
15611 .iterations(1)
15612 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15613 }
15614 }
15615 }
15616 }
15617
15618 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_div_16) {
15619 TEST_REQUIRES_ARM_NEON;
15620 for (size_t k = 32; k <= 160; k += 16) {
15621 GemmMicrokernelTester()
15622 .mr(2)
15623 .nr(8)
15624 .kr(4)
15625 .sr(1)
15626 .m(2)
15627 .n(8)
15628 .k(k)
15629 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15630 }
15631 }
15632
15633 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, k_div_16_subtile) {
15634 TEST_REQUIRES_ARM_NEON;
15635 for (size_t k = 32; k <= 160; k += 16) {
15636 for (uint32_t m = 1; m <= 2; m++) {
15637 for (uint32_t n = 1; n <= 8; n++) {
15638 GemmMicrokernelTester()
15639 .mr(2)
15640 .nr(8)
15641 .kr(4)
15642 .sr(1)
15643 .m(m)
15644 .n(n)
15645 .k(k)
15646 .iterations(1)
15647 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15648 }
15649 }
15650 }
15651 }
15652
15653 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, n_gt_8) {
15654 TEST_REQUIRES_ARM_NEON;
15655 for (uint32_t n = 9; n < 16; n++) {
15656 for (size_t k = 1; k <= 80; k += 17) {
15657 GemmMicrokernelTester()
15658 .mr(2)
15659 .nr(8)
15660 .kr(4)
15661 .sr(1)
15662 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015663 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015664 .k(k)
15665 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15666 }
15667 }
15668 }
15669
15670 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, n_gt_8_strided_cn) {
15671 TEST_REQUIRES_ARM_NEON;
15672 for (uint32_t n = 9; n < 16; n++) {
15673 for (size_t k = 1; k <= 80; k += 17) {
15674 GemmMicrokernelTester()
15675 .mr(2)
15676 .nr(8)
15677 .kr(4)
15678 .sr(1)
15679 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015680 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015681 .k(k)
15682 .cn_stride(11)
15683 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15684 }
15685 }
15686 }
15687
15688 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, n_gt_8_subtile) {
15689 TEST_REQUIRES_ARM_NEON;
15690 for (uint32_t n = 9; n < 16; n++) {
15691 for (size_t k = 1; k <= 80; k += 17) {
15692 for (uint32_t m = 1; m <= 2; m++) {
15693 GemmMicrokernelTester()
15694 .mr(2)
15695 .nr(8)
15696 .kr(4)
15697 .sr(1)
15698 .m(m)
15699 .n(n)
15700 .k(k)
15701 .iterations(1)
15702 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15703 }
15704 }
15705 }
15706 }
15707
15708 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, n_div_8) {
15709 TEST_REQUIRES_ARM_NEON;
15710 for (uint32_t n = 16; n <= 24; n += 8) {
15711 for (size_t k = 1; k <= 80; k += 17) {
15712 GemmMicrokernelTester()
15713 .mr(2)
15714 .nr(8)
15715 .kr(4)
15716 .sr(1)
15717 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015718 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015719 .k(k)
15720 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15721 }
15722 }
15723 }
15724
15725 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, n_div_8_strided_cn) {
15726 TEST_REQUIRES_ARM_NEON;
15727 for (uint32_t n = 16; n <= 24; n += 8) {
15728 for (size_t k = 1; k <= 80; k += 17) {
15729 GemmMicrokernelTester()
15730 .mr(2)
15731 .nr(8)
15732 .kr(4)
15733 .sr(1)
15734 .m(2)
15735 .n(n)
15736 .k(k)
15737 .cn_stride(11)
15738 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15739 }
15740 }
15741 }
15742
15743 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, n_div_8_subtile) {
15744 TEST_REQUIRES_ARM_NEON;
15745 for (uint32_t n = 16; n <= 24; n += 8) {
15746 for (size_t k = 1; k <= 80; k += 17) {
15747 for (uint32_t m = 1; m <= 2; m++) {
15748 GemmMicrokernelTester()
15749 .mr(2)
15750 .nr(8)
15751 .kr(4)
15752 .sr(1)
15753 .m(m)
15754 .n(n)
15755 .k(k)
15756 .iterations(1)
15757 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15758 }
15759 }
15760 }
15761 }
15762
15763 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, small_kernel) {
15764 TEST_REQUIRES_ARM_NEON;
15765 for (size_t k = 1; k <= 80; k += 17) {
15766 GemmMicrokernelTester()
15767 .mr(2)
15768 .nr(8)
15769 .kr(4)
15770 .sr(1)
15771 .m(2)
15772 .n(8)
15773 .k(k)
15774 .ks(3)
15775 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15776 }
15777 }
15778
15779 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, small_kernel_subtile) {
15780 TEST_REQUIRES_ARM_NEON;
15781 for (size_t k = 1; k <= 80; k += 17) {
15782 for (uint32_t m = 1; m <= 2; m++) {
15783 for (uint32_t n = 1; n <= 8; n++) {
15784 GemmMicrokernelTester()
15785 .mr(2)
15786 .nr(8)
15787 .kr(4)
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_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15795 }
15796 }
15797 }
15798 }
15799
15800 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, n_gt_8_small_kernel) {
15801 TEST_REQUIRES_ARM_NEON;
15802 for (uint32_t n = 9; n < 16; n++) {
15803 for (size_t k = 1; k <= 80; k += 17) {
15804 GemmMicrokernelTester()
15805 .mr(2)
15806 .nr(8)
15807 .kr(4)
15808 .sr(1)
15809 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015810 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015811 .k(k)
15812 .ks(3)
15813 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15814 }
15815 }
15816 }
15817
15818 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, n_div_8_small_kernel) {
15819 TEST_REQUIRES_ARM_NEON;
15820 for (uint32_t n = 16; n <= 24; n += 8) {
15821 for (size_t k = 1; k <= 80; k += 17) {
15822 GemmMicrokernelTester()
15823 .mr(2)
15824 .nr(8)
15825 .kr(4)
15826 .sr(1)
15827 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080015828 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080015829 .k(k)
15830 .ks(3)
15831 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15832 }
15833 }
15834 }
15835
15836 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, strided_cm_subtile) {
15837 TEST_REQUIRES_ARM_NEON;
15838 for (size_t k = 1; k <= 80; k += 17) {
15839 for (uint32_t m = 1; m <= 2; m++) {
15840 for (uint32_t n = 1; n <= 8; n++) {
15841 GemmMicrokernelTester()
15842 .mr(2)
15843 .nr(8)
15844 .kr(4)
15845 .sr(1)
15846 .m(m)
15847 .n(n)
15848 .k(k)
15849 .cm_stride(11)
15850 .iterations(1)
15851 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15852 }
15853 }
15854 }
15855 }
15856
15857 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, a_offset) {
15858 TEST_REQUIRES_ARM_NEON;
15859 for (size_t k = 1; k <= 80; k += 17) {
15860 GemmMicrokernelTester()
15861 .mr(2)
15862 .nr(8)
15863 .kr(4)
15864 .sr(1)
15865 .m(2)
15866 .n(8)
15867 .k(k)
15868 .ks(3)
15869 .a_offset(163)
15870 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15871 }
15872 }
15873
15874 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, zero) {
15875 TEST_REQUIRES_ARM_NEON;
15876 for (uint32_t mz = 0; mz < 2; mz++) {
15877 for (size_t k = 1; k <= 80; k += 17) {
15878 GemmMicrokernelTester()
15879 .mr(2)
15880 .nr(8)
15881 .kr(4)
15882 .sr(1)
15883 .m(2)
15884 .n(8)
15885 .k(k)
15886 .ks(3)
15887 .a_offset(163)
15888 .zero_index(mz)
15889 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15890 }
15891 }
15892 }
15893
15894 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, qmin) {
15895 TEST_REQUIRES_ARM_NEON;
15896 GemmMicrokernelTester()
15897 .mr(2)
15898 .nr(8)
15899 .kr(4)
15900 .sr(1)
15901 .m(2)
15902 .n(8)
15903 .k(16)
15904 .qmin(128)
15905 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15906 }
15907
15908 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, qmax) {
15909 TEST_REQUIRES_ARM_NEON;
15910 GemmMicrokernelTester()
15911 .mr(2)
15912 .nr(8)
15913 .kr(4)
15914 .sr(1)
15915 .m(2)
15916 .n(8)
15917 .k(16)
15918 .qmax(128)
15919 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15920 }
15921
15922 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_DUP, strided_cm) {
15923 TEST_REQUIRES_ARM_NEON;
15924 GemmMicrokernelTester()
15925 .mr(2)
15926 .nr(8)
15927 .kr(4)
15928 .sr(1)
15929 .m(2)
15930 .n(8)
15931 .k(16)
15932 .cm_stride(11)
15933 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015934 }
15935#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
15936
15937
15938#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -080015939 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_eq_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070015940 TEST_REQUIRES_ARM_NEON;
15941 GemmMicrokernelTester()
15942 .mr(4)
15943 .nr(8)
15944 .kr(4)
15945 .sr(1)
15946 .m(4)
15947 .n(8)
15948 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -080015949 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015950 }
15951
Frank Barcharde22685a2021-11-12 11:36:58 -080015952 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070015953 TEST_REQUIRES_ARM_NEON;
15954 GemmMicrokernelTester()
15955 .mr(4)
15956 .nr(8)
15957 .kr(4)
15958 .sr(1)
15959 .m(4)
15960 .n(8)
15961 .k(16)
15962 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080015963 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015964 }
15965
Frank Barcharde22685a2021-11-12 11:36:58 -080015966 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_eq_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070015967 TEST_REQUIRES_ARM_NEON;
15968 for (uint32_t m = 1; m <= 4; m++) {
15969 for (uint32_t n = 1; n <= 8; n++) {
15970 GemmMicrokernelTester()
15971 .mr(4)
15972 .nr(8)
15973 .kr(4)
15974 .sr(1)
15975 .m(m)
15976 .n(n)
15977 .k(16)
15978 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080015979 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015980 }
15981 }
15982 }
15983
Frank Barcharde22685a2021-11-12 11:36:58 -080015984 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_eq_16_subtile_m) {
Frank Barchard287952a2021-11-03 15:26:45 -070015985 TEST_REQUIRES_ARM_NEON;
15986 for (uint32_t m = 1; m <= 4; m++) {
15987 GemmMicrokernelTester()
15988 .mr(4)
15989 .nr(8)
15990 .kr(4)
15991 .sr(1)
15992 .m(m)
15993 .n(8)
15994 .k(16)
15995 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080015996 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070015997 }
15998 }
15999
Frank Barcharde22685a2021-11-12 11:36:58 -080016000 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_eq_16_subtile_n) {
Frank Barchard287952a2021-11-03 15:26:45 -070016001 TEST_REQUIRES_ARM_NEON;
16002 for (uint32_t n = 1; n <= 8; n++) {
16003 GemmMicrokernelTester()
16004 .mr(4)
16005 .nr(8)
16006 .kr(4)
16007 .sr(1)
16008 .m(4)
16009 .n(n)
16010 .k(16)
16011 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016012 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016013 }
16014 }
16015
Frank Barcharde22685a2021-11-12 11:36:58 -080016016 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_lt_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070016017 TEST_REQUIRES_ARM_NEON;
16018 for (size_t k = 1; k < 16; k++) {
16019 GemmMicrokernelTester()
16020 .mr(4)
16021 .nr(8)
16022 .kr(4)
16023 .sr(1)
16024 .m(4)
16025 .n(8)
16026 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080016027 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016028 }
16029 }
16030
Frank Barcharde22685a2021-11-12 11:36:58 -080016031 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_lt_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070016032 TEST_REQUIRES_ARM_NEON;
16033 for (size_t k = 1; k < 16; k++) {
16034 for (uint32_t m = 1; m <= 4; m++) {
16035 for (uint32_t n = 1; n <= 8; n++) {
16036 GemmMicrokernelTester()
16037 .mr(4)
16038 .nr(8)
16039 .kr(4)
16040 .sr(1)
16041 .m(m)
16042 .n(n)
16043 .k(k)
16044 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016045 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016046 }
16047 }
16048 }
16049 }
16050
Frank Barcharde22685a2021-11-12 11:36:58 -080016051 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_gt_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070016052 TEST_REQUIRES_ARM_NEON;
16053 for (size_t k = 17; k < 32; k++) {
16054 GemmMicrokernelTester()
16055 .mr(4)
16056 .nr(8)
16057 .kr(4)
16058 .sr(1)
16059 .m(4)
16060 .n(8)
16061 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080016062 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016063 }
16064 }
16065
Frank Barcharde22685a2021-11-12 11:36:58 -080016066 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_gt_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070016067 TEST_REQUIRES_ARM_NEON;
16068 for (size_t k = 17; k < 32; k++) {
16069 for (uint32_t m = 1; m <= 4; m++) {
16070 for (uint32_t n = 1; n <= 8; n++) {
16071 GemmMicrokernelTester()
16072 .mr(4)
16073 .nr(8)
16074 .kr(4)
16075 .sr(1)
16076 .m(m)
16077 .n(n)
16078 .k(k)
16079 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016080 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016081 }
16082 }
16083 }
16084 }
16085
Frank Barcharde22685a2021-11-12 11:36:58 -080016086 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_div_16) {
Frank Barchard287952a2021-11-03 15:26:45 -070016087 TEST_REQUIRES_ARM_NEON;
16088 for (size_t k = 32; k <= 160; k += 16) {
16089 GemmMicrokernelTester()
16090 .mr(4)
16091 .nr(8)
16092 .kr(4)
16093 .sr(1)
16094 .m(4)
16095 .n(8)
16096 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080016097 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016098 }
16099 }
16100
Frank Barcharde22685a2021-11-12 11:36:58 -080016101 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, k_div_16_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070016102 TEST_REQUIRES_ARM_NEON;
16103 for (size_t k = 32; k <= 160; k += 16) {
16104 for (uint32_t m = 1; m <= 4; m++) {
16105 for (uint32_t n = 1; n <= 8; n++) {
16106 GemmMicrokernelTester()
16107 .mr(4)
16108 .nr(8)
16109 .kr(4)
16110 .sr(1)
16111 .m(m)
16112 .n(n)
16113 .k(k)
16114 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016115 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016116 }
16117 }
16118 }
16119 }
16120
Frank Barcharde22685a2021-11-12 11:36:58 -080016121 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, n_gt_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070016122 TEST_REQUIRES_ARM_NEON;
16123 for (uint32_t n = 9; n < 16; n++) {
16124 for (size_t k = 1; k <= 80; k += 17) {
16125 GemmMicrokernelTester()
16126 .mr(4)
16127 .nr(8)
16128 .kr(4)
16129 .sr(1)
16130 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016131 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070016132 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080016133 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016134 }
16135 }
16136 }
16137
Frank Barcharde22685a2021-11-12 11:36:58 -080016138 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, n_gt_8_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070016139 TEST_REQUIRES_ARM_NEON;
16140 for (uint32_t n = 9; n < 16; n++) {
16141 for (size_t k = 1; k <= 80; k += 17) {
16142 GemmMicrokernelTester()
16143 .mr(4)
16144 .nr(8)
16145 .kr(4)
16146 .sr(1)
16147 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016148 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070016149 .k(k)
16150 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080016151 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016152 }
16153 }
16154 }
16155
Frank Barcharde22685a2021-11-12 11:36:58 -080016156 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, n_gt_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070016157 TEST_REQUIRES_ARM_NEON;
16158 for (uint32_t n = 9; n < 16; n++) {
16159 for (size_t k = 1; k <= 80; k += 17) {
16160 for (uint32_t m = 1; m <= 4; m++) {
16161 GemmMicrokernelTester()
16162 .mr(4)
16163 .nr(8)
16164 .kr(4)
16165 .sr(1)
16166 .m(m)
16167 .n(n)
16168 .k(k)
16169 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016170 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016171 }
16172 }
16173 }
16174 }
16175
Frank Barcharde22685a2021-11-12 11:36:58 -080016176 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, n_div_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070016177 TEST_REQUIRES_ARM_NEON;
16178 for (uint32_t n = 16; n <= 24; n += 8) {
16179 for (size_t k = 1; k <= 80; k += 17) {
16180 GemmMicrokernelTester()
16181 .mr(4)
16182 .nr(8)
16183 .kr(4)
16184 .sr(1)
16185 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016186 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070016187 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080016188 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016189 }
16190 }
16191 }
16192
Frank Barcharde22685a2021-11-12 11:36:58 -080016193 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, n_div_8_strided_cn) {
Frank Barchard287952a2021-11-03 15:26:45 -070016194 TEST_REQUIRES_ARM_NEON;
16195 for (uint32_t n = 16; n <= 24; n += 8) {
16196 for (size_t k = 1; k <= 80; k += 17) {
16197 GemmMicrokernelTester()
16198 .mr(4)
16199 .nr(8)
16200 .kr(4)
16201 .sr(1)
16202 .m(4)
16203 .n(n)
16204 .k(k)
16205 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080016206 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016207 }
16208 }
16209 }
16210
Frank Barcharde22685a2021-11-12 11:36:58 -080016211 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, n_div_8_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070016212 TEST_REQUIRES_ARM_NEON;
16213 for (uint32_t n = 16; n <= 24; n += 8) {
16214 for (size_t k = 1; k <= 80; k += 17) {
16215 for (uint32_t m = 1; m <= 4; m++) {
16216 GemmMicrokernelTester()
16217 .mr(4)
16218 .nr(8)
16219 .kr(4)
16220 .sr(1)
16221 .m(m)
16222 .n(n)
16223 .k(k)
16224 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016225 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016226 }
16227 }
16228 }
16229 }
16230
Frank Barcharde22685a2021-11-12 11:36:58 -080016231 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070016232 TEST_REQUIRES_ARM_NEON;
16233 for (size_t k = 1; k <= 80; k += 17) {
16234 GemmMicrokernelTester()
16235 .mr(4)
16236 .nr(8)
16237 .kr(4)
16238 .sr(1)
16239 .m(4)
16240 .n(8)
16241 .k(k)
16242 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080016243 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016244 }
16245 }
16246
Frank Barcharde22685a2021-11-12 11:36:58 -080016247 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, small_kernel_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070016248 TEST_REQUIRES_ARM_NEON;
16249 for (size_t k = 1; k <= 80; k += 17) {
16250 for (uint32_t m = 1; m <= 4; m++) {
16251 for (uint32_t n = 1; n <= 8; n++) {
16252 GemmMicrokernelTester()
16253 .mr(4)
16254 .nr(8)
16255 .kr(4)
16256 .sr(1)
16257 .m(m)
16258 .n(n)
16259 .k(k)
16260 .ks(3)
16261 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016262 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016263 }
16264 }
16265 }
16266 }
16267
Frank Barcharde22685a2021-11-12 11:36:58 -080016268 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, n_gt_8_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070016269 TEST_REQUIRES_ARM_NEON;
16270 for (uint32_t n = 9; n < 16; n++) {
16271 for (size_t k = 1; k <= 80; k += 17) {
16272 GemmMicrokernelTester()
16273 .mr(4)
16274 .nr(8)
16275 .kr(4)
16276 .sr(1)
16277 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016278 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070016279 .k(k)
16280 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080016281 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016282 }
16283 }
16284 }
16285
Frank Barcharde22685a2021-11-12 11:36:58 -080016286 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, n_div_8_small_kernel) {
Frank Barchard287952a2021-11-03 15:26:45 -070016287 TEST_REQUIRES_ARM_NEON;
16288 for (uint32_t n = 16; n <= 24; n += 8) {
16289 for (size_t k = 1; k <= 80; k += 17) {
16290 GemmMicrokernelTester()
16291 .mr(4)
16292 .nr(8)
16293 .kr(4)
16294 .sr(1)
16295 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016296 .n(n)
Frank Barchard287952a2021-11-03 15:26:45 -070016297 .k(k)
16298 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080016299 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016300 }
16301 }
16302 }
16303
Frank Barcharde22685a2021-11-12 11:36:58 -080016304 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, strided_cm_subtile) {
Frank Barchard287952a2021-11-03 15:26:45 -070016305 TEST_REQUIRES_ARM_NEON;
16306 for (size_t k = 1; k <= 80; k += 17) {
16307 for (uint32_t m = 1; m <= 4; m++) {
16308 for (uint32_t n = 1; n <= 8; n++) {
16309 GemmMicrokernelTester()
16310 .mr(4)
16311 .nr(8)
16312 .kr(4)
16313 .sr(1)
16314 .m(m)
16315 .n(n)
16316 .k(k)
16317 .cm_stride(11)
16318 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016319 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016320 }
16321 }
16322 }
16323 }
16324
Frank Barcharde22685a2021-11-12 11:36:58 -080016325 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, a_offset) {
Frank Barchard287952a2021-11-03 15:26:45 -070016326 TEST_REQUIRES_ARM_NEON;
16327 for (size_t k = 1; k <= 80; k += 17) {
16328 GemmMicrokernelTester()
16329 .mr(4)
16330 .nr(8)
16331 .kr(4)
16332 .sr(1)
16333 .m(4)
16334 .n(8)
16335 .k(k)
16336 .ks(3)
16337 .a_offset(331)
Marat Dukhan50323b82022-01-11 00:12:01 -080016338 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016339 }
16340 }
16341
Frank Barcharde22685a2021-11-12 11:36:58 -080016342 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, zero) {
Frank Barchard287952a2021-11-03 15:26:45 -070016343 TEST_REQUIRES_ARM_NEON;
16344 for (uint32_t mz = 0; mz < 4; mz++) {
16345 for (size_t k = 1; k <= 80; k += 17) {
16346 GemmMicrokernelTester()
16347 .mr(4)
16348 .nr(8)
16349 .kr(4)
16350 .sr(1)
16351 .m(4)
16352 .n(8)
16353 .k(k)
16354 .ks(3)
16355 .a_offset(331)
16356 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080016357 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016358 }
16359 }
16360 }
16361
Frank Barcharde22685a2021-11-12 11:36:58 -080016362 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, qmin) {
Frank Barchard287952a2021-11-03 15:26:45 -070016363 TEST_REQUIRES_ARM_NEON;
16364 GemmMicrokernelTester()
16365 .mr(4)
16366 .nr(8)
16367 .kr(4)
16368 .sr(1)
16369 .m(4)
16370 .n(8)
16371 .k(16)
16372 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080016373 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016374 }
16375
Frank Barcharde22685a2021-11-12 11:36:58 -080016376 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, qmax) {
Frank Barchard287952a2021-11-03 15:26:45 -070016377 TEST_REQUIRES_ARM_NEON;
16378 GemmMicrokernelTester()
16379 .mr(4)
16380 .nr(8)
16381 .kr(4)
16382 .sr(1)
16383 .m(4)
16384 .n(8)
16385 .k(16)
16386 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080016387 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016388 }
16389
Frank Barcharde22685a2021-11-12 11:36:58 -080016390 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MLAL_DUP, strided_cm) {
Frank Barchard287952a2021-11-03 15:26:45 -070016391 TEST_REQUIRES_ARM_NEON;
16392 GemmMicrokernelTester()
16393 .mr(4)
16394 .nr(8)
16395 .kr(4)
16396 .sr(1)
16397 .m(4)
16398 .n(8)
16399 .k(16)
16400 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080016401 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard287952a2021-11-03 15:26:45 -070016402 }
16403#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
16404
16405
16406#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016407 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_eq_8) {
Frank Barchard287952a2021-11-03 15:26:45 -070016408 TEST_REQUIRES_ARM_NEON;
16409 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016410 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016411 .nr(8)
16412 .kr(4)
16413 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016414 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016415 .n(8)
16416 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016417 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016418 }
16419
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016420 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016421 TEST_REQUIRES_ARM_NEON;
16422 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016423 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016424 .nr(8)
16425 .kr(4)
16426 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016427 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016428 .n(8)
16429 .k(8)
16430 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016431 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016432 }
16433
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016434 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_eq_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016435 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016436 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016437 for (uint32_t n = 1; n <= 8; n++) {
16438 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016439 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016440 .nr(8)
16441 .kr(4)
16442 .sr(1)
16443 .m(m)
16444 .n(n)
16445 .k(8)
16446 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016447 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016448 }
16449 }
16450 }
16451
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016452 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_eq_8_subtile_m) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016453 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016454 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016455 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016456 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016457 .nr(8)
16458 .kr(4)
16459 .sr(1)
16460 .m(m)
16461 .n(8)
16462 .k(8)
16463 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016464 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016465 }
16466 }
16467
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016468 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_eq_8_subtile_n) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016469 TEST_REQUIRES_ARM_NEON;
16470 for (uint32_t n = 1; n <= 8; n++) {
16471 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016472 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016473 .nr(8)
16474 .kr(4)
16475 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016476 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016477 .n(n)
16478 .k(8)
16479 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016480 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016481 }
16482 }
16483
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016484 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_lt_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016485 TEST_REQUIRES_ARM_NEON;
16486 for (size_t k = 1; k < 8; k++) {
16487 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016488 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016489 .nr(8)
16490 .kr(4)
16491 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016492 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016493 .n(8)
16494 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016495 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016496 }
16497 }
16498
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016499 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_lt_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016500 TEST_REQUIRES_ARM_NEON;
16501 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016502 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016503 for (uint32_t n = 1; n <= 8; n++) {
16504 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016505 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016506 .nr(8)
16507 .kr(4)
16508 .sr(1)
16509 .m(m)
16510 .n(n)
16511 .k(k)
16512 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016513 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016514 }
16515 }
16516 }
16517 }
16518
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016519 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_gt_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016520 TEST_REQUIRES_ARM_NEON;
16521 for (size_t k = 9; k < 16; k++) {
16522 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016523 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016524 .nr(8)
16525 .kr(4)
16526 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016527 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016528 .n(8)
16529 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016530 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016531 }
16532 }
16533
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016534 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_gt_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016535 TEST_REQUIRES_ARM_NEON;
16536 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016537 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016538 for (uint32_t n = 1; n <= 8; n++) {
16539 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016540 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016541 .nr(8)
16542 .kr(4)
16543 .sr(1)
16544 .m(m)
16545 .n(n)
16546 .k(k)
16547 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016548 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016549 }
16550 }
16551 }
16552 }
16553
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016554 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_div_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016555 TEST_REQUIRES_ARM_NEON;
16556 for (size_t k = 16; k <= 80; k += 8) {
16557 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016558 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016559 .nr(8)
16560 .kr(4)
16561 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016562 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016563 .n(8)
16564 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016565 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016566 }
16567 }
16568
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016569 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_div_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016570 TEST_REQUIRES_ARM_NEON;
16571 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016572 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016573 for (uint32_t n = 1; n <= 8; n++) {
16574 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016575 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016576 .nr(8)
16577 .kr(4)
16578 .sr(1)
16579 .m(m)
16580 .n(n)
16581 .k(k)
16582 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016583 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016584 }
16585 }
16586 }
16587 }
16588
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016589 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_gt_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016590 TEST_REQUIRES_ARM_NEON;
16591 for (uint32_t n = 9; n < 16; n++) {
16592 for (size_t k = 1; k <= 40; k += 9) {
16593 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016594 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016595 .nr(8)
16596 .kr(4)
16597 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016598 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016599 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016600 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016601 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016602 }
16603 }
16604 }
16605
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016606 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_gt_8_strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016607 TEST_REQUIRES_ARM_NEON;
16608 for (uint32_t n = 9; n < 16; n++) {
16609 for (size_t k = 1; k <= 40; k += 9) {
16610 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016611 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016612 .nr(8)
16613 .kr(4)
16614 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016615 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016616 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016617 .k(k)
16618 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016619 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016620 }
16621 }
16622 }
16623
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016624 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_gt_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016625 TEST_REQUIRES_ARM_NEON;
16626 for (uint32_t n = 9; n < 16; n++) {
16627 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016628 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016629 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016630 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016631 .nr(8)
16632 .kr(4)
16633 .sr(1)
16634 .m(m)
16635 .n(n)
16636 .k(k)
16637 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016638 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016639 }
16640 }
16641 }
16642 }
16643
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016644 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_div_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016645 TEST_REQUIRES_ARM_NEON;
16646 for (uint32_t n = 16; n <= 24; n += 8) {
16647 for (size_t k = 1; k <= 40; k += 9) {
16648 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016649 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016650 .nr(8)
16651 .kr(4)
16652 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016653 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016654 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016655 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016656 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016657 }
16658 }
16659 }
16660
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016661 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_div_8_strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016662 TEST_REQUIRES_ARM_NEON;
16663 for (uint32_t n = 16; n <= 24; n += 8) {
16664 for (size_t k = 1; k <= 40; k += 9) {
16665 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016666 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016667 .nr(8)
16668 .kr(4)
16669 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016670 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016671 .n(n)
16672 .k(k)
16673 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016674 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016675 }
16676 }
16677 }
16678
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016679 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_div_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016680 TEST_REQUIRES_ARM_NEON;
16681 for (uint32_t n = 16; n <= 24; n += 8) {
16682 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016683 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016684 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016685 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016686 .nr(8)
16687 .kr(4)
16688 .sr(1)
16689 .m(m)
16690 .n(n)
16691 .k(k)
16692 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016693 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016694 }
16695 }
16696 }
16697 }
16698
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016699 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016700 TEST_REQUIRES_ARM_NEON;
16701 for (size_t k = 1; k <= 40; k += 9) {
16702 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016703 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016704 .nr(8)
16705 .kr(4)
16706 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016707 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016708 .n(8)
16709 .k(k)
16710 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016711 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016712 }
16713 }
16714
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016715 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, small_kernel_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016716 TEST_REQUIRES_ARM_NEON;
16717 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016718 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016719 for (uint32_t n = 1; n <= 8; n++) {
16720 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016721 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016722 .nr(8)
16723 .kr(4)
16724 .sr(1)
16725 .m(m)
16726 .n(n)
16727 .k(k)
16728 .ks(3)
16729 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016730 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016731 }
16732 }
16733 }
16734 }
16735
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016736 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_gt_8_small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016737 TEST_REQUIRES_ARM_NEON;
16738 for (uint32_t n = 9; n < 16; n++) {
16739 for (size_t k = 1; k <= 40; k += 9) {
16740 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016741 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016742 .nr(8)
16743 .kr(4)
16744 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016745 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016746 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016747 .k(k)
16748 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016749 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016750 }
16751 }
16752 }
16753
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016754 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_div_8_small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016755 TEST_REQUIRES_ARM_NEON;
16756 for (uint32_t n = 16; n <= 24; n += 8) {
16757 for (size_t k = 1; k <= 40; k += 9) {
16758 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016759 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016760 .nr(8)
16761 .kr(4)
16762 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016763 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080016764 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016765 .k(k)
16766 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016767 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016768 }
16769 }
16770 }
16771
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016772 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, strided_cm_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016773 TEST_REQUIRES_ARM_NEON;
16774 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016775 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016776 for (uint32_t n = 1; n <= 8; n++) {
16777 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016778 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016779 .nr(8)
16780 .kr(4)
16781 .sr(1)
16782 .m(m)
16783 .n(n)
16784 .k(k)
16785 .cm_stride(11)
16786 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016787 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016788 }
16789 }
16790 }
16791 }
16792
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016793 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, a_offset) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016794 TEST_REQUIRES_ARM_NEON;
16795 for (size_t k = 1; k <= 40; k += 9) {
16796 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016797 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016798 .nr(8)
16799 .kr(4)
16800 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016801 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016802 .n(8)
16803 .k(k)
16804 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016805 .a_offset(43)
16806 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016807 }
16808 }
16809
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016810 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, zero) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016811 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016812 for (uint32_t mz = 0; mz < 1; mz++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016813 for (size_t k = 1; k <= 40; k += 9) {
16814 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016815 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016816 .nr(8)
16817 .kr(4)
16818 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016819 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016820 .n(8)
16821 .k(k)
16822 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016823 .a_offset(43)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016824 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016825 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016826 }
16827 }
16828 }
16829
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016830 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, qmin) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016831 TEST_REQUIRES_ARM_NEON;
16832 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016833 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016834 .nr(8)
16835 .kr(4)
16836 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016837 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016838 .n(8)
16839 .k(8)
16840 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016841 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016842 }
16843
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016844 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, qmax) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016845 TEST_REQUIRES_ARM_NEON;
16846 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016847 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016848 .nr(8)
16849 .kr(4)
16850 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016851 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016852 .n(8)
16853 .k(8)
16854 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016855 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016856 }
16857
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016858 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, strided_cm) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080016859 TEST_REQUIRES_ARM_NEON;
16860 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016861 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016862 .nr(8)
16863 .kr(4)
16864 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016865 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080016866 .n(8)
16867 .k(8)
16868 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080016869 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016870 }
16871#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
16872
16873
16874#if XNN_ARCH_ARM || XNN_ARCH_ARM64
16875 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, k_eq_8) {
16876 TEST_REQUIRES_ARM_NEON;
16877 GemmMicrokernelTester()
16878 .mr(3)
16879 .nr(8)
16880 .kr(4)
16881 .sr(1)
16882 .m(3)
16883 .n(8)
16884 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080016885 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016886 }
16887
16888 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, strided_cn) {
16889 TEST_REQUIRES_ARM_NEON;
16890 GemmMicrokernelTester()
16891 .mr(3)
16892 .nr(8)
16893 .kr(4)
16894 .sr(1)
16895 .m(3)
16896 .n(8)
16897 .k(8)
16898 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080016899 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016900 }
16901
16902 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, k_eq_8_subtile) {
16903 TEST_REQUIRES_ARM_NEON;
16904 for (uint32_t m = 1; m <= 3; m++) {
16905 for (uint32_t n = 1; n <= 8; n++) {
16906 GemmMicrokernelTester()
16907 .mr(3)
16908 .nr(8)
16909 .kr(4)
16910 .sr(1)
16911 .m(m)
16912 .n(n)
16913 .k(8)
16914 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016915 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016916 }
16917 }
16918 }
16919
16920 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
16927 .sr(1)
16928 .m(m)
16929 .n(8)
16930 .k(8)
16931 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016932 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016933 }
16934 }
16935
16936 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
16943 .sr(1)
16944 .m(3)
16945 .n(n)
16946 .k(8)
16947 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016948 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016949 }
16950 }
16951
16952 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
16959 .sr(1)
16960 .m(3)
16961 .n(8)
16962 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080016963 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016964 }
16965 }
16966
16967 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, k_lt_8_subtile) {
16968 TEST_REQUIRES_ARM_NEON;
16969 for (size_t k = 1; k < 8; k++) {
16970 for (uint32_t m = 1; m <= 3; m++) {
16971 for (uint32_t n = 1; n <= 8; n++) {
16972 GemmMicrokernelTester()
16973 .mr(3)
16974 .nr(8)
16975 .kr(4)
16976 .sr(1)
16977 .m(m)
16978 .n(n)
16979 .k(k)
16980 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080016981 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016982 }
16983 }
16984 }
16985 }
16986
16987 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
16994 .sr(1)
16995 .m(3)
16996 .n(8)
16997 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080016998 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080016999 }
17000 }
17001
17002 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, k_gt_8_subtile) {
17003 TEST_REQUIRES_ARM_NEON;
17004 for (size_t k = 9; k < 16; k++) {
17005 for (uint32_t m = 1; m <= 3; m++) {
17006 for (uint32_t n = 1; n <= 8; n++) {
17007 GemmMicrokernelTester()
17008 .mr(3)
17009 .nr(8)
17010 .kr(4)
17011 .sr(1)
17012 .m(m)
17013 .n(n)
17014 .k(k)
17015 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080017016 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017017 }
17018 }
17019 }
17020 }
17021
17022 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17029 .sr(1)
17030 .m(3)
17031 .n(8)
17032 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080017033 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017034 }
17035 }
17036
17037 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, k_div_8_subtile) {
17038 TEST_REQUIRES_ARM_NEON;
17039 for (size_t k = 16; k <= 80; k += 8) {
17040 for (uint32_t m = 1; m <= 3; m++) {
17041 for (uint32_t n = 1; n <= 8; n++) {
17042 GemmMicrokernelTester()
17043 .mr(3)
17044 .nr(8)
17045 .kr(4)
17046 .sr(1)
17047 .m(m)
17048 .n(n)
17049 .k(k)
17050 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080017051 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017052 }
17053 }
17054 }
17055 }
17056
17057 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17065 .sr(1)
17066 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017067 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017068 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080017069 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017070 }
17071 }
17072 }
17073
17074 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17082 .sr(1)
17083 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017084 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017085 .k(k)
17086 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080017087 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017088 }
17089 }
17090 }
17091
17092 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17101 .sr(1)
17102 .m(m)
17103 .n(n)
17104 .k(k)
17105 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080017106 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017107 }
17108 }
17109 }
17110 }
17111
17112 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17120 .sr(1)
17121 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017122 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017123 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080017124 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017125 }
17126 }
17127 }
17128
17129 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17137 .sr(1)
17138 .m(3)
17139 .n(n)
17140 .k(k)
17141 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080017142 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017143 }
17144 }
17145 }
17146
17147 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17156 .sr(1)
17157 .m(m)
17158 .n(n)
17159 .k(k)
17160 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080017161 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017162 }
17163 }
17164 }
17165 }
17166
17167 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17174 .sr(1)
17175 .m(3)
17176 .n(8)
17177 .k(k)
17178 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080017179 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017180 }
17181 }
17182
17183 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, small_kernel_subtile) {
17184 TEST_REQUIRES_ARM_NEON;
17185 for (size_t k = 1; k <= 40; k += 9) {
17186 for (uint32_t m = 1; m <= 3; m++) {
17187 for (uint32_t n = 1; n <= 8; n++) {
17188 GemmMicrokernelTester()
17189 .mr(3)
17190 .nr(8)
17191 .kr(4)
17192 .sr(1)
17193 .m(m)
17194 .n(n)
17195 .k(k)
17196 .ks(3)
17197 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080017198 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017199 }
17200 }
17201 }
17202 }
17203
17204 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17212 .sr(1)
17213 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017214 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017215 .k(k)
17216 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080017217 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017218 }
17219 }
17220 }
17221
17222 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17230 .sr(1)
17231 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017232 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017233 .k(k)
17234 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080017235 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017236 }
17237 }
17238 }
17239
17240 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, strided_cm_subtile) {
17241 TEST_REQUIRES_ARM_NEON;
17242 for (size_t k = 1; k <= 40; k += 9) {
17243 for (uint32_t m = 1; m <= 3; m++) {
17244 for (uint32_t n = 1; n <= 8; n++) {
17245 GemmMicrokernelTester()
17246 .mr(3)
17247 .nr(8)
17248 .kr(4)
17249 .sr(1)
17250 .m(m)
17251 .n(n)
17252 .k(k)
17253 .cm_stride(11)
17254 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080017255 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017256 }
17257 }
17258 }
17259 }
17260
17261 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, 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(4)
17268 .sr(1)
17269 .m(3)
17270 .n(8)
17271 .k(k)
17272 .ks(3)
17273 .a_offset(127)
Marat Dukhan50323b82022-01-11 00:12:01 -080017274 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017275 }
17276 }
17277
17278 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, zero) {
17279 TEST_REQUIRES_ARM_NEON;
17280 for (uint32_t mz = 0; mz < 3; mz++) {
17281 for (size_t k = 1; k <= 40; k += 9) {
17282 GemmMicrokernelTester()
17283 .mr(3)
17284 .nr(8)
17285 .kr(4)
17286 .sr(1)
17287 .m(3)
17288 .n(8)
17289 .k(k)
17290 .ks(3)
17291 .a_offset(127)
17292 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080017293 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017294 }
17295 }
17296 }
17297
17298 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, qmin) {
17299 TEST_REQUIRES_ARM_NEON;
17300 GemmMicrokernelTester()
17301 .mr(3)
17302 .nr(8)
17303 .kr(4)
17304 .sr(1)
17305 .m(3)
17306 .n(8)
17307 .k(8)
17308 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080017309 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017310 }
17311
17312 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, qmax) {
17313 TEST_REQUIRES_ARM_NEON;
17314 GemmMicrokernelTester()
17315 .mr(3)
17316 .nr(8)
17317 .kr(4)
17318 .sr(1)
17319 .m(3)
17320 .n(8)
17321 .k(8)
17322 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080017323 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017324 }
17325
17326 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD1R, strided_cm) {
17327 TEST_REQUIRES_ARM_NEON;
17328 GemmMicrokernelTester()
17329 .mr(3)
17330 .nr(8)
17331 .kr(4)
17332 .sr(1)
17333 .m(3)
17334 .n(8)
17335 .k(8)
17336 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080017337 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017338 }
17339#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
17340
17341
17342#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017343 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_eq_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017344 TEST_REQUIRES_ARM_NEON;
17345 GemmMicrokernelTester()
17346 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017347 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017348 .kr(4)
17349 .sr(1)
17350 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017351 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017352 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017353 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017354 }
17355
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017356 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017357 TEST_REQUIRES_ARM_NEON;
17358 GemmMicrokernelTester()
17359 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017360 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017361 .kr(4)
17362 .sr(1)
17363 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017364 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017365 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017366 .cn_stride(19)
17367 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017368 }
17369
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017370 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_eq_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017371 TEST_REQUIRES_ARM_NEON;
17372 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017373 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017374 GemmMicrokernelTester()
17375 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017376 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017377 .kr(4)
17378 .sr(1)
17379 .m(m)
17380 .n(n)
17381 .k(8)
17382 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017383 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017384 }
17385 }
17386 }
17387
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017388 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_eq_8_subtile_m) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017389 TEST_REQUIRES_ARM_NEON;
17390 for (uint32_t m = 1; m <= 4; m++) {
17391 GemmMicrokernelTester()
17392 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017393 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017394 .kr(4)
17395 .sr(1)
17396 .m(m)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017397 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017398 .k(8)
17399 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017400 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017401 }
17402 }
17403
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017404 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_eq_8_subtile_n) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017405 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017406 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017407 GemmMicrokernelTester()
17408 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017409 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017410 .kr(4)
17411 .sr(1)
17412 .m(4)
17413 .n(n)
17414 .k(8)
17415 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017416 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017417 }
17418 }
17419
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017420 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_lt_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017421 TEST_REQUIRES_ARM_NEON;
17422 for (size_t k = 1; k < 8; k++) {
17423 GemmMicrokernelTester()
17424 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017425 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017426 .kr(4)
17427 .sr(1)
17428 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017429 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017430 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017431 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017432 }
17433 }
17434
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017435 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_lt_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017436 TEST_REQUIRES_ARM_NEON;
17437 for (size_t k = 1; k < 8; k++) {
17438 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017439 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017440 GemmMicrokernelTester()
17441 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017442 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017443 .kr(4)
17444 .sr(1)
17445 .m(m)
17446 .n(n)
17447 .k(k)
17448 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017449 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017450 }
17451 }
17452 }
17453 }
17454
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017455 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_gt_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017456 TEST_REQUIRES_ARM_NEON;
17457 for (size_t k = 9; k < 16; k++) {
17458 GemmMicrokernelTester()
17459 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017460 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017461 .kr(4)
17462 .sr(1)
17463 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017464 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017465 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017466 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017467 }
17468 }
17469
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017470 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_gt_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017471 TEST_REQUIRES_ARM_NEON;
17472 for (size_t k = 9; k < 16; k++) {
17473 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017474 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017475 GemmMicrokernelTester()
17476 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017477 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017478 .kr(4)
17479 .sr(1)
17480 .m(m)
17481 .n(n)
17482 .k(k)
17483 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017484 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017485 }
17486 }
17487 }
17488 }
17489
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017490 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_div_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017491 TEST_REQUIRES_ARM_NEON;
17492 for (size_t k = 16; k <= 80; k += 8) {
17493 GemmMicrokernelTester()
17494 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017495 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017496 .kr(4)
17497 .sr(1)
17498 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017499 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017500 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017501 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017502 }
17503 }
17504
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017505 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, k_div_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017506 TEST_REQUIRES_ARM_NEON;
17507 for (size_t k = 16; k <= 80; k += 8) {
17508 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017509 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017510 GemmMicrokernelTester()
17511 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017512 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017513 .kr(4)
17514 .sr(1)
17515 .m(m)
17516 .n(n)
17517 .k(k)
17518 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017519 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017520 }
17521 }
17522 }
17523 }
17524
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017525 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, n_gt_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017526 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017527 for (uint32_t n = 17; n < 32; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017528 for (size_t k = 1; k <= 40; k += 9) {
17529 GemmMicrokernelTester()
17530 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017531 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017532 .kr(4)
17533 .sr(1)
17534 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017535 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017536 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017537 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017538 }
17539 }
17540 }
17541
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017542 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, n_gt_16_strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017543 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017544 for (uint32_t n = 17; n < 32; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017545 for (size_t k = 1; k <= 40; k += 9) {
17546 GemmMicrokernelTester()
17547 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017548 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017549 .kr(4)
17550 .sr(1)
17551 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017552 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017553 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017554 .cn_stride(19)
17555 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017556 }
17557 }
17558 }
17559
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017560 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, n_gt_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017561 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017562 for (uint32_t n = 17; n < 32; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017563 for (size_t k = 1; k <= 40; k += 9) {
17564 for (uint32_t m = 1; m <= 4; m++) {
17565 GemmMicrokernelTester()
17566 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017567 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017568 .kr(4)
17569 .sr(1)
17570 .m(m)
17571 .n(n)
17572 .k(k)
17573 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017574 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017575 }
17576 }
17577 }
17578 }
17579
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017580 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, n_div_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017581 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017582 for (uint32_t n = 32; n <= 48; n += 16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017583 for (size_t k = 1; k <= 40; k += 9) {
17584 GemmMicrokernelTester()
17585 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017586 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017587 .kr(4)
17588 .sr(1)
17589 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017590 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017591 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017592 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017593 }
17594 }
17595 }
17596
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017597 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, n_div_16_strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017598 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017599 for (uint32_t n = 32; n <= 48; n += 16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017600 for (size_t k = 1; k <= 40; k += 9) {
17601 GemmMicrokernelTester()
17602 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017603 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017604 .kr(4)
17605 .sr(1)
17606 .m(4)
17607 .n(n)
17608 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017609 .cn_stride(19)
17610 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017611 }
17612 }
17613 }
17614
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017615 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, n_div_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017616 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017617 for (uint32_t n = 32; n <= 48; n += 16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017618 for (size_t k = 1; k <= 40; k += 9) {
17619 for (uint32_t m = 1; m <= 4; m++) {
17620 GemmMicrokernelTester()
17621 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017622 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017623 .kr(4)
17624 .sr(1)
17625 .m(m)
17626 .n(n)
17627 .k(k)
17628 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017629 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017630 }
17631 }
17632 }
17633 }
17634
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017635 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017636 TEST_REQUIRES_ARM_NEON;
17637 for (size_t k = 1; k <= 40; k += 9) {
17638 GemmMicrokernelTester()
17639 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017640 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017641 .kr(4)
17642 .sr(1)
17643 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017644 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017645 .k(k)
17646 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017647 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017648 }
17649 }
17650
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017651 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, small_kernel_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017652 TEST_REQUIRES_ARM_NEON;
17653 for (size_t k = 1; k <= 40; k += 9) {
17654 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017655 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017656 GemmMicrokernelTester()
17657 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017658 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017659 .kr(4)
17660 .sr(1)
17661 .m(m)
17662 .n(n)
17663 .k(k)
17664 .ks(3)
17665 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017666 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017667 }
17668 }
17669 }
17670 }
17671
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017672 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, n_gt_16_small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017673 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017674 for (uint32_t n = 17; n < 32; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017675 for (size_t k = 1; k <= 40; k += 9) {
17676 GemmMicrokernelTester()
17677 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017678 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017679 .kr(4)
17680 .sr(1)
17681 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017682 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017683 .k(k)
17684 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017685 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017686 }
17687 }
17688 }
17689
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017690 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, n_div_16_small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017691 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017692 for (uint32_t n = 32; n <= 48; n += 16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017693 for (size_t k = 1; k <= 40; k += 9) {
17694 GemmMicrokernelTester()
17695 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017696 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017697 .kr(4)
17698 .sr(1)
17699 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080017700 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017701 .k(k)
17702 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017703 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017704 }
17705 }
17706 }
17707
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017708 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, strided_cm_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017709 TEST_REQUIRES_ARM_NEON;
17710 for (size_t k = 1; k <= 40; k += 9) {
17711 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017712 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017713 GemmMicrokernelTester()
17714 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017715 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017716 .kr(4)
17717 .sr(1)
17718 .m(m)
17719 .n(n)
17720 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017721 .cm_stride(19)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017722 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017723 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017724 }
17725 }
17726 }
17727 }
17728
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017729 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, a_offset) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017730 TEST_REQUIRES_ARM_NEON;
17731 for (size_t k = 1; k <= 40; k += 9) {
17732 GemmMicrokernelTester()
17733 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017734 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017735 .kr(4)
17736 .sr(1)
17737 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017738 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017739 .k(k)
17740 .ks(3)
17741 .a_offset(163)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017742 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017743 }
17744 }
17745
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017746 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, zero) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017747 TEST_REQUIRES_ARM_NEON;
17748 for (uint32_t mz = 0; mz < 4; mz++) {
17749 for (size_t k = 1; k <= 40; k += 9) {
17750 GemmMicrokernelTester()
17751 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017752 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017753 .kr(4)
17754 .sr(1)
17755 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017756 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017757 .k(k)
17758 .ks(3)
17759 .a_offset(163)
17760 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017761 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017762 }
17763 }
17764 }
17765
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017766 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, qmin) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017767 TEST_REQUIRES_ARM_NEON;
17768 GemmMicrokernelTester()
17769 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017770 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017771 .kr(4)
17772 .sr(1)
17773 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017774 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017775 .k(8)
17776 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017777 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017778 }
17779
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017780 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, qmax) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017781 TEST_REQUIRES_ARM_NEON;
17782 GemmMicrokernelTester()
17783 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017784 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017785 .kr(4)
17786 .sr(1)
17787 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017788 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017789 .k(8)
17790 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017791 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017792 }
17793
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017794 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MULL_LD1R, strided_cm) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017795 TEST_REQUIRES_ARM_NEON;
17796 GemmMicrokernelTester()
17797 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017798 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017799 .kr(4)
17800 .sr(1)
17801 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017802 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017803 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017804 .cm_stride(19)
17805 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017806 }
17807#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
17808
17809
17810#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017811 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_eq_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017812 TEST_REQUIRES_ARM_NEON;
17813 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017814 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017815 .nr(8)
17816 .kr(4)
17817 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017818 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017819 .n(8)
17820 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017821 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017822 }
17823
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017824 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017825 TEST_REQUIRES_ARM_NEON;
17826 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017827 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017828 .nr(8)
17829 .kr(4)
17830 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017831 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017832 .n(8)
17833 .k(16)
17834 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017835 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017836 }
17837
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017838 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_eq_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017839 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017840 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017841 for (uint32_t n = 1; n <= 8; n++) {
17842 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017843 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017844 .nr(8)
17845 .kr(4)
17846 .sr(1)
17847 .m(m)
17848 .n(n)
17849 .k(16)
17850 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017851 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017852 }
17853 }
17854 }
17855
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017856 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_eq_16_subtile_m) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017857 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017858 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017859 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017860 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017861 .nr(8)
17862 .kr(4)
17863 .sr(1)
17864 .m(m)
17865 .n(8)
17866 .k(16)
17867 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017868 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017869 }
17870 }
17871
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017872 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_eq_16_subtile_n) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017873 TEST_REQUIRES_ARM_NEON;
17874 for (uint32_t n = 1; n <= 8; n++) {
17875 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017876 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017877 .nr(8)
17878 .kr(4)
17879 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017880 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017881 .n(n)
17882 .k(16)
17883 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017884 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017885 }
17886 }
17887
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017888 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_lt_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017889 TEST_REQUIRES_ARM_NEON;
17890 for (size_t k = 1; k < 16; k++) {
17891 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017892 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017893 .nr(8)
17894 .kr(4)
17895 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017896 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017897 .n(8)
17898 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017899 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017900 }
17901 }
17902
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017903 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_lt_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017904 TEST_REQUIRES_ARM_NEON;
17905 for (size_t k = 1; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017906 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017907 for (uint32_t n = 1; n <= 8; n++) {
17908 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017909 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017910 .nr(8)
17911 .kr(4)
17912 .sr(1)
17913 .m(m)
17914 .n(n)
17915 .k(k)
17916 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017917 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017918 }
17919 }
17920 }
17921 }
17922
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017923 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_gt_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017924 TEST_REQUIRES_ARM_NEON;
17925 for (size_t k = 17; k < 32; k++) {
17926 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017927 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017928 .nr(8)
17929 .kr(4)
17930 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017931 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017932 .n(8)
17933 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017934 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017935 }
17936 }
17937
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017938 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_gt_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017939 TEST_REQUIRES_ARM_NEON;
17940 for (size_t k = 17; k < 32; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017941 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017942 for (uint32_t n = 1; n <= 8; n++) {
17943 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017944 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017945 .nr(8)
17946 .kr(4)
17947 .sr(1)
17948 .m(m)
17949 .n(n)
17950 .k(k)
17951 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017952 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017953 }
17954 }
17955 }
17956 }
17957
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017958 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_div_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017959 TEST_REQUIRES_ARM_NEON;
17960 for (size_t k = 32; k <= 160; k += 16) {
17961 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017962 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017963 .nr(8)
17964 .kr(4)
17965 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017966 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017967 .n(8)
17968 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017969 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017970 }
17971 }
17972
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017973 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_div_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017974 TEST_REQUIRES_ARM_NEON;
17975 for (size_t k = 32; k <= 160; k += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017976 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017977 for (uint32_t n = 1; n <= 8; n++) {
17978 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017979 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017980 .nr(8)
17981 .kr(4)
17982 .sr(1)
17983 .m(m)
17984 .n(n)
17985 .k(k)
17986 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017987 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080017988 }
17989 }
17990 }
17991 }
17992
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017993 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_gt_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080017994 TEST_REQUIRES_ARM_NEON;
17995 for (uint32_t n = 9; n < 16; n++) {
17996 for (size_t k = 1; k <= 80; k += 17) {
17997 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080017998 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080017999 .nr(8)
18000 .kr(4)
18001 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018002 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018003 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018004 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018005 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018006 }
18007 }
18008 }
18009
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018010 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_gt_8_strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018011 TEST_REQUIRES_ARM_NEON;
18012 for (uint32_t n = 9; n < 16; n++) {
18013 for (size_t k = 1; k <= 80; k += 17) {
18014 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018015 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018016 .nr(8)
18017 .kr(4)
18018 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018019 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018020 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018021 .k(k)
18022 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018023 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018024 }
18025 }
18026 }
18027
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018028 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_gt_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018029 TEST_REQUIRES_ARM_NEON;
18030 for (uint32_t n = 9; n < 16; n++) {
18031 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018032 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018033 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018034 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018035 .nr(8)
18036 .kr(4)
18037 .sr(1)
18038 .m(m)
18039 .n(n)
18040 .k(k)
18041 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018042 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018043 }
18044 }
18045 }
18046 }
18047
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018048 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_div_8) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018049 TEST_REQUIRES_ARM_NEON;
18050 for (uint32_t n = 16; n <= 24; n += 8) {
18051 for (size_t k = 1; k <= 80; k += 17) {
18052 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018053 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018054 .nr(8)
18055 .kr(4)
18056 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018057 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018058 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018059 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018060 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018061 }
18062 }
18063 }
18064
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018065 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_div_8_strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018066 TEST_REQUIRES_ARM_NEON;
18067 for (uint32_t n = 16; n <= 24; n += 8) {
18068 for (size_t k = 1; k <= 80; k += 17) {
18069 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018070 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018071 .nr(8)
18072 .kr(4)
18073 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018074 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018075 .n(n)
18076 .k(k)
18077 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018078 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018079 }
18080 }
18081 }
18082
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018083 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_div_8_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018084 TEST_REQUIRES_ARM_NEON;
18085 for (uint32_t n = 16; n <= 24; n += 8) {
18086 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018087 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018088 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018089 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018090 .nr(8)
18091 .kr(4)
18092 .sr(1)
18093 .m(m)
18094 .n(n)
18095 .k(k)
18096 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018097 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018098 }
18099 }
18100 }
18101 }
18102
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018103 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018104 TEST_REQUIRES_ARM_NEON;
18105 for (size_t k = 1; k <= 80; k += 17) {
18106 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018107 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018108 .nr(8)
18109 .kr(4)
18110 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018111 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018112 .n(8)
18113 .k(k)
18114 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018115 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018116 }
18117 }
18118
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018119 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, small_kernel_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018120 TEST_REQUIRES_ARM_NEON;
18121 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018122 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018123 for (uint32_t n = 1; n <= 8; n++) {
18124 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018125 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018126 .nr(8)
18127 .kr(4)
18128 .sr(1)
18129 .m(m)
18130 .n(n)
18131 .k(k)
18132 .ks(3)
18133 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018134 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018135 }
18136 }
18137 }
18138 }
18139
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018140 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_gt_8_small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018141 TEST_REQUIRES_ARM_NEON;
18142 for (uint32_t n = 9; n < 16; n++) {
18143 for (size_t k = 1; k <= 80; k += 17) {
18144 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018145 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018146 .nr(8)
18147 .kr(4)
18148 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018149 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018150 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018151 .k(k)
18152 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018153 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018154 }
18155 }
18156 }
18157
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018158 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_div_8_small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018159 TEST_REQUIRES_ARM_NEON;
18160 for (uint32_t n = 16; n <= 24; n += 8) {
18161 for (size_t k = 1; k <= 80; k += 17) {
18162 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018163 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018164 .nr(8)
18165 .kr(4)
18166 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018167 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018168 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018169 .k(k)
18170 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018171 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018172 }
18173 }
18174 }
18175
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018176 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, strided_cm_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018177 TEST_REQUIRES_ARM_NEON;
18178 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018179 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018180 for (uint32_t n = 1; n <= 8; n++) {
18181 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018182 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018183 .nr(8)
18184 .kr(4)
18185 .sr(1)
18186 .m(m)
18187 .n(n)
18188 .k(k)
18189 .cm_stride(11)
18190 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018191 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018192 }
18193 }
18194 }
18195 }
18196
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018197 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, a_offset) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018198 TEST_REQUIRES_ARM_NEON;
18199 for (size_t k = 1; k <= 80; k += 17) {
18200 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018201 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018202 .nr(8)
18203 .kr(4)
18204 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018205 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018206 .n(8)
18207 .k(k)
18208 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018209 .a_offset(83)
18210 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018211 }
18212 }
18213
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018214 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, zero) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018215 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018216 for (uint32_t mz = 0; mz < 1; mz++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018217 for (size_t k = 1; k <= 80; k += 17) {
18218 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018219 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018220 .nr(8)
18221 .kr(4)
18222 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018223 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018224 .n(8)
18225 .k(k)
18226 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018227 .a_offset(83)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018228 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018229 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018230 }
18231 }
18232 }
18233
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018234 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, qmin) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018235 TEST_REQUIRES_ARM_NEON;
18236 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018237 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018238 .nr(8)
18239 .kr(4)
18240 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018241 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018242 .n(8)
18243 .k(16)
18244 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018245 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018246 }
18247
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018248 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, qmax) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018249 TEST_REQUIRES_ARM_NEON;
18250 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018251 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018252 .nr(8)
18253 .kr(4)
18254 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018255 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018256 .n(8)
18257 .k(16)
18258 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018259 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018260 }
18261
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018262 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, strided_cm) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018263 TEST_REQUIRES_ARM_NEON;
18264 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018265 .mr(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018266 .nr(8)
18267 .kr(4)
18268 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018269 .m(1)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018270 .n(8)
18271 .k(16)
18272 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018273 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018274 }
18275#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
18276
18277
18278#if XNN_ARCH_ARM || XNN_ARCH_ARM64
18279 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_eq_16) {
18280 TEST_REQUIRES_ARM_NEON;
18281 GemmMicrokernelTester()
18282 .mr(3)
18283 .nr(8)
18284 .kr(4)
18285 .sr(1)
18286 .m(3)
18287 .n(8)
18288 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -080018289 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018290 }
18291
18292 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, strided_cn) {
18293 TEST_REQUIRES_ARM_NEON;
18294 GemmMicrokernelTester()
18295 .mr(3)
18296 .nr(8)
18297 .kr(4)
18298 .sr(1)
18299 .m(3)
18300 .n(8)
18301 .k(16)
18302 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080018303 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018304 }
18305
18306 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_eq_16_subtile) {
18307 TEST_REQUIRES_ARM_NEON;
18308 for (uint32_t m = 1; m <= 3; m++) {
18309 for (uint32_t n = 1; n <= 8; n++) {
18310 GemmMicrokernelTester()
18311 .mr(3)
18312 .nr(8)
18313 .kr(4)
18314 .sr(1)
18315 .m(m)
18316 .n(n)
18317 .k(16)
18318 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018319 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018320 }
18321 }
18322 }
18323
18324 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_eq_16_subtile_m) {
18325 TEST_REQUIRES_ARM_NEON;
18326 for (uint32_t m = 1; m <= 3; m++) {
18327 GemmMicrokernelTester()
18328 .mr(3)
18329 .nr(8)
18330 .kr(4)
18331 .sr(1)
18332 .m(m)
18333 .n(8)
18334 .k(16)
18335 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018336 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018337 }
18338 }
18339
18340 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_eq_16_subtile_n) {
18341 TEST_REQUIRES_ARM_NEON;
18342 for (uint32_t n = 1; n <= 8; n++) {
18343 GemmMicrokernelTester()
18344 .mr(3)
18345 .nr(8)
18346 .kr(4)
18347 .sr(1)
18348 .m(3)
18349 .n(n)
18350 .k(16)
18351 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018352 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018353 }
18354 }
18355
18356 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_lt_16) {
18357 TEST_REQUIRES_ARM_NEON;
18358 for (size_t k = 1; k < 16; k++) {
18359 GemmMicrokernelTester()
18360 .mr(3)
18361 .nr(8)
18362 .kr(4)
18363 .sr(1)
18364 .m(3)
18365 .n(8)
18366 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080018367 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018368 }
18369 }
18370
18371 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_lt_16_subtile) {
18372 TEST_REQUIRES_ARM_NEON;
18373 for (size_t k = 1; k < 16; k++) {
18374 for (uint32_t m = 1; m <= 3; m++) {
18375 for (uint32_t n = 1; n <= 8; n++) {
18376 GemmMicrokernelTester()
18377 .mr(3)
18378 .nr(8)
18379 .kr(4)
18380 .sr(1)
18381 .m(m)
18382 .n(n)
18383 .k(k)
18384 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018385 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018386 }
18387 }
18388 }
18389 }
18390
18391 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_gt_16) {
18392 TEST_REQUIRES_ARM_NEON;
18393 for (size_t k = 17; k < 32; k++) {
18394 GemmMicrokernelTester()
18395 .mr(3)
18396 .nr(8)
18397 .kr(4)
18398 .sr(1)
18399 .m(3)
18400 .n(8)
18401 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080018402 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018403 }
18404 }
18405
18406 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_gt_16_subtile) {
18407 TEST_REQUIRES_ARM_NEON;
18408 for (size_t k = 17; k < 32; k++) {
18409 for (uint32_t m = 1; m <= 3; m++) {
18410 for (uint32_t n = 1; n <= 8; n++) {
18411 GemmMicrokernelTester()
18412 .mr(3)
18413 .nr(8)
18414 .kr(4)
18415 .sr(1)
18416 .m(m)
18417 .n(n)
18418 .k(k)
18419 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018420 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018421 }
18422 }
18423 }
18424 }
18425
18426 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_div_16) {
18427 TEST_REQUIRES_ARM_NEON;
18428 for (size_t k = 32; k <= 160; k += 16) {
18429 GemmMicrokernelTester()
18430 .mr(3)
18431 .nr(8)
18432 .kr(4)
18433 .sr(1)
18434 .m(3)
18435 .n(8)
18436 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080018437 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018438 }
18439 }
18440
18441 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, k_div_16_subtile) {
18442 TEST_REQUIRES_ARM_NEON;
18443 for (size_t k = 32; k <= 160; k += 16) {
18444 for (uint32_t m = 1; m <= 3; m++) {
18445 for (uint32_t n = 1; n <= 8; n++) {
18446 GemmMicrokernelTester()
18447 .mr(3)
18448 .nr(8)
18449 .kr(4)
18450 .sr(1)
18451 .m(m)
18452 .n(n)
18453 .k(k)
18454 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018455 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018456 }
18457 }
18458 }
18459 }
18460
18461 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, n_gt_8) {
18462 TEST_REQUIRES_ARM_NEON;
18463 for (uint32_t n = 9; n < 16; n++) {
18464 for (size_t k = 1; k <= 80; k += 17) {
18465 GemmMicrokernelTester()
18466 .mr(3)
18467 .nr(8)
18468 .kr(4)
18469 .sr(1)
18470 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018471 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018472 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080018473 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018474 }
18475 }
18476 }
18477
18478 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, n_gt_8_strided_cn) {
18479 TEST_REQUIRES_ARM_NEON;
18480 for (uint32_t n = 9; n < 16; n++) {
18481 for (size_t k = 1; k <= 80; k += 17) {
18482 GemmMicrokernelTester()
18483 .mr(3)
18484 .nr(8)
18485 .kr(4)
18486 .sr(1)
18487 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018488 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018489 .k(k)
18490 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080018491 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018492 }
18493 }
18494 }
18495
18496 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, n_gt_8_subtile) {
18497 TEST_REQUIRES_ARM_NEON;
18498 for (uint32_t n = 9; n < 16; n++) {
18499 for (size_t k = 1; k <= 80; k += 17) {
18500 for (uint32_t m = 1; m <= 3; m++) {
18501 GemmMicrokernelTester()
18502 .mr(3)
18503 .nr(8)
18504 .kr(4)
18505 .sr(1)
18506 .m(m)
18507 .n(n)
18508 .k(k)
18509 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018510 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018511 }
18512 }
18513 }
18514 }
18515
18516 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, n_div_8) {
18517 TEST_REQUIRES_ARM_NEON;
18518 for (uint32_t n = 16; n <= 24; n += 8) {
18519 for (size_t k = 1; k <= 80; k += 17) {
18520 GemmMicrokernelTester()
18521 .mr(3)
18522 .nr(8)
18523 .kr(4)
18524 .sr(1)
18525 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018526 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018527 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080018528 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018529 }
18530 }
18531 }
18532
18533 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, n_div_8_strided_cn) {
18534 TEST_REQUIRES_ARM_NEON;
18535 for (uint32_t n = 16; n <= 24; n += 8) {
18536 for (size_t k = 1; k <= 80; k += 17) {
18537 GemmMicrokernelTester()
18538 .mr(3)
18539 .nr(8)
18540 .kr(4)
18541 .sr(1)
18542 .m(3)
18543 .n(n)
18544 .k(k)
18545 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080018546 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018547 }
18548 }
18549 }
18550
18551 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, n_div_8_subtile) {
18552 TEST_REQUIRES_ARM_NEON;
18553 for (uint32_t n = 16; n <= 24; n += 8) {
18554 for (size_t k = 1; k <= 80; k += 17) {
18555 for (uint32_t m = 1; m <= 3; m++) {
18556 GemmMicrokernelTester()
18557 .mr(3)
18558 .nr(8)
18559 .kr(4)
18560 .sr(1)
18561 .m(m)
18562 .n(n)
18563 .k(k)
18564 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018565 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018566 }
18567 }
18568 }
18569 }
18570
18571 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, small_kernel) {
18572 TEST_REQUIRES_ARM_NEON;
18573 for (size_t k = 1; k <= 80; k += 17) {
18574 GemmMicrokernelTester()
18575 .mr(3)
18576 .nr(8)
18577 .kr(4)
18578 .sr(1)
18579 .m(3)
18580 .n(8)
18581 .k(k)
18582 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080018583 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018584 }
18585 }
18586
18587 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, small_kernel_subtile) {
18588 TEST_REQUIRES_ARM_NEON;
18589 for (size_t k = 1; k <= 80; k += 17) {
18590 for (uint32_t m = 1; m <= 3; m++) {
18591 for (uint32_t n = 1; n <= 8; n++) {
18592 GemmMicrokernelTester()
18593 .mr(3)
18594 .nr(8)
18595 .kr(4)
18596 .sr(1)
18597 .m(m)
18598 .n(n)
18599 .k(k)
18600 .ks(3)
18601 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018602 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018603 }
18604 }
18605 }
18606 }
18607
18608 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, n_gt_8_small_kernel) {
18609 TEST_REQUIRES_ARM_NEON;
18610 for (uint32_t n = 9; n < 16; n++) {
18611 for (size_t k = 1; k <= 80; k += 17) {
18612 GemmMicrokernelTester()
18613 .mr(3)
18614 .nr(8)
18615 .kr(4)
18616 .sr(1)
18617 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018618 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018619 .k(k)
18620 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080018621 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018622 }
18623 }
18624 }
18625
18626 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, n_div_8_small_kernel) {
18627 TEST_REQUIRES_ARM_NEON;
18628 for (uint32_t n = 16; n <= 24; n += 8) {
18629 for (size_t k = 1; k <= 80; k += 17) {
18630 GemmMicrokernelTester()
18631 .mr(3)
18632 .nr(8)
18633 .kr(4)
18634 .sr(1)
18635 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018636 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018637 .k(k)
18638 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080018639 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018640 }
18641 }
18642 }
18643
18644 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, strided_cm_subtile) {
18645 TEST_REQUIRES_ARM_NEON;
18646 for (size_t k = 1; k <= 80; k += 17) {
18647 for (uint32_t m = 1; m <= 3; m++) {
18648 for (uint32_t n = 1; n <= 8; n++) {
18649 GemmMicrokernelTester()
18650 .mr(3)
18651 .nr(8)
18652 .kr(4)
18653 .sr(1)
18654 .m(m)
18655 .n(n)
18656 .k(k)
18657 .cm_stride(11)
18658 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080018659 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018660 }
18661 }
18662 }
18663 }
18664
18665 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, a_offset) {
18666 TEST_REQUIRES_ARM_NEON;
18667 for (size_t k = 1; k <= 80; k += 17) {
18668 GemmMicrokernelTester()
18669 .mr(3)
18670 .nr(8)
18671 .kr(4)
18672 .sr(1)
18673 .m(3)
18674 .n(8)
18675 .k(k)
18676 .ks(3)
18677 .a_offset(251)
Marat Dukhan50323b82022-01-11 00:12:01 -080018678 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018679 }
18680 }
18681
18682 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, zero) {
18683 TEST_REQUIRES_ARM_NEON;
18684 for (uint32_t mz = 0; mz < 3; mz++) {
18685 for (size_t k = 1; k <= 80; k += 17) {
18686 GemmMicrokernelTester()
18687 .mr(3)
18688 .nr(8)
18689 .kr(4)
18690 .sr(1)
18691 .m(3)
18692 .n(8)
18693 .k(k)
18694 .ks(3)
18695 .a_offset(251)
18696 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080018697 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018698 }
18699 }
18700 }
18701
18702 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, qmin) {
18703 TEST_REQUIRES_ARM_NEON;
18704 GemmMicrokernelTester()
18705 .mr(3)
18706 .nr(8)
18707 .kr(4)
18708 .sr(1)
18709 .m(3)
18710 .n(8)
18711 .k(16)
18712 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080018713 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018714 }
18715
18716 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, qmax) {
18717 TEST_REQUIRES_ARM_NEON;
18718 GemmMicrokernelTester()
18719 .mr(3)
18720 .nr(8)
18721 .kr(4)
18722 .sr(1)
18723 .m(3)
18724 .n(8)
18725 .k(16)
18726 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080018727 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018728 }
18729
18730 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_LD1R, strided_cm) {
18731 TEST_REQUIRES_ARM_NEON;
18732 GemmMicrokernelTester()
18733 .mr(3)
18734 .nr(8)
18735 .kr(4)
18736 .sr(1)
18737 .m(3)
18738 .n(8)
18739 .k(16)
18740 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080018741 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018742 }
18743#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
18744
18745
18746#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018747 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_eq_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018748 TEST_REQUIRES_ARM_NEON;
18749 GemmMicrokernelTester()
18750 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018751 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018752 .kr(4)
18753 .sr(1)
18754 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018755 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018756 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018757 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018758 }
18759
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018760 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018761 TEST_REQUIRES_ARM_NEON;
18762 GemmMicrokernelTester()
18763 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018764 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018765 .kr(4)
18766 .sr(1)
18767 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018768 .n(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018769 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018770 .cn_stride(19)
18771 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018772 }
18773
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018774 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_eq_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018775 TEST_REQUIRES_ARM_NEON;
18776 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018777 for (uint32_t n = 1; n <= 16; n++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018778 GemmMicrokernelTester()
18779 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018780 .nr(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018781 .kr(4)
18782 .sr(1)
18783 .m(m)
18784 .n(n)
18785 .k(16)
18786 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018787 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018788 }
18789 }
18790 }
18791
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018792 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_eq_16_subtile_m) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018793 TEST_REQUIRES_ARM_NEON;
18794 for (uint32_t m = 1; m <= 4; m++) {
18795 GemmMicrokernelTester()
18796 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018797 .nr(16)
18798 .kr(4)
18799 .sr(1)
18800 .m(m)
18801 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018802 .k(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018803 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018804 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018805 }
18806 }
18807
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018808 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_eq_16_subtile_n) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018809 TEST_REQUIRES_ARM_NEON;
18810 for (uint32_t n = 1; n <= 16; n++) {
18811 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018812 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018813 .nr(16)
18814 .kr(4)
18815 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018816 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018817 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018818 .k(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018819 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018820 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018821 }
18822 }
18823
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018824 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_lt_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018825 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018826 for (size_t k = 1; k < 16; k++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018827 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018828 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018829 .nr(16)
18830 .kr(4)
18831 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018832 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018833 .n(16)
18834 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018835 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018836 }
18837 }
18838
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018839 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_lt_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018840 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018841 for (size_t k = 1; k < 16; k++) {
18842 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018843 for (uint32_t n = 1; n <= 16; n++) {
18844 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018845 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018846 .nr(16)
18847 .kr(4)
18848 .sr(1)
18849 .m(m)
18850 .n(n)
18851 .k(k)
18852 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018853 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018854 }
18855 }
18856 }
18857 }
18858
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018859 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_gt_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018860 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018861 for (size_t k = 17; k < 32; k++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018862 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018863 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018864 .nr(16)
18865 .kr(4)
18866 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018867 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018868 .n(16)
18869 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018870 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018871 }
18872 }
18873
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018874 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_gt_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018875 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018876 for (size_t k = 17; k < 32; k++) {
18877 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018878 for (uint32_t n = 1; n <= 16; n++) {
18879 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018880 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018881 .nr(16)
18882 .kr(4)
18883 .sr(1)
18884 .m(m)
18885 .n(n)
18886 .k(k)
18887 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018888 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018889 }
18890 }
18891 }
18892 }
18893
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018894 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_div_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018895 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018896 for (size_t k = 32; k <= 160; k += 16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018897 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018898 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018899 .nr(16)
18900 .kr(4)
18901 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018902 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018903 .n(16)
18904 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018905 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018906 }
18907 }
18908
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018909 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, k_div_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018910 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018911 for (size_t k = 32; k <= 160; k += 16) {
18912 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018913 for (uint32_t n = 1; n <= 16; n++) {
18914 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018915 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018916 .nr(16)
18917 .kr(4)
18918 .sr(1)
18919 .m(m)
18920 .n(n)
18921 .k(k)
18922 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018923 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018924 }
18925 }
18926 }
18927 }
18928
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018929 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, n_gt_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018930 TEST_REQUIRES_ARM_NEON;
18931 for (uint32_t n = 17; n < 32; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018932 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018933 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018934 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018935 .nr(16)
18936 .kr(4)
18937 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018938 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018939 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018940 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018941 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018942 }
18943 }
18944 }
18945
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018946 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, n_gt_16_strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018947 TEST_REQUIRES_ARM_NEON;
18948 for (uint32_t n = 17; n < 32; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018949 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018950 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018951 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018952 .nr(16)
18953 .kr(4)
18954 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018955 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018956 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018957 .k(k)
18958 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018959 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018960 }
18961 }
18962 }
18963
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018964 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, n_gt_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018965 TEST_REQUIRES_ARM_NEON;
18966 for (uint32_t n = 17; n < 32; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018967 for (size_t k = 1; k <= 80; k += 17) {
18968 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018969 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018970 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018971 .nr(16)
18972 .kr(4)
18973 .sr(1)
18974 .m(m)
18975 .n(n)
18976 .k(k)
18977 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018978 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018979 }
18980 }
18981 }
18982 }
18983
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018984 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, n_div_16) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018985 TEST_REQUIRES_ARM_NEON;
18986 for (uint32_t n = 32; n <= 48; n += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018987 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080018988 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018989 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018990 .nr(16)
18991 .kr(4)
18992 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018993 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080018994 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080018995 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080018996 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080018997 }
18998 }
18999 }
19000
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019001 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, n_div_16_strided_cn) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019002 TEST_REQUIRES_ARM_NEON;
19003 for (uint32_t n = 32; n <= 48; n += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019004 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019005 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019006 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019007 .nr(16)
19008 .kr(4)
19009 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019010 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019011 .n(n)
19012 .k(k)
19013 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019014 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019015 }
19016 }
19017 }
19018
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019019 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, n_div_16_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019020 TEST_REQUIRES_ARM_NEON;
19021 for (uint32_t n = 32; n <= 48; n += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019022 for (size_t k = 1; k <= 80; k += 17) {
19023 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019024 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019025 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019026 .nr(16)
19027 .kr(4)
19028 .sr(1)
19029 .m(m)
19030 .n(n)
19031 .k(k)
19032 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019033 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019034 }
19035 }
19036 }
19037 }
19038
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019039 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019040 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019041 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019042 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019043 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019044 .nr(16)
19045 .kr(4)
19046 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019047 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019048 .n(16)
19049 .k(k)
19050 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019051 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019052 }
19053 }
19054
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019055 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, small_kernel_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019056 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019057 for (size_t k = 1; k <= 80; k += 17) {
19058 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019059 for (uint32_t n = 1; n <= 16; n++) {
19060 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019061 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019062 .nr(16)
19063 .kr(4)
19064 .sr(1)
19065 .m(m)
19066 .n(n)
19067 .k(k)
19068 .ks(3)
19069 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019070 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019071 }
19072 }
19073 }
19074 }
19075
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019076 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, n_gt_16_small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019077 TEST_REQUIRES_ARM_NEON;
19078 for (uint32_t n = 17; n < 32; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019079 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019080 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019081 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019082 .nr(16)
19083 .kr(4)
19084 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019085 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019086 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019087 .k(k)
19088 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019089 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019090 }
19091 }
19092 }
19093
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019094 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, n_div_16_small_kernel) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019095 TEST_REQUIRES_ARM_NEON;
19096 for (uint32_t n = 32; n <= 48; n += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019097 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019098 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019099 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019100 .nr(16)
19101 .kr(4)
19102 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019103 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019104 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019105 .k(k)
19106 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019107 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019108 }
19109 }
19110 }
19111
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019112 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, strided_cm_subtile) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019113 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019114 for (size_t k = 1; k <= 80; k += 17) {
19115 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019116 for (uint32_t n = 1; n <= 16; n++) {
19117 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019118 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019119 .nr(16)
19120 .kr(4)
19121 .sr(1)
19122 .m(m)
19123 .n(n)
19124 .k(k)
19125 .cm_stride(19)
19126 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019127 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019128 }
19129 }
19130 }
19131 }
19132
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019133 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, a_offset) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019134 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019135 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019136 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019137 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019138 .nr(16)
19139 .kr(4)
19140 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019141 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019142 .n(16)
19143 .k(k)
19144 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019145 .a_offset(331)
19146 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019147 }
19148 }
19149
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019150 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, zero) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019151 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019152 for (uint32_t mz = 0; mz < 4; mz++) {
19153 for (size_t k = 1; k <= 80; k += 17) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019154 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019155 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019156 .nr(16)
19157 .kr(4)
19158 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019159 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019160 .n(16)
19161 .k(k)
19162 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019163 .a_offset(331)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019164 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019165 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019166 }
19167 }
19168 }
19169
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019170 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, qmin) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019171 TEST_REQUIRES_ARM_NEON;
19172 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019173 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019174 .nr(16)
19175 .kr(4)
19176 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019177 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019178 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019179 .k(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019180 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019181 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019182 }
19183
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019184 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, qmax) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019185 TEST_REQUIRES_ARM_NEON;
19186 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019187 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019188 .nr(16)
19189 .kr(4)
19190 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019191 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019192 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019193 .k(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019194 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019195 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019196 }
19197
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019198 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD1R, strided_cm) {
Frank Barchard64ab1b72021-11-22 10:57:40 -080019199 TEST_REQUIRES_ARM_NEON;
19200 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019201 .mr(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019202 .nr(16)
19203 .kr(4)
19204 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019205 .m(4)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019206 .n(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019207 .k(16)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019208 .cm_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080019209 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019210 }
19211#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
19212
19213
19214#if XNN_ARCH_ARM || XNN_ARCH_ARM64
19215 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_eq_8) {
19216 TEST_REQUIRES_ARM_NEON;
19217 GemmMicrokernelTester()
19218 .mr(2)
19219 .nr(16)
19220 .kr(4)
19221 .sr(1)
19222 .m(2)
19223 .n(16)
19224 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080019225 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019226 }
19227
19228 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, strided_cn) {
19229 TEST_REQUIRES_ARM_NEON;
19230 GemmMicrokernelTester()
19231 .mr(2)
19232 .nr(16)
19233 .kr(4)
19234 .sr(1)
19235 .m(2)
19236 .n(16)
19237 .k(8)
19238 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080019239 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019240 }
19241
19242 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_eq_8_subtile) {
19243 TEST_REQUIRES_ARM_NEON;
19244 for (uint32_t m = 1; m <= 2; m++) {
19245 for (uint32_t n = 1; n <= 16; n++) {
19246 GemmMicrokernelTester()
19247 .mr(2)
19248 .nr(16)
19249 .kr(4)
19250 .sr(1)
19251 .m(m)
19252 .n(n)
19253 .k(8)
19254 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019255 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019256 }
19257 }
19258 }
19259
19260 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_eq_8_subtile_m) {
19261 TEST_REQUIRES_ARM_NEON;
19262 for (uint32_t m = 1; m <= 2; m++) {
19263 GemmMicrokernelTester()
19264 .mr(2)
19265 .nr(16)
19266 .kr(4)
19267 .sr(1)
19268 .m(m)
19269 .n(16)
19270 .k(8)
19271 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019272 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019273 }
19274 }
19275
19276 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_eq_8_subtile_n) {
19277 TEST_REQUIRES_ARM_NEON;
19278 for (uint32_t n = 1; n <= 16; n++) {
19279 GemmMicrokernelTester()
19280 .mr(2)
19281 .nr(16)
19282 .kr(4)
19283 .sr(1)
19284 .m(2)
19285 .n(n)
19286 .k(8)
19287 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019288 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019289 }
19290 }
19291
19292 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_lt_8) {
19293 TEST_REQUIRES_ARM_NEON;
19294 for (size_t k = 1; k < 8; k++) {
19295 GemmMicrokernelTester()
19296 .mr(2)
19297 .nr(16)
19298 .kr(4)
19299 .sr(1)
19300 .m(2)
19301 .n(16)
19302 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019303 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019304 }
19305 }
19306
19307 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_lt_8_subtile) {
19308 TEST_REQUIRES_ARM_NEON;
19309 for (size_t k = 1; k < 8; k++) {
19310 for (uint32_t m = 1; m <= 2; m++) {
19311 for (uint32_t n = 1; n <= 16; n++) {
19312 GemmMicrokernelTester()
19313 .mr(2)
19314 .nr(16)
19315 .kr(4)
19316 .sr(1)
19317 .m(m)
19318 .n(n)
19319 .k(k)
19320 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019321 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019322 }
19323 }
19324 }
19325 }
19326
19327 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_gt_8) {
19328 TEST_REQUIRES_ARM_NEON;
19329 for (size_t k = 9; k < 16; k++) {
19330 GemmMicrokernelTester()
19331 .mr(2)
19332 .nr(16)
19333 .kr(4)
19334 .sr(1)
19335 .m(2)
19336 .n(16)
19337 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019338 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019339 }
19340 }
19341
19342 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_gt_8_subtile) {
19343 TEST_REQUIRES_ARM_NEON;
19344 for (size_t k = 9; k < 16; k++) {
19345 for (uint32_t m = 1; m <= 2; m++) {
19346 for (uint32_t n = 1; n <= 16; n++) {
19347 GemmMicrokernelTester()
19348 .mr(2)
19349 .nr(16)
19350 .kr(4)
19351 .sr(1)
19352 .m(m)
19353 .n(n)
19354 .k(k)
19355 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019356 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019357 }
19358 }
19359 }
19360 }
19361
19362 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_div_8) {
19363 TEST_REQUIRES_ARM_NEON;
19364 for (size_t k = 16; k <= 80; k += 8) {
19365 GemmMicrokernelTester()
19366 .mr(2)
19367 .nr(16)
19368 .kr(4)
19369 .sr(1)
19370 .m(2)
19371 .n(16)
19372 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019373 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019374 }
19375 }
19376
19377 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, k_div_8_subtile) {
19378 TEST_REQUIRES_ARM_NEON;
19379 for (size_t k = 16; k <= 80; k += 8) {
19380 for (uint32_t m = 1; m <= 2; m++) {
19381 for (uint32_t n = 1; n <= 16; n++) {
19382 GemmMicrokernelTester()
19383 .mr(2)
19384 .nr(16)
19385 .kr(4)
19386 .sr(1)
19387 .m(m)
19388 .n(n)
19389 .k(k)
19390 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019391 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019392 }
19393 }
19394 }
19395 }
19396
19397 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, n_gt_16) {
19398 TEST_REQUIRES_ARM_NEON;
19399 for (uint32_t n = 17; n < 32; n++) {
19400 for (size_t k = 1; k <= 40; k += 9) {
19401 GemmMicrokernelTester()
19402 .mr(2)
19403 .nr(16)
19404 .kr(4)
19405 .sr(1)
19406 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019407 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019408 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019409 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019410 }
19411 }
19412 }
19413
19414 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, 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 <= 40; k += 9) {
19418 GemmMicrokernelTester()
19419 .mr(2)
19420 .nr(16)
19421 .kr(4)
19422 .sr(1)
19423 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019424 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019425 .k(k)
19426 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080019427 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019428 }
19429 }
19430 }
19431
19432 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, 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 <= 40; k += 9) {
19436 for (uint32_t m = 1; m <= 2; m++) {
19437 GemmMicrokernelTester()
19438 .mr(2)
19439 .nr(16)
19440 .kr(4)
19441 .sr(1)
19442 .m(m)
19443 .n(n)
19444 .k(k)
19445 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019446 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019447 }
19448 }
19449 }
19450 }
19451
19452 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, 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 <= 40; k += 9) {
19456 GemmMicrokernelTester()
19457 .mr(2)
19458 .nr(16)
19459 .kr(4)
19460 .sr(1)
19461 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019462 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019463 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019464 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019465 }
19466 }
19467 }
19468
19469 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, 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 <= 40; k += 9) {
19473 GemmMicrokernelTester()
19474 .mr(2)
19475 .nr(16)
19476 .kr(4)
19477 .sr(1)
19478 .m(2)
19479 .n(n)
19480 .k(k)
19481 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080019482 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019483 }
19484 }
19485 }
19486
19487 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, 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 <= 40; k += 9) {
19491 for (uint32_t m = 1; m <= 2; m++) {
19492 GemmMicrokernelTester()
19493 .mr(2)
19494 .nr(16)
19495 .kr(4)
19496 .sr(1)
19497 .m(m)
19498 .n(n)
19499 .k(k)
19500 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019501 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019502 }
19503 }
19504 }
19505 }
19506
19507 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, small_kernel) {
19508 TEST_REQUIRES_ARM_NEON;
19509 for (size_t k = 1; k <= 40; k += 9) {
19510 GemmMicrokernelTester()
19511 .mr(2)
19512 .nr(16)
19513 .kr(4)
19514 .sr(1)
19515 .m(2)
19516 .n(16)
19517 .k(k)
19518 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080019519 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019520 }
19521 }
19522
19523 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, small_kernel_subtile) {
19524 TEST_REQUIRES_ARM_NEON;
19525 for (size_t k = 1; k <= 40; k += 9) {
19526 for (uint32_t m = 1; m <= 2; m++) {
19527 for (uint32_t n = 1; n <= 16; n++) {
19528 GemmMicrokernelTester()
19529 .mr(2)
19530 .nr(16)
19531 .kr(4)
19532 .sr(1)
19533 .m(m)
19534 .n(n)
19535 .k(k)
19536 .ks(3)
19537 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019538 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019539 }
19540 }
19541 }
19542 }
19543
19544 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, 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 <= 40; k += 9) {
19548 GemmMicrokernelTester()
19549 .mr(2)
19550 .nr(16)
19551 .kr(4)
19552 .sr(1)
19553 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019554 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019555 .k(k)
19556 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080019557 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019558 }
19559 }
19560 }
19561
19562 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, 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 <= 40; k += 9) {
19566 GemmMicrokernelTester()
19567 .mr(2)
19568 .nr(16)
19569 .kr(4)
19570 .sr(1)
19571 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019572 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019573 .k(k)
19574 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080019575 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019576 }
19577 }
19578 }
19579
19580 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, strided_cm_subtile) {
19581 TEST_REQUIRES_ARM_NEON;
19582 for (size_t k = 1; k <= 40; k += 9) {
19583 for (uint32_t m = 1; m <= 2; m++) {
19584 for (uint32_t n = 1; n <= 16; n++) {
19585 GemmMicrokernelTester()
19586 .mr(2)
19587 .nr(16)
19588 .kr(4)
19589 .sr(1)
19590 .m(m)
19591 .n(n)
19592 .k(k)
19593 .cm_stride(19)
19594 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019595 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019596 }
19597 }
19598 }
19599 }
19600
19601 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, a_offset) {
19602 TEST_REQUIRES_ARM_NEON;
19603 for (size_t k = 1; k <= 40; k += 9) {
19604 GemmMicrokernelTester()
19605 .mr(2)
19606 .nr(16)
19607 .kr(4)
19608 .sr(1)
19609 .m(2)
19610 .n(16)
19611 .k(k)
19612 .ks(3)
19613 .a_offset(83)
Marat Dukhan50323b82022-01-11 00:12:01 -080019614 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019615 }
19616 }
19617
19618 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, zero) {
19619 TEST_REQUIRES_ARM_NEON;
19620 for (uint32_t mz = 0; mz < 2; mz++) {
19621 for (size_t k = 1; k <= 40; k += 9) {
19622 GemmMicrokernelTester()
19623 .mr(2)
19624 .nr(16)
19625 .kr(4)
19626 .sr(1)
19627 .m(2)
19628 .n(16)
19629 .k(k)
19630 .ks(3)
19631 .a_offset(83)
19632 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080019633 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019634 }
19635 }
19636 }
19637
19638 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, qmin) {
19639 TEST_REQUIRES_ARM_NEON;
19640 GemmMicrokernelTester()
19641 .mr(2)
19642 .nr(16)
19643 .kr(4)
19644 .sr(1)
19645 .m(2)
19646 .n(16)
19647 .k(8)
19648 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080019649 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019650 }
19651
19652 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, qmax) {
19653 TEST_REQUIRES_ARM_NEON;
19654 GemmMicrokernelTester()
19655 .mr(2)
19656 .nr(16)
19657 .kr(4)
19658 .sr(1)
19659 .m(2)
19660 .n(16)
19661 .k(8)
19662 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080019663 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019664 }
19665
19666 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_LD2R, strided_cm) {
19667 TEST_REQUIRES_ARM_NEON;
19668 GemmMicrokernelTester()
19669 .mr(2)
19670 .nr(16)
19671 .kr(4)
19672 .sr(1)
19673 .m(2)
19674 .n(16)
19675 .k(8)
19676 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080019677 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019678 }
19679#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
19680
19681
19682#if XNN_ARCH_ARM || XNN_ARCH_ARM64
19683 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_eq_8) {
19684 TEST_REQUIRES_ARM_NEON;
19685 GemmMicrokernelTester()
19686 .mr(3)
19687 .nr(16)
19688 .kr(4)
19689 .sr(1)
19690 .m(3)
19691 .n(16)
19692 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080019693 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019694 }
19695
19696 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, strided_cn) {
19697 TEST_REQUIRES_ARM_NEON;
19698 GemmMicrokernelTester()
19699 .mr(3)
19700 .nr(16)
19701 .kr(4)
19702 .sr(1)
19703 .m(3)
19704 .n(16)
19705 .k(8)
19706 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080019707 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019708 }
19709
19710 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_eq_8_subtile) {
19711 TEST_REQUIRES_ARM_NEON;
19712 for (uint32_t m = 1; m <= 3; m++) {
19713 for (uint32_t n = 1; n <= 16; n++) {
19714 GemmMicrokernelTester()
19715 .mr(3)
19716 .nr(16)
19717 .kr(4)
19718 .sr(1)
19719 .m(m)
19720 .n(n)
19721 .k(8)
19722 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019723 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019724 }
19725 }
19726 }
19727
19728 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_eq_8_subtile_m) {
19729 TEST_REQUIRES_ARM_NEON;
19730 for (uint32_t m = 1; m <= 3; m++) {
19731 GemmMicrokernelTester()
19732 .mr(3)
19733 .nr(16)
19734 .kr(4)
19735 .sr(1)
19736 .m(m)
19737 .n(16)
19738 .k(8)
19739 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019740 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019741 }
19742 }
19743
19744 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_eq_8_subtile_n) {
19745 TEST_REQUIRES_ARM_NEON;
19746 for (uint32_t n = 1; n <= 16; n++) {
19747 GemmMicrokernelTester()
19748 .mr(3)
19749 .nr(16)
19750 .kr(4)
19751 .sr(1)
19752 .m(3)
19753 .n(n)
19754 .k(8)
19755 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019756 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019757 }
19758 }
19759
19760 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_lt_8) {
19761 TEST_REQUIRES_ARM_NEON;
19762 for (size_t k = 1; k < 8; k++) {
19763 GemmMicrokernelTester()
19764 .mr(3)
19765 .nr(16)
19766 .kr(4)
19767 .sr(1)
19768 .m(3)
19769 .n(16)
19770 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019771 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019772 }
19773 }
19774
19775 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_lt_8_subtile) {
19776 TEST_REQUIRES_ARM_NEON;
19777 for (size_t k = 1; k < 8; k++) {
19778 for (uint32_t m = 1; m <= 3; m++) {
19779 for (uint32_t n = 1; n <= 16; n++) {
19780 GemmMicrokernelTester()
19781 .mr(3)
19782 .nr(16)
19783 .kr(4)
19784 .sr(1)
19785 .m(m)
19786 .n(n)
19787 .k(k)
19788 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019789 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019790 }
19791 }
19792 }
19793 }
19794
19795 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_gt_8) {
19796 TEST_REQUIRES_ARM_NEON;
19797 for (size_t k = 9; k < 16; k++) {
19798 GemmMicrokernelTester()
19799 .mr(3)
19800 .nr(16)
19801 .kr(4)
19802 .sr(1)
19803 .m(3)
19804 .n(16)
19805 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019806 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019807 }
19808 }
19809
19810 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_gt_8_subtile) {
19811 TEST_REQUIRES_ARM_NEON;
19812 for (size_t k = 9; k < 16; k++) {
19813 for (uint32_t m = 1; m <= 3; m++) {
19814 for (uint32_t n = 1; n <= 16; n++) {
19815 GemmMicrokernelTester()
19816 .mr(3)
19817 .nr(16)
19818 .kr(4)
19819 .sr(1)
19820 .m(m)
19821 .n(n)
19822 .k(k)
19823 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019824 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019825 }
19826 }
19827 }
19828 }
19829
19830 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_div_8) {
19831 TEST_REQUIRES_ARM_NEON;
19832 for (size_t k = 16; k <= 80; k += 8) {
19833 GemmMicrokernelTester()
19834 .mr(3)
19835 .nr(16)
19836 .kr(4)
19837 .sr(1)
19838 .m(3)
19839 .n(16)
19840 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019841 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019842 }
19843 }
19844
19845 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, k_div_8_subtile) {
19846 TEST_REQUIRES_ARM_NEON;
19847 for (size_t k = 16; k <= 80; k += 8) {
19848 for (uint32_t m = 1; m <= 3; m++) {
19849 for (uint32_t n = 1; n <= 16; n++) {
19850 GemmMicrokernelTester()
19851 .mr(3)
19852 .nr(16)
19853 .kr(4)
19854 .sr(1)
19855 .m(m)
19856 .n(n)
19857 .k(k)
19858 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019859 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019860 }
19861 }
19862 }
19863 }
19864
19865 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, n_gt_16) {
19866 TEST_REQUIRES_ARM_NEON;
19867 for (uint32_t n = 17; n < 32; n++) {
19868 for (size_t k = 1; k <= 40; k += 9) {
19869 GemmMicrokernelTester()
19870 .mr(3)
19871 .nr(16)
19872 .kr(4)
19873 .sr(1)
19874 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019875 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019876 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019877 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019878 }
19879 }
19880 }
19881
19882 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, n_gt_16_strided_cn) {
19883 TEST_REQUIRES_ARM_NEON;
19884 for (uint32_t n = 17; n < 32; n++) {
19885 for (size_t k = 1; k <= 40; k += 9) {
19886 GemmMicrokernelTester()
19887 .mr(3)
19888 .nr(16)
19889 .kr(4)
19890 .sr(1)
19891 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019892 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019893 .k(k)
19894 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080019895 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019896 }
19897 }
19898 }
19899
19900 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, n_gt_16_subtile) {
19901 TEST_REQUIRES_ARM_NEON;
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 <= 3; m++) {
19905 GemmMicrokernelTester()
19906 .mr(3)
19907 .nr(16)
19908 .kr(4)
19909 .sr(1)
19910 .m(m)
19911 .n(n)
19912 .k(k)
19913 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019914 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019915 }
19916 }
19917 }
19918 }
19919
19920 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, n_div_16) {
19921 TEST_REQUIRES_ARM_NEON;
19922 for (uint32_t n = 32; n <= 48; n += 16) {
19923 for (size_t k = 1; k <= 40; k += 9) {
19924 GemmMicrokernelTester()
19925 .mr(3)
19926 .nr(16)
19927 .kr(4)
19928 .sr(1)
19929 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080019930 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080019931 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080019932 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019933 }
19934 }
19935 }
19936
19937 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, n_div_16_strided_cn) {
19938 TEST_REQUIRES_ARM_NEON;
19939 for (uint32_t n = 32; n <= 48; n += 16) {
19940 for (size_t k = 1; k <= 40; k += 9) {
19941 GemmMicrokernelTester()
19942 .mr(3)
19943 .nr(16)
19944 .kr(4)
19945 .sr(1)
19946 .m(3)
19947 .n(n)
19948 .k(k)
19949 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080019950 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019951 }
19952 }
19953 }
19954
19955 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, n_div_16_subtile) {
19956 TEST_REQUIRES_ARM_NEON;
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 <= 3; m++) {
19960 GemmMicrokernelTester()
19961 .mr(3)
19962 .nr(16)
19963 .kr(4)
19964 .sr(1)
19965 .m(m)
19966 .n(n)
19967 .k(k)
19968 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080019969 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019970 }
19971 }
19972 }
19973 }
19974
19975 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, small_kernel) {
19976 TEST_REQUIRES_ARM_NEON;
19977 for (size_t k = 1; k <= 40; k += 9) {
19978 GemmMicrokernelTester()
19979 .mr(3)
19980 .nr(16)
19981 .kr(4)
19982 .sr(1)
19983 .m(3)
19984 .n(16)
19985 .k(k)
19986 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080019987 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080019988 }
19989 }
19990
19991 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, small_kernel_subtile) {
19992 TEST_REQUIRES_ARM_NEON;
19993 for (size_t k = 1; k <= 40; k += 9) {
19994 for (uint32_t m = 1; m <= 3; m++) {
19995 for (uint32_t n = 1; n <= 16; n++) {
19996 GemmMicrokernelTester()
19997 .mr(3)
19998 .nr(16)
19999 .kr(4)
20000 .sr(1)
20001 .m(m)
20002 .n(n)
20003 .k(k)
20004 .ks(3)
20005 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020006 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020007 }
20008 }
20009 }
20010 }
20011
20012 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, n_gt_16_small_kernel) {
20013 TEST_REQUIRES_ARM_NEON;
20014 for (uint32_t n = 17; n < 32; n++) {
20015 for (size_t k = 1; k <= 40; k += 9) {
20016 GemmMicrokernelTester()
20017 .mr(3)
20018 .nr(16)
20019 .kr(4)
20020 .sr(1)
20021 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020022 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020023 .k(k)
20024 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080020025 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020026 }
20027 }
20028 }
20029
20030 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, n_div_16_small_kernel) {
20031 TEST_REQUIRES_ARM_NEON;
20032 for (uint32_t n = 32; n <= 48; n += 16) {
20033 for (size_t k = 1; k <= 40; k += 9) {
20034 GemmMicrokernelTester()
20035 .mr(3)
20036 .nr(16)
20037 .kr(4)
20038 .sr(1)
20039 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020040 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020041 .k(k)
20042 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080020043 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020044 }
20045 }
20046 }
20047
20048 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, strided_cm_subtile) {
20049 TEST_REQUIRES_ARM_NEON;
20050 for (size_t k = 1; k <= 40; k += 9) {
20051 for (uint32_t m = 1; m <= 3; m++) {
20052 for (uint32_t n = 1; n <= 16; n++) {
20053 GemmMicrokernelTester()
20054 .mr(3)
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)
Marat Dukhan50323b82022-01-11 00:12:01 -080020063 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020064 }
20065 }
20066 }
20067 }
20068
20069 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, a_offset) {
20070 TEST_REQUIRES_ARM_NEON;
20071 for (size_t k = 1; k <= 40; k += 9) {
20072 GemmMicrokernelTester()
20073 .mr(3)
20074 .nr(16)
20075 .kr(4)
20076 .sr(1)
20077 .m(3)
20078 .n(16)
20079 .k(k)
20080 .ks(3)
20081 .a_offset(127)
Marat Dukhan50323b82022-01-11 00:12:01 -080020082 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020083 }
20084 }
20085
20086 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, zero) {
20087 TEST_REQUIRES_ARM_NEON;
20088 for (uint32_t mz = 0; mz < 3; mz++) {
20089 for (size_t k = 1; k <= 40; k += 9) {
20090 GemmMicrokernelTester()
20091 .mr(3)
20092 .nr(16)
20093 .kr(4)
20094 .sr(1)
20095 .m(3)
20096 .n(16)
20097 .k(k)
20098 .ks(3)
20099 .a_offset(127)
20100 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080020101 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020102 }
20103 }
20104 }
20105
20106 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, qmin) {
20107 TEST_REQUIRES_ARM_NEON;
20108 GemmMicrokernelTester()
20109 .mr(3)
20110 .nr(16)
20111 .kr(4)
20112 .sr(1)
20113 .m(3)
20114 .n(16)
20115 .k(8)
20116 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080020117 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020118 }
20119
20120 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, qmax) {
20121 TEST_REQUIRES_ARM_NEON;
20122 GemmMicrokernelTester()
20123 .mr(3)
20124 .nr(16)
20125 .kr(4)
20126 .sr(1)
20127 .m(3)
20128 .n(16)
20129 .k(8)
20130 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080020131 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020132 }
20133
20134 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD2R, strided_cm) {
20135 TEST_REQUIRES_ARM_NEON;
20136 GemmMicrokernelTester()
20137 .mr(3)
20138 .nr(16)
20139 .kr(4)
20140 .sr(1)
20141 .m(3)
20142 .n(16)
20143 .k(8)
20144 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080020145 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020146 }
20147#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
20148
20149
20150#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barchard64ab1b72021-11-22 10:57:40 -080020151 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_eq_16) {
20152 TEST_REQUIRES_ARM_NEON;
20153 GemmMicrokernelTester()
20154 .mr(3)
20155 .nr(16)
20156 .kr(4)
20157 .sr(1)
20158 .m(3)
20159 .n(16)
20160 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -080020161 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020162 }
20163
20164 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, strided_cn) {
20165 TEST_REQUIRES_ARM_NEON;
20166 GemmMicrokernelTester()
20167 .mr(3)
20168 .nr(16)
20169 .kr(4)
20170 .sr(1)
20171 .m(3)
20172 .n(16)
20173 .k(16)
20174 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080020175 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020176 }
20177
20178 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_eq_16_subtile) {
20179 TEST_REQUIRES_ARM_NEON;
20180 for (uint32_t m = 1; m <= 3; m++) {
20181 for (uint32_t n = 1; n <= 16; n++) {
20182 GemmMicrokernelTester()
20183 .mr(3)
20184 .nr(16)
20185 .kr(4)
20186 .sr(1)
20187 .m(m)
20188 .n(n)
20189 .k(16)
20190 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020191 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020192 }
20193 }
20194 }
20195
20196 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_eq_16_subtile_m) {
20197 TEST_REQUIRES_ARM_NEON;
20198 for (uint32_t m = 1; m <= 3; m++) {
20199 GemmMicrokernelTester()
20200 .mr(3)
20201 .nr(16)
20202 .kr(4)
20203 .sr(1)
20204 .m(m)
20205 .n(16)
20206 .k(16)
20207 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020208 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020209 }
20210 }
20211
20212 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_eq_16_subtile_n) {
20213 TEST_REQUIRES_ARM_NEON;
20214 for (uint32_t n = 1; n <= 16; n++) {
20215 GemmMicrokernelTester()
20216 .mr(3)
20217 .nr(16)
20218 .kr(4)
20219 .sr(1)
20220 .m(3)
20221 .n(n)
20222 .k(16)
20223 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020224 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020225 }
20226 }
20227
20228 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_lt_16) {
20229 TEST_REQUIRES_ARM_NEON;
20230 for (size_t k = 1; k < 16; k++) {
20231 GemmMicrokernelTester()
20232 .mr(3)
20233 .nr(16)
20234 .kr(4)
20235 .sr(1)
20236 .m(3)
20237 .n(16)
20238 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020239 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020240 }
20241 }
20242
20243 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_lt_16_subtile) {
20244 TEST_REQUIRES_ARM_NEON;
20245 for (size_t k = 1; k < 16; k++) {
20246 for (uint32_t m = 1; m <= 3; m++) {
20247 for (uint32_t n = 1; n <= 16; n++) {
20248 GemmMicrokernelTester()
20249 .mr(3)
20250 .nr(16)
20251 .kr(4)
20252 .sr(1)
20253 .m(m)
20254 .n(n)
20255 .k(k)
20256 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020257 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020258 }
20259 }
20260 }
20261 }
20262
20263 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_gt_16) {
20264 TEST_REQUIRES_ARM_NEON;
20265 for (size_t k = 17; k < 32; k++) {
20266 GemmMicrokernelTester()
20267 .mr(3)
20268 .nr(16)
20269 .kr(4)
20270 .sr(1)
20271 .m(3)
20272 .n(16)
20273 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020274 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020275 }
20276 }
20277
20278 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_gt_16_subtile) {
20279 TEST_REQUIRES_ARM_NEON;
20280 for (size_t k = 17; k < 32; k++) {
20281 for (uint32_t m = 1; m <= 3; m++) {
20282 for (uint32_t n = 1; n <= 16; n++) {
20283 GemmMicrokernelTester()
20284 .mr(3)
20285 .nr(16)
20286 .kr(4)
20287 .sr(1)
20288 .m(m)
20289 .n(n)
20290 .k(k)
20291 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020292 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020293 }
20294 }
20295 }
20296 }
20297
20298 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_div_16) {
20299 TEST_REQUIRES_ARM_NEON;
20300 for (size_t k = 32; k <= 160; k += 16) {
20301 GemmMicrokernelTester()
20302 .mr(3)
20303 .nr(16)
20304 .kr(4)
20305 .sr(1)
20306 .m(3)
20307 .n(16)
20308 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020309 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020310 }
20311 }
20312
20313 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, k_div_16_subtile) {
20314 TEST_REQUIRES_ARM_NEON;
20315 for (size_t k = 32; k <= 160; k += 16) {
20316 for (uint32_t m = 1; m <= 3; m++) {
20317 for (uint32_t n = 1; n <= 16; n++) {
20318 GemmMicrokernelTester()
20319 .mr(3)
20320 .nr(16)
20321 .kr(4)
20322 .sr(1)
20323 .m(m)
20324 .n(n)
20325 .k(k)
20326 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020327 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020328 }
20329 }
20330 }
20331 }
20332
20333 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, n_gt_16) {
20334 TEST_REQUIRES_ARM_NEON;
20335 for (uint32_t n = 17; n < 32; n++) {
20336 for (size_t k = 1; k <= 80; k += 17) {
20337 GemmMicrokernelTester()
20338 .mr(3)
20339 .nr(16)
20340 .kr(4)
20341 .sr(1)
20342 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020343 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020344 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020345 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020346 }
20347 }
20348 }
20349
20350 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, n_gt_16_strided_cn) {
20351 TEST_REQUIRES_ARM_NEON;
20352 for (uint32_t n = 17; n < 32; n++) {
20353 for (size_t k = 1; k <= 80; k += 17) {
20354 GemmMicrokernelTester()
20355 .mr(3)
20356 .nr(16)
20357 .kr(4)
20358 .sr(1)
20359 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020360 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020361 .k(k)
20362 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080020363 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020364 }
20365 }
20366 }
20367
20368 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, n_gt_16_subtile) {
20369 TEST_REQUIRES_ARM_NEON;
20370 for (uint32_t n = 17; n < 32; n++) {
20371 for (size_t k = 1; k <= 80; k += 17) {
20372 for (uint32_t m = 1; m <= 3; m++) {
20373 GemmMicrokernelTester()
20374 .mr(3)
20375 .nr(16)
20376 .kr(4)
20377 .sr(1)
20378 .m(m)
20379 .n(n)
20380 .k(k)
20381 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020382 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020383 }
20384 }
20385 }
20386 }
20387
20388 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, n_div_16) {
20389 TEST_REQUIRES_ARM_NEON;
20390 for (uint32_t n = 32; n <= 48; n += 16) {
20391 for (size_t k = 1; k <= 80; k += 17) {
20392 GemmMicrokernelTester()
20393 .mr(3)
20394 .nr(16)
20395 .kr(4)
20396 .sr(1)
20397 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020398 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020399 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020400 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020401 }
20402 }
20403 }
20404
20405 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, n_div_16_strided_cn) {
20406 TEST_REQUIRES_ARM_NEON;
20407 for (uint32_t n = 32; n <= 48; n += 16) {
20408 for (size_t k = 1; k <= 80; k += 17) {
20409 GemmMicrokernelTester()
20410 .mr(3)
20411 .nr(16)
20412 .kr(4)
20413 .sr(1)
20414 .m(3)
20415 .n(n)
20416 .k(k)
20417 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080020418 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020419 }
20420 }
20421 }
20422
20423 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, n_div_16_subtile) {
20424 TEST_REQUIRES_ARM_NEON;
20425 for (uint32_t n = 32; n <= 48; n += 16) {
20426 for (size_t k = 1; k <= 80; k += 17) {
20427 for (uint32_t m = 1; m <= 3; m++) {
20428 GemmMicrokernelTester()
20429 .mr(3)
20430 .nr(16)
20431 .kr(4)
20432 .sr(1)
20433 .m(m)
20434 .n(n)
20435 .k(k)
20436 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020437 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020438 }
20439 }
20440 }
20441 }
20442
20443 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, small_kernel) {
20444 TEST_REQUIRES_ARM_NEON;
20445 for (size_t k = 1; k <= 80; k += 17) {
20446 GemmMicrokernelTester()
20447 .mr(3)
20448 .nr(16)
20449 .kr(4)
20450 .sr(1)
20451 .m(3)
20452 .n(16)
20453 .k(k)
20454 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080020455 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020456 }
20457 }
20458
20459 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, small_kernel_subtile) {
20460 TEST_REQUIRES_ARM_NEON;
20461 for (size_t k = 1; k <= 80; k += 17) {
20462 for (uint32_t m = 1; m <= 3; m++) {
20463 for (uint32_t n = 1; n <= 16; n++) {
20464 GemmMicrokernelTester()
20465 .mr(3)
20466 .nr(16)
20467 .kr(4)
20468 .sr(1)
20469 .m(m)
20470 .n(n)
20471 .k(k)
20472 .ks(3)
20473 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020474 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020475 }
20476 }
20477 }
20478 }
20479
20480 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, n_gt_16_small_kernel) {
20481 TEST_REQUIRES_ARM_NEON;
20482 for (uint32_t n = 17; n < 32; n++) {
20483 for (size_t k = 1; k <= 80; k += 17) {
20484 GemmMicrokernelTester()
20485 .mr(3)
20486 .nr(16)
20487 .kr(4)
20488 .sr(1)
20489 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020490 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020491 .k(k)
20492 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080020493 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020494 }
20495 }
20496 }
20497
20498 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, n_div_16_small_kernel) {
20499 TEST_REQUIRES_ARM_NEON;
20500 for (uint32_t n = 32; n <= 48; n += 16) {
20501 for (size_t k = 1; k <= 80; k += 17) {
20502 GemmMicrokernelTester()
20503 .mr(3)
20504 .nr(16)
20505 .kr(4)
20506 .sr(1)
20507 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020508 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020509 .k(k)
20510 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080020511 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020512 }
20513 }
20514 }
20515
20516 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, strided_cm_subtile) {
20517 TEST_REQUIRES_ARM_NEON;
20518 for (size_t k = 1; k <= 80; k += 17) {
20519 for (uint32_t m = 1; m <= 3; m++) {
20520 for (uint32_t n = 1; n <= 16; n++) {
20521 GemmMicrokernelTester()
20522 .mr(3)
20523 .nr(16)
20524 .kr(4)
20525 .sr(1)
20526 .m(m)
20527 .n(n)
20528 .k(k)
20529 .cm_stride(19)
20530 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020531 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020532 }
20533 }
20534 }
20535 }
20536
20537 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, a_offset) {
20538 TEST_REQUIRES_ARM_NEON;
20539 for (size_t k = 1; k <= 80; k += 17) {
20540 GemmMicrokernelTester()
20541 .mr(3)
20542 .nr(16)
20543 .kr(4)
20544 .sr(1)
20545 .m(3)
20546 .n(16)
20547 .k(k)
20548 .ks(3)
20549 .a_offset(251)
Marat Dukhan50323b82022-01-11 00:12:01 -080020550 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020551 }
20552 }
20553
20554 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, zero) {
20555 TEST_REQUIRES_ARM_NEON;
20556 for (uint32_t mz = 0; mz < 3; mz++) {
20557 for (size_t k = 1; k <= 80; k += 17) {
20558 GemmMicrokernelTester()
20559 .mr(3)
20560 .nr(16)
20561 .kr(4)
20562 .sr(1)
20563 .m(3)
20564 .n(16)
20565 .k(k)
20566 .ks(3)
20567 .a_offset(251)
20568 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080020569 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020570 }
20571 }
20572 }
20573
20574 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, qmin) {
20575 TEST_REQUIRES_ARM_NEON;
20576 GemmMicrokernelTester()
20577 .mr(3)
20578 .nr(16)
20579 .kr(4)
20580 .sr(1)
20581 .m(3)
20582 .n(16)
20583 .k(16)
20584 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080020585 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020586 }
20587
20588 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, qmax) {
20589 TEST_REQUIRES_ARM_NEON;
20590 GemmMicrokernelTester()
20591 .mr(3)
20592 .nr(16)
20593 .kr(4)
20594 .sr(1)
20595 .m(3)
20596 .n(16)
20597 .k(16)
20598 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080020599 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020600 }
20601
20602 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C4__NEON_MLAL_LD2R, strided_cm) {
20603 TEST_REQUIRES_ARM_NEON;
20604 GemmMicrokernelTester()
20605 .mr(3)
20606 .nr(16)
20607 .kr(4)
20608 .sr(1)
20609 .m(3)
20610 .n(16)
20611 .k(16)
20612 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080020613 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020614 }
20615#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
20616
20617
20618#if XNN_ARCH_ARM || XNN_ARCH_ARM64
20619 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_eq_16) {
20620 TEST_REQUIRES_ARM_NEON;
20621 GemmMicrokernelTester()
20622 .mr(4)
20623 .nr(16)
20624 .kr(4)
20625 .sr(1)
20626 .m(4)
20627 .n(16)
20628 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -080020629 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020630 }
20631
20632 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, strided_cn) {
20633 TEST_REQUIRES_ARM_NEON;
20634 GemmMicrokernelTester()
20635 .mr(4)
20636 .nr(16)
20637 .kr(4)
20638 .sr(1)
20639 .m(4)
20640 .n(16)
20641 .k(16)
20642 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080020643 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020644 }
20645
20646 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_eq_16_subtile) {
20647 TEST_REQUIRES_ARM_NEON;
20648 for (uint32_t m = 1; m <= 4; m++) {
20649 for (uint32_t n = 1; n <= 16; n++) {
20650 GemmMicrokernelTester()
20651 .mr(4)
20652 .nr(16)
20653 .kr(4)
20654 .sr(1)
20655 .m(m)
20656 .n(n)
20657 .k(16)
20658 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020659 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020660 }
20661 }
20662 }
20663
20664 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_eq_16_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(4)
20671 .sr(1)
20672 .m(m)
20673 .n(16)
20674 .k(16)
20675 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020676 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020677 }
20678 }
20679
20680 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_eq_16_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(4)
20687 .sr(1)
20688 .m(4)
20689 .n(n)
20690 .k(16)
20691 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020692 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020693 }
20694 }
20695
20696 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_lt_16) {
20697 TEST_REQUIRES_ARM_NEON;
20698 for (size_t k = 1; k < 16; k++) {
20699 GemmMicrokernelTester()
20700 .mr(4)
20701 .nr(16)
20702 .kr(4)
20703 .sr(1)
20704 .m(4)
20705 .n(16)
20706 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020707 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020708 }
20709 }
20710
20711 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_lt_16_subtile) {
20712 TEST_REQUIRES_ARM_NEON;
20713 for (size_t k = 1; k < 16; k++) {
20714 for (uint32_t m = 1; m <= 4; m++) {
20715 for (uint32_t n = 1; n <= 16; n++) {
20716 GemmMicrokernelTester()
20717 .mr(4)
20718 .nr(16)
20719 .kr(4)
20720 .sr(1)
20721 .m(m)
20722 .n(n)
20723 .k(k)
20724 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020725 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020726 }
20727 }
20728 }
20729 }
20730
20731 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_gt_16) {
20732 TEST_REQUIRES_ARM_NEON;
20733 for (size_t k = 17; k < 32; k++) {
20734 GemmMicrokernelTester()
20735 .mr(4)
20736 .nr(16)
20737 .kr(4)
20738 .sr(1)
20739 .m(4)
20740 .n(16)
20741 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020742 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020743 }
20744 }
20745
20746 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_gt_16_subtile) {
20747 TEST_REQUIRES_ARM_NEON;
20748 for (size_t k = 17; k < 32; k++) {
20749 for (uint32_t m = 1; m <= 4; m++) {
20750 for (uint32_t n = 1; n <= 16; n++) {
20751 GemmMicrokernelTester()
20752 .mr(4)
20753 .nr(16)
20754 .kr(4)
20755 .sr(1)
20756 .m(m)
20757 .n(n)
20758 .k(k)
20759 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020760 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020761 }
20762 }
20763 }
20764 }
20765
20766 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_div_16) {
20767 TEST_REQUIRES_ARM_NEON;
20768 for (size_t k = 32; k <= 160; k += 16) {
20769 GemmMicrokernelTester()
20770 .mr(4)
20771 .nr(16)
20772 .kr(4)
20773 .sr(1)
20774 .m(4)
20775 .n(16)
20776 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020777 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020778 }
20779 }
20780
20781 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, k_div_16_subtile) {
20782 TEST_REQUIRES_ARM_NEON;
20783 for (size_t k = 32; k <= 160; k += 16) {
20784 for (uint32_t m = 1; m <= 4; m++) {
20785 for (uint32_t n = 1; n <= 16; n++) {
20786 GemmMicrokernelTester()
20787 .mr(4)
20788 .nr(16)
20789 .kr(4)
20790 .sr(1)
20791 .m(m)
20792 .n(n)
20793 .k(k)
20794 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020795 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020796 }
20797 }
20798 }
20799 }
20800
20801 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, n_gt_16) {
20802 TEST_REQUIRES_ARM_NEON;
20803 for (uint32_t n = 17; n < 32; n++) {
20804 for (size_t k = 1; k <= 80; k += 17) {
20805 GemmMicrokernelTester()
20806 .mr(4)
20807 .nr(16)
20808 .kr(4)
20809 .sr(1)
20810 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020811 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020812 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020813 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020814 }
20815 }
20816 }
20817
20818 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, 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 <= 80; k += 17) {
20822 GemmMicrokernelTester()
20823 .mr(4)
20824 .nr(16)
20825 .kr(4)
20826 .sr(1)
20827 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020828 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020829 .k(k)
20830 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080020831 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020832 }
20833 }
20834 }
20835
20836 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, 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 <= 80; k += 17) {
20840 for (uint32_t m = 1; m <= 4; m++) {
20841 GemmMicrokernelTester()
20842 .mr(4)
20843 .nr(16)
20844 .kr(4)
20845 .sr(1)
20846 .m(m)
20847 .n(n)
20848 .k(k)
20849 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020850 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020851 }
20852 }
20853 }
20854 }
20855
20856 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, 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 <= 80; k += 17) {
20860 GemmMicrokernelTester()
20861 .mr(4)
20862 .nr(16)
20863 .kr(4)
20864 .sr(1)
20865 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020866 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020867 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080020868 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020869 }
20870 }
20871 }
20872
20873 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, 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 <= 80; k += 17) {
20877 GemmMicrokernelTester()
20878 .mr(4)
20879 .nr(16)
20880 .kr(4)
20881 .sr(1)
20882 .m(4)
20883 .n(n)
20884 .k(k)
20885 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080020886 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020887 }
20888 }
20889 }
20890
20891 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, 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 <= 80; k += 17) {
20895 for (uint32_t m = 1; m <= 4; m++) {
20896 GemmMicrokernelTester()
20897 .mr(4)
20898 .nr(16)
20899 .kr(4)
20900 .sr(1)
20901 .m(m)
20902 .n(n)
20903 .k(k)
20904 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020905 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020906 }
20907 }
20908 }
20909 }
20910
20911 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, small_kernel) {
20912 TEST_REQUIRES_ARM_NEON;
20913 for (size_t k = 1; k <= 80; k += 17) {
20914 GemmMicrokernelTester()
20915 .mr(4)
20916 .nr(16)
20917 .kr(4)
20918 .sr(1)
20919 .m(4)
20920 .n(16)
20921 .k(k)
20922 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080020923 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020924 }
20925 }
20926
20927 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, small_kernel_subtile) {
20928 TEST_REQUIRES_ARM_NEON;
20929 for (size_t k = 1; k <= 80; k += 17) {
20930 for (uint32_t m = 1; m <= 4; m++) {
20931 for (uint32_t n = 1; n <= 16; n++) {
20932 GemmMicrokernelTester()
20933 .mr(4)
20934 .nr(16)
20935 .kr(4)
20936 .sr(1)
20937 .m(m)
20938 .n(n)
20939 .k(k)
20940 .ks(3)
20941 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020942 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020943 }
20944 }
20945 }
20946 }
20947
20948 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, 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 <= 80; k += 17) {
20952 GemmMicrokernelTester()
20953 .mr(4)
20954 .nr(16)
20955 .kr(4)
20956 .sr(1)
20957 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020958 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020959 .k(k)
20960 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080020961 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020962 }
20963 }
20964 }
20965
20966 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, 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 <= 80; k += 17) {
20970 GemmMicrokernelTester()
20971 .mr(4)
20972 .nr(16)
20973 .kr(4)
20974 .sr(1)
20975 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080020976 .n(n)
Frank Barchard64ab1b72021-11-22 10:57:40 -080020977 .k(k)
20978 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080020979 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080020980 }
20981 }
20982 }
20983
20984 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, strided_cm_subtile) {
20985 TEST_REQUIRES_ARM_NEON;
20986 for (size_t k = 1; k <= 80; k += 17) {
20987 for (uint32_t m = 1; m <= 4; m++) {
20988 for (uint32_t n = 1; n <= 16; n++) {
20989 GemmMicrokernelTester()
20990 .mr(4)
20991 .nr(16)
20992 .kr(4)
20993 .sr(1)
20994 .m(m)
20995 .n(n)
20996 .k(k)
20997 .cm_stride(19)
20998 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080020999 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080021000 }
21001 }
21002 }
21003 }
21004
21005 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, a_offset) {
21006 TEST_REQUIRES_ARM_NEON;
21007 for (size_t k = 1; k <= 80; k += 17) {
21008 GemmMicrokernelTester()
21009 .mr(4)
21010 .nr(16)
21011 .kr(4)
21012 .sr(1)
21013 .m(4)
21014 .n(16)
21015 .k(k)
21016 .ks(3)
21017 .a_offset(331)
Marat Dukhan50323b82022-01-11 00:12:01 -080021018 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080021019 }
21020 }
21021
21022 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, zero) {
21023 TEST_REQUIRES_ARM_NEON;
21024 for (uint32_t mz = 0; mz < 4; mz++) {
21025 for (size_t k = 1; k <= 80; k += 17) {
21026 GemmMicrokernelTester()
21027 .mr(4)
21028 .nr(16)
21029 .kr(4)
21030 .sr(1)
21031 .m(4)
21032 .n(16)
21033 .k(k)
21034 .ks(3)
21035 .a_offset(331)
21036 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080021037 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080021038 }
21039 }
21040 }
21041
21042 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, qmin) {
21043 TEST_REQUIRES_ARM_NEON;
21044 GemmMicrokernelTester()
21045 .mr(4)
21046 .nr(16)
21047 .kr(4)
21048 .sr(1)
21049 .m(4)
21050 .n(16)
21051 .k(16)
21052 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080021053 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080021054 }
21055
21056 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, qmax) {
21057 TEST_REQUIRES_ARM_NEON;
21058 GemmMicrokernelTester()
21059 .mr(4)
21060 .nr(16)
21061 .kr(4)
21062 .sr(1)
21063 .m(4)
21064 .n(16)
21065 .k(16)
21066 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080021067 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080021068 }
21069
21070 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEON_MLAL_LD2R, strided_cm) {
21071 TEST_REQUIRES_ARM_NEON;
21072 GemmMicrokernelTester()
21073 .mr(4)
21074 .nr(16)
21075 .kr(4)
21076 .sr(1)
21077 .m(4)
21078 .n(16)
21079 .k(16)
21080 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080021081 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard64ab1b72021-11-22 10:57:40 -080021082 }
21083#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
21084
21085
21086#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021087 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021088 TEST_REQUIRES_ARM_NEON;
21089 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021090 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021091 .nr(8)
21092 .kr(2)
21093 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021094 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021095 .n(8)
21096 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021097 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021098 }
21099
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021100 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070021101 TEST_REQUIRES_ARM_NEON;
21102 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021103 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021104 .nr(8)
21105 .kr(2)
21106 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021107 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021108 .n(8)
21109 .k(8)
21110 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021111 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021112 }
21113
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021114 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021115 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021116 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021117 for (uint32_t n = 1; n <= 8; n++) {
21118 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021119 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021120 .nr(8)
21121 .kr(2)
21122 .sr(1)
21123 .m(m)
21124 .n(n)
21125 .k(8)
21126 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021127 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021128 }
21129 }
21130 }
21131
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021132 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_subtile_m) {
Frank Barchard1d412472021-10-25 17:27:21 -070021133 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021134 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021135 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021136 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021137 .nr(8)
21138 .kr(2)
21139 .sr(1)
21140 .m(m)
21141 .n(8)
21142 .k(8)
21143 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021144 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021145 }
21146 }
21147
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021148 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_subtile_n) {
Frank Barchard1d412472021-10-25 17:27:21 -070021149 TEST_REQUIRES_ARM_NEON;
21150 for (uint32_t n = 1; n <= 8; n++) {
21151 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021152 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021153 .nr(8)
21154 .kr(2)
21155 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021156 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021157 .n(n)
21158 .k(8)
21159 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021160 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021161 }
21162 }
21163
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021164 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_lt_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021165 TEST_REQUIRES_ARM_NEON;
21166 for (size_t k = 1; k < 8; k++) {
21167 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021168 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021169 .nr(8)
21170 .kr(2)
21171 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021172 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021173 .n(8)
21174 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021175 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021176 }
21177 }
21178
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021179 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_lt_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021180 TEST_REQUIRES_ARM_NEON;
21181 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021182 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021183 for (uint32_t n = 1; n <= 8; n++) {
21184 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021185 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021186 .nr(8)
21187 .kr(2)
21188 .sr(1)
21189 .m(m)
21190 .n(n)
21191 .k(k)
21192 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021193 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021194 }
21195 }
21196 }
21197 }
21198
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021199 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_gt_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021200 TEST_REQUIRES_ARM_NEON;
21201 for (size_t k = 9; k < 16; k++) {
21202 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021203 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021204 .nr(8)
21205 .kr(2)
21206 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021207 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021208 .n(8)
21209 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021210 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021211 }
21212 }
21213
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021214 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_gt_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021215 TEST_REQUIRES_ARM_NEON;
21216 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021217 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021218 for (uint32_t n = 1; n <= 8; n++) {
21219 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021220 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021221 .nr(8)
21222 .kr(2)
21223 .sr(1)
21224 .m(m)
21225 .n(n)
21226 .k(k)
21227 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021228 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021229 }
21230 }
21231 }
21232 }
21233
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021234 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_div_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021235 TEST_REQUIRES_ARM_NEON;
21236 for (size_t k = 16; k <= 80; k += 8) {
21237 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021238 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021239 .nr(8)
21240 .kr(2)
21241 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021242 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021243 .n(8)
21244 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021245 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021246 }
21247 }
21248
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021249 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_div_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021250 TEST_REQUIRES_ARM_NEON;
21251 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021252 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021253 for (uint32_t n = 1; n <= 8; n++) {
21254 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021255 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021256 .nr(8)
21257 .kr(2)
21258 .sr(1)
21259 .m(m)
21260 .n(n)
21261 .k(k)
21262 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021263 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021264 }
21265 }
21266 }
21267 }
21268
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021269 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021270 TEST_REQUIRES_ARM_NEON;
21271 for (uint32_t n = 9; n < 16; n++) {
21272 for (size_t k = 1; k <= 40; k += 9) {
21273 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021274 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021275 .nr(8)
21276 .kr(2)
21277 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021278 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021279 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021280 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021281 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021282 }
21283 }
21284 }
21285
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021286 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070021287 TEST_REQUIRES_ARM_NEON;
21288 for (uint32_t n = 9; n < 16; n++) {
21289 for (size_t k = 1; k <= 40; k += 9) {
21290 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021291 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021292 .nr(8)
21293 .kr(2)
21294 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021295 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021296 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021297 .k(k)
21298 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021299 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021300 }
21301 }
21302 }
21303
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021304 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021305 TEST_REQUIRES_ARM_NEON;
21306 for (uint32_t n = 9; n < 16; n++) {
21307 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021308 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021309 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021310 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021311 .nr(8)
21312 .kr(2)
21313 .sr(1)
21314 .m(m)
21315 .n(n)
21316 .k(k)
21317 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021318 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021319 }
21320 }
21321 }
21322 }
21323
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021324 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021325 TEST_REQUIRES_ARM_NEON;
21326 for (uint32_t n = 16; n <= 24; n += 8) {
21327 for (size_t k = 1; k <= 40; k += 9) {
21328 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021329 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021330 .nr(8)
21331 .kr(2)
21332 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021333 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021334 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021335 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021336 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021337 }
21338 }
21339 }
21340
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021341 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070021342 TEST_REQUIRES_ARM_NEON;
21343 for (uint32_t n = 16; n <= 24; n += 8) {
21344 for (size_t k = 1; k <= 40; k += 9) {
21345 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021346 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021347 .nr(8)
21348 .kr(2)
21349 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021350 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021351 .n(n)
21352 .k(k)
21353 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021354 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021355 }
21356 }
21357 }
21358
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021359 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021360 TEST_REQUIRES_ARM_NEON;
21361 for (uint32_t n = 16; n <= 24; n += 8) {
21362 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021363 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021364 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021365 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021366 .nr(8)
21367 .kr(2)
21368 .sr(1)
21369 .m(m)
21370 .n(n)
21371 .k(k)
21372 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021373 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021374 }
21375 }
21376 }
21377 }
21378
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021379 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070021380 TEST_REQUIRES_ARM_NEON;
21381 for (size_t k = 1; k <= 40; k += 9) {
21382 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021383 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021384 .nr(8)
21385 .kr(2)
21386 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021387 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021388 .n(8)
21389 .k(k)
21390 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021391 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021392 }
21393 }
21394
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021395 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, small_kernel_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021396 TEST_REQUIRES_ARM_NEON;
21397 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021398 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021399 for (uint32_t n = 1; n <= 8; n++) {
21400 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021401 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021402 .nr(8)
21403 .kr(2)
21404 .sr(1)
21405 .m(m)
21406 .n(n)
21407 .k(k)
21408 .ks(3)
21409 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021410 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021411 }
21412 }
21413 }
21414 }
21415
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021416 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070021417 TEST_REQUIRES_ARM_NEON;
21418 for (uint32_t n = 9; n < 16; n++) {
21419 for (size_t k = 1; k <= 40; k += 9) {
21420 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021421 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021422 .nr(8)
21423 .kr(2)
21424 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021425 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021426 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021427 .k(k)
21428 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021429 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021430 }
21431 }
21432 }
21433
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021434 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070021435 TEST_REQUIRES_ARM_NEON;
21436 for (uint32_t n = 16; n <= 24; n += 8) {
21437 for (size_t k = 1; k <= 40; k += 9) {
21438 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021439 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021440 .nr(8)
21441 .kr(2)
21442 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021443 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021444 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021445 .k(k)
21446 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021447 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021448 }
21449 }
21450 }
21451
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021452 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, strided_cm_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021453 TEST_REQUIRES_ARM_NEON;
21454 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021455 for (uint32_t m = 1; m <= 4; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021456 for (uint32_t n = 1; n <= 8; n++) {
21457 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021458 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021459 .nr(8)
21460 .kr(2)
21461 .sr(1)
21462 .m(m)
21463 .n(n)
21464 .k(k)
21465 .cm_stride(11)
21466 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021467 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021468 }
21469 }
21470 }
21471 }
21472
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021473 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, a_offset) {
Frank Barchard1d412472021-10-25 17:27:21 -070021474 TEST_REQUIRES_ARM_NEON;
21475 for (size_t k = 1; k <= 40; k += 9) {
21476 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021477 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021478 .nr(8)
21479 .kr(2)
21480 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021481 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021482 .n(8)
21483 .k(k)
21484 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021485 .a_offset(163)
21486 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021487 }
21488 }
21489
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021490 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, zero) {
Frank Barchard1d412472021-10-25 17:27:21 -070021491 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021492 for (uint32_t mz = 0; mz < 4; mz++) {
Frank Barchard1d412472021-10-25 17:27:21 -070021493 for (size_t k = 1; k <= 40; k += 9) {
21494 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021495 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021496 .nr(8)
21497 .kr(2)
21498 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021499 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021500 .n(8)
21501 .k(k)
21502 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021503 .a_offset(163)
Frank Barchard1d412472021-10-25 17:27:21 -070021504 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021505 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021506 }
21507 }
21508 }
21509
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021510 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, qmin) {
Frank Barchard1d412472021-10-25 17:27:21 -070021511 TEST_REQUIRES_ARM_NEON;
21512 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021513 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021514 .nr(8)
21515 .kr(2)
21516 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021517 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021518 .n(8)
21519 .k(8)
21520 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021521 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021522 }
21523
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021524 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, qmax) {
Frank Barchard1d412472021-10-25 17:27:21 -070021525 TEST_REQUIRES_ARM_NEON;
21526 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021527 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021528 .nr(8)
21529 .kr(2)
21530 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021531 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021532 .n(8)
21533 .k(8)
21534 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021536 }
21537
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021538 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, strided_cm) {
Frank Barchard1d412472021-10-25 17:27:21 -070021539 TEST_REQUIRES_ARM_NEON;
21540 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021541 .mr(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021542 .nr(8)
21543 .kr(2)
21544 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021545 .m(4)
Frank Barchard1d412472021-10-25 17:27:21 -070021546 .n(8)
21547 .k(8)
21548 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080021549 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021550 }
21551#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
21552
21553
21554#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barcharde22685a2021-11-12 11:36:58 -080021555 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_eq_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021556 TEST_REQUIRES_ARM_NEON;
21557 GemmMicrokernelTester()
21558 .mr(1)
21559 .nr(16)
21560 .kr(2)
21561 .sr(1)
21562 .m(1)
21563 .n(16)
21564 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080021565 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021566 }
21567
Frank Barcharde22685a2021-11-12 11:36:58 -080021568 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070021569 TEST_REQUIRES_ARM_NEON;
21570 GemmMicrokernelTester()
21571 .mr(1)
21572 .nr(16)
21573 .kr(2)
21574 .sr(1)
21575 .m(1)
21576 .n(16)
21577 .k(8)
21578 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080021579 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021580 }
21581
Frank Barcharde22685a2021-11-12 11:36:58 -080021582 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_eq_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021583 TEST_REQUIRES_ARM_NEON;
21584 for (uint32_t m = 1; m <= 1; m++) {
21585 for (uint32_t n = 1; n <= 16; n++) {
21586 GemmMicrokernelTester()
21587 .mr(1)
21588 .nr(16)
21589 .kr(2)
21590 .sr(1)
21591 .m(m)
21592 .n(n)
21593 .k(8)
21594 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021595 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021596 }
21597 }
21598 }
21599
Frank Barcharde22685a2021-11-12 11:36:58 -080021600 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_eq_8_subtile_m) {
Frank Barchard1d412472021-10-25 17:27:21 -070021601 TEST_REQUIRES_ARM_NEON;
21602 for (uint32_t m = 1; m <= 1; m++) {
21603 GemmMicrokernelTester()
21604 .mr(1)
21605 .nr(16)
21606 .kr(2)
21607 .sr(1)
21608 .m(m)
21609 .n(16)
21610 .k(8)
21611 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021612 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021613 }
21614 }
21615
Frank Barcharde22685a2021-11-12 11:36:58 -080021616 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_eq_8_subtile_n) {
Frank Barchard1d412472021-10-25 17:27:21 -070021617 TEST_REQUIRES_ARM_NEON;
21618 for (uint32_t n = 1; n <= 16; n++) {
21619 GemmMicrokernelTester()
21620 .mr(1)
21621 .nr(16)
21622 .kr(2)
21623 .sr(1)
21624 .m(1)
21625 .n(n)
21626 .k(8)
21627 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021628 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021629 }
21630 }
21631
Frank Barcharde22685a2021-11-12 11:36:58 -080021632 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_lt_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021633 TEST_REQUIRES_ARM_NEON;
21634 for (size_t k = 1; k < 8; k++) {
21635 GemmMicrokernelTester()
21636 .mr(1)
21637 .nr(16)
21638 .kr(2)
21639 .sr(1)
21640 .m(1)
21641 .n(16)
21642 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080021643 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021644 }
21645 }
21646
Frank Barcharde22685a2021-11-12 11:36:58 -080021647 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_lt_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021648 TEST_REQUIRES_ARM_NEON;
21649 for (size_t k = 1; k < 8; k++) {
21650 for (uint32_t m = 1; m <= 1; m++) {
21651 for (uint32_t n = 1; n <= 16; n++) {
21652 GemmMicrokernelTester()
21653 .mr(1)
21654 .nr(16)
21655 .kr(2)
21656 .sr(1)
21657 .m(m)
21658 .n(n)
21659 .k(k)
21660 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021661 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021662 }
21663 }
21664 }
21665 }
21666
Frank Barcharde22685a2021-11-12 11:36:58 -080021667 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_gt_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021668 TEST_REQUIRES_ARM_NEON;
21669 for (size_t k = 9; k < 16; k++) {
21670 GemmMicrokernelTester()
21671 .mr(1)
21672 .nr(16)
21673 .kr(2)
21674 .sr(1)
21675 .m(1)
21676 .n(16)
21677 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080021678 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021679 }
21680 }
21681
Frank Barcharde22685a2021-11-12 11:36:58 -080021682 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_gt_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021683 TEST_REQUIRES_ARM_NEON;
21684 for (size_t k = 9; k < 16; k++) {
21685 for (uint32_t m = 1; m <= 1; m++) {
21686 for (uint32_t n = 1; n <= 16; n++) {
21687 GemmMicrokernelTester()
21688 .mr(1)
21689 .nr(16)
21690 .kr(2)
21691 .sr(1)
21692 .m(m)
21693 .n(n)
21694 .k(k)
21695 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021696 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021697 }
21698 }
21699 }
21700 }
21701
Frank Barcharde22685a2021-11-12 11:36:58 -080021702 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_div_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070021703 TEST_REQUIRES_ARM_NEON;
21704 for (size_t k = 16; k <= 80; k += 8) {
21705 GemmMicrokernelTester()
21706 .mr(1)
21707 .nr(16)
21708 .kr(2)
21709 .sr(1)
21710 .m(1)
21711 .n(16)
21712 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080021713 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021714 }
21715 }
21716
Frank Barcharde22685a2021-11-12 11:36:58 -080021717 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, k_div_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021718 TEST_REQUIRES_ARM_NEON;
21719 for (size_t k = 16; k <= 80; k += 8) {
21720 for (uint32_t m = 1; m <= 1; m++) {
21721 for (uint32_t n = 1; n <= 16; n++) {
21722 GemmMicrokernelTester()
21723 .mr(1)
21724 .nr(16)
21725 .kr(2)
21726 .sr(1)
21727 .m(m)
21728 .n(n)
21729 .k(k)
21730 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021731 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021732 }
21733 }
21734 }
21735 }
21736
Frank Barcharde22685a2021-11-12 11:36:58 -080021737 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, n_gt_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070021738 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(1)
21743 .nr(16)
21744 .kr(2)
21745 .sr(1)
21746 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021747 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021748 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080021749 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021750 }
21751 }
21752 }
21753
Frank Barcharde22685a2021-11-12 11:36:58 -080021754 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, n_gt_16_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070021755 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(1)
21760 .nr(16)
21761 .kr(2)
21762 .sr(1)
21763 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021764 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021765 .k(k)
21766 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080021767 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021768 }
21769 }
21770 }
21771
Frank Barcharde22685a2021-11-12 11:36:58 -080021772 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, n_gt_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021773 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 <= 1; m++) {
21777 GemmMicrokernelTester()
21778 .mr(1)
21779 .nr(16)
21780 .kr(2)
21781 .sr(1)
21782 .m(m)
21783 .n(n)
21784 .k(k)
21785 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021786 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021787 }
21788 }
21789 }
21790 }
21791
Frank Barcharde22685a2021-11-12 11:36:58 -080021792 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, n_div_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070021793 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(1)
21798 .nr(16)
21799 .kr(2)
21800 .sr(1)
21801 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021802 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021803 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080021804 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021805 }
21806 }
21807 }
21808
Frank Barcharde22685a2021-11-12 11:36:58 -080021809 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, n_div_16_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070021810 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(1)
21815 .nr(16)
21816 .kr(2)
21817 .sr(1)
21818 .m(1)
21819 .n(n)
21820 .k(k)
21821 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080021822 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021823 }
21824 }
21825 }
21826
Frank Barcharde22685a2021-11-12 11:36:58 -080021827 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, n_div_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021828 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 <= 1; m++) {
21832 GemmMicrokernelTester()
21833 .mr(1)
21834 .nr(16)
21835 .kr(2)
21836 .sr(1)
21837 .m(m)
21838 .n(n)
21839 .k(k)
21840 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021841 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021842 }
21843 }
21844 }
21845 }
21846
Frank Barcharde22685a2021-11-12 11:36:58 -080021847 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070021848 TEST_REQUIRES_ARM_NEON;
21849 for (size_t k = 1; k <= 40; k += 9) {
21850 GemmMicrokernelTester()
21851 .mr(1)
21852 .nr(16)
21853 .kr(2)
21854 .sr(1)
21855 .m(1)
21856 .n(16)
21857 .k(k)
21858 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080021859 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021860 }
21861 }
21862
Frank Barcharde22685a2021-11-12 11:36:58 -080021863 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, small_kernel_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021864 TEST_REQUIRES_ARM_NEON;
21865 for (size_t k = 1; k <= 40; k += 9) {
21866 for (uint32_t m = 1; m <= 1; m++) {
21867 for (uint32_t n = 1; n <= 16; n++) {
21868 GemmMicrokernelTester()
21869 .mr(1)
21870 .nr(16)
21871 .kr(2)
21872 .sr(1)
21873 .m(m)
21874 .n(n)
21875 .k(k)
21876 .ks(3)
21877 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021878 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021879 }
21880 }
21881 }
21882 }
21883
Frank Barcharde22685a2021-11-12 11:36:58 -080021884 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, n_gt_16_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070021885 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(1)
21890 .nr(16)
21891 .kr(2)
21892 .sr(1)
21893 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021894 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021895 .k(k)
21896 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080021897 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021898 }
21899 }
21900 }
21901
Frank Barcharde22685a2021-11-12 11:36:58 -080021902 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, n_div_16_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070021903 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(1)
21908 .nr(16)
21909 .kr(2)
21910 .sr(1)
21911 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080021912 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070021913 .k(k)
21914 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080021915 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021916 }
21917 }
21918 }
21919
Frank Barcharde22685a2021-11-12 11:36:58 -080021920 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, strided_cm_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070021921 TEST_REQUIRES_ARM_NEON;
21922 for (size_t k = 1; k <= 40; k += 9) {
21923 for (uint32_t m = 1; m <= 1; m++) {
21924 for (uint32_t n = 1; n <= 16; n++) {
21925 GemmMicrokernelTester()
21926 .mr(1)
21927 .nr(16)
21928 .kr(2)
21929 .sr(1)
21930 .m(m)
21931 .n(n)
21932 .k(k)
21933 .cm_stride(19)
21934 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080021935 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021936 }
21937 }
21938 }
21939 }
21940
Frank Barcharde22685a2021-11-12 11:36:58 -080021941 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, a_offset) {
Frank Barchard1d412472021-10-25 17:27:21 -070021942 TEST_REQUIRES_ARM_NEON;
21943 for (size_t k = 1; k <= 40; k += 9) {
21944 GemmMicrokernelTester()
21945 .mr(1)
21946 .nr(16)
21947 .kr(2)
21948 .sr(1)
21949 .m(1)
21950 .n(16)
21951 .k(k)
21952 .ks(3)
21953 .a_offset(43)
Marat Dukhan50323b82022-01-11 00:12:01 -080021954 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021955 }
21956 }
21957
Frank Barcharde22685a2021-11-12 11:36:58 -080021958 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, zero) {
Frank Barchard1d412472021-10-25 17:27:21 -070021959 TEST_REQUIRES_ARM_NEON;
21960 for (uint32_t mz = 0; mz < 1; mz++) {
21961 for (size_t k = 1; k <= 40; k += 9) {
21962 GemmMicrokernelTester()
21963 .mr(1)
21964 .nr(16)
21965 .kr(2)
21966 .sr(1)
21967 .m(1)
21968 .n(16)
21969 .k(k)
21970 .ks(3)
21971 .a_offset(43)
21972 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080021973 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021974 }
21975 }
21976 }
21977
Frank Barcharde22685a2021-11-12 11:36:58 -080021978 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, qmin) {
Frank Barchard1d412472021-10-25 17:27:21 -070021979 TEST_REQUIRES_ARM_NEON;
21980 GemmMicrokernelTester()
21981 .mr(1)
21982 .nr(16)
21983 .kr(2)
21984 .sr(1)
21985 .m(1)
21986 .n(16)
21987 .k(8)
21988 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080021989 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070021990 }
21991
Frank Barcharde22685a2021-11-12 11:36:58 -080021992 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, qmax) {
Frank Barchard1d412472021-10-25 17:27:21 -070021993 TEST_REQUIRES_ARM_NEON;
21994 GemmMicrokernelTester()
21995 .mr(1)
21996 .nr(16)
21997 .kr(2)
21998 .sr(1)
21999 .m(1)
22000 .n(16)
22001 .k(8)
22002 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080022003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022004 }
22005
Frank Barcharde22685a2021-11-12 11:36:58 -080022006 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_DUP, strided_cm) {
Frank Barchard1d412472021-10-25 17:27:21 -070022007 TEST_REQUIRES_ARM_NEON;
22008 GemmMicrokernelTester()
22009 .mr(1)
22010 .nr(16)
22011 .kr(2)
22012 .sr(1)
22013 .m(1)
22014 .n(16)
22015 .k(8)
22016 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080022017 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022018 }
22019#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
22020
22021
22022#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022023 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_eq_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070022024 TEST_REQUIRES_ARM_NEON;
22025 GemmMicrokernelTester()
22026 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022027 .nr(16)
Frank Barchard1d412472021-10-25 17:27:21 -070022028 .kr(2)
22029 .sr(1)
22030 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022031 .n(16)
22032 .k(8)
22033 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022034 }
22035
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022036 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070022037 TEST_REQUIRES_ARM_NEON;
22038 GemmMicrokernelTester()
22039 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022040 .nr(16)
Frank Barchard1d412472021-10-25 17:27:21 -070022041 .kr(2)
22042 .sr(1)
22043 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022044 .n(16)
22045 .k(8)
22046 .cn_stride(19)
22047 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22048 }
22049
22050 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_eq_8_subtile) {
22051 TEST_REQUIRES_ARM_NEON;
22052 for (uint32_t m = 1; m <= 4; m++) {
22053 for (uint32_t n = 1; n <= 16; n++) {
22054 GemmMicrokernelTester()
22055 .mr(4)
22056 .nr(16)
22057 .kr(2)
22058 .sr(1)
22059 .m(m)
22060 .n(n)
22061 .k(8)
22062 .iterations(1)
22063 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22064 }
22065 }
22066 }
22067
22068 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_eq_8_subtile_m) {
22069 TEST_REQUIRES_ARM_NEON;
22070 for (uint32_t m = 1; m <= 4; m++) {
22071 GemmMicrokernelTester()
22072 .mr(4)
22073 .nr(16)
22074 .kr(2)
22075 .sr(1)
22076 .m(m)
22077 .n(16)
22078 .k(8)
22079 .iterations(1)
22080 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22081 }
22082 }
22083
22084 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_eq_8_subtile_n) {
22085 TEST_REQUIRES_ARM_NEON;
22086 for (uint32_t n = 1; n <= 16; n++) {
22087 GemmMicrokernelTester()
22088 .mr(4)
22089 .nr(16)
22090 .kr(2)
22091 .sr(1)
22092 .m(4)
22093 .n(n)
22094 .k(8)
22095 .iterations(1)
22096 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22097 }
22098 }
22099
22100 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_lt_8) {
22101 TEST_REQUIRES_ARM_NEON;
22102 for (size_t k = 1; k < 8; k++) {
22103 GemmMicrokernelTester()
22104 .mr(4)
22105 .nr(16)
22106 .kr(2)
22107 .sr(1)
22108 .m(4)
22109 .n(16)
22110 .k(k)
22111 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22112 }
22113 }
22114
22115 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_lt_8_subtile) {
22116 TEST_REQUIRES_ARM_NEON;
22117 for (size_t k = 1; k < 8; k++) {
22118 for (uint32_t m = 1; m <= 4; m++) {
22119 for (uint32_t n = 1; n <= 16; n++) {
22120 GemmMicrokernelTester()
22121 .mr(4)
22122 .nr(16)
22123 .kr(2)
22124 .sr(1)
22125 .m(m)
22126 .n(n)
22127 .k(k)
22128 .iterations(1)
22129 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22130 }
22131 }
22132 }
22133 }
22134
22135 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_gt_8) {
22136 TEST_REQUIRES_ARM_NEON;
22137 for (size_t k = 9; k < 16; k++) {
22138 GemmMicrokernelTester()
22139 .mr(4)
22140 .nr(16)
22141 .kr(2)
22142 .sr(1)
22143 .m(4)
22144 .n(16)
22145 .k(k)
22146 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22147 }
22148 }
22149
22150 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_gt_8_subtile) {
22151 TEST_REQUIRES_ARM_NEON;
22152 for (size_t k = 9; k < 16; k++) {
22153 for (uint32_t m = 1; m <= 4; m++) {
22154 for (uint32_t n = 1; n <= 16; n++) {
22155 GemmMicrokernelTester()
22156 .mr(4)
22157 .nr(16)
22158 .kr(2)
22159 .sr(1)
22160 .m(m)
22161 .n(n)
22162 .k(k)
22163 .iterations(1)
22164 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22165 }
22166 }
22167 }
22168 }
22169
22170 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_div_8) {
22171 TEST_REQUIRES_ARM_NEON;
22172 for (size_t k = 16; k <= 80; k += 8) {
22173 GemmMicrokernelTester()
22174 .mr(4)
22175 .nr(16)
22176 .kr(2)
22177 .sr(1)
22178 .m(4)
22179 .n(16)
22180 .k(k)
22181 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22182 }
22183 }
22184
22185 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_div_8_subtile) {
22186 TEST_REQUIRES_ARM_NEON;
22187 for (size_t k = 16; k <= 80; k += 8) {
22188 for (uint32_t m = 1; m <= 4; m++) {
22189 for (uint32_t n = 1; n <= 16; n++) {
22190 GemmMicrokernelTester()
22191 .mr(4)
22192 .nr(16)
22193 .kr(2)
22194 .sr(1)
22195 .m(m)
22196 .n(n)
22197 .k(k)
22198 .iterations(1)
22199 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22200 }
22201 }
22202 }
22203 }
22204
22205 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_gt_16) {
22206 TEST_REQUIRES_ARM_NEON;
22207 for (uint32_t n = 17; n < 32; n++) {
22208 for (size_t k = 1; k <= 40; k += 9) {
22209 GemmMicrokernelTester()
22210 .mr(4)
22211 .nr(16)
22212 .kr(2)
22213 .sr(1)
22214 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022215 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022216 .k(k)
22217 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22218 }
22219 }
22220 }
22221
22222 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_gt_16_strided_cn) {
22223 TEST_REQUIRES_ARM_NEON;
22224 for (uint32_t n = 17; n < 32; n++) {
22225 for (size_t k = 1; k <= 40; k += 9) {
22226 GemmMicrokernelTester()
22227 .mr(4)
22228 .nr(16)
22229 .kr(2)
22230 .sr(1)
22231 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022232 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022233 .k(k)
22234 .cn_stride(19)
22235 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22236 }
22237 }
22238 }
22239
22240 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_gt_16_subtile) {
22241 TEST_REQUIRES_ARM_NEON;
22242 for (uint32_t n = 17; n < 32; 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(16)
22248 .kr(2)
22249 .sr(1)
22250 .m(m)
22251 .n(n)
22252 .k(k)
22253 .iterations(1)
22254 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22255 }
22256 }
22257 }
22258 }
22259
22260 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_div_16) {
22261 TEST_REQUIRES_ARM_NEON;
22262 for (uint32_t n = 32; n <= 48; n += 16) {
22263 for (size_t k = 1; k <= 40; k += 9) {
22264 GemmMicrokernelTester()
22265 .mr(4)
22266 .nr(16)
22267 .kr(2)
22268 .sr(1)
22269 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022270 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022271 .k(k)
22272 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22273 }
22274 }
22275 }
22276
22277 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_div_16_strided_cn) {
22278 TEST_REQUIRES_ARM_NEON;
22279 for (uint32_t n = 32; n <= 48; n += 16) {
22280 for (size_t k = 1; k <= 40; k += 9) {
22281 GemmMicrokernelTester()
22282 .mr(4)
22283 .nr(16)
22284 .kr(2)
22285 .sr(1)
22286 .m(4)
22287 .n(n)
22288 .k(k)
22289 .cn_stride(19)
22290 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22291 }
22292 }
22293 }
22294
22295 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_div_16_subtile) {
22296 TEST_REQUIRES_ARM_NEON;
22297 for (uint32_t n = 32; n <= 48; n += 16) {
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(16)
22303 .kr(2)
22304 .sr(1)
22305 .m(m)
22306 .n(n)
22307 .k(k)
22308 .iterations(1)
22309 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22310 }
22311 }
22312 }
22313 }
22314
22315 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, small_kernel) {
22316 TEST_REQUIRES_ARM_NEON;
22317 for (size_t k = 1; k <= 40; k += 9) {
22318 GemmMicrokernelTester()
22319 .mr(4)
22320 .nr(16)
22321 .kr(2)
22322 .sr(1)
22323 .m(4)
22324 .n(16)
22325 .k(k)
22326 .ks(3)
22327 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22328 }
22329 }
22330
22331 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, small_kernel_subtile) {
22332 TEST_REQUIRES_ARM_NEON;
22333 for (size_t k = 1; k <= 40; k += 9) {
22334 for (uint32_t m = 1; m <= 4; m++) {
22335 for (uint32_t n = 1; n <= 16; n++) {
22336 GemmMicrokernelTester()
22337 .mr(4)
22338 .nr(16)
22339 .kr(2)
22340 .sr(1)
22341 .m(m)
22342 .n(n)
22343 .k(k)
22344 .ks(3)
22345 .iterations(1)
22346 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22347 }
22348 }
22349 }
22350 }
22351
22352 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_gt_16_small_kernel) {
22353 TEST_REQUIRES_ARM_NEON;
22354 for (uint32_t n = 17; n < 32; n++) {
22355 for (size_t k = 1; k <= 40; k += 9) {
22356 GemmMicrokernelTester()
22357 .mr(4)
22358 .nr(16)
22359 .kr(2)
22360 .sr(1)
22361 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022362 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022363 .k(k)
22364 .ks(3)
22365 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22366 }
22367 }
22368 }
22369
22370 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_div_16_small_kernel) {
22371 TEST_REQUIRES_ARM_NEON;
22372 for (uint32_t n = 32; n <= 48; n += 16) {
22373 for (size_t k = 1; k <= 40; k += 9) {
22374 GemmMicrokernelTester()
22375 .mr(4)
22376 .nr(16)
22377 .kr(2)
22378 .sr(1)
22379 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022380 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022381 .k(k)
22382 .ks(3)
22383 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22384 }
22385 }
22386 }
22387
22388 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, strided_cm_subtile) {
22389 TEST_REQUIRES_ARM_NEON;
22390 for (size_t k = 1; k <= 40; k += 9) {
22391 for (uint32_t m = 1; m <= 4; m++) {
22392 for (uint32_t n = 1; n <= 16; n++) {
22393 GemmMicrokernelTester()
22394 .mr(4)
22395 .nr(16)
22396 .kr(2)
22397 .sr(1)
22398 .m(m)
22399 .n(n)
22400 .k(k)
22401 .cm_stride(19)
22402 .iterations(1)
22403 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22404 }
22405 }
22406 }
22407 }
22408
22409 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, a_offset) {
22410 TEST_REQUIRES_ARM_NEON;
22411 for (size_t k = 1; k <= 40; k += 9) {
22412 GemmMicrokernelTester()
22413 .mr(4)
22414 .nr(16)
22415 .kr(2)
22416 .sr(1)
22417 .m(4)
22418 .n(16)
22419 .k(k)
22420 .ks(3)
22421 .a_offset(163)
22422 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22423 }
22424 }
22425
22426 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, zero) {
22427 TEST_REQUIRES_ARM_NEON;
22428 for (uint32_t mz = 0; mz < 4; mz++) {
22429 for (size_t k = 1; k <= 40; k += 9) {
22430 GemmMicrokernelTester()
22431 .mr(4)
22432 .nr(16)
22433 .kr(2)
22434 .sr(1)
22435 .m(4)
22436 .n(16)
22437 .k(k)
22438 .ks(3)
22439 .a_offset(163)
22440 .zero_index(mz)
22441 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22442 }
22443 }
22444 }
22445
22446 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, qmin) {
22447 TEST_REQUIRES_ARM_NEON;
22448 GemmMicrokernelTester()
22449 .mr(4)
22450 .nr(16)
22451 .kr(2)
22452 .sr(1)
22453 .m(4)
22454 .n(16)
22455 .k(8)
22456 .qmin(128)
22457 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22458 }
22459
22460 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, qmax) {
22461 TEST_REQUIRES_ARM_NEON;
22462 GemmMicrokernelTester()
22463 .mr(4)
22464 .nr(16)
22465 .kr(2)
22466 .sr(1)
22467 .m(4)
22468 .n(16)
22469 .k(8)
22470 .qmax(128)
22471 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22472 }
22473
22474 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, strided_cm) {
22475 TEST_REQUIRES_ARM_NEON;
22476 GemmMicrokernelTester()
22477 .mr(4)
22478 .nr(16)
22479 .kr(2)
22480 .sr(1)
22481 .m(4)
22482 .n(16)
22483 .k(8)
22484 .cm_stride(19)
22485 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22486 }
22487#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
22488
22489
22490#if XNN_ARCH_ARM || XNN_ARCH_ARM64
22491 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_eq_16) {
22492 TEST_REQUIRES_ARM_NEON;
22493 GemmMicrokernelTester()
22494 .mr(1)
22495 .nr(8)
22496 .kr(2)
22497 .sr(1)
22498 .m(1)
22499 .n(8)
22500 .k(16)
22501 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22502 }
22503
22504 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, strided_cn) {
22505 TEST_REQUIRES_ARM_NEON;
22506 GemmMicrokernelTester()
22507 .mr(1)
22508 .nr(8)
22509 .kr(2)
22510 .sr(1)
22511 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022512 .n(8)
22513 .k(16)
22514 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022515 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022516 }
22517
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022518 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_eq_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070022519 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022520 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022521 for (uint32_t n = 1; n <= 8; n++) {
22522 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022523 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022524 .nr(8)
22525 .kr(2)
22526 .sr(1)
22527 .m(m)
22528 .n(n)
22529 .k(16)
22530 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022531 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022532 }
22533 }
22534 }
22535
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022536 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_eq_16_subtile_m) {
Frank Barchard1d412472021-10-25 17:27:21 -070022537 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022538 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022539 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022540 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022541 .nr(8)
22542 .kr(2)
22543 .sr(1)
22544 .m(m)
22545 .n(8)
22546 .k(16)
22547 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022548 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022549 }
22550 }
22551
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022552 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_eq_16_subtile_n) {
Frank Barchard1d412472021-10-25 17:27:21 -070022553 TEST_REQUIRES_ARM_NEON;
22554 for (uint32_t n = 1; n <= 8; n++) {
22555 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022556 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022557 .nr(8)
22558 .kr(2)
22559 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022560 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022561 .n(n)
22562 .k(16)
22563 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022564 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022565 }
22566 }
22567
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022568 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_lt_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070022569 TEST_REQUIRES_ARM_NEON;
22570 for (size_t k = 1; k < 16; k++) {
22571 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022572 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022573 .nr(8)
22574 .kr(2)
22575 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022576 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022577 .n(8)
22578 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022579 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022580 }
22581 }
22582
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022583 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_lt_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070022584 TEST_REQUIRES_ARM_NEON;
22585 for (size_t k = 1; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022586 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022587 for (uint32_t n = 1; n <= 8; n++) {
22588 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022589 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022590 .nr(8)
22591 .kr(2)
22592 .sr(1)
22593 .m(m)
22594 .n(n)
22595 .k(k)
22596 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022597 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022598 }
22599 }
22600 }
22601 }
22602
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022603 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_gt_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070022604 TEST_REQUIRES_ARM_NEON;
22605 for (size_t k = 17; k < 32; k++) {
22606 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022607 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022608 .nr(8)
22609 .kr(2)
22610 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022611 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022612 .n(8)
22613 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022614 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022615 }
22616 }
22617
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022618 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_gt_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070022619 TEST_REQUIRES_ARM_NEON;
22620 for (size_t k = 17; k < 32; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022621 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022622 for (uint32_t n = 1; n <= 8; n++) {
22623 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022624 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022625 .nr(8)
22626 .kr(2)
22627 .sr(1)
22628 .m(m)
22629 .n(n)
22630 .k(k)
22631 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022632 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022633 }
22634 }
22635 }
22636 }
22637
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022638 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_div_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070022639 TEST_REQUIRES_ARM_NEON;
22640 for (size_t k = 32; k <= 160; k += 16) {
22641 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022642 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022643 .nr(8)
22644 .kr(2)
22645 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022646 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022647 .n(8)
22648 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022649 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022650 }
22651 }
22652
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022653 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_div_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070022654 TEST_REQUIRES_ARM_NEON;
22655 for (size_t k = 32; k <= 160; k += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022656 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022657 for (uint32_t n = 1; n <= 8; n++) {
22658 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022659 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022660 .nr(8)
22661 .kr(2)
22662 .sr(1)
22663 .m(m)
22664 .n(n)
22665 .k(k)
22666 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022667 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022668 }
22669 }
22670 }
22671 }
22672
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022673 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_gt_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070022674 TEST_REQUIRES_ARM_NEON;
22675 for (uint32_t n = 9; n < 16; n++) {
22676 for (size_t k = 1; k <= 80; k += 17) {
22677 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022678 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022679 .nr(8)
22680 .kr(2)
22681 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022682 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022683 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070022684 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022685 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022686 }
22687 }
22688 }
22689
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022690 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_gt_8_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070022691 TEST_REQUIRES_ARM_NEON;
22692 for (uint32_t n = 9; n < 16; n++) {
22693 for (size_t k = 1; k <= 80; k += 17) {
22694 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022695 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022696 .nr(8)
22697 .kr(2)
22698 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022699 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022700 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070022701 .k(k)
22702 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022703 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022704 }
22705 }
22706 }
22707
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022708 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_gt_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070022709 TEST_REQUIRES_ARM_NEON;
22710 for (uint32_t n = 9; n < 16; n++) {
22711 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022712 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022713 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022714 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022715 .nr(8)
22716 .kr(2)
22717 .sr(1)
22718 .m(m)
22719 .n(n)
22720 .k(k)
22721 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022722 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022723 }
22724 }
22725 }
22726 }
22727
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022728 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_div_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070022729 TEST_REQUIRES_ARM_NEON;
22730 for (uint32_t n = 16; n <= 24; n += 8) {
22731 for (size_t k = 1; k <= 80; k += 17) {
22732 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022733 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022734 .nr(8)
22735 .kr(2)
22736 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022737 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022738 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070022739 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022740 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022741 }
22742 }
22743 }
22744
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022745 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_div_8_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070022746 TEST_REQUIRES_ARM_NEON;
22747 for (uint32_t n = 16; n <= 24; n += 8) {
22748 for (size_t k = 1; k <= 80; k += 17) {
22749 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022750 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022751 .nr(8)
22752 .kr(2)
22753 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022754 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022755 .n(n)
22756 .k(k)
22757 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022758 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022759 }
22760 }
22761 }
22762
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022763 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_div_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070022764 TEST_REQUIRES_ARM_NEON;
22765 for (uint32_t n = 16; n <= 24; n += 8) {
22766 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022767 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022768 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022769 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022770 .nr(8)
22771 .kr(2)
22772 .sr(1)
22773 .m(m)
22774 .n(n)
22775 .k(k)
22776 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022777 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022778 }
22779 }
22780 }
22781 }
22782
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022783 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070022784 TEST_REQUIRES_ARM_NEON;
22785 for (size_t k = 1; k <= 80; k += 17) {
22786 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022787 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022788 .nr(8)
22789 .kr(2)
22790 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022791 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022792 .n(8)
22793 .k(k)
22794 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022795 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022796 }
22797 }
22798
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022799 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, small_kernel_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070022800 TEST_REQUIRES_ARM_NEON;
22801 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022802 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022803 for (uint32_t n = 1; n <= 8; n++) {
22804 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022805 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022806 .nr(8)
22807 .kr(2)
22808 .sr(1)
22809 .m(m)
22810 .n(n)
22811 .k(k)
22812 .ks(3)
22813 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022814 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022815 }
22816 }
22817 }
22818 }
22819
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022820 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_gt_8_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070022821 TEST_REQUIRES_ARM_NEON;
22822 for (uint32_t n = 9; n < 16; n++) {
22823 for (size_t k = 1; k <= 80; k += 17) {
22824 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022825 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022826 .nr(8)
22827 .kr(2)
22828 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022829 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022830 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070022831 .k(k)
22832 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022833 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022834 }
22835 }
22836 }
22837
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022838 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_div_8_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070022839 TEST_REQUIRES_ARM_NEON;
22840 for (uint32_t n = 16; n <= 24; n += 8) {
22841 for (size_t k = 1; k <= 80; k += 17) {
22842 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022843 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022844 .nr(8)
22845 .kr(2)
22846 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022847 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080022848 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070022849 .k(k)
22850 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022851 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022852 }
22853 }
22854 }
22855
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022856 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, strided_cm_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070022857 TEST_REQUIRES_ARM_NEON;
22858 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022859 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022860 for (uint32_t n = 1; n <= 8; n++) {
22861 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022862 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022863 .nr(8)
22864 .kr(2)
22865 .sr(1)
22866 .m(m)
22867 .n(n)
22868 .k(k)
22869 .cm_stride(11)
22870 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022871 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022872 }
22873 }
22874 }
22875 }
22876
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022877 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, a_offset) {
Frank Barchard1d412472021-10-25 17:27:21 -070022878 TEST_REQUIRES_ARM_NEON;
22879 for (size_t k = 1; k <= 80; k += 17) {
22880 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022881 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022882 .nr(8)
22883 .kr(2)
22884 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022885 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022886 .n(8)
22887 .k(k)
22888 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022889 .a_offset(83)
22890 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022891 }
22892 }
22893
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022894 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, zero) {
Frank Barchard1d412472021-10-25 17:27:21 -070022895 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022896 for (uint32_t mz = 0; mz < 1; mz++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022897 for (size_t k = 1; k <= 80; k += 17) {
22898 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022899 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022900 .nr(8)
22901 .kr(2)
22902 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022903 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022904 .n(8)
22905 .k(k)
22906 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022907 .a_offset(83)
Frank Barchard1d412472021-10-25 17:27:21 -070022908 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022909 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022910 }
22911 }
22912 }
22913
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022914 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, qmin) {
Frank Barchard1d412472021-10-25 17:27:21 -070022915 TEST_REQUIRES_ARM_NEON;
22916 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022917 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022918 .nr(8)
22919 .kr(2)
22920 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022921 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022922 .n(8)
22923 .k(16)
22924 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022925 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022926 }
22927
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022928 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, qmax) {
Frank Barchard1d412472021-10-25 17:27:21 -070022929 TEST_REQUIRES_ARM_NEON;
22930 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022931 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022932 .nr(8)
22933 .kr(2)
22934 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022935 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022936 .n(8)
22937 .k(16)
22938 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022939 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022940 }
22941
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022942 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, strided_cm) {
Frank Barchard1d412472021-10-25 17:27:21 -070022943 TEST_REQUIRES_ARM_NEON;
22944 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022945 .mr(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022946 .nr(8)
22947 .kr(2)
22948 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022949 .m(1)
Frank Barchard1d412472021-10-25 17:27:21 -070022950 .n(8)
22951 .k(16)
22952 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022953 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022954 }
22955#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
22956
22957
22958#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022959 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_eq_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070022960 TEST_REQUIRES_ARM_NEON;
22961 GemmMicrokernelTester()
22962 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022963 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070022964 .kr(2)
22965 .sr(1)
22966 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022967 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070022968 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022969 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022970 }
22971
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022972 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070022973 TEST_REQUIRES_ARM_NEON;
22974 GemmMicrokernelTester()
22975 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022976 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070022977 .kr(2)
22978 .sr(1)
22979 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022980 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070022981 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022982 .cn_stride(11)
22983 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070022984 }
22985
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022986 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_eq_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070022987 TEST_REQUIRES_ARM_NEON;
22988 for (uint32_t m = 1; m <= 2; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022989 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070022990 GemmMicrokernelTester()
22991 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022992 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070022993 .kr(2)
22994 .sr(1)
22995 .m(m)
22996 .n(n)
22997 .k(16)
22998 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080022999 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023000 }
23001 }
23002 }
23003
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023004 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_eq_16_subtile_m) {
Frank Barchard1d412472021-10-25 17:27:21 -070023005 TEST_REQUIRES_ARM_NEON;
23006 for (uint32_t m = 1; m <= 2; m++) {
23007 GemmMicrokernelTester()
23008 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023009 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023010 .kr(2)
23011 .sr(1)
23012 .m(m)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023013 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023014 .k(16)
23015 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023016 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023017 }
23018 }
23019
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023020 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_eq_16_subtile_n) {
Frank Barchard1d412472021-10-25 17:27:21 -070023021 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023022 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023023 GemmMicrokernelTester()
23024 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023025 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023026 .kr(2)
23027 .sr(1)
23028 .m(2)
23029 .n(n)
23030 .k(16)
23031 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023032 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023033 }
23034 }
23035
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023036 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_lt_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070023037 TEST_REQUIRES_ARM_NEON;
23038 for (size_t k = 1; k < 16; k++) {
23039 GemmMicrokernelTester()
23040 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023041 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023042 .kr(2)
23043 .sr(1)
23044 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023045 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023046 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023047 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023048 }
23049 }
23050
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023051 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_lt_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023052 TEST_REQUIRES_ARM_NEON;
23053 for (size_t k = 1; k < 16; k++) {
23054 for (uint32_t m = 1; m <= 2; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023055 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023056 GemmMicrokernelTester()
23057 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023058 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023059 .kr(2)
23060 .sr(1)
23061 .m(m)
23062 .n(n)
23063 .k(k)
23064 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023065 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023066 }
23067 }
23068 }
23069 }
23070
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023071 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_gt_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070023072 TEST_REQUIRES_ARM_NEON;
23073 for (size_t k = 17; k < 32; k++) {
23074 GemmMicrokernelTester()
23075 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023076 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023077 .kr(2)
23078 .sr(1)
23079 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023080 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023081 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023082 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023083 }
23084 }
23085
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023086 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_gt_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023087 TEST_REQUIRES_ARM_NEON;
23088 for (size_t k = 17; k < 32; k++) {
23089 for (uint32_t m = 1; m <= 2; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023090 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023091 GemmMicrokernelTester()
23092 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023093 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023094 .kr(2)
23095 .sr(1)
23096 .m(m)
23097 .n(n)
23098 .k(k)
23099 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023100 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023101 }
23102 }
23103 }
23104 }
23105
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023106 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_div_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070023107 TEST_REQUIRES_ARM_NEON;
23108 for (size_t k = 32; k <= 160; k += 16) {
23109 GemmMicrokernelTester()
23110 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023111 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023112 .kr(2)
23113 .sr(1)
23114 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023115 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023116 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023117 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023118 }
23119 }
23120
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023121 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_div_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023122 TEST_REQUIRES_ARM_NEON;
23123 for (size_t k = 32; k <= 160; k += 16) {
23124 for (uint32_t m = 1; m <= 2; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023125 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023126 GemmMicrokernelTester()
23127 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023128 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023129 .kr(2)
23130 .sr(1)
23131 .m(m)
23132 .n(n)
23133 .k(k)
23134 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023135 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023136 }
23137 }
23138 }
23139 }
23140
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023141 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_gt_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023142 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023143 for (uint32_t n = 9; n < 16; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023144 for (size_t k = 1; k <= 80; k += 17) {
23145 GemmMicrokernelTester()
23146 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023147 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023148 .kr(2)
23149 .sr(1)
23150 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023151 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023152 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023153 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023154 }
23155 }
23156 }
23157
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023158 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_gt_8_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070023159 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023160 for (uint32_t n = 9; n < 16; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023161 for (size_t k = 1; k <= 80; k += 17) {
23162 GemmMicrokernelTester()
23163 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023164 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023165 .kr(2)
23166 .sr(1)
23167 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023168 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023169 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023170 .cn_stride(11)
23171 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023172 }
23173 }
23174 }
23175
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023176 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_gt_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023177 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023178 for (uint32_t n = 9; n < 16; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023179 for (size_t k = 1; k <= 80; k += 17) {
23180 for (uint32_t m = 1; m <= 2; m++) {
23181 GemmMicrokernelTester()
23182 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023183 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023184 .kr(2)
23185 .sr(1)
23186 .m(m)
23187 .n(n)
23188 .k(k)
23189 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023190 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023191 }
23192 }
23193 }
23194 }
23195
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023196 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_div_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023197 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023198 for (uint32_t n = 16; n <= 24; n += 8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023199 for (size_t k = 1; k <= 80; k += 17) {
23200 GemmMicrokernelTester()
23201 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023202 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023203 .kr(2)
23204 .sr(1)
23205 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023206 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023207 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023208 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023209 }
23210 }
23211 }
23212
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023213 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_div_8_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070023214 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023215 for (uint32_t n = 16; n <= 24; n += 8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023216 for (size_t k = 1; k <= 80; k += 17) {
23217 GemmMicrokernelTester()
23218 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023219 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023220 .kr(2)
23221 .sr(1)
23222 .m(2)
23223 .n(n)
23224 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023225 .cn_stride(11)
23226 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023227 }
23228 }
23229 }
23230
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023231 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_div_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023232 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023233 for (uint32_t n = 16; n <= 24; n += 8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023234 for (size_t k = 1; k <= 80; k += 17) {
23235 for (uint32_t m = 1; m <= 2; m++) {
23236 GemmMicrokernelTester()
23237 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023238 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023239 .kr(2)
23240 .sr(1)
23241 .m(m)
23242 .n(n)
23243 .k(k)
23244 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023245 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023246 }
23247 }
23248 }
23249 }
23250
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023251 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070023252 TEST_REQUIRES_ARM_NEON;
23253 for (size_t k = 1; k <= 80; k += 17) {
23254 GemmMicrokernelTester()
23255 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023256 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023257 .kr(2)
23258 .sr(1)
23259 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023260 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023261 .k(k)
23262 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023263 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023264 }
23265 }
23266
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023267 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, small_kernel_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023268 TEST_REQUIRES_ARM_NEON;
23269 for (size_t k = 1; k <= 80; k += 17) {
23270 for (uint32_t m = 1; m <= 2; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023271 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023272 GemmMicrokernelTester()
23273 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023274 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023275 .kr(2)
23276 .sr(1)
23277 .m(m)
23278 .n(n)
23279 .k(k)
23280 .ks(3)
23281 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023282 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023283 }
23284 }
23285 }
23286 }
23287
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023288 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_gt_8_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070023289 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023290 for (uint32_t n = 9; n < 16; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023291 for (size_t k = 1; k <= 80; k += 17) {
23292 GemmMicrokernelTester()
23293 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023294 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023295 .kr(2)
23296 .sr(1)
23297 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023298 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023299 .k(k)
23300 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023301 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023302 }
23303 }
23304 }
23305
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023306 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_div_8_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070023307 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023308 for (uint32_t n = 16; n <= 24; n += 8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023309 for (size_t k = 1; k <= 80; k += 17) {
23310 GemmMicrokernelTester()
23311 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023312 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023313 .kr(2)
23314 .sr(1)
23315 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023316 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023317 .k(k)
23318 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023319 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023320 }
23321 }
23322 }
23323
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023324 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, strided_cm_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023325 TEST_REQUIRES_ARM_NEON;
23326 for (size_t k = 1; k <= 80; k += 17) {
23327 for (uint32_t m = 1; m <= 2; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023328 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023329 GemmMicrokernelTester()
23330 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023331 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023332 .kr(2)
23333 .sr(1)
23334 .m(m)
23335 .n(n)
23336 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023337 .cm_stride(11)
Frank Barchard1d412472021-10-25 17:27:21 -070023338 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023339 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023340 }
23341 }
23342 }
23343 }
23344
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023345 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, a_offset) {
Frank Barchard1d412472021-10-25 17:27:21 -070023346 TEST_REQUIRES_ARM_NEON;
23347 for (size_t k = 1; k <= 80; k += 17) {
23348 GemmMicrokernelTester()
23349 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023350 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023351 .kr(2)
23352 .sr(1)
23353 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023354 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023355 .k(k)
23356 .ks(3)
23357 .a_offset(163)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023358 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023359 }
23360 }
23361
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023362 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, zero) {
Frank Barchard1d412472021-10-25 17:27:21 -070023363 TEST_REQUIRES_ARM_NEON;
23364 for (uint32_t mz = 0; mz < 2; mz++) {
23365 for (size_t k = 1; k <= 80; k += 17) {
23366 GemmMicrokernelTester()
23367 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023368 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023369 .kr(2)
23370 .sr(1)
23371 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023372 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023373 .k(k)
23374 .ks(3)
23375 .a_offset(163)
23376 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023377 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023378 }
23379 }
23380 }
23381
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023382 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, qmin) {
Frank Barchard1d412472021-10-25 17:27:21 -070023383 TEST_REQUIRES_ARM_NEON;
23384 GemmMicrokernelTester()
23385 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023386 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023387 .kr(2)
23388 .sr(1)
23389 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023390 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023391 .k(16)
23392 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023393 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023394 }
23395
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023396 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, qmax) {
Frank Barchard1d412472021-10-25 17:27:21 -070023397 TEST_REQUIRES_ARM_NEON;
23398 GemmMicrokernelTester()
23399 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023400 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023401 .kr(2)
23402 .sr(1)
23403 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023404 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023405 .k(16)
23406 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023408 }
23409
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023410 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, strided_cm) {
Frank Barchard1d412472021-10-25 17:27:21 -070023411 TEST_REQUIRES_ARM_NEON;
23412 GemmMicrokernelTester()
23413 .mr(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023414 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023415 .kr(2)
23416 .sr(1)
23417 .m(2)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023418 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023419 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023420 .cm_stride(11)
23421 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023422 }
23423#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
23424
23425
23426#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023427 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070023428 TEST_REQUIRES_ARM_NEON;
23429 GemmMicrokernelTester()
23430 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023431 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023432 .kr(2)
23433 .sr(1)
23434 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023435 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023436 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023437 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023438 }
23439
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023440 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070023441 TEST_REQUIRES_ARM_NEON;
23442 GemmMicrokernelTester()
23443 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023444 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023445 .kr(2)
23446 .sr(1)
23447 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023448 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023449 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023450 .cn_stride(11)
23451 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023452 }
23453
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023454 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023455 TEST_REQUIRES_ARM_NEON;
23456 for (uint32_t m = 1; m <= 3; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023457 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023458 GemmMicrokernelTester()
23459 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023460 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023461 .kr(2)
23462 .sr(1)
23463 .m(m)
23464 .n(n)
23465 .k(16)
23466 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023467 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023468 }
23469 }
23470 }
23471
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023472 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_subtile_m) {
Frank Barchard1d412472021-10-25 17:27:21 -070023473 TEST_REQUIRES_ARM_NEON;
23474 for (uint32_t m = 1; m <= 3; m++) {
23475 GemmMicrokernelTester()
23476 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023477 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023478 .kr(2)
23479 .sr(1)
23480 .m(m)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023481 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023482 .k(16)
23483 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023484 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023485 }
23486 }
23487
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023488 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_subtile_n) {
Frank Barchard1d412472021-10-25 17:27:21 -070023489 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023490 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023491 GemmMicrokernelTester()
23492 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023493 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023494 .kr(2)
23495 .sr(1)
23496 .m(3)
23497 .n(n)
23498 .k(16)
23499 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023500 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023501 }
23502 }
23503
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023504 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_lt_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070023505 TEST_REQUIRES_ARM_NEON;
23506 for (size_t k = 1; k < 16; k++) {
23507 GemmMicrokernelTester()
23508 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023509 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023510 .kr(2)
23511 .sr(1)
23512 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023513 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023514 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023515 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023516 }
23517 }
23518
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023519 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_lt_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023520 TEST_REQUIRES_ARM_NEON;
23521 for (size_t k = 1; k < 16; k++) {
23522 for (uint32_t m = 1; m <= 3; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023523 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023524 GemmMicrokernelTester()
23525 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023526 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023527 .kr(2)
23528 .sr(1)
23529 .m(m)
23530 .n(n)
23531 .k(k)
23532 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023533 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023534 }
23535 }
23536 }
23537 }
23538
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023539 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_gt_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070023540 TEST_REQUIRES_ARM_NEON;
23541 for (size_t k = 17; k < 32; k++) {
23542 GemmMicrokernelTester()
23543 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023544 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023545 .kr(2)
23546 .sr(1)
23547 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023548 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023549 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023550 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023551 }
23552 }
23553
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023554 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_gt_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023555 TEST_REQUIRES_ARM_NEON;
23556 for (size_t k = 17; k < 32; k++) {
23557 for (uint32_t m = 1; m <= 3; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023558 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023559 GemmMicrokernelTester()
23560 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023561 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023562 .kr(2)
23563 .sr(1)
23564 .m(m)
23565 .n(n)
23566 .k(k)
23567 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023568 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023569 }
23570 }
23571 }
23572 }
23573
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023574 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_div_16) {
Frank Barchard1d412472021-10-25 17:27:21 -070023575 TEST_REQUIRES_ARM_NEON;
23576 for (size_t k = 32; k <= 160; k += 16) {
23577 GemmMicrokernelTester()
23578 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023579 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023580 .kr(2)
23581 .sr(1)
23582 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023583 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023584 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023585 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023586 }
23587 }
23588
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023589 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_div_16_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023590 TEST_REQUIRES_ARM_NEON;
23591 for (size_t k = 32; k <= 160; k += 16) {
23592 for (uint32_t m = 1; m <= 3; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023593 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023594 GemmMicrokernelTester()
23595 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023596 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023597 .kr(2)
23598 .sr(1)
23599 .m(m)
23600 .n(n)
23601 .k(k)
23602 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023603 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023604 }
23605 }
23606 }
23607 }
23608
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023609 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023610 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023611 for (uint32_t n = 9; n < 16; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023612 for (size_t k = 1; k <= 80; k += 17) {
23613 GemmMicrokernelTester()
23614 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023615 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023616 .kr(2)
23617 .sr(1)
23618 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023619 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023620 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023621 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023622 }
23623 }
23624 }
23625
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023626 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070023627 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023628 for (uint32_t n = 9; n < 16; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023629 for (size_t k = 1; k <= 80; k += 17) {
23630 GemmMicrokernelTester()
23631 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023632 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023633 .kr(2)
23634 .sr(1)
23635 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023636 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023637 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023638 .cn_stride(11)
23639 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023640 }
23641 }
23642 }
23643
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023644 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023645 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023646 for (uint32_t n = 9; n < 16; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023647 for (size_t k = 1; k <= 80; k += 17) {
23648 for (uint32_t m = 1; m <= 3; m++) {
23649 GemmMicrokernelTester()
23650 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023651 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023652 .kr(2)
23653 .sr(1)
23654 .m(m)
23655 .n(n)
23656 .k(k)
23657 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023658 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023659 }
23660 }
23661 }
23662 }
23663
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023664 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023665 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023666 for (uint32_t n = 16; n <= 24; n += 8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023667 for (size_t k = 1; k <= 80; k += 17) {
23668 GemmMicrokernelTester()
23669 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023670 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023671 .kr(2)
23672 .sr(1)
23673 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023674 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023675 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023676 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023677 }
23678 }
23679 }
23680
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023681 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8_strided_cn) {
Frank Barchard1d412472021-10-25 17:27:21 -070023682 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023683 for (uint32_t n = 16; n <= 24; n += 8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023684 for (size_t k = 1; k <= 80; k += 17) {
23685 GemmMicrokernelTester()
23686 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023687 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023688 .kr(2)
23689 .sr(1)
23690 .m(3)
23691 .n(n)
23692 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023693 .cn_stride(11)
23694 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023695 }
23696 }
23697 }
23698
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023699 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023700 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023701 for (uint32_t n = 16; n <= 24; n += 8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023702 for (size_t k = 1; k <= 80; k += 17) {
23703 for (uint32_t m = 1; m <= 3; m++) {
23704 GemmMicrokernelTester()
23705 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023706 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023707 .kr(2)
23708 .sr(1)
23709 .m(m)
23710 .n(n)
23711 .k(k)
23712 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023713 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023714 }
23715 }
23716 }
23717 }
23718
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023719 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070023720 TEST_REQUIRES_ARM_NEON;
23721 for (size_t k = 1; k <= 80; k += 17) {
23722 GemmMicrokernelTester()
23723 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023724 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023725 .kr(2)
23726 .sr(1)
23727 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023728 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023729 .k(k)
23730 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023731 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023732 }
23733 }
23734
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023735 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, small_kernel_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023736 TEST_REQUIRES_ARM_NEON;
23737 for (size_t k = 1; k <= 80; k += 17) {
23738 for (uint32_t m = 1; m <= 3; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023739 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023740 GemmMicrokernelTester()
23741 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023742 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023743 .kr(2)
23744 .sr(1)
23745 .m(m)
23746 .n(n)
23747 .k(k)
23748 .ks(3)
23749 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023750 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023751 }
23752 }
23753 }
23754 }
23755
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023756 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070023757 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023758 for (uint32_t n = 9; n < 16; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023759 for (size_t k = 1; k <= 80; k += 17) {
23760 GemmMicrokernelTester()
23761 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023762 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023763 .kr(2)
23764 .sr(1)
23765 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023766 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023767 .k(k)
23768 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023769 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023770 }
23771 }
23772 }
23773
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023774 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8_small_kernel) {
Frank Barchard1d412472021-10-25 17:27:21 -070023775 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023776 for (uint32_t n = 16; n <= 24; n += 8) {
Frank Barchard1d412472021-10-25 17:27:21 -070023777 for (size_t k = 1; k <= 80; k += 17) {
23778 GemmMicrokernelTester()
23779 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023780 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023781 .kr(2)
23782 .sr(1)
23783 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080023784 .n(n)
Frank Barchard1d412472021-10-25 17:27:21 -070023785 .k(k)
23786 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023787 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023788 }
23789 }
23790 }
23791
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023792 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, strided_cm_subtile) {
Frank Barchard1d412472021-10-25 17:27:21 -070023793 TEST_REQUIRES_ARM_NEON;
23794 for (size_t k = 1; k <= 80; k += 17) {
23795 for (uint32_t m = 1; m <= 3; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023796 for (uint32_t n = 1; n <= 8; n++) {
Frank Barchard1d412472021-10-25 17:27:21 -070023797 GemmMicrokernelTester()
23798 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023799 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023800 .kr(2)
23801 .sr(1)
23802 .m(m)
23803 .n(n)
23804 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023805 .cm_stride(11)
Frank Barchard1d412472021-10-25 17:27:21 -070023806 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023807 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023808 }
23809 }
23810 }
23811 }
23812
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023813 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, a_offset) {
Frank Barchard1d412472021-10-25 17:27:21 -070023814 TEST_REQUIRES_ARM_NEON;
23815 for (size_t k = 1; k <= 80; k += 17) {
23816 GemmMicrokernelTester()
23817 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023818 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023819 .kr(2)
23820 .sr(1)
23821 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023822 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023823 .k(k)
23824 .ks(3)
23825 .a_offset(251)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023826 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023827 }
23828 }
23829
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023830 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, zero) {
Frank Barchard1d412472021-10-25 17:27:21 -070023831 TEST_REQUIRES_ARM_NEON;
23832 for (uint32_t mz = 0; mz < 3; mz++) {
23833 for (size_t k = 1; k <= 80; k += 17) {
23834 GemmMicrokernelTester()
23835 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023836 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023837 .kr(2)
23838 .sr(1)
23839 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023840 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023841 .k(k)
23842 .ks(3)
23843 .a_offset(251)
23844 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023845 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023846 }
23847 }
23848 }
23849
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023850 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, qmin) {
Frank Barchard1d412472021-10-25 17:27:21 -070023851 TEST_REQUIRES_ARM_NEON;
23852 GemmMicrokernelTester()
23853 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023854 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023855 .kr(2)
23856 .sr(1)
23857 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023858 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023859 .k(16)
23860 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023861 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023862 }
23863
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023864 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, qmax) {
Frank Barchard1d412472021-10-25 17:27:21 -070023865 TEST_REQUIRES_ARM_NEON;
23866 GemmMicrokernelTester()
23867 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023868 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023869 .kr(2)
23870 .sr(1)
23871 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023872 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023873 .k(16)
23874 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023876 }
23877
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023878 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, strided_cm) {
Frank Barchard1d412472021-10-25 17:27:21 -070023879 TEST_REQUIRES_ARM_NEON;
23880 GemmMicrokernelTester()
23881 .mr(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023882 .nr(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023883 .kr(2)
23884 .sr(1)
23885 .m(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023886 .n(8)
Frank Barchard1d412472021-10-25 17:27:21 -070023887 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023888 .cm_stride(11)
23889 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard1d412472021-10-25 17:27:21 -070023890 }
23891#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
23892
23893
Frank Barcharde31f29e2021-12-21 15:57:10 -080023894#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023895 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_eq_16) {
Frank Barchard13db60f2021-07-20 14:34:35 -070023896 TEST_REQUIRES_ARM_NEON;
23897 GemmMicrokernelTester()
23898 .mr(2)
23899 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023900 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070023901 .sr(1)
23902 .m(2)
23903 .n(8)
23904 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023905 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070023906 }
23907
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023908 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, strided_cn) {
Frank Barchard13db60f2021-07-20 14:34:35 -070023909 TEST_REQUIRES_ARM_NEON;
23910 GemmMicrokernelTester()
23911 .mr(2)
23912 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023913 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070023914 .sr(1)
23915 .m(2)
23916 .n(8)
23917 .k(16)
23918 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023919 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070023920 }
23921
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023922 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_eq_16_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070023923 TEST_REQUIRES_ARM_NEON;
23924 for (uint32_t m = 1; m <= 2; m++) {
23925 for (uint32_t n = 1; n <= 8; n++) {
23926 GemmMicrokernelTester()
23927 .mr(2)
23928 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023929 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070023930 .sr(1)
23931 .m(m)
23932 .n(n)
23933 .k(16)
23934 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023935 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070023936 }
23937 }
23938 }
23939
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023940 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_eq_16_subtile_m) {
Frank Barchard13db60f2021-07-20 14:34:35 -070023941 TEST_REQUIRES_ARM_NEON;
23942 for (uint32_t m = 1; m <= 2; m++) {
23943 GemmMicrokernelTester()
23944 .mr(2)
23945 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023946 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070023947 .sr(1)
23948 .m(m)
23949 .n(8)
23950 .k(16)
23951 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023952 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070023953 }
23954 }
23955
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023956 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_eq_16_subtile_n) {
Frank Barchard13db60f2021-07-20 14:34:35 -070023957 TEST_REQUIRES_ARM_NEON;
23958 for (uint32_t n = 1; n <= 8; n++) {
23959 GemmMicrokernelTester()
23960 .mr(2)
23961 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023962 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070023963 .sr(1)
23964 .m(2)
23965 .n(n)
23966 .k(16)
23967 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023968 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070023969 }
23970 }
23971
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023972 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_lt_16) {
Frank Barchard13db60f2021-07-20 14:34:35 -070023973 TEST_REQUIRES_ARM_NEON;
23974 for (size_t k = 1; k < 16; k++) {
23975 GemmMicrokernelTester()
23976 .mr(2)
23977 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023978 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070023979 .sr(1)
23980 .m(2)
23981 .n(8)
23982 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023983 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070023984 }
23985 }
23986
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023987 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_lt_16_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070023988 TEST_REQUIRES_ARM_NEON;
23989 for (size_t k = 1; k < 16; k++) {
23990 for (uint32_t m = 1; m <= 2; m++) {
23991 for (uint32_t n = 1; n <= 8; n++) {
23992 GemmMicrokernelTester()
23993 .mr(2)
23994 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080023995 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070023996 .sr(1)
23997 .m(m)
23998 .n(n)
23999 .k(k)
24000 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024001 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024002 }
24003 }
24004 }
24005 }
24006
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024007 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_gt_16) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024008 TEST_REQUIRES_ARM_NEON;
24009 for (size_t k = 17; k < 32; k++) {
24010 GemmMicrokernelTester()
24011 .mr(2)
24012 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024013 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024014 .sr(1)
24015 .m(2)
24016 .n(8)
24017 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024018 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024019 }
24020 }
24021
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024022 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_gt_16_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024023 TEST_REQUIRES_ARM_NEON;
24024 for (size_t k = 17; k < 32; k++) {
24025 for (uint32_t m = 1; m <= 2; m++) {
24026 for (uint32_t n = 1; n <= 8; n++) {
24027 GemmMicrokernelTester()
24028 .mr(2)
24029 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024030 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024031 .sr(1)
24032 .m(m)
24033 .n(n)
24034 .k(k)
24035 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024036 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024037 }
24038 }
24039 }
24040 }
24041
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024042 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_div_16) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024043 TEST_REQUIRES_ARM_NEON;
24044 for (size_t k = 32; k <= 160; k += 16) {
24045 GemmMicrokernelTester()
24046 .mr(2)
24047 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024048 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024049 .sr(1)
24050 .m(2)
24051 .n(8)
24052 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024053 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024054 }
24055 }
24056
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024057 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_div_16_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024058 TEST_REQUIRES_ARM_NEON;
24059 for (size_t k = 32; k <= 160; k += 16) {
24060 for (uint32_t m = 1; m <= 2; m++) {
24061 for (uint32_t n = 1; n <= 8; n++) {
24062 GemmMicrokernelTester()
24063 .mr(2)
24064 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024065 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024066 .sr(1)
24067 .m(m)
24068 .n(n)
24069 .k(k)
24070 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024071 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024072 }
24073 }
24074 }
24075 }
24076
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024077 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_gt_8) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024078 TEST_REQUIRES_ARM_NEON;
24079 for (uint32_t n = 9; n < 16; n++) {
24080 for (size_t k = 1; k <= 80; k += 17) {
24081 GemmMicrokernelTester()
24082 .mr(2)
24083 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024084 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024085 .sr(1)
24086 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024087 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070024088 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024089 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024090 }
24091 }
24092 }
24093
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024094 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_gt_8_strided_cn) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024095 TEST_REQUIRES_ARM_NEON;
24096 for (uint32_t n = 9; n < 16; n++) {
24097 for (size_t k = 1; k <= 80; k += 17) {
24098 GemmMicrokernelTester()
24099 .mr(2)
24100 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024101 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024102 .sr(1)
24103 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024104 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070024105 .k(k)
24106 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024107 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024108 }
24109 }
24110 }
24111
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024112 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_gt_8_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024113 TEST_REQUIRES_ARM_NEON;
24114 for (uint32_t n = 9; n < 16; n++) {
24115 for (size_t k = 1; k <= 80; k += 17) {
24116 for (uint32_t m = 1; m <= 2; m++) {
24117 GemmMicrokernelTester()
24118 .mr(2)
24119 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024120 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024121 .sr(1)
24122 .m(m)
24123 .n(n)
24124 .k(k)
24125 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024126 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024127 }
24128 }
24129 }
24130 }
24131
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024132 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_div_8) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024133 TEST_REQUIRES_ARM_NEON;
24134 for (uint32_t n = 16; n <= 24; n += 8) {
24135 for (size_t k = 1; k <= 80; k += 17) {
24136 GemmMicrokernelTester()
24137 .mr(2)
24138 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024139 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024140 .sr(1)
24141 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024142 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070024143 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024144 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024145 }
24146 }
24147 }
24148
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024149 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_div_8_strided_cn) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024150 TEST_REQUIRES_ARM_NEON;
24151 for (uint32_t n = 16; n <= 24; n += 8) {
24152 for (size_t k = 1; k <= 80; k += 17) {
24153 GemmMicrokernelTester()
24154 .mr(2)
24155 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024156 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024157 .sr(1)
24158 .m(2)
24159 .n(n)
24160 .k(k)
24161 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024162 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024163 }
24164 }
24165 }
24166
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024167 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_div_8_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024168 TEST_REQUIRES_ARM_NEON;
24169 for (uint32_t n = 16; n <= 24; n += 8) {
24170 for (size_t k = 1; k <= 80; k += 17) {
24171 for (uint32_t m = 1; m <= 2; m++) {
24172 GemmMicrokernelTester()
24173 .mr(2)
24174 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024175 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024176 .sr(1)
24177 .m(m)
24178 .n(n)
24179 .k(k)
24180 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024181 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024182 }
24183 }
24184 }
24185 }
24186
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024187 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, small_kernel) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024188 TEST_REQUIRES_ARM_NEON;
24189 for (size_t k = 1; k <= 80; k += 17) {
24190 GemmMicrokernelTester()
24191 .mr(2)
24192 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024193 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024194 .sr(1)
24195 .m(2)
24196 .n(8)
24197 .k(k)
24198 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024199 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024200 }
24201 }
24202
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024203 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, small_kernel_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024204 TEST_REQUIRES_ARM_NEON;
24205 for (size_t k = 1; k <= 80; k += 17) {
24206 for (uint32_t m = 1; m <= 2; m++) {
24207 for (uint32_t n = 1; n <= 8; n++) {
24208 GemmMicrokernelTester()
24209 .mr(2)
24210 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024211 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024212 .sr(1)
24213 .m(m)
24214 .n(n)
24215 .k(k)
24216 .ks(3)
24217 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024218 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024219 }
24220 }
24221 }
24222 }
24223
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024224 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_gt_8_small_kernel) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024225 TEST_REQUIRES_ARM_NEON;
24226 for (uint32_t n = 9; n < 16; n++) {
24227 for (size_t k = 1; k <= 80; k += 17) {
24228 GemmMicrokernelTester()
24229 .mr(2)
24230 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024231 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024232 .sr(1)
24233 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024234 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070024235 .k(k)
24236 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024237 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024238 }
24239 }
24240 }
24241
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024242 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_div_8_small_kernel) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024243 TEST_REQUIRES_ARM_NEON;
24244 for (uint32_t n = 16; n <= 24; n += 8) {
24245 for (size_t k = 1; k <= 80; k += 17) {
24246 GemmMicrokernelTester()
24247 .mr(2)
24248 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024249 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024250 .sr(1)
24251 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024252 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070024253 .k(k)
24254 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024255 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024256 }
24257 }
24258 }
24259
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024260 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, strided_cm_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024261 TEST_REQUIRES_ARM_NEON;
24262 for (size_t k = 1; k <= 80; k += 17) {
24263 for (uint32_t m = 1; m <= 2; m++) {
24264 for (uint32_t n = 1; n <= 8; n++) {
24265 GemmMicrokernelTester()
24266 .mr(2)
24267 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024268 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024269 .sr(1)
24270 .m(m)
24271 .n(n)
24272 .k(k)
24273 .cm_stride(11)
24274 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024275 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024276 }
24277 }
24278 }
24279 }
24280
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024281 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, a_offset) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024282 TEST_REQUIRES_ARM_NEON;
24283 for (size_t k = 1; k <= 80; k += 17) {
24284 GemmMicrokernelTester()
24285 .mr(2)
24286 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024287 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024288 .sr(1)
24289 .m(2)
24290 .n(8)
24291 .k(k)
24292 .ks(3)
24293 .a_offset(163)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024294 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024295 }
24296 }
24297
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024298 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, zero) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024299 TEST_REQUIRES_ARM_NEON;
24300 for (uint32_t mz = 0; mz < 2; mz++) {
24301 for (size_t k = 1; k <= 80; k += 17) {
24302 GemmMicrokernelTester()
24303 .mr(2)
24304 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024305 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024306 .sr(1)
24307 .m(2)
24308 .n(8)
24309 .k(k)
24310 .ks(3)
24311 .a_offset(163)
24312 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024313 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024314 }
24315 }
24316 }
24317
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024318 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, qmin) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024319 TEST_REQUIRES_ARM_NEON;
24320 GemmMicrokernelTester()
24321 .mr(2)
24322 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024323 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024324 .sr(1)
24325 .m(2)
24326 .n(8)
24327 .k(16)
24328 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024329 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024330 }
24331
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024332 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, qmax) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024333 TEST_REQUIRES_ARM_NEON;
24334 GemmMicrokernelTester()
24335 .mr(2)
24336 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024337 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024338 .sr(1)
24339 .m(2)
24340 .n(8)
24341 .k(16)
24342 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024344 }
24345
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024346 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_CORTEX_A53, strided_cm) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024347 TEST_REQUIRES_ARM_NEON;
24348 GemmMicrokernelTester()
24349 .mr(2)
24350 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024351 .kr(8)
Frank Barchard13db60f2021-07-20 14:34:35 -070024352 .sr(1)
24353 .m(2)
24354 .n(8)
24355 .k(16)
24356 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024357 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24358 }
24359#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
24360
24361
24362#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
24363 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_eq_16) {
24364 TEST_REQUIRES_ARM_NEON;
24365 GemmMicrokernelTester()
24366 .mr(2)
24367 .nr(8)
24368 .kr(8)
24369 .sr(1)
24370 .m(2)
24371 .n(8)
24372 .k(16)
24373 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24374 }
24375
24376 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, strided_cn) {
24377 TEST_REQUIRES_ARM_NEON;
24378 GemmMicrokernelTester()
24379 .mr(2)
24380 .nr(8)
24381 .kr(8)
24382 .sr(1)
24383 .m(2)
24384 .n(8)
24385 .k(16)
24386 .cn_stride(11)
24387 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24388 }
24389
24390 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_eq_16_subtile) {
24391 TEST_REQUIRES_ARM_NEON;
24392 for (uint32_t m = 1; m <= 2; m++) {
24393 for (uint32_t n = 1; n <= 8; n++) {
24394 GemmMicrokernelTester()
24395 .mr(2)
24396 .nr(8)
24397 .kr(8)
24398 .sr(1)
24399 .m(m)
24400 .n(n)
24401 .k(16)
24402 .iterations(1)
24403 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24404 }
24405 }
24406 }
24407
24408 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_eq_16_subtile_m) {
24409 TEST_REQUIRES_ARM_NEON;
24410 for (uint32_t m = 1; m <= 2; m++) {
24411 GemmMicrokernelTester()
24412 .mr(2)
24413 .nr(8)
24414 .kr(8)
24415 .sr(1)
24416 .m(m)
24417 .n(8)
24418 .k(16)
24419 .iterations(1)
24420 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24421 }
24422 }
24423
24424 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_eq_16_subtile_n) {
24425 TEST_REQUIRES_ARM_NEON;
24426 for (uint32_t n = 1; n <= 8; n++) {
24427 GemmMicrokernelTester()
24428 .mr(2)
24429 .nr(8)
24430 .kr(8)
24431 .sr(1)
24432 .m(2)
24433 .n(n)
24434 .k(16)
24435 .iterations(1)
24436 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24437 }
24438 }
24439
24440 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_lt_16) {
24441 TEST_REQUIRES_ARM_NEON;
24442 for (size_t k = 1; k < 16; k++) {
24443 GemmMicrokernelTester()
24444 .mr(2)
24445 .nr(8)
24446 .kr(8)
24447 .sr(1)
24448 .m(2)
24449 .n(8)
24450 .k(k)
24451 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24452 }
24453 }
24454
24455 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_lt_16_subtile) {
24456 TEST_REQUIRES_ARM_NEON;
24457 for (size_t k = 1; k < 16; k++) {
24458 for (uint32_t m = 1; m <= 2; m++) {
24459 for (uint32_t n = 1; n <= 8; n++) {
24460 GemmMicrokernelTester()
24461 .mr(2)
24462 .nr(8)
24463 .kr(8)
24464 .sr(1)
24465 .m(m)
24466 .n(n)
24467 .k(k)
24468 .iterations(1)
24469 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24470 }
24471 }
24472 }
24473 }
24474
24475 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_gt_16) {
24476 TEST_REQUIRES_ARM_NEON;
24477 for (size_t k = 17; k < 32; k++) {
24478 GemmMicrokernelTester()
24479 .mr(2)
24480 .nr(8)
24481 .kr(8)
24482 .sr(1)
24483 .m(2)
24484 .n(8)
24485 .k(k)
24486 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24487 }
24488 }
24489
24490 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_gt_16_subtile) {
24491 TEST_REQUIRES_ARM_NEON;
24492 for (size_t k = 17; k < 32; k++) {
24493 for (uint32_t m = 1; m <= 2; m++) {
24494 for (uint32_t n = 1; n <= 8; n++) {
24495 GemmMicrokernelTester()
24496 .mr(2)
24497 .nr(8)
24498 .kr(8)
24499 .sr(1)
24500 .m(m)
24501 .n(n)
24502 .k(k)
24503 .iterations(1)
24504 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24505 }
24506 }
24507 }
24508 }
24509
24510 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_div_16) {
24511 TEST_REQUIRES_ARM_NEON;
24512 for (size_t k = 32; k <= 160; k += 16) {
24513 GemmMicrokernelTester()
24514 .mr(2)
24515 .nr(8)
24516 .kr(8)
24517 .sr(1)
24518 .m(2)
24519 .n(8)
24520 .k(k)
24521 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24522 }
24523 }
24524
24525 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, k_div_16_subtile) {
24526 TEST_REQUIRES_ARM_NEON;
24527 for (size_t k = 32; k <= 160; k += 16) {
24528 for (uint32_t m = 1; m <= 2; m++) {
24529 for (uint32_t n = 1; n <= 8; n++) {
24530 GemmMicrokernelTester()
24531 .mr(2)
24532 .nr(8)
24533 .kr(8)
24534 .sr(1)
24535 .m(m)
24536 .n(n)
24537 .k(k)
24538 .iterations(1)
24539 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24540 }
24541 }
24542 }
24543 }
24544
24545 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, n_gt_8) {
24546 TEST_REQUIRES_ARM_NEON;
24547 for (uint32_t n = 9; n < 16; n++) {
24548 for (size_t k = 1; k <= 80; k += 17) {
24549 GemmMicrokernelTester()
24550 .mr(2)
24551 .nr(8)
24552 .kr(8)
24553 .sr(1)
24554 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024555 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024556 .k(k)
24557 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24558 }
24559 }
24560 }
24561
24562 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, n_gt_8_strided_cn) {
24563 TEST_REQUIRES_ARM_NEON;
24564 for (uint32_t n = 9; n < 16; n++) {
24565 for (size_t k = 1; k <= 80; k += 17) {
24566 GemmMicrokernelTester()
24567 .mr(2)
24568 .nr(8)
24569 .kr(8)
24570 .sr(1)
24571 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024572 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024573 .k(k)
24574 .cn_stride(11)
24575 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24576 }
24577 }
24578 }
24579
24580 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, n_gt_8_subtile) {
24581 TEST_REQUIRES_ARM_NEON;
24582 for (uint32_t n = 9; n < 16; n++) {
24583 for (size_t k = 1; k <= 80; k += 17) {
24584 for (uint32_t m = 1; m <= 2; m++) {
24585 GemmMicrokernelTester()
24586 .mr(2)
24587 .nr(8)
24588 .kr(8)
24589 .sr(1)
24590 .m(m)
24591 .n(n)
24592 .k(k)
24593 .iterations(1)
24594 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24595 }
24596 }
24597 }
24598 }
24599
24600 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, n_div_8) {
24601 TEST_REQUIRES_ARM_NEON;
24602 for (uint32_t n = 16; n <= 24; n += 8) {
24603 for (size_t k = 1; k <= 80; k += 17) {
24604 GemmMicrokernelTester()
24605 .mr(2)
24606 .nr(8)
24607 .kr(8)
24608 .sr(1)
24609 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024610 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024611 .k(k)
24612 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24613 }
24614 }
24615 }
24616
24617 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, n_div_8_strided_cn) {
24618 TEST_REQUIRES_ARM_NEON;
24619 for (uint32_t n = 16; n <= 24; n += 8) {
24620 for (size_t k = 1; k <= 80; k += 17) {
24621 GemmMicrokernelTester()
24622 .mr(2)
24623 .nr(8)
24624 .kr(8)
24625 .sr(1)
24626 .m(2)
24627 .n(n)
24628 .k(k)
24629 .cn_stride(11)
24630 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24631 }
24632 }
24633 }
24634
24635 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, n_div_8_subtile) {
24636 TEST_REQUIRES_ARM_NEON;
24637 for (uint32_t n = 16; n <= 24; n += 8) {
24638 for (size_t k = 1; k <= 80; k += 17) {
24639 for (uint32_t m = 1; m <= 2; m++) {
24640 GemmMicrokernelTester()
24641 .mr(2)
24642 .nr(8)
24643 .kr(8)
24644 .sr(1)
24645 .m(m)
24646 .n(n)
24647 .k(k)
24648 .iterations(1)
24649 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24650 }
24651 }
24652 }
24653 }
24654
24655 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, small_kernel) {
24656 TEST_REQUIRES_ARM_NEON;
24657 for (size_t k = 1; k <= 80; k += 17) {
24658 GemmMicrokernelTester()
24659 .mr(2)
24660 .nr(8)
24661 .kr(8)
24662 .sr(1)
24663 .m(2)
24664 .n(8)
24665 .k(k)
24666 .ks(3)
24667 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24668 }
24669 }
24670
24671 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, small_kernel_subtile) {
24672 TEST_REQUIRES_ARM_NEON;
24673 for (size_t k = 1; k <= 80; k += 17) {
24674 for (uint32_t m = 1; m <= 2; m++) {
24675 for (uint32_t n = 1; n <= 8; n++) {
24676 GemmMicrokernelTester()
24677 .mr(2)
24678 .nr(8)
24679 .kr(8)
24680 .sr(1)
24681 .m(m)
24682 .n(n)
24683 .k(k)
24684 .ks(3)
24685 .iterations(1)
24686 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24687 }
24688 }
24689 }
24690 }
24691
24692 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, n_gt_8_small_kernel) {
24693 TEST_REQUIRES_ARM_NEON;
24694 for (uint32_t n = 9; n < 16; n++) {
24695 for (size_t k = 1; k <= 80; k += 17) {
24696 GemmMicrokernelTester()
24697 .mr(2)
24698 .nr(8)
24699 .kr(8)
24700 .sr(1)
24701 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024702 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024703 .k(k)
24704 .ks(3)
24705 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24706 }
24707 }
24708 }
24709
24710 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, n_div_8_small_kernel) {
24711 TEST_REQUIRES_ARM_NEON;
24712 for (uint32_t n = 16; n <= 24; n += 8) {
24713 for (size_t k = 1; k <= 80; k += 17) {
24714 GemmMicrokernelTester()
24715 .mr(2)
24716 .nr(8)
24717 .kr(8)
24718 .sr(1)
24719 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080024720 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080024721 .k(k)
24722 .ks(3)
24723 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24724 }
24725 }
24726 }
24727
24728 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, strided_cm_subtile) {
24729 TEST_REQUIRES_ARM_NEON;
24730 for (size_t k = 1; k <= 80; k += 17) {
24731 for (uint32_t m = 1; m <= 2; m++) {
24732 for (uint32_t n = 1; n <= 8; n++) {
24733 GemmMicrokernelTester()
24734 .mr(2)
24735 .nr(8)
24736 .kr(8)
24737 .sr(1)
24738 .m(m)
24739 .n(n)
24740 .k(k)
24741 .cm_stride(11)
24742 .iterations(1)
24743 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24744 }
24745 }
24746 }
24747 }
24748
24749 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, a_offset) {
24750 TEST_REQUIRES_ARM_NEON;
24751 for (size_t k = 1; k <= 80; k += 17) {
24752 GemmMicrokernelTester()
24753 .mr(2)
24754 .nr(8)
24755 .kr(8)
24756 .sr(1)
24757 .m(2)
24758 .n(8)
24759 .k(k)
24760 .ks(3)
24761 .a_offset(163)
24762 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24763 }
24764 }
24765
24766 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, zero) {
24767 TEST_REQUIRES_ARM_NEON;
24768 for (uint32_t mz = 0; mz < 2; mz++) {
24769 for (size_t k = 1; k <= 80; k += 17) {
24770 GemmMicrokernelTester()
24771 .mr(2)
24772 .nr(8)
24773 .kr(8)
24774 .sr(1)
24775 .m(2)
24776 .n(8)
24777 .k(k)
24778 .ks(3)
24779 .a_offset(163)
24780 .zero_index(mz)
24781 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24782 }
24783 }
24784 }
24785
24786 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, qmin) {
24787 TEST_REQUIRES_ARM_NEON;
24788 GemmMicrokernelTester()
24789 .mr(2)
24790 .nr(8)
24791 .kr(8)
24792 .sr(1)
24793 .m(2)
24794 .n(8)
24795 .k(16)
24796 .qmin(128)
24797 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24798 }
24799
24800 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, qmax) {
24801 TEST_REQUIRES_ARM_NEON;
24802 GemmMicrokernelTester()
24803 .mr(2)
24804 .nr(8)
24805 .kr(8)
24806 .sr(1)
24807 .m(2)
24808 .n(8)
24809 .k(16)
24810 .qmax(128)
24811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
24812 }
24813
24814 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM_CORTEX_A53, strided_cm) {
24815 TEST_REQUIRES_ARM_NEON;
24816 GemmMicrokernelTester()
24817 .mr(2)
24818 .nr(8)
24819 .kr(8)
24820 .sr(1)
24821 .m(2)
24822 .n(8)
24823 .k(16)
24824 .cm_stride(11)
24825 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024826 }
Frank Barcharde31f29e2021-12-21 15:57:10 -080024827#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
Frank Barchard13db60f2021-07-20 14:34:35 -070024828
24829
Frank Barcharde31f29e2021-12-21 15:57:10 -080024830#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
Frank Barcharde22685a2021-11-12 11:36:58 -080024831 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_eq_16) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024832 TEST_REQUIRES_ARM_NEON;
24833 GemmMicrokernelTester()
24834 .mr(1)
24835 .nr(8)
24836 .kr(8)
24837 .sr(1)
24838 .m(1)
24839 .n(8)
24840 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -080024841 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024842 }
24843
Frank Barcharde22685a2021-11-12 11:36:58 -080024844 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, strided_cn) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024845 TEST_REQUIRES_ARM_NEON;
24846 GemmMicrokernelTester()
24847 .mr(1)
24848 .nr(8)
24849 .kr(8)
24850 .sr(1)
24851 .m(1)
24852 .n(8)
24853 .k(16)
24854 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080024855 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024856 }
24857
Frank Barcharde22685a2021-11-12 11:36:58 -080024858 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_eq_16_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024859 TEST_REQUIRES_ARM_NEON;
24860 for (uint32_t m = 1; m <= 1; m++) {
24861 for (uint32_t n = 1; n <= 8; n++) {
24862 GemmMicrokernelTester()
24863 .mr(1)
24864 .nr(8)
24865 .kr(8)
24866 .sr(1)
24867 .m(m)
24868 .n(n)
24869 .k(16)
24870 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080024871 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024872 }
24873 }
24874 }
24875
Frank Barcharde22685a2021-11-12 11:36:58 -080024876 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_eq_16_subtile_m) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024877 TEST_REQUIRES_ARM_NEON;
24878 for (uint32_t m = 1; m <= 1; m++) {
24879 GemmMicrokernelTester()
24880 .mr(1)
24881 .nr(8)
24882 .kr(8)
24883 .sr(1)
24884 .m(m)
24885 .n(8)
24886 .k(16)
24887 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080024888 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024889 }
24890 }
24891
Frank Barcharde22685a2021-11-12 11:36:58 -080024892 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_eq_16_subtile_n) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024893 TEST_REQUIRES_ARM_NEON;
24894 for (uint32_t n = 1; n <= 8; n++) {
24895 GemmMicrokernelTester()
24896 .mr(1)
24897 .nr(8)
24898 .kr(8)
24899 .sr(1)
24900 .m(1)
24901 .n(n)
24902 .k(16)
24903 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080024904 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024905 }
24906 }
24907
Frank Barcharde22685a2021-11-12 11:36:58 -080024908 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_lt_16) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024909 TEST_REQUIRES_ARM_NEON;
24910 for (size_t k = 1; k < 16; k++) {
24911 GemmMicrokernelTester()
24912 .mr(1)
24913 .nr(8)
24914 .kr(8)
24915 .sr(1)
24916 .m(1)
24917 .n(8)
24918 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080024919 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024920 }
24921 }
24922
Frank Barcharde22685a2021-11-12 11:36:58 -080024923 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_lt_16_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024924 TEST_REQUIRES_ARM_NEON;
24925 for (size_t k = 1; k < 16; k++) {
24926 for (uint32_t m = 1; m <= 1; m++) {
24927 for (uint32_t n = 1; n <= 8; n++) {
24928 GemmMicrokernelTester()
24929 .mr(1)
24930 .nr(8)
24931 .kr(8)
24932 .sr(1)
24933 .m(m)
24934 .n(n)
24935 .k(k)
24936 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080024937 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024938 }
24939 }
24940 }
24941 }
24942
Frank Barcharde22685a2021-11-12 11:36:58 -080024943 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_gt_16) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024944 TEST_REQUIRES_ARM_NEON;
24945 for (size_t k = 17; k < 32; k++) {
24946 GemmMicrokernelTester()
24947 .mr(1)
24948 .nr(8)
24949 .kr(8)
24950 .sr(1)
24951 .m(1)
24952 .n(8)
24953 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080024954 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024955 }
24956 }
24957
Frank Barcharde22685a2021-11-12 11:36:58 -080024958 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_gt_16_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024959 TEST_REQUIRES_ARM_NEON;
24960 for (size_t k = 17; k < 32; k++) {
24961 for (uint32_t m = 1; m <= 1; m++) {
24962 for (uint32_t n = 1; n <= 8; n++) {
24963 GemmMicrokernelTester()
24964 .mr(1)
24965 .nr(8)
24966 .kr(8)
24967 .sr(1)
24968 .m(m)
24969 .n(n)
24970 .k(k)
24971 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080024972 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024973 }
24974 }
24975 }
24976 }
24977
Frank Barcharde22685a2021-11-12 11:36:58 -080024978 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_div_16) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024979 TEST_REQUIRES_ARM_NEON;
24980 for (size_t k = 32; k <= 160; k += 16) {
24981 GemmMicrokernelTester()
24982 .mr(1)
24983 .nr(8)
24984 .kr(8)
24985 .sr(1)
24986 .m(1)
24987 .n(8)
24988 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080024989 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070024990 }
24991 }
24992
Frank Barcharde22685a2021-11-12 11:36:58 -080024993 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, k_div_16_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070024994 TEST_REQUIRES_ARM_NEON;
24995 for (size_t k = 32; k <= 160; k += 16) {
24996 for (uint32_t m = 1; m <= 1; m++) {
24997 for (uint32_t n = 1; n <= 8; n++) {
24998 GemmMicrokernelTester()
24999 .mr(1)
25000 .nr(8)
25001 .kr(8)
25002 .sr(1)
25003 .m(m)
25004 .n(n)
25005 .k(k)
25006 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025007 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025008 }
25009 }
25010 }
25011 }
25012
Frank Barcharde22685a2021-11-12 11:36:58 -080025013 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_gt_8) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025014 TEST_REQUIRES_ARM_NEON;
25015 for (uint32_t n = 9; n < 16; n++) {
25016 for (size_t k = 1; k <= 80; k += 17) {
25017 GemmMicrokernelTester()
25018 .mr(1)
25019 .nr(8)
25020 .kr(8)
25021 .sr(1)
25022 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025023 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070025024 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025025 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025026 }
25027 }
25028 }
25029
Frank Barcharde22685a2021-11-12 11:36:58 -080025030 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_gt_8_strided_cn) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025031 TEST_REQUIRES_ARM_NEON;
25032 for (uint32_t n = 9; n < 16; n++) {
25033 for (size_t k = 1; k <= 80; k += 17) {
25034 GemmMicrokernelTester()
25035 .mr(1)
25036 .nr(8)
25037 .kr(8)
25038 .sr(1)
25039 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025040 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070025041 .k(k)
25042 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080025043 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025044 }
25045 }
25046 }
25047
Frank Barcharde22685a2021-11-12 11:36:58 -080025048 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_gt_8_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025049 TEST_REQUIRES_ARM_NEON;
25050 for (uint32_t n = 9; n < 16; n++) {
25051 for (size_t k = 1; k <= 80; k += 17) {
25052 for (uint32_t m = 1; m <= 1; m++) {
25053 GemmMicrokernelTester()
25054 .mr(1)
25055 .nr(8)
25056 .kr(8)
25057 .sr(1)
25058 .m(m)
25059 .n(n)
25060 .k(k)
25061 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025062 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025063 }
25064 }
25065 }
25066 }
25067
Frank Barcharde22685a2021-11-12 11:36:58 -080025068 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_div_8) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025069 TEST_REQUIRES_ARM_NEON;
25070 for (uint32_t n = 16; n <= 24; n += 8) {
25071 for (size_t k = 1; k <= 80; k += 17) {
25072 GemmMicrokernelTester()
25073 .mr(1)
25074 .nr(8)
25075 .kr(8)
25076 .sr(1)
25077 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025078 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070025079 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025080 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025081 }
25082 }
25083 }
25084
Frank Barcharde22685a2021-11-12 11:36:58 -080025085 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_div_8_strided_cn) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025086 TEST_REQUIRES_ARM_NEON;
25087 for (uint32_t n = 16; n <= 24; n += 8) {
25088 for (size_t k = 1; k <= 80; k += 17) {
25089 GemmMicrokernelTester()
25090 .mr(1)
25091 .nr(8)
25092 .kr(8)
25093 .sr(1)
25094 .m(1)
25095 .n(n)
25096 .k(k)
25097 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080025098 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025099 }
25100 }
25101 }
25102
Frank Barcharde22685a2021-11-12 11:36:58 -080025103 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_div_8_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025104 TEST_REQUIRES_ARM_NEON;
25105 for (uint32_t n = 16; n <= 24; n += 8) {
25106 for (size_t k = 1; k <= 80; k += 17) {
25107 for (uint32_t m = 1; m <= 1; m++) {
25108 GemmMicrokernelTester()
25109 .mr(1)
25110 .nr(8)
25111 .kr(8)
25112 .sr(1)
25113 .m(m)
25114 .n(n)
25115 .k(k)
25116 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025117 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025118 }
25119 }
25120 }
25121 }
25122
Frank Barcharde22685a2021-11-12 11:36:58 -080025123 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, small_kernel) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025124 TEST_REQUIRES_ARM_NEON;
25125 for (size_t k = 1; k <= 80; k += 17) {
25126 GemmMicrokernelTester()
25127 .mr(1)
25128 .nr(8)
25129 .kr(8)
25130 .sr(1)
25131 .m(1)
25132 .n(8)
25133 .k(k)
25134 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080025135 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025136 }
25137 }
25138
Frank Barcharde22685a2021-11-12 11:36:58 -080025139 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, small_kernel_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025140 TEST_REQUIRES_ARM_NEON;
25141 for (size_t k = 1; k <= 80; k += 17) {
25142 for (uint32_t m = 1; m <= 1; m++) {
25143 for (uint32_t n = 1; n <= 8; n++) {
25144 GemmMicrokernelTester()
25145 .mr(1)
25146 .nr(8)
25147 .kr(8)
25148 .sr(1)
25149 .m(m)
25150 .n(n)
25151 .k(k)
25152 .ks(3)
25153 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025154 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025155 }
25156 }
25157 }
25158 }
25159
Frank Barcharde22685a2021-11-12 11:36:58 -080025160 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_gt_8_small_kernel) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025161 TEST_REQUIRES_ARM_NEON;
25162 for (uint32_t n = 9; n < 16; n++) {
25163 for (size_t k = 1; k <= 80; k += 17) {
25164 GemmMicrokernelTester()
25165 .mr(1)
25166 .nr(8)
25167 .kr(8)
25168 .sr(1)
25169 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025170 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070025171 .k(k)
25172 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080025173 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025174 }
25175 }
25176 }
25177
Frank Barcharde22685a2021-11-12 11:36:58 -080025178 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, n_div_8_small_kernel) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025179 TEST_REQUIRES_ARM_NEON;
25180 for (uint32_t n = 16; n <= 24; n += 8) {
25181 for (size_t k = 1; k <= 80; k += 17) {
25182 GemmMicrokernelTester()
25183 .mr(1)
25184 .nr(8)
25185 .kr(8)
25186 .sr(1)
25187 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025188 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070025189 .k(k)
25190 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080025191 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025192 }
25193 }
25194 }
25195
Frank Barcharde22685a2021-11-12 11:36:58 -080025196 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, strided_cm_subtile) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025197 TEST_REQUIRES_ARM_NEON;
25198 for (size_t k = 1; k <= 80; k += 17) {
25199 for (uint32_t m = 1; m <= 1; m++) {
25200 for (uint32_t n = 1; n <= 8; n++) {
25201 GemmMicrokernelTester()
25202 .mr(1)
25203 .nr(8)
25204 .kr(8)
25205 .sr(1)
25206 .m(m)
25207 .n(n)
25208 .k(k)
25209 .cm_stride(11)
25210 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025211 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025212 }
25213 }
25214 }
25215 }
25216
Frank Barcharde22685a2021-11-12 11:36:58 -080025217 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, a_offset) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025218 TEST_REQUIRES_ARM_NEON;
25219 for (size_t k = 1; k <= 80; k += 17) {
25220 GemmMicrokernelTester()
25221 .mr(1)
25222 .nr(8)
25223 .kr(8)
25224 .sr(1)
25225 .m(1)
25226 .n(8)
25227 .k(k)
25228 .ks(3)
25229 .a_offset(83)
Marat Dukhan50323b82022-01-11 00:12:01 -080025230 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025231 }
25232 }
25233
Frank Barcharde22685a2021-11-12 11:36:58 -080025234 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, zero) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025235 TEST_REQUIRES_ARM_NEON;
25236 for (uint32_t mz = 0; mz < 1; mz++) {
25237 for (size_t k = 1; k <= 80; k += 17) {
25238 GemmMicrokernelTester()
25239 .mr(1)
25240 .nr(8)
25241 .kr(8)
25242 .sr(1)
25243 .m(1)
25244 .n(8)
25245 .k(k)
25246 .ks(3)
25247 .a_offset(83)
25248 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080025249 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025250 }
25251 }
25252 }
25253
Frank Barcharde22685a2021-11-12 11:36:58 -080025254 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, qmin) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025255 TEST_REQUIRES_ARM_NEON;
25256 GemmMicrokernelTester()
25257 .mr(1)
25258 .nr(8)
25259 .kr(8)
25260 .sr(1)
25261 .m(1)
25262 .n(8)
25263 .k(16)
25264 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080025265 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025266 }
25267
Frank Barcharde22685a2021-11-12 11:36:58 -080025268 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, qmax) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025269 TEST_REQUIRES_ARM_NEON;
25270 GemmMicrokernelTester()
25271 .mr(1)
25272 .nr(8)
25273 .kr(8)
25274 .sr(1)
25275 .m(1)
25276 .n(8)
25277 .k(16)
25278 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080025279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025280 }
25281
Frank Barcharde22685a2021-11-12 11:36:58 -080025282 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL_CORTEX_A53, strided_cm) {
Frank Barchard13db60f2021-07-20 14:34:35 -070025283 TEST_REQUIRES_ARM_NEON;
25284 GemmMicrokernelTester()
25285 .mr(1)
25286 .nr(8)
25287 .kr(8)
25288 .sr(1)
25289 .m(1)
25290 .n(8)
25291 .k(16)
25292 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080025293 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025294 }
Frank Barcharde31f29e2021-12-21 15:57:10 -080025295#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
Frank Barchard13db60f2021-07-20 14:34:35 -070025296
25297
Frank Barcharde31f29e2021-12-21 15:57:10 -080025298#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
Frank Barchard5cffb642021-11-22 13:59:43 -080025299 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_eq_8) {
25300 TEST_REQUIRES_ARM_NEON;
25301 GemmMicrokernelTester()
25302 .mr(4)
25303 .nr(16)
25304 .kr(1)
25305 .sr(1)
25306 .m(4)
25307 .n(16)
25308 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080025309 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025310 }
25311
25312 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, strided_cn) {
25313 TEST_REQUIRES_ARM_NEON;
25314 GemmMicrokernelTester()
25315 .mr(4)
25316 .nr(16)
25317 .kr(1)
25318 .sr(1)
25319 .m(4)
25320 .n(16)
25321 .k(8)
25322 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080025323 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025324 }
25325
25326 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_subtile) {
25327 TEST_REQUIRES_ARM_NEON;
25328 for (uint32_t m = 1; m <= 4; m++) {
25329 for (uint32_t n = 1; n <= 16; n++) {
25330 GemmMicrokernelTester()
25331 .mr(4)
25332 .nr(16)
25333 .kr(1)
25334 .sr(1)
25335 .m(m)
25336 .n(n)
25337 .k(8)
25338 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025339 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025340 }
25341 }
25342 }
25343
25344 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_subtile_m) {
25345 TEST_REQUIRES_ARM_NEON;
25346 for (uint32_t m = 1; m <= 4; m++) {
25347 GemmMicrokernelTester()
25348 .mr(4)
25349 .nr(16)
25350 .kr(1)
25351 .sr(1)
25352 .m(m)
25353 .n(16)
25354 .k(8)
25355 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025356 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025357 }
25358 }
25359
25360 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_subtile_n) {
25361 TEST_REQUIRES_ARM_NEON;
25362 for (uint32_t n = 1; n <= 16; n++) {
25363 GemmMicrokernelTester()
25364 .mr(4)
25365 .nr(16)
25366 .kr(1)
25367 .sr(1)
25368 .m(4)
25369 .n(n)
25370 .k(8)
25371 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025372 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025373 }
25374 }
25375
25376 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_lt_8) {
25377 TEST_REQUIRES_ARM_NEON;
25378 for (size_t k = 1; k < 8; k++) {
25379 GemmMicrokernelTester()
25380 .mr(4)
25381 .nr(16)
25382 .kr(1)
25383 .sr(1)
25384 .m(4)
25385 .n(16)
25386 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025387 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025388 }
25389 }
25390
25391 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_lt_8_subtile) {
25392 TEST_REQUIRES_ARM_NEON;
25393 for (size_t k = 1; k < 8; k++) {
25394 for (uint32_t m = 1; m <= 4; m++) {
25395 for (uint32_t n = 1; n <= 16; n++) {
25396 GemmMicrokernelTester()
25397 .mr(4)
25398 .nr(16)
25399 .kr(1)
25400 .sr(1)
25401 .m(m)
25402 .n(n)
25403 .k(k)
25404 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025405 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025406 }
25407 }
25408 }
25409 }
25410
25411 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_gt_8) {
25412 TEST_REQUIRES_ARM_NEON;
25413 for (size_t k = 9; k < 16; k++) {
25414 GemmMicrokernelTester()
25415 .mr(4)
25416 .nr(16)
25417 .kr(1)
25418 .sr(1)
25419 .m(4)
25420 .n(16)
25421 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025422 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025423 }
25424 }
25425
25426 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_gt_8_subtile) {
25427 TEST_REQUIRES_ARM_NEON;
25428 for (size_t k = 9; k < 16; k++) {
25429 for (uint32_t m = 1; m <= 4; m++) {
25430 for (uint32_t n = 1; n <= 16; n++) {
25431 GemmMicrokernelTester()
25432 .mr(4)
25433 .nr(16)
25434 .kr(1)
25435 .sr(1)
25436 .m(m)
25437 .n(n)
25438 .k(k)
25439 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025440 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025441 }
25442 }
25443 }
25444 }
25445
25446 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_div_8) {
25447 TEST_REQUIRES_ARM_NEON;
25448 for (size_t k = 16; k <= 80; k += 8) {
25449 GemmMicrokernelTester()
25450 .mr(4)
25451 .nr(16)
25452 .kr(1)
25453 .sr(1)
25454 .m(4)
25455 .n(16)
25456 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025457 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025458 }
25459 }
25460
25461 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, k_div_8_subtile) {
25462 TEST_REQUIRES_ARM_NEON;
25463 for (size_t k = 16; k <= 80; k += 8) {
25464 for (uint32_t m = 1; m <= 4; m++) {
25465 for (uint32_t n = 1; n <= 16; n++) {
25466 GemmMicrokernelTester()
25467 .mr(4)
25468 .nr(16)
25469 .kr(1)
25470 .sr(1)
25471 .m(m)
25472 .n(n)
25473 .k(k)
25474 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025475 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025476 }
25477 }
25478 }
25479 }
25480
25481 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, n_gt_16) {
25482 TEST_REQUIRES_ARM_NEON;
25483 for (uint32_t n = 17; n < 32; n++) {
25484 for (size_t k = 1; k <= 40; k += 9) {
25485 GemmMicrokernelTester()
25486 .mr(4)
25487 .nr(16)
25488 .kr(1)
25489 .sr(1)
25490 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025491 .n(n)
Frank Barchard5cffb642021-11-22 13:59:43 -080025492 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025493 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025494 }
25495 }
25496 }
25497
25498 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, n_gt_16_strided_cn) {
25499 TEST_REQUIRES_ARM_NEON;
25500 for (uint32_t n = 17; n < 32; n++) {
25501 for (size_t k = 1; k <= 40; k += 9) {
25502 GemmMicrokernelTester()
25503 .mr(4)
25504 .nr(16)
25505 .kr(1)
25506 .sr(1)
25507 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025508 .n(n)
Frank Barchard5cffb642021-11-22 13:59:43 -080025509 .k(k)
25510 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080025511 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025512 }
25513 }
25514 }
25515
25516 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, n_gt_16_subtile) {
25517 TEST_REQUIRES_ARM_NEON;
25518 for (uint32_t n = 17; n < 32; n++) {
25519 for (size_t k = 1; k <= 40; k += 9) {
25520 for (uint32_t m = 1; m <= 4; m++) {
25521 GemmMicrokernelTester()
25522 .mr(4)
25523 .nr(16)
25524 .kr(1)
25525 .sr(1)
25526 .m(m)
25527 .n(n)
25528 .k(k)
25529 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025530 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025531 }
25532 }
25533 }
25534 }
25535
25536 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, n_div_16) {
25537 TEST_REQUIRES_ARM_NEON;
25538 for (uint32_t n = 32; n <= 48; n += 16) {
25539 for (size_t k = 1; k <= 40; k += 9) {
25540 GemmMicrokernelTester()
25541 .mr(4)
25542 .nr(16)
25543 .kr(1)
25544 .sr(1)
25545 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025546 .n(n)
Frank Barchard5cffb642021-11-22 13:59:43 -080025547 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025548 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025549 }
25550 }
25551 }
25552
25553 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, n_div_16_strided_cn) {
25554 TEST_REQUIRES_ARM_NEON;
25555 for (uint32_t n = 32; n <= 48; n += 16) {
25556 for (size_t k = 1; k <= 40; k += 9) {
25557 GemmMicrokernelTester()
25558 .mr(4)
25559 .nr(16)
25560 .kr(1)
25561 .sr(1)
25562 .m(4)
25563 .n(n)
25564 .k(k)
25565 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080025566 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025567 }
25568 }
25569 }
25570
25571 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, n_div_16_subtile) {
25572 TEST_REQUIRES_ARM_NEON;
25573 for (uint32_t n = 32; n <= 48; n += 16) {
25574 for (size_t k = 1; k <= 40; k += 9) {
25575 for (uint32_t m = 1; m <= 4; m++) {
25576 GemmMicrokernelTester()
25577 .mr(4)
25578 .nr(16)
25579 .kr(1)
25580 .sr(1)
25581 .m(m)
25582 .n(n)
25583 .k(k)
25584 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025585 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025586 }
25587 }
25588 }
25589 }
25590
25591 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, small_kernel) {
25592 TEST_REQUIRES_ARM_NEON;
25593 for (size_t k = 1; k <= 40; k += 9) {
25594 GemmMicrokernelTester()
25595 .mr(4)
25596 .nr(16)
25597 .kr(1)
25598 .sr(1)
25599 .m(4)
25600 .n(16)
25601 .k(k)
25602 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080025603 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025604 }
25605 }
25606
25607 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, small_kernel_subtile) {
25608 TEST_REQUIRES_ARM_NEON;
25609 for (size_t k = 1; k <= 40; k += 9) {
25610 for (uint32_t m = 1; m <= 4; m++) {
25611 for (uint32_t n = 1; n <= 16; n++) {
25612 GemmMicrokernelTester()
25613 .mr(4)
25614 .nr(16)
25615 .kr(1)
25616 .sr(1)
25617 .m(m)
25618 .n(n)
25619 .k(k)
25620 .ks(3)
25621 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025622 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025623 }
25624 }
25625 }
25626 }
25627
25628 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, n_gt_16_small_kernel) {
25629 TEST_REQUIRES_ARM_NEON;
25630 for (uint32_t n = 17; n < 32; n++) {
25631 for (size_t k = 1; k <= 40; k += 9) {
25632 GemmMicrokernelTester()
25633 .mr(4)
25634 .nr(16)
25635 .kr(1)
25636 .sr(1)
25637 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025638 .n(n)
Frank Barchard5cffb642021-11-22 13:59:43 -080025639 .k(k)
25640 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080025641 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025642 }
25643 }
25644 }
25645
25646 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, n_div_16_small_kernel) {
25647 TEST_REQUIRES_ARM_NEON;
25648 for (uint32_t n = 32; n <= 48; n += 16) {
25649 for (size_t k = 1; k <= 40; k += 9) {
25650 GemmMicrokernelTester()
25651 .mr(4)
25652 .nr(16)
25653 .kr(1)
25654 .sr(1)
25655 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025656 .n(n)
Frank Barchard5cffb642021-11-22 13:59:43 -080025657 .k(k)
25658 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080025659 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025660 }
25661 }
25662 }
25663
25664 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, strided_cm_subtile) {
25665 TEST_REQUIRES_ARM_NEON;
25666 for (size_t k = 1; k <= 40; k += 9) {
25667 for (uint32_t m = 1; m <= 4; m++) {
25668 for (uint32_t n = 1; n <= 16; n++) {
25669 GemmMicrokernelTester()
25670 .mr(4)
25671 .nr(16)
25672 .kr(1)
25673 .sr(1)
25674 .m(m)
25675 .n(n)
25676 .k(k)
25677 .cm_stride(19)
25678 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025679 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025680 }
25681 }
25682 }
25683 }
25684
25685 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, a_offset) {
25686 TEST_REQUIRES_ARM_NEON;
25687 for (size_t k = 1; k <= 40; k += 9) {
25688 GemmMicrokernelTester()
25689 .mr(4)
25690 .nr(16)
25691 .kr(1)
25692 .sr(1)
25693 .m(4)
25694 .n(16)
25695 .k(k)
25696 .ks(3)
25697 .a_offset(163)
Marat Dukhan50323b82022-01-11 00:12:01 -080025698 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025699 }
25700 }
25701
25702 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, zero) {
25703 TEST_REQUIRES_ARM_NEON;
25704 for (uint32_t mz = 0; mz < 4; mz++) {
25705 for (size_t k = 1; k <= 40; k += 9) {
25706 GemmMicrokernelTester()
25707 .mr(4)
25708 .nr(16)
25709 .kr(1)
25710 .sr(1)
25711 .m(4)
25712 .n(16)
25713 .k(k)
25714 .ks(3)
25715 .a_offset(163)
25716 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080025717 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025718 }
25719 }
25720 }
25721
25722 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, qmin) {
25723 TEST_REQUIRES_ARM_NEON;
25724 GemmMicrokernelTester()
25725 .mr(4)
25726 .nr(16)
25727 .kr(1)
25728 .sr(1)
25729 .m(4)
25730 .n(16)
25731 .k(8)
25732 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080025733 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025734 }
25735
25736 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, qmax) {
25737 TEST_REQUIRES_ARM_NEON;
25738 GemmMicrokernelTester()
25739 .mr(4)
25740 .nr(16)
25741 .kr(1)
25742 .sr(1)
25743 .m(4)
25744 .n(16)
25745 .k(8)
25746 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080025747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025748 }
25749
25750 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_LD64, strided_cm) {
25751 TEST_REQUIRES_ARM_NEON;
25752 GemmMicrokernelTester()
25753 .mr(4)
25754 .nr(16)
25755 .kr(1)
25756 .sr(1)
25757 .m(4)
25758 .n(16)
25759 .k(8)
25760 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080025761 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard5cffb642021-11-22 13:59:43 -080025762 }
Frank Barcharde31f29e2021-12-21 15:57:10 -080025763#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
Frank Barchard5cffb642021-11-22 13:59:43 -080025764
25765
Frank Barcharde31f29e2021-12-21 15:57:10 -080025766#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
Frank Barchard13db60f2021-07-20 14:34:35 -070025767 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16) {
25768 TEST_REQUIRES_ARM_NEON_DOT;
25769 GemmMicrokernelTester()
25770 .mr(4)
25771 .nr(16)
25772 .kr(4)
25773 .sr(1)
25774 .m(4)
25775 .n(16)
25776 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -080025777 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025778 }
25779
25780 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, strided_cn) {
25781 TEST_REQUIRES_ARM_NEON_DOT;
25782 GemmMicrokernelTester()
25783 .mr(4)
25784 .nr(16)
25785 .kr(4)
25786 .sr(1)
25787 .m(4)
25788 .n(16)
25789 .k(16)
25790 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080025791 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025792 }
25793
25794 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_subtile) {
25795 TEST_REQUIRES_ARM_NEON_DOT;
25796 for (uint32_t m = 1; m <= 4; m++) {
25797 for (uint32_t n = 1; n <= 16; n++) {
25798 GemmMicrokernelTester()
25799 .mr(4)
25800 .nr(16)
25801 .kr(4)
25802 .sr(1)
25803 .m(m)
25804 .n(n)
25805 .k(16)
25806 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025807 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025808 }
25809 }
25810 }
25811
25812 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_subtile_m) {
25813 TEST_REQUIRES_ARM_NEON_DOT;
25814 for (uint32_t m = 1; m <= 4; m++) {
25815 GemmMicrokernelTester()
25816 .mr(4)
25817 .nr(16)
25818 .kr(4)
25819 .sr(1)
25820 .m(m)
25821 .n(16)
25822 .k(16)
25823 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025824 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025825 }
25826 }
25827
25828 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_subtile_n) {
25829 TEST_REQUIRES_ARM_NEON_DOT;
25830 for (uint32_t n = 1; n <= 16; n++) {
25831 GemmMicrokernelTester()
25832 .mr(4)
25833 .nr(16)
25834 .kr(4)
25835 .sr(1)
25836 .m(4)
25837 .n(n)
25838 .k(16)
25839 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025840 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025841 }
25842 }
25843
25844 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_lt_16) {
25845 TEST_REQUIRES_ARM_NEON_DOT;
25846 for (size_t k = 1; k < 16; k++) {
25847 GemmMicrokernelTester()
25848 .mr(4)
25849 .nr(16)
25850 .kr(4)
25851 .sr(1)
25852 .m(4)
25853 .n(16)
25854 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025855 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025856 }
25857 }
25858
25859 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_lt_16_subtile) {
25860 TEST_REQUIRES_ARM_NEON_DOT;
25861 for (size_t k = 1; k < 16; k++) {
25862 for (uint32_t m = 1; m <= 4; m++) {
25863 for (uint32_t n = 1; n <= 16; n++) {
25864 GemmMicrokernelTester()
25865 .mr(4)
25866 .nr(16)
25867 .kr(4)
25868 .sr(1)
25869 .m(m)
25870 .n(n)
25871 .k(k)
25872 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025873 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025874 }
25875 }
25876 }
25877 }
25878
25879 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_gt_16) {
25880 TEST_REQUIRES_ARM_NEON_DOT;
25881 for (size_t k = 17; k < 32; k++) {
25882 GemmMicrokernelTester()
25883 .mr(4)
25884 .nr(16)
25885 .kr(4)
25886 .sr(1)
25887 .m(4)
25888 .n(16)
25889 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025890 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025891 }
25892 }
25893
25894 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_gt_16_subtile) {
25895 TEST_REQUIRES_ARM_NEON_DOT;
25896 for (size_t k = 17; k < 32; k++) {
25897 for (uint32_t m = 1; m <= 4; m++) {
25898 for (uint32_t n = 1; n <= 16; n++) {
25899 GemmMicrokernelTester()
25900 .mr(4)
25901 .nr(16)
25902 .kr(4)
25903 .sr(1)
25904 .m(m)
25905 .n(n)
25906 .k(k)
25907 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025908 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025909 }
25910 }
25911 }
25912 }
25913
25914 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_div_16) {
25915 TEST_REQUIRES_ARM_NEON_DOT;
25916 for (size_t k = 32; k <= 160; k += 16) {
25917 GemmMicrokernelTester()
25918 .mr(4)
25919 .nr(16)
25920 .kr(4)
25921 .sr(1)
25922 .m(4)
25923 .n(16)
25924 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025925 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025926 }
25927 }
25928
25929 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_div_16_subtile) {
25930 TEST_REQUIRES_ARM_NEON_DOT;
25931 for (size_t k = 32; k <= 160; k += 16) {
25932 for (uint32_t m = 1; m <= 4; m++) {
25933 for (uint32_t n = 1; n <= 16; n++) {
25934 GemmMicrokernelTester()
25935 .mr(4)
25936 .nr(16)
25937 .kr(4)
25938 .sr(1)
25939 .m(m)
25940 .n(n)
25941 .k(k)
25942 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025943 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025944 }
25945 }
25946 }
25947 }
25948
25949 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16) {
25950 TEST_REQUIRES_ARM_NEON_DOT;
25951 for (uint32_t n = 17; n < 32; n++) {
25952 for (size_t k = 1; k <= 80; k += 17) {
25953 GemmMicrokernelTester()
25954 .mr(4)
25955 .nr(16)
25956 .kr(4)
25957 .sr(1)
25958 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025959 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070025960 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080025961 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025962 }
25963 }
25964 }
25965
25966 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16_strided_cn) {
25967 TEST_REQUIRES_ARM_NEON_DOT;
25968 for (uint32_t n = 17; n < 32; n++) {
25969 for (size_t k = 1; k <= 80; k += 17) {
25970 GemmMicrokernelTester()
25971 .mr(4)
25972 .nr(16)
25973 .kr(4)
25974 .sr(1)
25975 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080025976 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070025977 .k(k)
25978 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080025979 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025980 }
25981 }
25982 }
25983
25984 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16_subtile) {
25985 TEST_REQUIRES_ARM_NEON_DOT;
25986 for (uint32_t n = 17; n < 32; n++) {
25987 for (size_t k = 1; k <= 80; k += 17) {
25988 for (uint32_t m = 1; m <= 4; m++) {
25989 GemmMicrokernelTester()
25990 .mr(4)
25991 .nr(16)
25992 .kr(4)
25993 .sr(1)
25994 .m(m)
25995 .n(n)
25996 .k(k)
25997 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080025998 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070025999 }
26000 }
26001 }
26002 }
26003
26004 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16) {
26005 TEST_REQUIRES_ARM_NEON_DOT;
26006 for (uint32_t n = 32; n <= 48; n += 16) {
26007 for (size_t k = 1; k <= 80; k += 17) {
26008 GemmMicrokernelTester()
26009 .mr(4)
26010 .nr(16)
26011 .kr(4)
26012 .sr(1)
26013 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026014 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070026015 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080026016 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026017 }
26018 }
26019 }
26020
26021 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16_strided_cn) {
26022 TEST_REQUIRES_ARM_NEON_DOT;
26023 for (uint32_t n = 32; n <= 48; n += 16) {
26024 for (size_t k = 1; k <= 80; k += 17) {
26025 GemmMicrokernelTester()
26026 .mr(4)
26027 .nr(16)
26028 .kr(4)
26029 .sr(1)
26030 .m(4)
26031 .n(n)
26032 .k(k)
26033 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080026034 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026035 }
26036 }
26037 }
26038
26039 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16_subtile) {
26040 TEST_REQUIRES_ARM_NEON_DOT;
26041 for (uint32_t n = 32; n <= 48; n += 16) {
26042 for (size_t k = 1; k <= 80; k += 17) {
26043 for (uint32_t m = 1; m <= 4; m++) {
26044 GemmMicrokernelTester()
26045 .mr(4)
26046 .nr(16)
26047 .kr(4)
26048 .sr(1)
26049 .m(m)
26050 .n(n)
26051 .k(k)
26052 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026053 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026054 }
26055 }
26056 }
26057 }
26058
26059 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, small_kernel) {
26060 TEST_REQUIRES_ARM_NEON_DOT;
26061 for (size_t k = 1; k <= 80; k += 17) {
26062 GemmMicrokernelTester()
26063 .mr(4)
26064 .nr(16)
26065 .kr(4)
26066 .sr(1)
26067 .m(4)
26068 .n(16)
26069 .k(k)
26070 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080026071 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026072 }
26073 }
26074
26075 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, small_kernel_subtile) {
26076 TEST_REQUIRES_ARM_NEON_DOT;
26077 for (size_t k = 1; k <= 80; k += 17) {
26078 for (uint32_t m = 1; m <= 4; m++) {
26079 for (uint32_t n = 1; n <= 16; n++) {
26080 GemmMicrokernelTester()
26081 .mr(4)
26082 .nr(16)
26083 .kr(4)
26084 .sr(1)
26085 .m(m)
26086 .n(n)
26087 .k(k)
26088 .ks(3)
26089 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026090 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026091 }
26092 }
26093 }
26094 }
26095
26096 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16_small_kernel) {
26097 TEST_REQUIRES_ARM_NEON_DOT;
26098 for (uint32_t n = 17; n < 32; n++) {
26099 for (size_t k = 1; k <= 80; k += 17) {
26100 GemmMicrokernelTester()
26101 .mr(4)
26102 .nr(16)
26103 .kr(4)
26104 .sr(1)
26105 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026106 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070026107 .k(k)
26108 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080026109 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026110 }
26111 }
26112 }
26113
26114 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16_small_kernel) {
26115 TEST_REQUIRES_ARM_NEON_DOT;
26116 for (uint32_t n = 32; n <= 48; n += 16) {
26117 for (size_t k = 1; k <= 80; k += 17) {
26118 GemmMicrokernelTester()
26119 .mr(4)
26120 .nr(16)
26121 .kr(4)
26122 .sr(1)
26123 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026124 .n(n)
Frank Barchard13db60f2021-07-20 14:34:35 -070026125 .k(k)
26126 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080026127 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026128 }
26129 }
26130 }
26131
26132 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, strided_cm_subtile) {
26133 TEST_REQUIRES_ARM_NEON_DOT;
26134 for (size_t k = 1; k <= 80; k += 17) {
26135 for (uint32_t m = 1; m <= 4; m++) {
26136 for (uint32_t n = 1; n <= 16; n++) {
26137 GemmMicrokernelTester()
26138 .mr(4)
26139 .nr(16)
26140 .kr(4)
26141 .sr(1)
26142 .m(m)
26143 .n(n)
26144 .k(k)
26145 .cm_stride(19)
26146 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026147 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026148 }
26149 }
26150 }
26151 }
26152
26153 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, a_offset) {
26154 TEST_REQUIRES_ARM_NEON_DOT;
26155 for (size_t k = 1; k <= 80; k += 17) {
26156 GemmMicrokernelTester()
26157 .mr(4)
26158 .nr(16)
26159 .kr(4)
26160 .sr(1)
26161 .m(4)
26162 .n(16)
26163 .k(k)
26164 .ks(3)
26165 .a_offset(331)
Marat Dukhan50323b82022-01-11 00:12:01 -080026166 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026167 }
26168 }
26169
26170 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, zero) {
26171 TEST_REQUIRES_ARM_NEON_DOT;
26172 for (uint32_t mz = 0; mz < 4; mz++) {
26173 for (size_t k = 1; k <= 80; k += 17) {
26174 GemmMicrokernelTester()
26175 .mr(4)
26176 .nr(16)
26177 .kr(4)
26178 .sr(1)
26179 .m(4)
26180 .n(16)
26181 .k(k)
26182 .ks(3)
26183 .a_offset(331)
26184 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080026185 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026186 }
26187 }
26188 }
26189
26190 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, qmin) {
26191 TEST_REQUIRES_ARM_NEON_DOT;
26192 GemmMicrokernelTester()
26193 .mr(4)
26194 .nr(16)
26195 .kr(4)
26196 .sr(1)
26197 .m(4)
26198 .n(16)
26199 .k(16)
26200 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080026201 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026202 }
26203
26204 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, qmax) {
26205 TEST_REQUIRES_ARM_NEON_DOT;
26206 GemmMicrokernelTester()
26207 .mr(4)
26208 .nr(16)
26209 .kr(4)
26210 .sr(1)
26211 .m(4)
26212 .n(16)
26213 .k(16)
26214 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080026215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026216 }
26217
26218 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, strided_cm) {
26219 TEST_REQUIRES_ARM_NEON_DOT;
26220 GemmMicrokernelTester()
26221 .mr(4)
26222 .nr(16)
26223 .kr(4)
26224 .sr(1)
26225 .m(4)
26226 .n(16)
26227 .k(16)
26228 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080026229 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard13db60f2021-07-20 14:34:35 -070026230 }
Frank Barcharde31f29e2021-12-21 15:57:10 -080026231#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
Frank Barchard13db60f2021-07-20 14:34:35 -070026232
26233
Marat Dukhane903dff2021-07-16 19:43:41 -070026234#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Marat Dukhan89991902021-12-06 00:54:36 -080026235 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_eq_8) {
26236 TEST_REQUIRES_ARM_NEON;
26237 GemmMicrokernelTester()
26238 .mr(2)
26239 .nr(8)
26240 .kr(8)
26241 .sr(1)
26242 .m(2)
26243 .n(8)
26244 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080026245 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026246 }
26247
26248 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, strided_cn) {
26249 TEST_REQUIRES_ARM_NEON;
26250 GemmMicrokernelTester()
26251 .mr(2)
26252 .nr(8)
26253 .kr(8)
26254 .sr(1)
26255 .m(2)
26256 .n(8)
26257 .k(8)
26258 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080026259 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026260 }
26261
26262 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_eq_8_subtile) {
26263 TEST_REQUIRES_ARM_NEON;
26264 for (uint32_t m = 1; m <= 2; m++) {
26265 for (uint32_t n = 1; n <= 8; n++) {
26266 GemmMicrokernelTester()
26267 .mr(2)
26268 .nr(8)
26269 .kr(8)
26270 .sr(1)
26271 .m(m)
26272 .n(n)
26273 .k(8)
26274 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026275 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026276 }
26277 }
26278 }
26279
26280 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_eq_8_subtile_m) {
26281 TEST_REQUIRES_ARM_NEON;
26282 for (uint32_t m = 1; m <= 2; m++) {
26283 GemmMicrokernelTester()
26284 .mr(2)
26285 .nr(8)
26286 .kr(8)
26287 .sr(1)
26288 .m(m)
26289 .n(8)
26290 .k(8)
26291 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026292 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026293 }
26294 }
26295
26296 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_eq_8_subtile_n) {
26297 TEST_REQUIRES_ARM_NEON;
26298 for (uint32_t n = 1; n <= 8; n++) {
26299 GemmMicrokernelTester()
26300 .mr(2)
26301 .nr(8)
26302 .kr(8)
26303 .sr(1)
26304 .m(2)
26305 .n(n)
26306 .k(8)
26307 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026308 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026309 }
26310 }
26311
26312 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_lt_8) {
26313 TEST_REQUIRES_ARM_NEON;
26314 for (size_t k = 1; k < 8; k++) {
26315 GemmMicrokernelTester()
26316 .mr(2)
26317 .nr(8)
26318 .kr(8)
26319 .sr(1)
26320 .m(2)
26321 .n(8)
26322 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080026323 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026324 }
26325 }
26326
26327 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_lt_8_subtile) {
26328 TEST_REQUIRES_ARM_NEON;
26329 for (size_t k = 1; k < 8; k++) {
26330 for (uint32_t m = 1; m <= 2; m++) {
26331 for (uint32_t n = 1; n <= 8; n++) {
26332 GemmMicrokernelTester()
26333 .mr(2)
26334 .nr(8)
26335 .kr(8)
26336 .sr(1)
26337 .m(m)
26338 .n(n)
26339 .k(k)
26340 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026341 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026342 }
26343 }
26344 }
26345 }
26346
26347 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_gt_8) {
26348 TEST_REQUIRES_ARM_NEON;
26349 for (size_t k = 9; k < 16; k++) {
26350 GemmMicrokernelTester()
26351 .mr(2)
26352 .nr(8)
26353 .kr(8)
26354 .sr(1)
26355 .m(2)
26356 .n(8)
26357 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080026358 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026359 }
26360 }
26361
26362 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_gt_8_subtile) {
26363 TEST_REQUIRES_ARM_NEON;
26364 for (size_t k = 9; k < 16; k++) {
26365 for (uint32_t m = 1; m <= 2; m++) {
26366 for (uint32_t n = 1; n <= 8; n++) {
26367 GemmMicrokernelTester()
26368 .mr(2)
26369 .nr(8)
26370 .kr(8)
26371 .sr(1)
26372 .m(m)
26373 .n(n)
26374 .k(k)
26375 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026376 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026377 }
26378 }
26379 }
26380 }
26381
26382 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_div_8) {
26383 TEST_REQUIRES_ARM_NEON;
26384 for (size_t k = 16; k <= 80; k += 8) {
26385 GemmMicrokernelTester()
26386 .mr(2)
26387 .nr(8)
26388 .kr(8)
26389 .sr(1)
26390 .m(2)
26391 .n(8)
26392 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080026393 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026394 }
26395 }
26396
26397 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_div_8_subtile) {
26398 TEST_REQUIRES_ARM_NEON;
26399 for (size_t k = 16; k <= 80; k += 8) {
26400 for (uint32_t m = 1; m <= 2; m++) {
26401 for (uint32_t n = 1; n <= 8; n++) {
26402 GemmMicrokernelTester()
26403 .mr(2)
26404 .nr(8)
26405 .kr(8)
26406 .sr(1)
26407 .m(m)
26408 .n(n)
26409 .k(k)
26410 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026411 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026412 }
26413 }
26414 }
26415 }
26416
26417 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_gt_8) {
26418 TEST_REQUIRES_ARM_NEON;
26419 for (uint32_t n = 9; n < 16; n++) {
26420 for (size_t k = 1; k <= 40; k += 9) {
26421 GemmMicrokernelTester()
26422 .mr(2)
26423 .nr(8)
26424 .kr(8)
26425 .sr(1)
26426 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026427 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080026428 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080026429 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026430 }
26431 }
26432 }
26433
26434 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_gt_8_strided_cn) {
26435 TEST_REQUIRES_ARM_NEON;
26436 for (uint32_t n = 9; n < 16; n++) {
26437 for (size_t k = 1; k <= 40; k += 9) {
26438 GemmMicrokernelTester()
26439 .mr(2)
26440 .nr(8)
26441 .kr(8)
26442 .sr(1)
26443 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026444 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080026445 .k(k)
26446 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080026447 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026448 }
26449 }
26450 }
26451
26452 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_gt_8_subtile) {
26453 TEST_REQUIRES_ARM_NEON;
26454 for (uint32_t n = 9; n < 16; n++) {
26455 for (size_t k = 1; k <= 40; k += 9) {
26456 for (uint32_t m = 1; m <= 2; m++) {
26457 GemmMicrokernelTester()
26458 .mr(2)
26459 .nr(8)
26460 .kr(8)
26461 .sr(1)
26462 .m(m)
26463 .n(n)
26464 .k(k)
26465 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026466 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026467 }
26468 }
26469 }
26470 }
26471
26472 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_div_8) {
26473 TEST_REQUIRES_ARM_NEON;
26474 for (uint32_t n = 16; n <= 24; n += 8) {
26475 for (size_t k = 1; k <= 40; k += 9) {
26476 GemmMicrokernelTester()
26477 .mr(2)
26478 .nr(8)
26479 .kr(8)
26480 .sr(1)
26481 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026482 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080026483 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080026484 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026485 }
26486 }
26487 }
26488
26489 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_div_8_strided_cn) {
26490 TEST_REQUIRES_ARM_NEON;
26491 for (uint32_t n = 16; n <= 24; n += 8) {
26492 for (size_t k = 1; k <= 40; k += 9) {
26493 GemmMicrokernelTester()
26494 .mr(2)
26495 .nr(8)
26496 .kr(8)
26497 .sr(1)
26498 .m(2)
26499 .n(n)
26500 .k(k)
26501 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080026502 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026503 }
26504 }
26505 }
26506
26507 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_div_8_subtile) {
26508 TEST_REQUIRES_ARM_NEON;
26509 for (uint32_t n = 16; n <= 24; n += 8) {
26510 for (size_t k = 1; k <= 40; k += 9) {
26511 for (uint32_t m = 1; m <= 2; m++) {
26512 GemmMicrokernelTester()
26513 .mr(2)
26514 .nr(8)
26515 .kr(8)
26516 .sr(1)
26517 .m(m)
26518 .n(n)
26519 .k(k)
26520 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026521 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026522 }
26523 }
26524 }
26525 }
26526
26527 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, small_kernel) {
26528 TEST_REQUIRES_ARM_NEON;
26529 for (size_t k = 1; k <= 40; k += 9) {
26530 GemmMicrokernelTester()
26531 .mr(2)
26532 .nr(8)
26533 .kr(8)
26534 .sr(1)
26535 .m(2)
26536 .n(8)
26537 .k(k)
26538 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080026539 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026540 }
26541 }
26542
26543 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, small_kernel_subtile) {
26544 TEST_REQUIRES_ARM_NEON;
26545 for (size_t k = 1; k <= 40; k += 9) {
26546 for (uint32_t m = 1; m <= 2; m++) {
26547 for (uint32_t n = 1; n <= 8; n++) {
26548 GemmMicrokernelTester()
26549 .mr(2)
26550 .nr(8)
26551 .kr(8)
26552 .sr(1)
26553 .m(m)
26554 .n(n)
26555 .k(k)
26556 .ks(3)
26557 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026558 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026559 }
26560 }
26561 }
26562 }
26563
26564 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_gt_8_small_kernel) {
26565 TEST_REQUIRES_ARM_NEON;
26566 for (uint32_t n = 9; n < 16; n++) {
26567 for (size_t k = 1; k <= 40; k += 9) {
26568 GemmMicrokernelTester()
26569 .mr(2)
26570 .nr(8)
26571 .kr(8)
26572 .sr(1)
26573 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026574 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080026575 .k(k)
26576 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080026577 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026578 }
26579 }
26580 }
26581
26582 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_div_8_small_kernel) {
26583 TEST_REQUIRES_ARM_NEON;
26584 for (uint32_t n = 16; n <= 24; n += 8) {
26585 for (size_t k = 1; k <= 40; k += 9) {
26586 GemmMicrokernelTester()
26587 .mr(2)
26588 .nr(8)
26589 .kr(8)
26590 .sr(1)
26591 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026592 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080026593 .k(k)
26594 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080026595 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026596 }
26597 }
26598 }
26599
26600 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, strided_cm_subtile) {
26601 TEST_REQUIRES_ARM_NEON;
26602 for (size_t k = 1; k <= 40; k += 9) {
26603 for (uint32_t m = 1; m <= 2; m++) {
26604 for (uint32_t n = 1; n <= 8; n++) {
26605 GemmMicrokernelTester()
26606 .mr(2)
26607 .nr(8)
26608 .kr(8)
26609 .sr(1)
26610 .m(m)
26611 .n(n)
26612 .k(k)
26613 .cm_stride(11)
26614 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080026615 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026616 }
26617 }
26618 }
26619 }
26620
26621 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, a_offset) {
26622 TEST_REQUIRES_ARM_NEON;
26623 for (size_t k = 1; k <= 40; k += 9) {
26624 GemmMicrokernelTester()
26625 .mr(2)
26626 .nr(8)
26627 .kr(8)
26628 .sr(1)
26629 .m(2)
26630 .n(8)
26631 .k(k)
26632 .ks(3)
26633 .a_offset(83)
Marat Dukhan50323b82022-01-11 00:12:01 -080026634 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026635 }
26636 }
26637
26638 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, zero) {
26639 TEST_REQUIRES_ARM_NEON;
26640 for (uint32_t mz = 0; mz < 2; mz++) {
26641 for (size_t k = 1; k <= 40; k += 9) {
26642 GemmMicrokernelTester()
26643 .mr(2)
26644 .nr(8)
26645 .kr(8)
26646 .sr(1)
26647 .m(2)
26648 .n(8)
26649 .k(k)
26650 .ks(3)
26651 .a_offset(83)
26652 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080026653 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026654 }
26655 }
26656 }
26657
26658 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, qmin) {
26659 TEST_REQUIRES_ARM_NEON;
26660 GemmMicrokernelTester()
26661 .mr(2)
26662 .nr(8)
26663 .kr(8)
26664 .sr(1)
26665 .m(2)
26666 .n(8)
26667 .k(8)
26668 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080026669 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026670 }
26671
26672 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, qmax) {
26673 TEST_REQUIRES_ARM_NEON;
26674 GemmMicrokernelTester()
26675 .mr(2)
26676 .nr(8)
26677 .kr(8)
26678 .sr(1)
26679 .m(2)
26680 .n(8)
26681 .k(8)
26682 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080026683 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026684 }
26685
26686 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, strided_cm) {
26687 TEST_REQUIRES_ARM_NEON;
26688 GemmMicrokernelTester()
26689 .mr(2)
26690 .nr(8)
26691 .kr(8)
26692 .sr(1)
26693 .m(2)
26694 .n(8)
26695 .k(8)
26696 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080026697 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026698 }
26699#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
26700
26701
26702#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026703 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080026704 TEST_REQUIRES_ARM_NEON;
26705 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026706 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026707 .nr(8)
26708 .kr(8)
26709 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026710 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026711 .n(8)
26712 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026713 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026714 }
26715
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026716 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, strided_cn) {
Marat Dukhan89991902021-12-06 00:54:36 -080026717 TEST_REQUIRES_ARM_NEON;
26718 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026719 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026720 .nr(8)
26721 .kr(8)
26722 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026723 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026724 .n(8)
26725 .k(8)
26726 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026727 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026728 }
26729
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026730 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080026731 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026732 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080026733 for (uint32_t n = 1; n <= 8; n++) {
26734 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026735 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026736 .nr(8)
26737 .kr(8)
26738 .sr(1)
26739 .m(m)
26740 .n(n)
26741 .k(8)
26742 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026743 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026744 }
26745 }
26746 }
26747
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026748 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_subtile_m) {
Marat Dukhan89991902021-12-06 00:54:36 -080026749 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026750 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080026751 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026752 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026753 .nr(8)
26754 .kr(8)
26755 .sr(1)
26756 .m(m)
26757 .n(8)
26758 .k(8)
26759 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026760 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026761 }
26762 }
26763
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026764 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_subtile_n) {
Marat Dukhan89991902021-12-06 00:54:36 -080026765 TEST_REQUIRES_ARM_NEON;
26766 for (uint32_t n = 1; n <= 8; n++) {
26767 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026768 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026769 .nr(8)
26770 .kr(8)
26771 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026772 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026773 .n(n)
26774 .k(8)
26775 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026776 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026777 }
26778 }
26779
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026780 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_lt_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080026781 TEST_REQUIRES_ARM_NEON;
26782 for (size_t k = 1; k < 8; k++) {
26783 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026784 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026785 .nr(8)
26786 .kr(8)
26787 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026788 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026789 .n(8)
26790 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026791 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026792 }
26793 }
26794
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026795 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_lt_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080026796 TEST_REQUIRES_ARM_NEON;
26797 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026798 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080026799 for (uint32_t n = 1; n <= 8; n++) {
26800 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026801 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026802 .nr(8)
26803 .kr(8)
26804 .sr(1)
26805 .m(m)
26806 .n(n)
26807 .k(k)
26808 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026809 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026810 }
26811 }
26812 }
26813 }
26814
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026815 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_gt_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080026816 TEST_REQUIRES_ARM_NEON;
26817 for (size_t k = 9; k < 16; k++) {
26818 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026819 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026820 .nr(8)
26821 .kr(8)
26822 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026823 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026824 .n(8)
26825 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026826 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026827 }
26828 }
26829
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026830 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_gt_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080026831 TEST_REQUIRES_ARM_NEON;
26832 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026833 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080026834 for (uint32_t n = 1; n <= 8; n++) {
26835 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026836 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026837 .nr(8)
26838 .kr(8)
26839 .sr(1)
26840 .m(m)
26841 .n(n)
26842 .k(k)
26843 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026844 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026845 }
26846 }
26847 }
26848 }
26849
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026850 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_div_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080026851 TEST_REQUIRES_ARM_NEON;
26852 for (size_t k = 16; k <= 80; k += 8) {
26853 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026854 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026855 .nr(8)
26856 .kr(8)
26857 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026858 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026859 .n(8)
26860 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026861 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026862 }
26863 }
26864
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026865 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_div_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080026866 TEST_REQUIRES_ARM_NEON;
26867 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026868 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080026869 for (uint32_t n = 1; n <= 8; n++) {
26870 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026871 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026872 .nr(8)
26873 .kr(8)
26874 .sr(1)
26875 .m(m)
26876 .n(n)
26877 .k(k)
26878 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026879 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026880 }
26881 }
26882 }
26883 }
26884
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026885 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080026886 TEST_REQUIRES_ARM_NEON;
26887 for (uint32_t n = 9; n < 16; n++) {
26888 for (size_t k = 1; k <= 40; k += 9) {
26889 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026890 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026891 .nr(8)
26892 .kr(8)
26893 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026894 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026895 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080026896 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026897 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026898 }
26899 }
26900 }
26901
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026902 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8_strided_cn) {
Marat Dukhan89991902021-12-06 00:54:36 -080026903 TEST_REQUIRES_ARM_NEON;
26904 for (uint32_t n = 9; n < 16; n++) {
26905 for (size_t k = 1; k <= 40; k += 9) {
26906 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026907 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026908 .nr(8)
26909 .kr(8)
26910 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026911 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026912 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080026913 .k(k)
26914 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026915 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026916 }
26917 }
26918 }
26919
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026920 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080026921 TEST_REQUIRES_ARM_NEON;
26922 for (uint32_t n = 9; n < 16; n++) {
26923 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026924 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080026925 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026926 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026927 .nr(8)
26928 .kr(8)
26929 .sr(1)
26930 .m(m)
26931 .n(n)
26932 .k(k)
26933 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026934 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026935 }
26936 }
26937 }
26938 }
26939
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026940 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080026941 TEST_REQUIRES_ARM_NEON;
26942 for (uint32_t n = 16; n <= 24; n += 8) {
26943 for (size_t k = 1; k <= 40; k += 9) {
26944 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026945 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026946 .nr(8)
26947 .kr(8)
26948 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026949 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080026950 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080026951 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026952 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026953 }
26954 }
26955 }
26956
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026957 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8_strided_cn) {
Marat Dukhan89991902021-12-06 00:54:36 -080026958 TEST_REQUIRES_ARM_NEON;
26959 for (uint32_t n = 16; n <= 24; n += 8) {
26960 for (size_t k = 1; k <= 40; k += 9) {
26961 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026962 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026963 .nr(8)
26964 .kr(8)
26965 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026966 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026967 .n(n)
26968 .k(k)
26969 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026970 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026971 }
26972 }
26973 }
26974
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026975 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080026976 TEST_REQUIRES_ARM_NEON;
26977 for (uint32_t n = 16; n <= 24; n += 8) {
26978 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026979 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080026980 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026981 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080026982 .nr(8)
26983 .kr(8)
26984 .sr(1)
26985 .m(m)
26986 .n(n)
26987 .k(k)
26988 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026989 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080026990 }
26991 }
26992 }
26993 }
26994
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026995 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, small_kernel) {
Marat Dukhan89991902021-12-06 00:54:36 -080026996 TEST_REQUIRES_ARM_NEON;
26997 for (size_t k = 1; k <= 40; k += 9) {
26998 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080026999 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027000 .nr(8)
27001 .kr(8)
27002 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027003 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027004 .n(8)
27005 .k(k)
27006 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027007 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027008 }
27009 }
27010
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027011 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, small_kernel_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027012 TEST_REQUIRES_ARM_NEON;
27013 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027014 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027015 for (uint32_t n = 1; n <= 8; n++) {
27016 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027017 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027018 .nr(8)
27019 .kr(8)
27020 .sr(1)
27021 .m(m)
27022 .n(n)
27023 .k(k)
27024 .ks(3)
27025 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027026 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027027 }
27028 }
27029 }
27030 }
27031
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027032 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8_small_kernel) {
Marat Dukhan89991902021-12-06 00:54:36 -080027033 TEST_REQUIRES_ARM_NEON;
27034 for (uint32_t n = 9; n < 16; n++) {
27035 for (size_t k = 1; k <= 40; k += 9) {
27036 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027037 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027038 .nr(8)
27039 .kr(8)
27040 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027041 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027042 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080027043 .k(k)
27044 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027045 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027046 }
27047 }
27048 }
27049
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027050 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8_small_kernel) {
Marat Dukhan89991902021-12-06 00:54:36 -080027051 TEST_REQUIRES_ARM_NEON;
27052 for (uint32_t n = 16; n <= 24; n += 8) {
27053 for (size_t k = 1; k <= 40; k += 9) {
27054 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027055 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027056 .nr(8)
27057 .kr(8)
27058 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027059 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027060 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080027061 .k(k)
27062 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027063 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027064 }
27065 }
27066 }
27067
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027068 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, strided_cm_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027069 TEST_REQUIRES_ARM_NEON;
27070 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027071 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027072 for (uint32_t n = 1; n <= 8; n++) {
27073 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027074 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027075 .nr(8)
27076 .kr(8)
27077 .sr(1)
27078 .m(m)
27079 .n(n)
27080 .k(k)
27081 .cm_stride(11)
27082 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027083 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027084 }
27085 }
27086 }
27087 }
27088
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027089 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, a_offset) {
Marat Dukhan89991902021-12-06 00:54:36 -080027090 TEST_REQUIRES_ARM_NEON;
27091 for (size_t k = 1; k <= 40; k += 9) {
27092 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027093 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027094 .nr(8)
27095 .kr(8)
27096 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027097 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027098 .n(8)
27099 .k(k)
27100 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027101 .a_offset(127)
27102 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027103 }
27104 }
27105
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027106 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, zero) {
Marat Dukhan89991902021-12-06 00:54:36 -080027107 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027108 for (uint32_t mz = 0; mz < 3; mz++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027109 for (size_t k = 1; k <= 40; k += 9) {
27110 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027111 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027112 .nr(8)
27113 .kr(8)
27114 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027115 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027116 .n(8)
27117 .k(k)
27118 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027119 .a_offset(127)
Marat Dukhan89991902021-12-06 00:54:36 -080027120 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027121 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027122 }
27123 }
27124 }
27125
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027126 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, qmin) {
Marat Dukhan89991902021-12-06 00:54:36 -080027127 TEST_REQUIRES_ARM_NEON;
27128 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027129 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027130 .nr(8)
27131 .kr(8)
27132 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027133 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027134 .n(8)
27135 .k(8)
27136 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027137 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027138 }
27139
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027140 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, qmax) {
Marat Dukhan89991902021-12-06 00:54:36 -080027141 TEST_REQUIRES_ARM_NEON;
27142 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027143 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027144 .nr(8)
27145 .kr(8)
27146 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027147 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027148 .n(8)
27149 .k(8)
27150 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027151 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027152 }
27153
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027154 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, strided_cm) {
Marat Dukhan89991902021-12-06 00:54:36 -080027155 TEST_REQUIRES_ARM_NEON;
27156 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027157 .mr(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027158 .nr(8)
27159 .kr(8)
27160 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027161 .m(3)
Marat Dukhan89991902021-12-06 00:54:36 -080027162 .n(8)
27163 .k(8)
27164 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027165 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027166 }
27167#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
27168
27169
27170#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027171 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_eq_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080027172 TEST_REQUIRES_ARM_NEON;
27173 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027174 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027175 .nr(16)
27176 .kr(8)
27177 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027178 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027179 .n(16)
27180 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027181 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027182 }
27183
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027184 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, strided_cn) {
Marat Dukhan89991902021-12-06 00:54:36 -080027185 TEST_REQUIRES_ARM_NEON;
27186 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027187 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027188 .nr(16)
27189 .kr(8)
27190 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027191 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027192 .n(16)
27193 .k(8)
27194 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027195 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027196 }
27197
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027198 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_eq_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027199 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027200 for (uint32_t m = 1; m <= 1; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027201 for (uint32_t n = 1; n <= 16; n++) {
27202 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027203 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027204 .nr(16)
27205 .kr(8)
27206 .sr(1)
27207 .m(m)
27208 .n(n)
27209 .k(8)
27210 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027211 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027212 }
27213 }
27214 }
27215
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027216 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_eq_8_subtile_m) {
Marat Dukhan89991902021-12-06 00:54:36 -080027217 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027218 for (uint32_t m = 1; m <= 1; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027219 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027220 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027221 .nr(16)
27222 .kr(8)
27223 .sr(1)
27224 .m(m)
27225 .n(16)
27226 .k(8)
27227 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027228 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027229 }
27230 }
27231
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027232 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_eq_8_subtile_n) {
Marat Dukhan89991902021-12-06 00:54:36 -080027233 TEST_REQUIRES_ARM_NEON;
27234 for (uint32_t n = 1; n <= 16; n++) {
27235 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027236 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027237 .nr(16)
27238 .kr(8)
27239 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027240 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027241 .n(n)
27242 .k(8)
27243 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027244 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027245 }
27246 }
27247
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027248 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_lt_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080027249 TEST_REQUIRES_ARM_NEON;
27250 for (size_t k = 1; k < 8; k++) {
27251 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027252 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027253 .nr(16)
27254 .kr(8)
27255 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027256 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027257 .n(16)
27258 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027259 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027260 }
27261 }
27262
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027263 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_lt_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027264 TEST_REQUIRES_ARM_NEON;
27265 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027266 for (uint32_t m = 1; m <= 1; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027267 for (uint32_t n = 1; n <= 16; n++) {
27268 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027269 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027270 .nr(16)
27271 .kr(8)
27272 .sr(1)
27273 .m(m)
27274 .n(n)
27275 .k(k)
27276 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027277 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027278 }
27279 }
27280 }
27281 }
27282
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027283 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_gt_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080027284 TEST_REQUIRES_ARM_NEON;
27285 for (size_t k = 9; k < 16; k++) {
27286 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027287 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027288 .nr(16)
27289 .kr(8)
27290 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027291 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027292 .n(16)
27293 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027294 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027295 }
27296 }
27297
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027298 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_gt_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027299 TEST_REQUIRES_ARM_NEON;
27300 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027301 for (uint32_t m = 1; m <= 1; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027302 for (uint32_t n = 1; n <= 16; n++) {
27303 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027304 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027305 .nr(16)
27306 .kr(8)
27307 .sr(1)
27308 .m(m)
27309 .n(n)
27310 .k(k)
27311 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027312 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027313 }
27314 }
27315 }
27316 }
27317
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027318 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_div_8) {
Marat Dukhan89991902021-12-06 00:54:36 -080027319 TEST_REQUIRES_ARM_NEON;
27320 for (size_t k = 16; k <= 80; k += 8) {
27321 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027322 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027323 .nr(16)
27324 .kr(8)
27325 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027326 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027327 .n(16)
27328 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027329 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027330 }
27331 }
27332
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027333 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_div_8_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027334 TEST_REQUIRES_ARM_NEON;
27335 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027336 for (uint32_t m = 1; m <= 1; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027337 for (uint32_t n = 1; n <= 16; n++) {
27338 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027339 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027340 .nr(16)
27341 .kr(8)
27342 .sr(1)
27343 .m(m)
27344 .n(n)
27345 .k(k)
27346 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027347 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027348 }
27349 }
27350 }
27351 }
27352
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027353 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_gt_16) {
Marat Dukhan89991902021-12-06 00:54:36 -080027354 TEST_REQUIRES_ARM_NEON;
27355 for (uint32_t n = 17; n < 32; n++) {
27356 for (size_t k = 1; k <= 40; k += 9) {
27357 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027358 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027359 .nr(16)
27360 .kr(8)
27361 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027362 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027363 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080027364 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027365 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027366 }
27367 }
27368 }
27369
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027370 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_gt_16_strided_cn) {
Marat Dukhan89991902021-12-06 00:54:36 -080027371 TEST_REQUIRES_ARM_NEON;
27372 for (uint32_t n = 17; n < 32; n++) {
27373 for (size_t k = 1; k <= 40; k += 9) {
27374 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027375 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027376 .nr(16)
27377 .kr(8)
27378 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027379 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027380 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080027381 .k(k)
27382 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027383 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027384 }
27385 }
27386 }
27387
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027388 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_gt_16_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027389 TEST_REQUIRES_ARM_NEON;
27390 for (uint32_t n = 17; n < 32; n++) {
27391 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027392 for (uint32_t m = 1; m <= 1; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027393 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027394 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027395 .nr(16)
27396 .kr(8)
27397 .sr(1)
27398 .m(m)
27399 .n(n)
27400 .k(k)
27401 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027402 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027403 }
27404 }
27405 }
27406 }
27407
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027408 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_div_16) {
Marat Dukhan89991902021-12-06 00:54:36 -080027409 TEST_REQUIRES_ARM_NEON;
27410 for (uint32_t n = 32; n <= 48; n += 16) {
27411 for (size_t k = 1; k <= 40; k += 9) {
27412 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027413 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027414 .nr(16)
27415 .kr(8)
27416 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027417 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027418 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080027419 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027420 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027421 }
27422 }
27423 }
27424
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027425 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_div_16_strided_cn) {
Marat Dukhan89991902021-12-06 00:54:36 -080027426 TEST_REQUIRES_ARM_NEON;
27427 for (uint32_t n = 32; n <= 48; n += 16) {
27428 for (size_t k = 1; k <= 40; k += 9) {
27429 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027430 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027431 .nr(16)
27432 .kr(8)
27433 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027434 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027435 .n(n)
27436 .k(k)
27437 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027438 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027439 }
27440 }
27441 }
27442
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027443 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_div_16_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027444 TEST_REQUIRES_ARM_NEON;
27445 for (uint32_t n = 32; n <= 48; n += 16) {
27446 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027447 for (uint32_t m = 1; m <= 1; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027448 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027449 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027450 .nr(16)
27451 .kr(8)
27452 .sr(1)
27453 .m(m)
27454 .n(n)
27455 .k(k)
27456 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027457 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027458 }
27459 }
27460 }
27461 }
27462
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027463 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, small_kernel) {
Marat Dukhan89991902021-12-06 00:54:36 -080027464 TEST_REQUIRES_ARM_NEON;
27465 for (size_t k = 1; k <= 40; k += 9) {
27466 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027467 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027468 .nr(16)
27469 .kr(8)
27470 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027471 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027472 .n(16)
27473 .k(k)
27474 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027475 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027476 }
27477 }
27478
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027479 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, small_kernel_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027480 TEST_REQUIRES_ARM_NEON;
27481 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027482 for (uint32_t m = 1; m <= 1; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027483 for (uint32_t n = 1; n <= 16; n++) {
27484 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027485 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027486 .nr(16)
27487 .kr(8)
27488 .sr(1)
27489 .m(m)
27490 .n(n)
27491 .k(k)
27492 .ks(3)
27493 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027494 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027495 }
27496 }
27497 }
27498 }
27499
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027500 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_gt_16_small_kernel) {
Marat Dukhan89991902021-12-06 00:54:36 -080027501 TEST_REQUIRES_ARM_NEON;
27502 for (uint32_t n = 17; n < 32; n++) {
27503 for (size_t k = 1; k <= 40; k += 9) {
27504 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027505 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027506 .nr(16)
27507 .kr(8)
27508 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027509 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027510 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080027511 .k(k)
27512 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027513 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027514 }
27515 }
27516 }
27517
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027518 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_div_16_small_kernel) {
Marat Dukhan89991902021-12-06 00:54:36 -080027519 TEST_REQUIRES_ARM_NEON;
27520 for (uint32_t n = 32; n <= 48; n += 16) {
27521 for (size_t k = 1; k <= 40; k += 9) {
27522 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027523 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027524 .nr(16)
27525 .kr(8)
27526 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027527 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027528 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080027529 .k(k)
27530 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027531 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027532 }
27533 }
27534 }
27535
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027536 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, strided_cm_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080027537 TEST_REQUIRES_ARM_NEON;
27538 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027539 for (uint32_t m = 1; m <= 1; m++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027540 for (uint32_t n = 1; n <= 16; n++) {
27541 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027542 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027543 .nr(16)
27544 .kr(8)
27545 .sr(1)
27546 .m(m)
27547 .n(n)
27548 .k(k)
27549 .cm_stride(19)
27550 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027551 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027552 }
27553 }
27554 }
27555 }
27556
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027557 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, a_offset) {
Marat Dukhan89991902021-12-06 00:54:36 -080027558 TEST_REQUIRES_ARM_NEON;
27559 for (size_t k = 1; k <= 40; k += 9) {
27560 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027561 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027562 .nr(16)
27563 .kr(8)
27564 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027565 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027566 .n(16)
27567 .k(k)
27568 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027569 .a_offset(43)
27570 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027571 }
27572 }
27573
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027574 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, zero) {
Marat Dukhan89991902021-12-06 00:54:36 -080027575 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027576 for (uint32_t mz = 0; mz < 1; mz++) {
Marat Dukhan89991902021-12-06 00:54:36 -080027577 for (size_t k = 1; k <= 40; k += 9) {
27578 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027579 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027580 .nr(16)
27581 .kr(8)
27582 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027583 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027584 .n(16)
27585 .k(k)
27586 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027587 .a_offset(43)
Marat Dukhan89991902021-12-06 00:54:36 -080027588 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027589 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027590 }
27591 }
27592 }
27593
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027594 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, qmin) {
Marat Dukhan89991902021-12-06 00:54:36 -080027595 TEST_REQUIRES_ARM_NEON;
27596 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027597 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027598 .nr(16)
27599 .kr(8)
27600 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027601 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027602 .n(16)
27603 .k(8)
27604 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027605 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027606 }
27607
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027608 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, qmax) {
Marat Dukhan89991902021-12-06 00:54:36 -080027609 TEST_REQUIRES_ARM_NEON;
27610 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027611 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027612 .nr(16)
27613 .kr(8)
27614 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027615 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027616 .n(16)
27617 .k(8)
27618 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027619 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027620 }
27621
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027622 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, strided_cm) {
Marat Dukhan89991902021-12-06 00:54:36 -080027623 TEST_REQUIRES_ARM_NEON;
27624 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027625 .mr(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027626 .nr(16)
27627 .kr(8)
27628 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027629 .m(1)
Marat Dukhan89991902021-12-06 00:54:36 -080027630 .n(16)
27631 .k(8)
27632 .cm_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027633 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080027634 }
27635#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
27636
27637
27638#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027639 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_eq_16) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027640 TEST_REQUIRES_ARM_NEON;
27641 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027642 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027643 .nr(8)
27644 .kr(8)
27645 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027646 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027647 .n(8)
27648 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027649 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027650 }
27651
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027652 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, strided_cn) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027653 TEST_REQUIRES_ARM_NEON;
27654 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027655 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027656 .nr(8)
27657 .kr(8)
27658 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027659 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027660 .n(8)
27661 .k(16)
27662 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027663 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027664 }
27665
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027666 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_eq_16_subtile) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027667 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027668 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027669 for (uint32_t n = 1; n <= 8; n++) {
27670 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027671 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027672 .nr(8)
27673 .kr(8)
27674 .sr(1)
27675 .m(m)
27676 .n(n)
27677 .k(16)
27678 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027679 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027680 }
27681 }
27682 }
27683
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027684 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_eq_16_subtile_m) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027685 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027686 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027687 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027688 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027689 .nr(8)
27690 .kr(8)
27691 .sr(1)
27692 .m(m)
27693 .n(8)
27694 .k(16)
27695 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027696 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027697 }
27698 }
27699
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027700 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_eq_16_subtile_n) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027701 TEST_REQUIRES_ARM_NEON;
27702 for (uint32_t n = 1; n <= 8; n++) {
27703 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027704 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027705 .nr(8)
27706 .kr(8)
27707 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027708 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027709 .n(n)
27710 .k(16)
27711 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027712 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027713 }
27714 }
27715
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027716 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_lt_16) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027717 TEST_REQUIRES_ARM_NEON;
27718 for (size_t k = 1; k < 16; k++) {
27719 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027720 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027721 .nr(8)
27722 .kr(8)
27723 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027724 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027725 .n(8)
27726 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027727 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027728 }
27729 }
27730
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027731 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_lt_16_subtile) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027732 TEST_REQUIRES_ARM_NEON;
27733 for (size_t k = 1; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027734 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027735 for (uint32_t n = 1; n <= 8; n++) {
27736 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027737 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027738 .nr(8)
27739 .kr(8)
27740 .sr(1)
27741 .m(m)
27742 .n(n)
27743 .k(k)
27744 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027745 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027746 }
27747 }
27748 }
27749 }
27750
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027751 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_gt_16) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027752 TEST_REQUIRES_ARM_NEON;
27753 for (size_t k = 17; k < 32; k++) {
27754 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027755 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027756 .nr(8)
27757 .kr(8)
27758 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027759 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027760 .n(8)
27761 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027762 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027763 }
27764 }
27765
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027766 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_gt_16_subtile) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027767 TEST_REQUIRES_ARM_NEON;
27768 for (size_t k = 17; k < 32; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027769 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027770 for (uint32_t n = 1; n <= 8; n++) {
27771 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027772 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027773 .nr(8)
27774 .kr(8)
27775 .sr(1)
27776 .m(m)
27777 .n(n)
27778 .k(k)
27779 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027780 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027781 }
27782 }
27783 }
27784 }
27785
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027786 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_div_16) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027787 TEST_REQUIRES_ARM_NEON;
27788 for (size_t k = 32; k <= 160; k += 16) {
27789 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027790 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027791 .nr(8)
27792 .kr(8)
27793 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027794 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027795 .n(8)
27796 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027797 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027798 }
27799 }
27800
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027801 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, k_div_16_subtile) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027802 TEST_REQUIRES_ARM_NEON;
27803 for (size_t k = 32; k <= 160; k += 16) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027804 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027805 for (uint32_t n = 1; n <= 8; n++) {
27806 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027807 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027808 .nr(8)
27809 .kr(8)
27810 .sr(1)
27811 .m(m)
27812 .n(n)
27813 .k(k)
27814 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027815 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027816 }
27817 }
27818 }
27819 }
27820
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027821 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, n_gt_8) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027822 TEST_REQUIRES_ARM_NEON;
27823 for (uint32_t n = 9; n < 16; n++) {
27824 for (size_t k = 1; k <= 80; k += 17) {
27825 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027826 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027827 .nr(8)
27828 .kr(8)
27829 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027830 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027831 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070027832 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027833 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027834 }
27835 }
27836 }
27837
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027838 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, n_gt_8_strided_cn) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027839 TEST_REQUIRES_ARM_NEON;
27840 for (uint32_t n = 9; n < 16; n++) {
27841 for (size_t k = 1; k <= 80; k += 17) {
27842 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027843 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027844 .nr(8)
27845 .kr(8)
27846 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027847 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027848 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070027849 .k(k)
27850 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027851 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027852 }
27853 }
27854 }
27855
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027856 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, n_gt_8_subtile) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027857 TEST_REQUIRES_ARM_NEON;
27858 for (uint32_t n = 9; n < 16; n++) {
27859 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027860 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027861 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027862 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027863 .nr(8)
27864 .kr(8)
27865 .sr(1)
27866 .m(m)
27867 .n(n)
27868 .k(k)
27869 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027870 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027871 }
27872 }
27873 }
27874 }
27875
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027876 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, n_div_8) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027877 TEST_REQUIRES_ARM_NEON;
27878 for (uint32_t n = 16; n <= 24; n += 8) {
27879 for (size_t k = 1; k <= 80; k += 17) {
27880 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027881 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027882 .nr(8)
27883 .kr(8)
27884 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027885 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027886 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070027887 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027888 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027889 }
27890 }
27891 }
27892
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027893 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, n_div_8_strided_cn) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027894 TEST_REQUIRES_ARM_NEON;
27895 for (uint32_t n = 16; n <= 24; n += 8) {
27896 for (size_t k = 1; k <= 80; k += 17) {
27897 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027898 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027899 .nr(8)
27900 .kr(8)
27901 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027902 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027903 .n(n)
27904 .k(k)
27905 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027906 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027907 }
27908 }
27909 }
27910
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027911 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, n_div_8_subtile) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027912 TEST_REQUIRES_ARM_NEON;
27913 for (uint32_t n = 16; n <= 24; n += 8) {
27914 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027915 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027916 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027917 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027918 .nr(8)
27919 .kr(8)
27920 .sr(1)
27921 .m(m)
27922 .n(n)
27923 .k(k)
27924 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027925 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027926 }
27927 }
27928 }
27929 }
27930
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027931 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, small_kernel) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027932 TEST_REQUIRES_ARM_NEON;
27933 for (size_t k = 1; k <= 80; k += 17) {
27934 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027935 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027936 .nr(8)
27937 .kr(8)
27938 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027939 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027940 .n(8)
27941 .k(k)
27942 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027943 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027944 }
27945 }
27946
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027947 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, small_kernel_subtile) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027948 TEST_REQUIRES_ARM_NEON;
27949 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027950 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027951 for (uint32_t n = 1; n <= 8; n++) {
27952 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027953 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027954 .nr(8)
27955 .kr(8)
27956 .sr(1)
27957 .m(m)
27958 .n(n)
27959 .k(k)
27960 .ks(3)
27961 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027962 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027963 }
27964 }
27965 }
27966 }
27967
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027968 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, n_gt_8_small_kernel) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027969 TEST_REQUIRES_ARM_NEON;
27970 for (uint32_t n = 9; n < 16; n++) {
27971 for (size_t k = 1; k <= 80; k += 17) {
27972 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027973 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027974 .nr(8)
27975 .kr(8)
27976 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027977 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027978 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070027979 .k(k)
27980 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027981 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070027982 }
27983 }
27984 }
27985
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027986 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, n_div_8_small_kernel) {
Marat Dukhane903dff2021-07-16 19:43:41 -070027987 TEST_REQUIRES_ARM_NEON;
27988 for (uint32_t n = 16; n <= 24; n += 8) {
27989 for (size_t k = 1; k <= 80; k += 17) {
27990 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027991 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070027992 .nr(8)
27993 .kr(8)
27994 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027995 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080027996 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070027997 .k(k)
27998 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080027999 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070028000 }
28001 }
28002 }
28003
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028004 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, strided_cm_subtile) {
Marat Dukhane903dff2021-07-16 19:43:41 -070028005 TEST_REQUIRES_ARM_NEON;
28006 for (size_t k = 1; k <= 80; k += 17) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028007 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070028008 for (uint32_t n = 1; n <= 8; n++) {
28009 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028010 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028011 .nr(8)
28012 .kr(8)
28013 .sr(1)
28014 .m(m)
28015 .n(n)
28016 .k(k)
28017 .cm_stride(11)
28018 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028019 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070028020 }
28021 }
28022 }
28023 }
28024
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028025 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, a_offset) {
Marat Dukhane903dff2021-07-16 19:43:41 -070028026 TEST_REQUIRES_ARM_NEON;
28027 for (size_t k = 1; k <= 80; k += 17) {
28028 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028029 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028030 .nr(8)
28031 .kr(8)
28032 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028033 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028034 .n(8)
28035 .k(k)
28036 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028037 .a_offset(251)
28038 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070028039 }
28040 }
28041
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028042 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, zero) {
Marat Dukhane903dff2021-07-16 19:43:41 -070028043 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028044 for (uint32_t mz = 0; mz < 3; mz++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070028045 for (size_t k = 1; k <= 80; k += 17) {
28046 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028047 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028048 .nr(8)
28049 .kr(8)
28050 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028051 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028052 .n(8)
28053 .k(k)
28054 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028055 .a_offset(251)
Marat Dukhane903dff2021-07-16 19:43:41 -070028056 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028057 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070028058 }
28059 }
28060 }
28061
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028062 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, qmin) {
Marat Dukhane903dff2021-07-16 19:43:41 -070028063 TEST_REQUIRES_ARM_NEON;
28064 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028065 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028066 .nr(8)
28067 .kr(8)
28068 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028069 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028070 .n(8)
28071 .k(16)
28072 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028073 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070028074 }
28075
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028076 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, qmax) {
Marat Dukhane903dff2021-07-16 19:43:41 -070028077 TEST_REQUIRES_ARM_NEON;
28078 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028079 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028080 .nr(8)
28081 .kr(8)
28082 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028083 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028084 .n(8)
28085 .k(16)
28086 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028087 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070028088 }
28089
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028090 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MLAL, strided_cm) {
Marat Dukhane903dff2021-07-16 19:43:41 -070028091 TEST_REQUIRES_ARM_NEON;
28092 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028093 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028094 .nr(8)
28095 .kr(8)
28096 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028097 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070028098 .n(8)
28099 .k(16)
28100 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028101 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028102 }
28103#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
28104
28105
28106#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Marat Dukhan89991902021-12-06 00:54:36 -080028107 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_eq_16) {
28108 TEST_REQUIRES_ARM_NEON;
28109 GemmMicrokernelTester()
28110 .mr(3)
28111 .nr(16)
28112 .kr(8)
28113 .sr(1)
28114 .m(3)
28115 .n(16)
28116 .k(16)
Marat Dukhan50323b82022-01-11 00:12:01 -080028117 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028118 }
28119
28120 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, strided_cn) {
28121 TEST_REQUIRES_ARM_NEON;
28122 GemmMicrokernelTester()
28123 .mr(3)
28124 .nr(16)
28125 .kr(8)
28126 .sr(1)
28127 .m(3)
28128 .n(16)
28129 .k(16)
28130 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080028131 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028132 }
28133
28134 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_eq_16_subtile) {
28135 TEST_REQUIRES_ARM_NEON;
28136 for (uint32_t m = 1; m <= 3; m++) {
28137 for (uint32_t n = 1; n <= 16; n++) {
28138 GemmMicrokernelTester()
28139 .mr(3)
28140 .nr(16)
28141 .kr(8)
28142 .sr(1)
28143 .m(m)
28144 .n(n)
28145 .k(16)
28146 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028147 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028148 }
28149 }
28150 }
28151
28152 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_eq_16_subtile_m) {
28153 TEST_REQUIRES_ARM_NEON;
28154 for (uint32_t m = 1; m <= 3; m++) {
28155 GemmMicrokernelTester()
28156 .mr(3)
28157 .nr(16)
28158 .kr(8)
28159 .sr(1)
28160 .m(m)
28161 .n(16)
28162 .k(16)
28163 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028164 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028165 }
28166 }
28167
28168 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_eq_16_subtile_n) {
28169 TEST_REQUIRES_ARM_NEON;
28170 for (uint32_t n = 1; n <= 16; n++) {
28171 GemmMicrokernelTester()
28172 .mr(3)
28173 .nr(16)
28174 .kr(8)
28175 .sr(1)
28176 .m(3)
28177 .n(n)
28178 .k(16)
28179 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028180 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028181 }
28182 }
28183
28184 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_lt_16) {
28185 TEST_REQUIRES_ARM_NEON;
28186 for (size_t k = 1; k < 16; k++) {
28187 GemmMicrokernelTester()
28188 .mr(3)
28189 .nr(16)
28190 .kr(8)
28191 .sr(1)
28192 .m(3)
28193 .n(16)
28194 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080028195 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028196 }
28197 }
28198
28199 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_lt_16_subtile) {
28200 TEST_REQUIRES_ARM_NEON;
28201 for (size_t k = 1; k < 16; k++) {
28202 for (uint32_t m = 1; m <= 3; m++) {
28203 for (uint32_t n = 1; n <= 16; n++) {
28204 GemmMicrokernelTester()
28205 .mr(3)
28206 .nr(16)
28207 .kr(8)
28208 .sr(1)
28209 .m(m)
28210 .n(n)
28211 .k(k)
28212 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028213 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028214 }
28215 }
28216 }
28217 }
28218
28219 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_gt_16) {
28220 TEST_REQUIRES_ARM_NEON;
28221 for (size_t k = 17; k < 32; k++) {
28222 GemmMicrokernelTester()
28223 .mr(3)
28224 .nr(16)
28225 .kr(8)
28226 .sr(1)
28227 .m(3)
28228 .n(16)
28229 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080028230 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028231 }
28232 }
28233
28234 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_gt_16_subtile) {
28235 TEST_REQUIRES_ARM_NEON;
28236 for (size_t k = 17; k < 32; k++) {
28237 for (uint32_t m = 1; m <= 3; m++) {
28238 for (uint32_t n = 1; n <= 16; n++) {
28239 GemmMicrokernelTester()
28240 .mr(3)
28241 .nr(16)
28242 .kr(8)
28243 .sr(1)
28244 .m(m)
28245 .n(n)
28246 .k(k)
28247 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028248 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028249 }
28250 }
28251 }
28252 }
28253
28254 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_div_16) {
28255 TEST_REQUIRES_ARM_NEON;
28256 for (size_t k = 32; k <= 160; k += 16) {
28257 GemmMicrokernelTester()
28258 .mr(3)
28259 .nr(16)
28260 .kr(8)
28261 .sr(1)
28262 .m(3)
28263 .n(16)
28264 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080028265 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028266 }
28267 }
28268
28269 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, k_div_16_subtile) {
28270 TEST_REQUIRES_ARM_NEON;
28271 for (size_t k = 32; k <= 160; k += 16) {
28272 for (uint32_t m = 1; m <= 3; m++) {
28273 for (uint32_t n = 1; n <= 16; n++) {
28274 GemmMicrokernelTester()
28275 .mr(3)
28276 .nr(16)
28277 .kr(8)
28278 .sr(1)
28279 .m(m)
28280 .n(n)
28281 .k(k)
28282 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028283 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028284 }
28285 }
28286 }
28287 }
28288
28289 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, n_gt_16) {
28290 TEST_REQUIRES_ARM_NEON;
28291 for (uint32_t n = 17; n < 32; n++) {
28292 for (size_t k = 1; k <= 80; k += 17) {
28293 GemmMicrokernelTester()
28294 .mr(3)
28295 .nr(16)
28296 .kr(8)
28297 .sr(1)
28298 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028299 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028300 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080028301 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028302 }
28303 }
28304 }
28305
28306 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, n_gt_16_strided_cn) {
28307 TEST_REQUIRES_ARM_NEON;
28308 for (uint32_t n = 17; n < 32; n++) {
28309 for (size_t k = 1; k <= 80; k += 17) {
28310 GemmMicrokernelTester()
28311 .mr(3)
28312 .nr(16)
28313 .kr(8)
28314 .sr(1)
28315 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028316 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028317 .k(k)
28318 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080028319 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028320 }
28321 }
28322 }
28323
28324 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, n_gt_16_subtile) {
28325 TEST_REQUIRES_ARM_NEON;
28326 for (uint32_t n = 17; n < 32; n++) {
28327 for (size_t k = 1; k <= 80; k += 17) {
28328 for (uint32_t m = 1; m <= 3; m++) {
28329 GemmMicrokernelTester()
28330 .mr(3)
28331 .nr(16)
28332 .kr(8)
28333 .sr(1)
28334 .m(m)
28335 .n(n)
28336 .k(k)
28337 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028338 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028339 }
28340 }
28341 }
28342 }
28343
28344 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, n_div_16) {
28345 TEST_REQUIRES_ARM_NEON;
28346 for (uint32_t n = 32; n <= 48; n += 16) {
28347 for (size_t k = 1; k <= 80; k += 17) {
28348 GemmMicrokernelTester()
28349 .mr(3)
28350 .nr(16)
28351 .kr(8)
28352 .sr(1)
28353 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028354 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028355 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080028356 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028357 }
28358 }
28359 }
28360
28361 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, n_div_16_strided_cn) {
28362 TEST_REQUIRES_ARM_NEON;
28363 for (uint32_t n = 32; n <= 48; n += 16) {
28364 for (size_t k = 1; k <= 80; k += 17) {
28365 GemmMicrokernelTester()
28366 .mr(3)
28367 .nr(16)
28368 .kr(8)
28369 .sr(1)
28370 .m(3)
28371 .n(n)
28372 .k(k)
28373 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080028374 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028375 }
28376 }
28377 }
28378
28379 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, n_div_16_subtile) {
28380 TEST_REQUIRES_ARM_NEON;
28381 for (uint32_t n = 32; n <= 48; n += 16) {
28382 for (size_t k = 1; k <= 80; k += 17) {
28383 for (uint32_t m = 1; m <= 3; m++) {
28384 GemmMicrokernelTester()
28385 .mr(3)
28386 .nr(16)
28387 .kr(8)
28388 .sr(1)
28389 .m(m)
28390 .n(n)
28391 .k(k)
28392 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028393 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028394 }
28395 }
28396 }
28397 }
28398
28399 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, small_kernel) {
28400 TEST_REQUIRES_ARM_NEON;
28401 for (size_t k = 1; k <= 80; k += 17) {
28402 GemmMicrokernelTester()
28403 .mr(3)
28404 .nr(16)
28405 .kr(8)
28406 .sr(1)
28407 .m(3)
28408 .n(16)
28409 .k(k)
28410 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080028411 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028412 }
28413 }
28414
28415 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, small_kernel_subtile) {
28416 TEST_REQUIRES_ARM_NEON;
28417 for (size_t k = 1; k <= 80; k += 17) {
28418 for (uint32_t m = 1; m <= 3; m++) {
28419 for (uint32_t n = 1; n <= 16; n++) {
28420 GemmMicrokernelTester()
28421 .mr(3)
28422 .nr(16)
28423 .kr(8)
28424 .sr(1)
28425 .m(m)
28426 .n(n)
28427 .k(k)
28428 .ks(3)
28429 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028430 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028431 }
28432 }
28433 }
28434 }
28435
28436 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, n_gt_16_small_kernel) {
28437 TEST_REQUIRES_ARM_NEON;
28438 for (uint32_t n = 17; n < 32; n++) {
28439 for (size_t k = 1; k <= 80; k += 17) {
28440 GemmMicrokernelTester()
28441 .mr(3)
28442 .nr(16)
28443 .kr(8)
28444 .sr(1)
28445 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028446 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028447 .k(k)
28448 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080028449 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028450 }
28451 }
28452 }
28453
28454 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, n_div_16_small_kernel) {
28455 TEST_REQUIRES_ARM_NEON;
28456 for (uint32_t n = 32; n <= 48; n += 16) {
28457 for (size_t k = 1; k <= 80; k += 17) {
28458 GemmMicrokernelTester()
28459 .mr(3)
28460 .nr(16)
28461 .kr(8)
28462 .sr(1)
28463 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028464 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028465 .k(k)
28466 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080028467 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028468 }
28469 }
28470 }
28471
28472 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, strided_cm_subtile) {
28473 TEST_REQUIRES_ARM_NEON;
28474 for (size_t k = 1; k <= 80; k += 17) {
28475 for (uint32_t m = 1; m <= 3; m++) {
28476 for (uint32_t n = 1; n <= 16; n++) {
28477 GemmMicrokernelTester()
28478 .mr(3)
28479 .nr(16)
28480 .kr(8)
28481 .sr(1)
28482 .m(m)
28483 .n(n)
28484 .k(k)
28485 .cm_stride(19)
28486 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080028487 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028488 }
28489 }
28490 }
28491 }
28492
28493 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, a_offset) {
28494 TEST_REQUIRES_ARM_NEON;
28495 for (size_t k = 1; k <= 80; k += 17) {
28496 GemmMicrokernelTester()
28497 .mr(3)
28498 .nr(16)
28499 .kr(8)
28500 .sr(1)
28501 .m(3)
28502 .n(16)
28503 .k(k)
28504 .ks(3)
28505 .a_offset(251)
Marat Dukhan50323b82022-01-11 00:12:01 -080028506 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028507 }
28508 }
28509
28510 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, zero) {
28511 TEST_REQUIRES_ARM_NEON;
28512 for (uint32_t mz = 0; mz < 3; mz++) {
28513 for (size_t k = 1; k <= 80; k += 17) {
28514 GemmMicrokernelTester()
28515 .mr(3)
28516 .nr(16)
28517 .kr(8)
28518 .sr(1)
28519 .m(3)
28520 .n(16)
28521 .k(k)
28522 .ks(3)
28523 .a_offset(251)
28524 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080028525 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028526 }
28527 }
28528 }
28529
28530 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, qmin) {
28531 TEST_REQUIRES_ARM_NEON;
28532 GemmMicrokernelTester()
28533 .mr(3)
28534 .nr(16)
28535 .kr(8)
28536 .sr(1)
28537 .m(3)
28538 .n(16)
28539 .k(16)
28540 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080028541 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028542 }
28543
28544 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, qmax) {
28545 TEST_REQUIRES_ARM_NEON;
28546 GemmMicrokernelTester()
28547 .mr(3)
28548 .nr(16)
28549 .kr(8)
28550 .sr(1)
28551 .m(3)
28552 .n(16)
28553 .k(16)
28554 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080028555 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028556 }
28557
28558 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C8__NEON_MLAL, strided_cm) {
28559 TEST_REQUIRES_ARM_NEON;
28560 GemmMicrokernelTester()
28561 .mr(3)
28562 .nr(16)
28563 .kr(8)
28564 .sr(1)
28565 .m(3)
28566 .n(16)
28567 .k(16)
28568 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080028569 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028570 }
28571#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
28572
28573
28574#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028575 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_eq_16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028576 TEST_REQUIRES_ARM_NEON;
28577 GemmMicrokernelTester()
28578 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028579 .nr(16)
28580 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028581 .sr(1)
28582 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028583 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080028584 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028585 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028586 }
28587
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028588 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, strided_cn) {
Marat Dukhan89991902021-12-06 00:54:36 -080028589 TEST_REQUIRES_ARM_NEON;
28590 GemmMicrokernelTester()
28591 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028592 .nr(16)
28593 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028594 .sr(1)
28595 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028596 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080028597 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028598 .cn_stride(19)
28599 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028600 }
28601
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028602 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_eq_16_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080028603 TEST_REQUIRES_ARM_NEON;
28604 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028605 for (uint32_t n = 1; n <= 16; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028606 GemmMicrokernelTester()
28607 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028608 .nr(16)
28609 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028610 .sr(1)
28611 .m(m)
28612 .n(n)
28613 .k(16)
28614 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028615 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028616 }
28617 }
28618 }
28619
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028620 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_eq_16_subtile_m) {
Marat Dukhan89991902021-12-06 00:54:36 -080028621 TEST_REQUIRES_ARM_NEON;
28622 for (uint32_t m = 1; m <= 4; m++) {
28623 GemmMicrokernelTester()
28624 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028625 .nr(16)
28626 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028627 .sr(1)
28628 .m(m)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028629 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080028630 .k(16)
28631 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028632 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028633 }
28634 }
28635
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028636 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_eq_16_subtile_n) {
Marat Dukhan89991902021-12-06 00:54:36 -080028637 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028638 for (uint32_t n = 1; n <= 16; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028639 GemmMicrokernelTester()
28640 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028641 .nr(16)
28642 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028643 .sr(1)
28644 .m(4)
28645 .n(n)
28646 .k(16)
28647 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028648 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028649 }
28650 }
28651
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028652 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_lt_16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028653 TEST_REQUIRES_ARM_NEON;
28654 for (size_t k = 1; k < 16; k++) {
28655 GemmMicrokernelTester()
28656 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028657 .nr(16)
28658 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028659 .sr(1)
28660 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028661 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080028662 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028663 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028664 }
28665 }
28666
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028667 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_lt_16_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080028668 TEST_REQUIRES_ARM_NEON;
28669 for (size_t k = 1; k < 16; k++) {
28670 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028671 for (uint32_t n = 1; n <= 16; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028672 GemmMicrokernelTester()
28673 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028674 .nr(16)
28675 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028676 .sr(1)
28677 .m(m)
28678 .n(n)
28679 .k(k)
28680 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028681 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028682 }
28683 }
28684 }
28685 }
28686
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028687 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_gt_16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028688 TEST_REQUIRES_ARM_NEON;
28689 for (size_t k = 17; k < 32; k++) {
28690 GemmMicrokernelTester()
28691 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028692 .nr(16)
28693 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028694 .sr(1)
28695 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028696 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080028697 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028698 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028699 }
28700 }
28701
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028702 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_gt_16_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080028703 TEST_REQUIRES_ARM_NEON;
28704 for (size_t k = 17; k < 32; k++) {
28705 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028706 for (uint32_t n = 1; n <= 16; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028707 GemmMicrokernelTester()
28708 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028709 .nr(16)
28710 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028711 .sr(1)
28712 .m(m)
28713 .n(n)
28714 .k(k)
28715 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028716 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028717 }
28718 }
28719 }
28720 }
28721
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028722 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_div_16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028723 TEST_REQUIRES_ARM_NEON;
28724 for (size_t k = 32; k <= 160; k += 16) {
28725 GemmMicrokernelTester()
28726 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028727 .nr(16)
28728 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028729 .sr(1)
28730 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028731 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080028732 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028733 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028734 }
28735 }
28736
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028737 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_div_16_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080028738 TEST_REQUIRES_ARM_NEON;
28739 for (size_t k = 32; k <= 160; k += 16) {
28740 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028741 for (uint32_t n = 1; n <= 16; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028742 GemmMicrokernelTester()
28743 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028744 .nr(16)
28745 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028746 .sr(1)
28747 .m(m)
28748 .n(n)
28749 .k(k)
28750 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028751 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028752 }
28753 }
28754 }
28755 }
28756
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028757 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_gt_16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028758 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028759 for (uint32_t n = 17; n < 32; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028760 for (size_t k = 1; k <= 80; k += 17) {
28761 GemmMicrokernelTester()
28762 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028763 .nr(16)
28764 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028765 .sr(1)
28766 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028767 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028768 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028769 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028770 }
28771 }
28772 }
28773
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028774 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_gt_16_strided_cn) {
Marat Dukhan89991902021-12-06 00:54:36 -080028775 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028776 for (uint32_t n = 17; n < 32; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028777 for (size_t k = 1; k <= 80; k += 17) {
28778 GemmMicrokernelTester()
28779 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028780 .nr(16)
28781 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028782 .sr(1)
28783 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028784 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028785 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028786 .cn_stride(19)
28787 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028788 }
28789 }
28790 }
28791
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028792 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_gt_16_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080028793 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028794 for (uint32_t n = 17; n < 32; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028795 for (size_t k = 1; k <= 80; k += 17) {
28796 for (uint32_t m = 1; m <= 4; m++) {
28797 GemmMicrokernelTester()
28798 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028799 .nr(16)
28800 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028801 .sr(1)
28802 .m(m)
28803 .n(n)
28804 .k(k)
28805 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028806 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028807 }
28808 }
28809 }
28810 }
28811
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028812 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_div_16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028813 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028814 for (uint32_t n = 32; n <= 48; n += 16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028815 for (size_t k = 1; k <= 80; k += 17) {
28816 GemmMicrokernelTester()
28817 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028818 .nr(16)
28819 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028820 .sr(1)
28821 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028822 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028823 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028824 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028825 }
28826 }
28827 }
28828
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028829 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_div_16_strided_cn) {
Marat Dukhan89991902021-12-06 00:54:36 -080028830 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028831 for (uint32_t n = 32; n <= 48; n += 16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028832 for (size_t k = 1; k <= 80; k += 17) {
28833 GemmMicrokernelTester()
28834 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028835 .nr(16)
28836 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028837 .sr(1)
28838 .m(4)
28839 .n(n)
28840 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028841 .cn_stride(19)
28842 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028843 }
28844 }
28845 }
28846
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028847 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_div_16_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080028848 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028849 for (uint32_t n = 32; n <= 48; n += 16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028850 for (size_t k = 1; k <= 80; k += 17) {
28851 for (uint32_t m = 1; m <= 4; m++) {
28852 GemmMicrokernelTester()
28853 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028854 .nr(16)
28855 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028856 .sr(1)
28857 .m(m)
28858 .n(n)
28859 .k(k)
28860 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028861 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028862 }
28863 }
28864 }
28865 }
28866
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028867 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, small_kernel) {
Marat Dukhan89991902021-12-06 00:54:36 -080028868 TEST_REQUIRES_ARM_NEON;
28869 for (size_t k = 1; k <= 80; k += 17) {
28870 GemmMicrokernelTester()
28871 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028872 .nr(16)
28873 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028874 .sr(1)
28875 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028876 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080028877 .k(k)
28878 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028879 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028880 }
28881 }
28882
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028883 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, small_kernel_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080028884 TEST_REQUIRES_ARM_NEON;
28885 for (size_t k = 1; k <= 80; k += 17) {
28886 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028887 for (uint32_t n = 1; n <= 16; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028888 GemmMicrokernelTester()
28889 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028890 .nr(16)
28891 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028892 .sr(1)
28893 .m(m)
28894 .n(n)
28895 .k(k)
28896 .ks(3)
28897 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028898 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028899 }
28900 }
28901 }
28902 }
28903
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028904 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_gt_16_small_kernel) {
Marat Dukhan89991902021-12-06 00:54:36 -080028905 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028906 for (uint32_t n = 17; n < 32; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028907 for (size_t k = 1; k <= 80; k += 17) {
28908 GemmMicrokernelTester()
28909 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028910 .nr(16)
28911 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028912 .sr(1)
28913 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028914 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028915 .k(k)
28916 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028917 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028918 }
28919 }
28920 }
28921
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028922 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_div_16_small_kernel) {
Marat Dukhan89991902021-12-06 00:54:36 -080028923 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028924 for (uint32_t n = 32; n <= 48; n += 16) {
Marat Dukhan89991902021-12-06 00:54:36 -080028925 for (size_t k = 1; k <= 80; k += 17) {
28926 GemmMicrokernelTester()
28927 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028928 .nr(16)
28929 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028930 .sr(1)
28931 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080028932 .n(n)
Marat Dukhan89991902021-12-06 00:54:36 -080028933 .k(k)
28934 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028935 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028936 }
28937 }
28938 }
28939
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028940 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, strided_cm_subtile) {
Marat Dukhan89991902021-12-06 00:54:36 -080028941 TEST_REQUIRES_ARM_NEON;
28942 for (size_t k = 1; k <= 80; k += 17) {
28943 for (uint32_t m = 1; m <= 4; m++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028944 for (uint32_t n = 1; n <= 16; n++) {
Marat Dukhan89991902021-12-06 00:54:36 -080028945 GemmMicrokernelTester()
28946 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028947 .nr(16)
28948 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028949 .sr(1)
28950 .m(m)
28951 .n(n)
28952 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028953 .cm_stride(19)
Marat Dukhan89991902021-12-06 00:54:36 -080028954 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028955 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028956 }
28957 }
28958 }
28959 }
28960
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028961 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, a_offset) {
Marat Dukhan89991902021-12-06 00:54:36 -080028962 TEST_REQUIRES_ARM_NEON;
28963 for (size_t k = 1; k <= 80; k += 17) {
28964 GemmMicrokernelTester()
28965 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028966 .nr(16)
28967 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028968 .sr(1)
28969 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028970 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080028971 .k(k)
28972 .ks(3)
28973 .a_offset(331)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028974 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028975 }
28976 }
28977
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028978 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, zero) {
Marat Dukhan89991902021-12-06 00:54:36 -080028979 TEST_REQUIRES_ARM_NEON;
28980 for (uint32_t mz = 0; mz < 4; mz++) {
28981 for (size_t k = 1; k <= 80; k += 17) {
28982 GemmMicrokernelTester()
28983 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028984 .nr(16)
28985 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080028986 .sr(1)
28987 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028988 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080028989 .k(k)
28990 .ks(3)
28991 .a_offset(331)
28992 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028993 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080028994 }
28995 }
28996 }
28997
Zhi An Ngc27f04b2022-01-11 09:34:07 -080028998 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, qmin) {
Marat Dukhan89991902021-12-06 00:54:36 -080028999 TEST_REQUIRES_ARM_NEON;
29000 GemmMicrokernelTester()
29001 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029002 .nr(16)
29003 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080029004 .sr(1)
29005 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029006 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080029007 .k(16)
29008 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029009 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080029010 }
29011
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029012 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, qmax) {
Marat Dukhan89991902021-12-06 00:54:36 -080029013 TEST_REQUIRES_ARM_NEON;
29014 GemmMicrokernelTester()
29015 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029016 .nr(16)
29017 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080029018 .sr(1)
29019 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029020 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080029021 .k(16)
29022 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029023 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080029024 }
29025
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029026 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, strided_cm) {
Marat Dukhan89991902021-12-06 00:54:36 -080029027 TEST_REQUIRES_ARM_NEON;
29028 GemmMicrokernelTester()
29029 .mr(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029030 .nr(16)
29031 .kr(8)
Marat Dukhan89991902021-12-06 00:54:36 -080029032 .sr(1)
29033 .m(4)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029034 .n(16)
Marat Dukhan89991902021-12-06 00:54:36 -080029035 .k(16)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029036 .cm_stride(19)
29037 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan89991902021-12-06 00:54:36 -080029038 }
29039#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
29040
29041
29042#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029043 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_eq_16) {
Marat Dukhan89991902021-12-06 00:54:36 -080029044 TEST_REQUIRES_ARM_NEON;
29045 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029046 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029047 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029048 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029049 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029050 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029051 .n(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029052 .k(16)
29053 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029054 }
29055
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029056 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, strided_cn) {
29057 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029058 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029059 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029060 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029061 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029062 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029063 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029064 .n(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029065 .k(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029066 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029067 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029068 }
29069
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029070 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_eq_16_subtile) {
29071 TEST_REQUIRES_ARM_NEON;
29072 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029073 for (uint32_t n = 1; n <= 8; n++) {
29074 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029075 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029076 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029077 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029078 .sr(1)
29079 .m(m)
29080 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029081 .k(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029082 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029083 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029084 }
29085 }
29086 }
29087
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029088 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_eq_16_subtile_m) {
29089 TEST_REQUIRES_ARM_NEON;
29090 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029091 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029092 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029093 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029094 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029095 .sr(1)
29096 .m(m)
29097 .n(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029098 .k(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029099 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029100 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029101 }
29102 }
29103
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029104 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_eq_16_subtile_n) {
29105 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029106 for (uint32_t n = 1; n <= 8; n++) {
29107 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029108 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029109 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029110 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029111 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029112 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029113 .n(n)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029114 .k(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029115 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029116 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029117 }
29118 }
29119
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029120 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_lt_16) {
29121 TEST_REQUIRES_ARM_NEON;
29122 for (size_t k = 1; k < 16; k++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029123 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029124 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029125 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029126 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029127 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029128 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029129 .n(8)
29130 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029131 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029132 }
29133 }
29134
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029135 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_lt_16_subtile) {
29136 TEST_REQUIRES_ARM_NEON;
29137 for (size_t k = 1; k < 16; k++) {
29138 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029139 for (uint32_t n = 1; n <= 8; n++) {
29140 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029141 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029142 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029143 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029144 .sr(1)
29145 .m(m)
29146 .n(n)
29147 .k(k)
29148 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029149 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029150 }
29151 }
29152 }
29153 }
29154
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029155 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_gt_16) {
29156 TEST_REQUIRES_ARM_NEON;
29157 for (size_t k = 17; k < 32; k++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029158 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029159 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029160 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029161 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029162 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029163 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029164 .n(8)
29165 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029166 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029167 }
29168 }
29169
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029170 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_gt_16_subtile) {
29171 TEST_REQUIRES_ARM_NEON;
29172 for (size_t k = 17; k < 32; k++) {
29173 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029174 for (uint32_t n = 1; n <= 8; n++) {
29175 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029176 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029177 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029178 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029179 .sr(1)
29180 .m(m)
29181 .n(n)
29182 .k(k)
29183 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029184 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029185 }
29186 }
29187 }
29188 }
29189
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029190 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_div_16) {
29191 TEST_REQUIRES_ARM_NEON;
29192 for (size_t k = 32; k <= 160; k += 16) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029193 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029194 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029195 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029196 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029197 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029198 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029199 .n(8)
29200 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029201 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029202 }
29203 }
29204
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029205 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, k_div_16_subtile) {
29206 TEST_REQUIRES_ARM_NEON;
29207 for (size_t k = 32; k <= 160; k += 16) {
29208 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029209 for (uint32_t n = 1; n <= 8; n++) {
29210 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029211 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029212 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029213 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029214 .sr(1)
29215 .m(m)
29216 .n(n)
29217 .k(k)
29218 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029219 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029220 }
29221 }
29222 }
29223 }
29224
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029225 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, n_gt_8) {
29226 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029227 for (uint32_t n = 9; n < 16; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029228 for (size_t k = 1; k <= 80; k += 17) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029229 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029230 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029231 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029232 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029233 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029234 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029235 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070029236 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029237 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029238 }
29239 }
29240 }
29241
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029242 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, n_gt_8_strided_cn) {
29243 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029244 for (uint32_t n = 9; n < 16; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029245 for (size_t k = 1; k <= 80; k += 17) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029246 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029247 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029248 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029249 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029250 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029251 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029252 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070029253 .k(k)
29254 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029255 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029256 }
29257 }
29258 }
29259
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029260 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, n_gt_8_subtile) {
29261 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029262 for (uint32_t n = 9; n < 16; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029263 for (size_t k = 1; k <= 80; k += 17) {
29264 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029265 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029266 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029267 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029268 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029269 .sr(1)
29270 .m(m)
29271 .n(n)
29272 .k(k)
29273 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029274 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029275 }
29276 }
29277 }
29278 }
29279
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029280 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, n_div_8) {
29281 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029282 for (uint32_t n = 16; n <= 24; n += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029283 for (size_t k = 1; k <= 80; k += 17) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029284 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029285 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029286 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029287 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029288 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029289 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029290 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070029291 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029292 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029293 }
29294 }
29295 }
29296
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029297 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, n_div_8_strided_cn) {
29298 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029299 for (uint32_t n = 16; n <= 24; n += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029300 for (size_t k = 1; k <= 80; k += 17) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029301 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029302 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029303 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029304 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029305 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029306 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029307 .n(n)
29308 .k(k)
29309 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029310 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029311 }
29312 }
29313 }
29314
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029315 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, n_div_8_subtile) {
29316 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029317 for (uint32_t n = 16; n <= 24; n += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029318 for (size_t k = 1; k <= 80; k += 17) {
29319 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029320 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029321 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029322 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029323 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029324 .sr(1)
29325 .m(m)
29326 .n(n)
29327 .k(k)
29328 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029329 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029330 }
29331 }
29332 }
29333 }
29334
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029335 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, small_kernel) {
29336 TEST_REQUIRES_ARM_NEON;
29337 for (size_t k = 1; k <= 80; k += 17) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029338 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029339 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029340 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029341 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029342 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029343 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029344 .n(8)
29345 .k(k)
29346 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029347 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029348 }
29349 }
29350
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029351 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, small_kernel_subtile) {
29352 TEST_REQUIRES_ARM_NEON;
29353 for (size_t k = 1; k <= 80; k += 17) {
29354 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029355 for (uint32_t n = 1; n <= 8; n++) {
29356 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029357 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029358 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029359 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029360 .sr(1)
29361 .m(m)
29362 .n(n)
29363 .k(k)
29364 .ks(3)
29365 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029366 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029367 }
29368 }
29369 }
29370 }
29371
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029372 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, n_gt_8_small_kernel) {
29373 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029374 for (uint32_t n = 9; n < 16; n++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029375 for (size_t k = 1; k <= 80; k += 17) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029376 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029377 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029378 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029379 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029380 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029381 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029382 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070029383 .k(k)
29384 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029385 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029386 }
29387 }
29388 }
29389
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029390 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, n_div_8_small_kernel) {
29391 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029392 for (uint32_t n = 16; n <= 24; n += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029393 for (size_t k = 1; k <= 80; k += 17) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029394 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029395 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029396 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029397 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029398 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029399 .m(3)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029400 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070029401 .k(k)
29402 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029403 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029404 }
29405 }
29406 }
29407
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029408 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, strided_cm_subtile) {
29409 TEST_REQUIRES_ARM_NEON;
29410 for (size_t k = 1; k <= 80; k += 17) {
29411 for (uint32_t m = 1; m <= 3; m++) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029412 for (uint32_t n = 1; n <= 8; n++) {
29413 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029414 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029415 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029416 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029417 .sr(1)
29418 .m(m)
29419 .n(n)
29420 .k(k)
29421 .cm_stride(11)
29422 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029423 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029424 }
29425 }
29426 }
29427 }
29428
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029429 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, a_offset) {
29430 TEST_REQUIRES_ARM_NEON;
29431 for (size_t k = 1; k <= 80; k += 17) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029432 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029433 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029434 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029435 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029436 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029437 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029438 .n(8)
29439 .k(k)
29440 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029441 .a_offset(251)
29442 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029443 }
29444 }
29445
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029446 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, zero) {
29447 TEST_REQUIRES_ARM_NEON;
29448 for (uint32_t mz = 0; mz < 3; mz++) {
29449 for (size_t k = 1; k <= 80; k += 17) {
Marat Dukhane903dff2021-07-16 19:43:41 -070029450 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029451 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029452 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029453 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029454 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029455 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029456 .n(8)
29457 .k(k)
29458 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029459 .a_offset(251)
Marat Dukhane903dff2021-07-16 19:43:41 -070029460 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029461 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029462 }
29463 }
29464 }
29465
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029466 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, qmin) {
29467 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029468 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029469 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029470 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029471 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029472 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029473 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029474 .n(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029475 .k(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029476 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029477 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029478 }
29479
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029480 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, qmax) {
29481 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029482 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029483 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029484 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029485 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029486 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029487 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029488 .n(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029489 .k(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029490 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029491 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029492 }
29493
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029494 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C16__NEON_MLAL, strided_cm) {
29495 TEST_REQUIRES_ARM_NEON;
Marat Dukhane903dff2021-07-16 19:43:41 -070029496 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029497 .mr(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029498 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029499 .kr(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029500 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029501 .m(3)
Marat Dukhane903dff2021-07-16 19:43:41 -070029502 .n(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029503 .k(16)
Marat Dukhane903dff2021-07-16 19:43:41 -070029504 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029505 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070029506 }
Zhi An Ngc27f04b2022-01-11 09:34:07 -080029507#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
Marat Dukhane903dff2021-07-16 19:43:41 -070029508
29509
29510#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
Marat Dukhan4486f872021-08-07 15:22:50 -070029511 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_eq_8) {
29512 TEST_REQUIRES_ARM_NEON_DOT;
29513 GemmMicrokernelTester()
29514 .mr(6)
29515 .nr(8)
29516 .kr(4)
29517 .sr(1)
29518 .m(6)
29519 .n(8)
29520 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080029521 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029522 }
29523
29524 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, strided_cn) {
29525 TEST_REQUIRES_ARM_NEON_DOT;
29526 GemmMicrokernelTester()
29527 .mr(6)
29528 .nr(8)
29529 .kr(4)
29530 .sr(1)
29531 .m(6)
29532 .n(8)
29533 .k(8)
29534 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080029535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029536 }
29537
29538 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_eq_8_subtile) {
29539 TEST_REQUIRES_ARM_NEON_DOT;
29540 for (uint32_t m = 1; m <= 6; m++) {
29541 for (uint32_t n = 1; n <= 8; n++) {
29542 GemmMicrokernelTester()
29543 .mr(6)
29544 .nr(8)
29545 .kr(4)
29546 .sr(1)
29547 .m(m)
29548 .n(n)
29549 .k(8)
29550 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029551 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029552 }
29553 }
29554 }
29555
29556 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_eq_8_subtile_m) {
29557 TEST_REQUIRES_ARM_NEON_DOT;
29558 for (uint32_t m = 1; m <= 6; m++) {
29559 GemmMicrokernelTester()
29560 .mr(6)
29561 .nr(8)
29562 .kr(4)
29563 .sr(1)
29564 .m(m)
29565 .n(8)
29566 .k(8)
29567 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029568 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029569 }
29570 }
29571
29572 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_eq_8_subtile_n) {
29573 TEST_REQUIRES_ARM_NEON_DOT;
29574 for (uint32_t n = 1; n <= 8; n++) {
29575 GemmMicrokernelTester()
29576 .mr(6)
29577 .nr(8)
29578 .kr(4)
29579 .sr(1)
29580 .m(6)
29581 .n(n)
29582 .k(8)
29583 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029584 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029585 }
29586 }
29587
29588 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_lt_8) {
29589 TEST_REQUIRES_ARM_NEON_DOT;
29590 for (size_t k = 1; k < 8; k++) {
29591 GemmMicrokernelTester()
29592 .mr(6)
29593 .nr(8)
29594 .kr(4)
29595 .sr(1)
29596 .m(6)
29597 .n(8)
29598 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080029599 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029600 }
29601 }
29602
29603 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_lt_8_subtile) {
29604 TEST_REQUIRES_ARM_NEON_DOT;
29605 for (size_t k = 1; k < 8; k++) {
29606 for (uint32_t m = 1; m <= 6; m++) {
29607 for (uint32_t n = 1; n <= 8; n++) {
29608 GemmMicrokernelTester()
29609 .mr(6)
29610 .nr(8)
29611 .kr(4)
29612 .sr(1)
29613 .m(m)
29614 .n(n)
29615 .k(k)
29616 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029617 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029618 }
29619 }
29620 }
29621 }
29622
29623 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_gt_8) {
29624 TEST_REQUIRES_ARM_NEON_DOT;
29625 for (size_t k = 9; k < 16; k++) {
29626 GemmMicrokernelTester()
29627 .mr(6)
29628 .nr(8)
29629 .kr(4)
29630 .sr(1)
29631 .m(6)
29632 .n(8)
29633 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080029634 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029635 }
29636 }
29637
29638 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_gt_8_subtile) {
29639 TEST_REQUIRES_ARM_NEON_DOT;
29640 for (size_t k = 9; k < 16; k++) {
29641 for (uint32_t m = 1; m <= 6; m++) {
29642 for (uint32_t n = 1; n <= 8; n++) {
29643 GemmMicrokernelTester()
29644 .mr(6)
29645 .nr(8)
29646 .kr(4)
29647 .sr(1)
29648 .m(m)
29649 .n(n)
29650 .k(k)
29651 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029652 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029653 }
29654 }
29655 }
29656 }
29657
29658 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_div_8) {
29659 TEST_REQUIRES_ARM_NEON_DOT;
29660 for (size_t k = 16; k <= 80; k += 8) {
29661 GemmMicrokernelTester()
29662 .mr(6)
29663 .nr(8)
29664 .kr(4)
29665 .sr(1)
29666 .m(6)
29667 .n(8)
29668 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080029669 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029670 }
29671 }
29672
29673 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, k_div_8_subtile) {
29674 TEST_REQUIRES_ARM_NEON_DOT;
29675 for (size_t k = 16; k <= 80; k += 8) {
29676 for (uint32_t m = 1; m <= 6; m++) {
29677 for (uint32_t n = 1; n <= 8; n++) {
29678 GemmMicrokernelTester()
29679 .mr(6)
29680 .nr(8)
29681 .kr(4)
29682 .sr(1)
29683 .m(m)
29684 .n(n)
29685 .k(k)
29686 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029687 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029688 }
29689 }
29690 }
29691 }
29692
29693 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, n_gt_8) {
29694 TEST_REQUIRES_ARM_NEON_DOT;
29695 for (uint32_t n = 9; n < 16; n++) {
29696 for (size_t k = 1; k <= 40; k += 9) {
29697 GemmMicrokernelTester()
29698 .mr(6)
29699 .nr(8)
29700 .kr(4)
29701 .sr(1)
29702 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029703 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070029704 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080029705 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029706 }
29707 }
29708 }
29709
29710 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, n_gt_8_strided_cn) {
29711 TEST_REQUIRES_ARM_NEON_DOT;
29712 for (uint32_t n = 9; n < 16; n++) {
29713 for (size_t k = 1; k <= 40; k += 9) {
29714 GemmMicrokernelTester()
29715 .mr(6)
29716 .nr(8)
29717 .kr(4)
29718 .sr(1)
29719 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029720 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070029721 .k(k)
29722 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080029723 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029724 }
29725 }
29726 }
29727
29728 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, n_gt_8_subtile) {
29729 TEST_REQUIRES_ARM_NEON_DOT;
29730 for (uint32_t n = 9; n < 16; n++) {
29731 for (size_t k = 1; k <= 40; k += 9) {
29732 for (uint32_t m = 1; m <= 6; m++) {
29733 GemmMicrokernelTester()
29734 .mr(6)
29735 .nr(8)
29736 .kr(4)
29737 .sr(1)
29738 .m(m)
29739 .n(n)
29740 .k(k)
29741 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029742 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029743 }
29744 }
29745 }
29746 }
29747
29748 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, n_div_8) {
29749 TEST_REQUIRES_ARM_NEON_DOT;
29750 for (uint32_t n = 16; n <= 24; n += 8) {
29751 for (size_t k = 1; k <= 40; k += 9) {
29752 GemmMicrokernelTester()
29753 .mr(6)
29754 .nr(8)
29755 .kr(4)
29756 .sr(1)
29757 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029758 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070029759 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080029760 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029761 }
29762 }
29763 }
29764
29765 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, n_div_8_strided_cn) {
29766 TEST_REQUIRES_ARM_NEON_DOT;
29767 for (uint32_t n = 16; n <= 24; n += 8) {
29768 for (size_t k = 1; k <= 40; k += 9) {
29769 GemmMicrokernelTester()
29770 .mr(6)
29771 .nr(8)
29772 .kr(4)
29773 .sr(1)
29774 .m(6)
29775 .n(n)
29776 .k(k)
29777 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080029778 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029779 }
29780 }
29781 }
29782
29783 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, n_div_8_subtile) {
29784 TEST_REQUIRES_ARM_NEON_DOT;
29785 for (uint32_t n = 16; n <= 24; n += 8) {
29786 for (size_t k = 1; k <= 40; k += 9) {
29787 for (uint32_t m = 1; m <= 6; m++) {
29788 GemmMicrokernelTester()
29789 .mr(6)
29790 .nr(8)
29791 .kr(4)
29792 .sr(1)
29793 .m(m)
29794 .n(n)
29795 .k(k)
29796 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029797 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029798 }
29799 }
29800 }
29801 }
29802
29803 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, small_kernel) {
29804 TEST_REQUIRES_ARM_NEON_DOT;
29805 for (size_t k = 1; k <= 40; k += 9) {
29806 GemmMicrokernelTester()
29807 .mr(6)
29808 .nr(8)
29809 .kr(4)
29810 .sr(1)
29811 .m(6)
29812 .n(8)
29813 .k(k)
29814 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080029815 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029816 }
29817 }
29818
29819 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, small_kernel_subtile) {
29820 TEST_REQUIRES_ARM_NEON_DOT;
29821 for (size_t k = 1; k <= 40; k += 9) {
29822 for (uint32_t m = 1; m <= 6; m++) {
29823 for (uint32_t n = 1; n <= 8; n++) {
29824 GemmMicrokernelTester()
29825 .mr(6)
29826 .nr(8)
29827 .kr(4)
29828 .sr(1)
29829 .m(m)
29830 .n(n)
29831 .k(k)
29832 .ks(3)
29833 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029834 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029835 }
29836 }
29837 }
29838 }
29839
29840 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, n_gt_8_small_kernel) {
29841 TEST_REQUIRES_ARM_NEON_DOT;
29842 for (uint32_t n = 9; n < 16; n++) {
29843 for (size_t k = 1; k <= 40; k += 9) {
29844 GemmMicrokernelTester()
29845 .mr(6)
29846 .nr(8)
29847 .kr(4)
29848 .sr(1)
29849 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029850 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070029851 .k(k)
29852 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080029853 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029854 }
29855 }
29856 }
29857
29858 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, n_div_8_small_kernel) {
29859 TEST_REQUIRES_ARM_NEON_DOT;
29860 for (uint32_t n = 16; n <= 24; n += 8) {
29861 for (size_t k = 1; k <= 40; k += 9) {
29862 GemmMicrokernelTester()
29863 .mr(6)
29864 .nr(8)
29865 .kr(4)
29866 .sr(1)
29867 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080029868 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070029869 .k(k)
29870 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080029871 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029872 }
29873 }
29874 }
29875
29876 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, strided_cm_subtile) {
29877 TEST_REQUIRES_ARM_NEON_DOT;
29878 for (size_t k = 1; k <= 40; k += 9) {
29879 for (uint32_t m = 1; m <= 6; m++) {
29880 for (uint32_t n = 1; n <= 8; n++) {
29881 GemmMicrokernelTester()
29882 .mr(6)
29883 .nr(8)
29884 .kr(4)
29885 .sr(1)
29886 .m(m)
29887 .n(n)
29888 .k(k)
29889 .cm_stride(11)
29890 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080029891 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029892 }
29893 }
29894 }
29895 }
29896
29897 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, a_offset) {
29898 TEST_REQUIRES_ARM_NEON_DOT;
29899 for (size_t k = 1; k <= 40; k += 9) {
29900 GemmMicrokernelTester()
29901 .mr(6)
29902 .nr(8)
29903 .kr(4)
29904 .sr(1)
29905 .m(6)
29906 .n(8)
29907 .k(k)
29908 .ks(3)
29909 .a_offset(251)
Marat Dukhan50323b82022-01-11 00:12:01 -080029910 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029911 }
29912 }
29913
29914 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, zero) {
29915 TEST_REQUIRES_ARM_NEON_DOT;
29916 for (uint32_t mz = 0; mz < 6; mz++) {
29917 for (size_t k = 1; k <= 40; k += 9) {
29918 GemmMicrokernelTester()
29919 .mr(6)
29920 .nr(8)
29921 .kr(4)
29922 .sr(1)
29923 .m(6)
29924 .n(8)
29925 .k(k)
29926 .ks(3)
29927 .a_offset(251)
29928 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080029929 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029930 }
29931 }
29932 }
29933
29934 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, qmin) {
29935 TEST_REQUIRES_ARM_NEON_DOT;
29936 GemmMicrokernelTester()
29937 .mr(6)
29938 .nr(8)
29939 .kr(4)
29940 .sr(1)
29941 .m(6)
29942 .n(8)
29943 .k(8)
29944 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080029945 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029946 }
29947
29948 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, qmax) {
29949 TEST_REQUIRES_ARM_NEON_DOT;
29950 GemmMicrokernelTester()
29951 .mr(6)
29952 .nr(8)
29953 .kr(4)
29954 .sr(1)
29955 .m(6)
29956 .n(8)
29957 .k(8)
29958 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080029959 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029960 }
29961
29962 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8C4__NEONDOT, strided_cm) {
29963 TEST_REQUIRES_ARM_NEON_DOT;
29964 GemmMicrokernelTester()
29965 .mr(6)
29966 .nr(8)
29967 .kr(4)
29968 .sr(1)
29969 .m(6)
29970 .n(8)
29971 .k(8)
29972 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080029973 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029974 }
29975#endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
29976
29977
29978#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
29979 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_eq_8) {
29980 TEST_REQUIRES_ARM_NEON_DOT;
29981 GemmMicrokernelTester()
29982 .mr(8)
29983 .nr(8)
29984 .kr(4)
29985 .sr(1)
29986 .m(8)
29987 .n(8)
29988 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080029989 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070029990 }
29991
29992 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, strided_cn) {
29993 TEST_REQUIRES_ARM_NEON_DOT;
29994 GemmMicrokernelTester()
29995 .mr(8)
29996 .nr(8)
29997 .kr(4)
29998 .sr(1)
29999 .m(8)
30000 .n(8)
30001 .k(8)
30002 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080030003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030004 }
30005
30006 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_eq_8_subtile) {
30007 TEST_REQUIRES_ARM_NEON_DOT;
30008 for (uint32_t m = 1; m <= 8; m++) {
30009 for (uint32_t n = 1; n <= 8; n++) {
30010 GemmMicrokernelTester()
30011 .mr(8)
30012 .nr(8)
30013 .kr(4)
30014 .sr(1)
30015 .m(m)
30016 .n(n)
30017 .k(8)
30018 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030019 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030020 }
30021 }
30022 }
30023
30024 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_eq_8_subtile_m) {
30025 TEST_REQUIRES_ARM_NEON_DOT;
30026 for (uint32_t m = 1; m <= 8; m++) {
30027 GemmMicrokernelTester()
30028 .mr(8)
30029 .nr(8)
30030 .kr(4)
30031 .sr(1)
30032 .m(m)
30033 .n(8)
30034 .k(8)
30035 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030036 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030037 }
30038 }
30039
30040 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_eq_8_subtile_n) {
30041 TEST_REQUIRES_ARM_NEON_DOT;
30042 for (uint32_t n = 1; n <= 8; n++) {
30043 GemmMicrokernelTester()
30044 .mr(8)
30045 .nr(8)
30046 .kr(4)
30047 .sr(1)
30048 .m(8)
30049 .n(n)
30050 .k(8)
30051 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030052 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030053 }
30054 }
30055
30056 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_lt_8) {
30057 TEST_REQUIRES_ARM_NEON_DOT;
30058 for (size_t k = 1; k < 8; k++) {
30059 GemmMicrokernelTester()
30060 .mr(8)
30061 .nr(8)
30062 .kr(4)
30063 .sr(1)
30064 .m(8)
30065 .n(8)
30066 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030067 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030068 }
30069 }
30070
30071 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_lt_8_subtile) {
30072 TEST_REQUIRES_ARM_NEON_DOT;
30073 for (size_t k = 1; k < 8; k++) {
30074 for (uint32_t m = 1; m <= 8; m++) {
30075 for (uint32_t n = 1; n <= 8; n++) {
30076 GemmMicrokernelTester()
30077 .mr(8)
30078 .nr(8)
30079 .kr(4)
30080 .sr(1)
30081 .m(m)
30082 .n(n)
30083 .k(k)
30084 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030085 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030086 }
30087 }
30088 }
30089 }
30090
30091 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_gt_8) {
30092 TEST_REQUIRES_ARM_NEON_DOT;
30093 for (size_t k = 9; k < 16; k++) {
30094 GemmMicrokernelTester()
30095 .mr(8)
30096 .nr(8)
30097 .kr(4)
30098 .sr(1)
30099 .m(8)
30100 .n(8)
30101 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030102 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030103 }
30104 }
30105
30106 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_gt_8_subtile) {
30107 TEST_REQUIRES_ARM_NEON_DOT;
30108 for (size_t k = 9; k < 16; k++) {
30109 for (uint32_t m = 1; m <= 8; m++) {
30110 for (uint32_t n = 1; n <= 8; n++) {
30111 GemmMicrokernelTester()
30112 .mr(8)
30113 .nr(8)
30114 .kr(4)
30115 .sr(1)
30116 .m(m)
30117 .n(n)
30118 .k(k)
30119 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030120 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030121 }
30122 }
30123 }
30124 }
30125
30126 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_div_8) {
30127 TEST_REQUIRES_ARM_NEON_DOT;
30128 for (size_t k = 16; k <= 80; k += 8) {
30129 GemmMicrokernelTester()
30130 .mr(8)
30131 .nr(8)
30132 .kr(4)
30133 .sr(1)
30134 .m(8)
30135 .n(8)
30136 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030137 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030138 }
30139 }
30140
30141 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, k_div_8_subtile) {
30142 TEST_REQUIRES_ARM_NEON_DOT;
30143 for (size_t k = 16; k <= 80; k += 8) {
30144 for (uint32_t m = 1; m <= 8; m++) {
30145 for (uint32_t n = 1; n <= 8; n++) {
30146 GemmMicrokernelTester()
30147 .mr(8)
30148 .nr(8)
30149 .kr(4)
30150 .sr(1)
30151 .m(m)
30152 .n(n)
30153 .k(k)
30154 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030155 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030156 }
30157 }
30158 }
30159 }
30160
30161 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, n_gt_8) {
30162 TEST_REQUIRES_ARM_NEON_DOT;
30163 for (uint32_t n = 9; n < 16; n++) {
30164 for (size_t k = 1; k <= 40; k += 9) {
30165 GemmMicrokernelTester()
30166 .mr(8)
30167 .nr(8)
30168 .kr(4)
30169 .sr(1)
30170 .m(8)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030171 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070030172 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030173 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030174 }
30175 }
30176 }
30177
30178 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, n_gt_8_strided_cn) {
30179 TEST_REQUIRES_ARM_NEON_DOT;
30180 for (uint32_t n = 9; n < 16; n++) {
30181 for (size_t k = 1; k <= 40; k += 9) {
30182 GemmMicrokernelTester()
30183 .mr(8)
30184 .nr(8)
30185 .kr(4)
30186 .sr(1)
30187 .m(8)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030188 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070030189 .k(k)
30190 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080030191 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030192 }
30193 }
30194 }
30195
30196 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, n_gt_8_subtile) {
30197 TEST_REQUIRES_ARM_NEON_DOT;
30198 for (uint32_t n = 9; n < 16; n++) {
30199 for (size_t k = 1; k <= 40; k += 9) {
30200 for (uint32_t m = 1; m <= 8; m++) {
30201 GemmMicrokernelTester()
30202 .mr(8)
30203 .nr(8)
30204 .kr(4)
30205 .sr(1)
30206 .m(m)
30207 .n(n)
30208 .k(k)
30209 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030210 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030211 }
30212 }
30213 }
30214 }
30215
30216 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, n_div_8) {
30217 TEST_REQUIRES_ARM_NEON_DOT;
30218 for (uint32_t n = 16; n <= 24; n += 8) {
30219 for (size_t k = 1; k <= 40; k += 9) {
30220 GemmMicrokernelTester()
30221 .mr(8)
30222 .nr(8)
30223 .kr(4)
30224 .sr(1)
30225 .m(8)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030226 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070030227 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030228 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030229 }
30230 }
30231 }
30232
30233 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, n_div_8_strided_cn) {
30234 TEST_REQUIRES_ARM_NEON_DOT;
30235 for (uint32_t n = 16; n <= 24; n += 8) {
30236 for (size_t k = 1; k <= 40; k += 9) {
30237 GemmMicrokernelTester()
30238 .mr(8)
30239 .nr(8)
30240 .kr(4)
30241 .sr(1)
30242 .m(8)
30243 .n(n)
30244 .k(k)
30245 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080030246 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030247 }
30248 }
30249 }
30250
30251 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, n_div_8_subtile) {
30252 TEST_REQUIRES_ARM_NEON_DOT;
30253 for (uint32_t n = 16; n <= 24; n += 8) {
30254 for (size_t k = 1; k <= 40; k += 9) {
30255 for (uint32_t m = 1; m <= 8; m++) {
30256 GemmMicrokernelTester()
30257 .mr(8)
30258 .nr(8)
30259 .kr(4)
30260 .sr(1)
30261 .m(m)
30262 .n(n)
30263 .k(k)
30264 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030265 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030266 }
30267 }
30268 }
30269 }
30270
30271 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, small_kernel) {
30272 TEST_REQUIRES_ARM_NEON_DOT;
30273 for (size_t k = 1; k <= 40; k += 9) {
30274 GemmMicrokernelTester()
30275 .mr(8)
30276 .nr(8)
30277 .kr(4)
30278 .sr(1)
30279 .m(8)
30280 .n(8)
30281 .k(k)
30282 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080030283 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030284 }
30285 }
30286
30287 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, small_kernel_subtile) {
30288 TEST_REQUIRES_ARM_NEON_DOT;
30289 for (size_t k = 1; k <= 40; k += 9) {
30290 for (uint32_t m = 1; m <= 8; m++) {
30291 for (uint32_t n = 1; n <= 8; n++) {
30292 GemmMicrokernelTester()
30293 .mr(8)
30294 .nr(8)
30295 .kr(4)
30296 .sr(1)
30297 .m(m)
30298 .n(n)
30299 .k(k)
30300 .ks(3)
30301 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030302 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030303 }
30304 }
30305 }
30306 }
30307
30308 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, n_gt_8_small_kernel) {
30309 TEST_REQUIRES_ARM_NEON_DOT;
30310 for (uint32_t n = 9; n < 16; n++) {
30311 for (size_t k = 1; k <= 40; k += 9) {
30312 GemmMicrokernelTester()
30313 .mr(8)
30314 .nr(8)
30315 .kr(4)
30316 .sr(1)
30317 .m(8)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030318 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070030319 .k(k)
30320 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080030321 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030322 }
30323 }
30324 }
30325
30326 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, n_div_8_small_kernel) {
30327 TEST_REQUIRES_ARM_NEON_DOT;
30328 for (uint32_t n = 16; n <= 24; n += 8) {
30329 for (size_t k = 1; k <= 40; k += 9) {
30330 GemmMicrokernelTester()
30331 .mr(8)
30332 .nr(8)
30333 .kr(4)
30334 .sr(1)
30335 .m(8)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030336 .n(n)
Marat Dukhan4486f872021-08-07 15:22:50 -070030337 .k(k)
30338 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080030339 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030340 }
30341 }
30342 }
30343
30344 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, strided_cm_subtile) {
30345 TEST_REQUIRES_ARM_NEON_DOT;
30346 for (size_t k = 1; k <= 40; k += 9) {
30347 for (uint32_t m = 1; m <= 8; m++) {
30348 for (uint32_t n = 1; n <= 8; n++) {
30349 GemmMicrokernelTester()
30350 .mr(8)
30351 .nr(8)
30352 .kr(4)
30353 .sr(1)
30354 .m(m)
30355 .n(n)
30356 .k(k)
30357 .cm_stride(11)
30358 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030359 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030360 }
30361 }
30362 }
30363 }
30364
30365 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, a_offset) {
30366 TEST_REQUIRES_ARM_NEON_DOT;
30367 for (size_t k = 1; k <= 40; k += 9) {
30368 GemmMicrokernelTester()
30369 .mr(8)
30370 .nr(8)
30371 .kr(4)
30372 .sr(1)
30373 .m(8)
30374 .n(8)
30375 .k(k)
30376 .ks(3)
30377 .a_offset(331)
Marat Dukhan50323b82022-01-11 00:12:01 -080030378 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030379 }
30380 }
30381
30382 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, zero) {
30383 TEST_REQUIRES_ARM_NEON_DOT;
30384 for (uint32_t mz = 0; mz < 8; mz++) {
30385 for (size_t k = 1; k <= 40; k += 9) {
30386 GemmMicrokernelTester()
30387 .mr(8)
30388 .nr(8)
30389 .kr(4)
30390 .sr(1)
30391 .m(8)
30392 .n(8)
30393 .k(k)
30394 .ks(3)
30395 .a_offset(331)
30396 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080030397 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030398 }
30399 }
30400 }
30401
30402 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, qmin) {
30403 TEST_REQUIRES_ARM_NEON_DOT;
30404 GemmMicrokernelTester()
30405 .mr(8)
30406 .nr(8)
30407 .kr(4)
30408 .sr(1)
30409 .m(8)
30410 .n(8)
30411 .k(8)
30412 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080030413 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030414 }
30415
30416 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, qmax) {
30417 TEST_REQUIRES_ARM_NEON_DOT;
30418 GemmMicrokernelTester()
30419 .mr(8)
30420 .nr(8)
30421 .kr(4)
30422 .sr(1)
30423 .m(8)
30424 .n(8)
30425 .k(8)
30426 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080030427 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030428 }
30429
30430 TEST(QS8_IGEMM_MINMAX_RNDNU_8X8C4__NEONDOT, strided_cm) {
30431 TEST_REQUIRES_ARM_NEON_DOT;
30432 GemmMicrokernelTester()
30433 .mr(8)
30434 .nr(8)
30435 .kr(4)
30436 .sr(1)
30437 .m(8)
30438 .n(8)
30439 .k(8)
30440 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080030441 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhan4486f872021-08-07 15:22:50 -070030442 }
30443#endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
30444
30445
30446#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
Marat Dukhane903dff2021-07-16 19:43:41 -070030447 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8) {
30448 TEST_REQUIRES_ARM_NEON_DOT;
30449 GemmMicrokernelTester()
30450 .mr(4)
30451 .nr(16)
30452 .kr(4)
30453 .sr(1)
30454 .m(4)
30455 .n(16)
30456 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080030457 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030458 }
30459
30460 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, strided_cn) {
30461 TEST_REQUIRES_ARM_NEON_DOT;
30462 GemmMicrokernelTester()
30463 .mr(4)
30464 .nr(16)
30465 .kr(4)
30466 .sr(1)
30467 .m(4)
30468 .n(16)
30469 .k(8)
30470 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080030471 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030472 }
30473
30474 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8_subtile) {
30475 TEST_REQUIRES_ARM_NEON_DOT;
30476 for (uint32_t m = 1; m <= 4; m++) {
30477 for (uint32_t n = 1; n <= 16; n++) {
30478 GemmMicrokernelTester()
30479 .mr(4)
30480 .nr(16)
30481 .kr(4)
30482 .sr(1)
30483 .m(m)
30484 .n(n)
30485 .k(8)
30486 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030487 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030488 }
30489 }
30490 }
30491
30492 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8_subtile_m) {
30493 TEST_REQUIRES_ARM_NEON_DOT;
30494 for (uint32_t m = 1; m <= 4; m++) {
30495 GemmMicrokernelTester()
30496 .mr(4)
30497 .nr(16)
30498 .kr(4)
30499 .sr(1)
30500 .m(m)
30501 .n(16)
30502 .k(8)
30503 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030504 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030505 }
30506 }
30507
30508 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8_subtile_n) {
30509 TEST_REQUIRES_ARM_NEON_DOT;
30510 for (uint32_t n = 1; n <= 16; n++) {
30511 GemmMicrokernelTester()
30512 .mr(4)
30513 .nr(16)
30514 .kr(4)
30515 .sr(1)
30516 .m(4)
30517 .n(n)
30518 .k(8)
30519 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030520 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030521 }
30522 }
30523
30524 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_lt_8) {
30525 TEST_REQUIRES_ARM_NEON_DOT;
30526 for (size_t k = 1; k < 8; k++) {
30527 GemmMicrokernelTester()
30528 .mr(4)
30529 .nr(16)
30530 .kr(4)
30531 .sr(1)
30532 .m(4)
30533 .n(16)
30534 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030536 }
30537 }
30538
30539 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_lt_8_subtile) {
30540 TEST_REQUIRES_ARM_NEON_DOT;
30541 for (size_t k = 1; k < 8; k++) {
30542 for (uint32_t m = 1; m <= 4; m++) {
30543 for (uint32_t n = 1; n <= 16; n++) {
30544 GemmMicrokernelTester()
30545 .mr(4)
30546 .nr(16)
30547 .kr(4)
30548 .sr(1)
30549 .m(m)
30550 .n(n)
30551 .k(k)
30552 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030553 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030554 }
30555 }
30556 }
30557 }
30558
30559 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_gt_8) {
30560 TEST_REQUIRES_ARM_NEON_DOT;
30561 for (size_t k = 9; k < 16; k++) {
30562 GemmMicrokernelTester()
30563 .mr(4)
30564 .nr(16)
30565 .kr(4)
30566 .sr(1)
30567 .m(4)
30568 .n(16)
30569 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030570 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030571 }
30572 }
30573
30574 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_gt_8_subtile) {
30575 TEST_REQUIRES_ARM_NEON_DOT;
30576 for (size_t k = 9; k < 16; k++) {
30577 for (uint32_t m = 1; m <= 4; m++) {
30578 for (uint32_t n = 1; n <= 16; n++) {
30579 GemmMicrokernelTester()
30580 .mr(4)
30581 .nr(16)
30582 .kr(4)
30583 .sr(1)
30584 .m(m)
30585 .n(n)
30586 .k(k)
30587 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030588 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030589 }
30590 }
30591 }
30592 }
30593
30594 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_div_8) {
30595 TEST_REQUIRES_ARM_NEON_DOT;
30596 for (size_t k = 16; k <= 80; k += 8) {
30597 GemmMicrokernelTester()
30598 .mr(4)
30599 .nr(16)
30600 .kr(4)
30601 .sr(1)
30602 .m(4)
30603 .n(16)
30604 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030605 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030606 }
30607 }
30608
30609 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_div_8_subtile) {
30610 TEST_REQUIRES_ARM_NEON_DOT;
30611 for (size_t k = 16; k <= 80; k += 8) {
30612 for (uint32_t m = 1; m <= 4; m++) {
30613 for (uint32_t n = 1; n <= 16; n++) {
30614 GemmMicrokernelTester()
30615 .mr(4)
30616 .nr(16)
30617 .kr(4)
30618 .sr(1)
30619 .m(m)
30620 .n(n)
30621 .k(k)
30622 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030623 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030624 }
30625 }
30626 }
30627 }
30628
30629 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_gt_16) {
30630 TEST_REQUIRES_ARM_NEON_DOT;
30631 for (uint32_t n = 17; n < 32; n++) {
30632 for (size_t k = 1; k <= 40; k += 9) {
30633 GemmMicrokernelTester()
30634 .mr(4)
30635 .nr(16)
30636 .kr(4)
30637 .sr(1)
30638 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030639 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070030640 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030641 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030642 }
30643 }
30644 }
30645
30646 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_gt_16_strided_cn) {
30647 TEST_REQUIRES_ARM_NEON_DOT;
30648 for (uint32_t n = 17; n < 32; n++) {
30649 for (size_t k = 1; k <= 40; k += 9) {
30650 GemmMicrokernelTester()
30651 .mr(4)
30652 .nr(16)
30653 .kr(4)
30654 .sr(1)
30655 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030656 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070030657 .k(k)
30658 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080030659 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030660 }
30661 }
30662 }
30663
30664 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_gt_16_subtile) {
30665 TEST_REQUIRES_ARM_NEON_DOT;
30666 for (uint32_t n = 17; n < 32; n++) {
30667 for (size_t k = 1; k <= 40; k += 9) {
30668 for (uint32_t m = 1; m <= 4; m++) {
30669 GemmMicrokernelTester()
30670 .mr(4)
30671 .nr(16)
30672 .kr(4)
30673 .sr(1)
30674 .m(m)
30675 .n(n)
30676 .k(k)
30677 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030678 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030679 }
30680 }
30681 }
30682 }
30683
30684 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_div_16) {
30685 TEST_REQUIRES_ARM_NEON_DOT;
30686 for (uint32_t n = 32; n <= 48; n += 16) {
30687 for (size_t k = 1; k <= 40; k += 9) {
30688 GemmMicrokernelTester()
30689 .mr(4)
30690 .nr(16)
30691 .kr(4)
30692 .sr(1)
30693 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030694 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070030695 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080030696 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030697 }
30698 }
30699 }
30700
30701 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_div_16_strided_cn) {
30702 TEST_REQUIRES_ARM_NEON_DOT;
30703 for (uint32_t n = 32; n <= 48; n += 16) {
30704 for (size_t k = 1; k <= 40; k += 9) {
30705 GemmMicrokernelTester()
30706 .mr(4)
30707 .nr(16)
30708 .kr(4)
30709 .sr(1)
30710 .m(4)
30711 .n(n)
30712 .k(k)
30713 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080030714 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030715 }
30716 }
30717 }
30718
30719 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_div_16_subtile) {
30720 TEST_REQUIRES_ARM_NEON_DOT;
30721 for (uint32_t n = 32; n <= 48; n += 16) {
30722 for (size_t k = 1; k <= 40; k += 9) {
30723 for (uint32_t m = 1; m <= 4; m++) {
30724 GemmMicrokernelTester()
30725 .mr(4)
30726 .nr(16)
30727 .kr(4)
30728 .sr(1)
30729 .m(m)
30730 .n(n)
30731 .k(k)
30732 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030733 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030734 }
30735 }
30736 }
30737 }
30738
30739 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, small_kernel) {
30740 TEST_REQUIRES_ARM_NEON_DOT;
30741 for (size_t k = 1; k <= 40; k += 9) {
30742 GemmMicrokernelTester()
30743 .mr(4)
30744 .nr(16)
30745 .kr(4)
30746 .sr(1)
30747 .m(4)
30748 .n(16)
30749 .k(k)
30750 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080030751 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030752 }
30753 }
30754
30755 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, small_kernel_subtile) {
30756 TEST_REQUIRES_ARM_NEON_DOT;
30757 for (size_t k = 1; k <= 40; k += 9) {
30758 for (uint32_t m = 1; m <= 4; m++) {
30759 for (uint32_t n = 1; n <= 16; n++) {
30760 GemmMicrokernelTester()
30761 .mr(4)
30762 .nr(16)
30763 .kr(4)
30764 .sr(1)
30765 .m(m)
30766 .n(n)
30767 .k(k)
30768 .ks(3)
30769 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030770 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030771 }
30772 }
30773 }
30774 }
30775
30776 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_gt_16_small_kernel) {
30777 TEST_REQUIRES_ARM_NEON_DOT;
30778 for (uint32_t n = 17; n < 32; n++) {
30779 for (size_t k = 1; k <= 40; k += 9) {
30780 GemmMicrokernelTester()
30781 .mr(4)
30782 .nr(16)
30783 .kr(4)
30784 .sr(1)
30785 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030786 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070030787 .k(k)
30788 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080030789 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030790 }
30791 }
30792 }
30793
30794 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_div_16_small_kernel) {
30795 TEST_REQUIRES_ARM_NEON_DOT;
30796 for (uint32_t n = 32; n <= 48; n += 16) {
30797 for (size_t k = 1; k <= 40; k += 9) {
30798 GemmMicrokernelTester()
30799 .mr(4)
30800 .nr(16)
30801 .kr(4)
30802 .sr(1)
30803 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080030804 .n(n)
Marat Dukhane903dff2021-07-16 19:43:41 -070030805 .k(k)
30806 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080030807 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030808 }
30809 }
30810 }
30811
30812 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, strided_cm_subtile) {
30813 TEST_REQUIRES_ARM_NEON_DOT;
30814 for (size_t k = 1; k <= 40; k += 9) {
30815 for (uint32_t m = 1; m <= 4; m++) {
30816 for (uint32_t n = 1; n <= 16; n++) {
30817 GemmMicrokernelTester()
30818 .mr(4)
30819 .nr(16)
30820 .kr(4)
30821 .sr(1)
30822 .m(m)
30823 .n(n)
30824 .k(k)
30825 .cm_stride(19)
30826 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030827 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030828 }
30829 }
30830 }
30831 }
30832
30833 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, a_offset) {
30834 TEST_REQUIRES_ARM_NEON_DOT;
30835 for (size_t k = 1; k <= 40; k += 9) {
30836 GemmMicrokernelTester()
30837 .mr(4)
30838 .nr(16)
30839 .kr(4)
30840 .sr(1)
30841 .m(4)
30842 .n(16)
30843 .k(k)
30844 .ks(3)
30845 .a_offset(163)
Marat Dukhan50323b82022-01-11 00:12:01 -080030846 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030847 }
30848 }
30849
30850 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, zero) {
30851 TEST_REQUIRES_ARM_NEON_DOT;
30852 for (uint32_t mz = 0; mz < 4; mz++) {
30853 for (size_t k = 1; k <= 40; k += 9) {
30854 GemmMicrokernelTester()
30855 .mr(4)
30856 .nr(16)
30857 .kr(4)
30858 .sr(1)
30859 .m(4)
30860 .n(16)
30861 .k(k)
30862 .ks(3)
30863 .a_offset(163)
30864 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080030865 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030866 }
30867 }
30868 }
30869
30870 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, qmin) {
30871 TEST_REQUIRES_ARM_NEON_DOT;
30872 GemmMicrokernelTester()
30873 .mr(4)
30874 .nr(16)
30875 .kr(4)
30876 .sr(1)
30877 .m(4)
30878 .n(16)
30879 .k(8)
30880 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080030881 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030882 }
30883
30884 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, qmax) {
30885 TEST_REQUIRES_ARM_NEON_DOT;
30886 GemmMicrokernelTester()
30887 .mr(4)
30888 .nr(16)
30889 .kr(4)
30890 .sr(1)
30891 .m(4)
30892 .n(16)
30893 .k(8)
30894 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080030895 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030896 }
30897
30898 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__NEONDOT, strided_cm) {
30899 TEST_REQUIRES_ARM_NEON_DOT;
30900 GemmMicrokernelTester()
30901 .mr(4)
30902 .nr(16)
30903 .kr(4)
30904 .sr(1)
30905 .m(4)
30906 .n(16)
30907 .k(8)
30908 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080030909 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Marat Dukhane903dff2021-07-16 19:43:41 -070030910 }
30911#endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
Frank Barchard22fbe772021-07-20 15:56:32 -070030912
30913
30914#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barchard27bf92c2021-11-24 15:47:52 -080030915 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_eq_8) {
30916 TEST_REQUIRES_ARM_NEON;
30917 GemmMicrokernelTester()
30918 .mr(1)
30919 .nr(8)
30920 .kr(1)
30921 .sr(1)
30922 .m(1)
30923 .n(8)
30924 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080030925 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080030926 }
30927
30928 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, strided_cn) {
30929 TEST_REQUIRES_ARM_NEON;
30930 GemmMicrokernelTester()
30931 .mr(1)
30932 .nr(8)
30933 .kr(1)
30934 .sr(1)
30935 .m(1)
30936 .n(8)
30937 .k(8)
30938 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080030939 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080030940 }
30941
30942 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_eq_8_subtile) {
30943 TEST_REQUIRES_ARM_NEON;
30944 for (uint32_t m = 1; m <= 1; m++) {
30945 for (uint32_t n = 1; n <= 8; n++) {
30946 GemmMicrokernelTester()
30947 .mr(1)
30948 .nr(8)
30949 .kr(1)
30950 .sr(1)
30951 .m(m)
30952 .n(n)
30953 .k(8)
30954 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030955 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080030956 }
30957 }
30958 }
30959
30960 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_eq_8_subtile_m) {
30961 TEST_REQUIRES_ARM_NEON;
30962 for (uint32_t m = 1; m <= 1; m++) {
30963 GemmMicrokernelTester()
30964 .mr(1)
30965 .nr(8)
30966 .kr(1)
30967 .sr(1)
30968 .m(m)
30969 .n(8)
30970 .k(8)
30971 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030972 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080030973 }
30974 }
30975
30976 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_eq_8_subtile_n) {
30977 TEST_REQUIRES_ARM_NEON;
30978 for (uint32_t n = 1; n <= 8; n++) {
30979 GemmMicrokernelTester()
30980 .mr(1)
30981 .nr(8)
30982 .kr(1)
30983 .sr(1)
30984 .m(1)
30985 .n(n)
30986 .k(8)
30987 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080030988 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080030989 }
30990 }
30991
30992 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_lt_8) {
30993 TEST_REQUIRES_ARM_NEON;
30994 for (size_t k = 1; k < 8; k++) {
30995 GemmMicrokernelTester()
30996 .mr(1)
30997 .nr(8)
30998 .kr(1)
30999 .sr(1)
31000 .m(1)
31001 .n(8)
31002 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031004 }
31005 }
31006
31007 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_lt_8_subtile) {
31008 TEST_REQUIRES_ARM_NEON;
31009 for (size_t k = 1; k < 8; k++) {
31010 for (uint32_t m = 1; m <= 1; m++) {
31011 for (uint32_t n = 1; n <= 8; n++) {
31012 GemmMicrokernelTester()
31013 .mr(1)
31014 .nr(8)
31015 .kr(1)
31016 .sr(1)
31017 .m(m)
31018 .n(n)
31019 .k(k)
31020 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031021 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031022 }
31023 }
31024 }
31025 }
31026
31027 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_gt_8) {
31028 TEST_REQUIRES_ARM_NEON;
31029 for (size_t k = 9; k < 16; k++) {
31030 GemmMicrokernelTester()
31031 .mr(1)
31032 .nr(8)
31033 .kr(1)
31034 .sr(1)
31035 .m(1)
31036 .n(8)
31037 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031038 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031039 }
31040 }
31041
31042 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_gt_8_subtile) {
31043 TEST_REQUIRES_ARM_NEON;
31044 for (size_t k = 9; k < 16; k++) {
31045 for (uint32_t m = 1; m <= 1; m++) {
31046 for (uint32_t n = 1; n <= 8; n++) {
31047 GemmMicrokernelTester()
31048 .mr(1)
31049 .nr(8)
31050 .kr(1)
31051 .sr(1)
31052 .m(m)
31053 .n(n)
31054 .k(k)
31055 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031056 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031057 }
31058 }
31059 }
31060 }
31061
31062 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_div_8) {
31063 TEST_REQUIRES_ARM_NEON;
31064 for (size_t k = 16; k <= 80; k += 8) {
31065 GemmMicrokernelTester()
31066 .mr(1)
31067 .nr(8)
31068 .kr(1)
31069 .sr(1)
31070 .m(1)
31071 .n(8)
31072 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031073 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031074 }
31075 }
31076
31077 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, k_div_8_subtile) {
31078 TEST_REQUIRES_ARM_NEON;
31079 for (size_t k = 16; k <= 80; k += 8) {
31080 for (uint32_t m = 1; m <= 1; m++) {
31081 for (uint32_t n = 1; n <= 8; n++) {
31082 GemmMicrokernelTester()
31083 .mr(1)
31084 .nr(8)
31085 .kr(1)
31086 .sr(1)
31087 .m(m)
31088 .n(n)
31089 .k(k)
31090 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031091 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031092 }
31093 }
31094 }
31095 }
31096
31097 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, n_gt_8) {
31098 TEST_REQUIRES_ARM_NEON;
31099 for (uint32_t n = 9; n < 16; n++) {
31100 for (size_t k = 1; k <= 40; k += 9) {
31101 GemmMicrokernelTester()
31102 .mr(1)
31103 .nr(8)
31104 .kr(1)
31105 .sr(1)
31106 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031107 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031108 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031109 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031110 }
31111 }
31112 }
31113
31114 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, n_gt_8_strided_cn) {
31115 TEST_REQUIRES_ARM_NEON;
31116 for (uint32_t n = 9; n < 16; n++) {
31117 for (size_t k = 1; k <= 40; k += 9) {
31118 GemmMicrokernelTester()
31119 .mr(1)
31120 .nr(8)
31121 .kr(1)
31122 .sr(1)
31123 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031124 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031125 .k(k)
31126 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080031127 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031128 }
31129 }
31130 }
31131
31132 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, n_gt_8_subtile) {
31133 TEST_REQUIRES_ARM_NEON;
31134 for (uint32_t n = 9; n < 16; n++) {
31135 for (size_t k = 1; k <= 40; k += 9) {
31136 for (uint32_t m = 1; m <= 1; m++) {
31137 GemmMicrokernelTester()
31138 .mr(1)
31139 .nr(8)
31140 .kr(1)
31141 .sr(1)
31142 .m(m)
31143 .n(n)
31144 .k(k)
31145 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031146 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031147 }
31148 }
31149 }
31150 }
31151
31152 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, n_div_8) {
31153 TEST_REQUIRES_ARM_NEON;
31154 for (uint32_t n = 16; n <= 24; n += 8) {
31155 for (size_t k = 1; k <= 40; k += 9) {
31156 GemmMicrokernelTester()
31157 .mr(1)
31158 .nr(8)
31159 .kr(1)
31160 .sr(1)
31161 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031162 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031163 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031164 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031165 }
31166 }
31167 }
31168
31169 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, n_div_8_strided_cn) {
31170 TEST_REQUIRES_ARM_NEON;
31171 for (uint32_t n = 16; n <= 24; n += 8) {
31172 for (size_t k = 1; k <= 40; k += 9) {
31173 GemmMicrokernelTester()
31174 .mr(1)
31175 .nr(8)
31176 .kr(1)
31177 .sr(1)
31178 .m(1)
31179 .n(n)
31180 .k(k)
31181 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080031182 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031183 }
31184 }
31185 }
31186
31187 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, n_div_8_subtile) {
31188 TEST_REQUIRES_ARM_NEON;
31189 for (uint32_t n = 16; n <= 24; n += 8) {
31190 for (size_t k = 1; k <= 40; k += 9) {
31191 for (uint32_t m = 1; m <= 1; m++) {
31192 GemmMicrokernelTester()
31193 .mr(1)
31194 .nr(8)
31195 .kr(1)
31196 .sr(1)
31197 .m(m)
31198 .n(n)
31199 .k(k)
31200 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031201 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031202 }
31203 }
31204 }
31205 }
31206
31207 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, small_kernel) {
31208 TEST_REQUIRES_ARM_NEON;
31209 for (size_t k = 1; k <= 40; k += 9) {
31210 GemmMicrokernelTester()
31211 .mr(1)
31212 .nr(8)
31213 .kr(1)
31214 .sr(1)
31215 .m(1)
31216 .n(8)
31217 .k(k)
31218 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080031219 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031220 }
31221 }
31222
31223 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, small_kernel_subtile) {
31224 TEST_REQUIRES_ARM_NEON;
31225 for (size_t k = 1; k <= 40; k += 9) {
31226 for (uint32_t m = 1; m <= 1; m++) {
31227 for (uint32_t n = 1; n <= 8; n++) {
31228 GemmMicrokernelTester()
31229 .mr(1)
31230 .nr(8)
31231 .kr(1)
31232 .sr(1)
31233 .m(m)
31234 .n(n)
31235 .k(k)
31236 .ks(3)
31237 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031238 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031239 }
31240 }
31241 }
31242 }
31243
31244 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, n_gt_8_small_kernel) {
31245 TEST_REQUIRES_ARM_NEON;
31246 for (uint32_t n = 9; n < 16; n++) {
31247 for (size_t k = 1; k <= 40; k += 9) {
31248 GemmMicrokernelTester()
31249 .mr(1)
31250 .nr(8)
31251 .kr(1)
31252 .sr(1)
31253 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031254 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031255 .k(k)
31256 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080031257 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031258 }
31259 }
31260 }
31261
31262 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, n_div_8_small_kernel) {
31263 TEST_REQUIRES_ARM_NEON;
31264 for (uint32_t n = 16; n <= 24; n += 8) {
31265 for (size_t k = 1; k <= 40; k += 9) {
31266 GemmMicrokernelTester()
31267 .mr(1)
31268 .nr(8)
31269 .kr(1)
31270 .sr(1)
31271 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031272 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031273 .k(k)
31274 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080031275 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031276 }
31277 }
31278 }
31279
31280 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, strided_cm_subtile) {
31281 TEST_REQUIRES_ARM_NEON;
31282 for (size_t k = 1; k <= 40; k += 9) {
31283 for (uint32_t m = 1; m <= 1; m++) {
31284 for (uint32_t n = 1; n <= 8; n++) {
31285 GemmMicrokernelTester()
31286 .mr(1)
31287 .nr(8)
31288 .kr(1)
31289 .sr(1)
31290 .m(m)
31291 .n(n)
31292 .k(k)
31293 .cm_stride(11)
31294 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031295 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031296 }
31297 }
31298 }
31299 }
31300
31301 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, a_offset) {
31302 TEST_REQUIRES_ARM_NEON;
31303 for (size_t k = 1; k <= 40; k += 9) {
31304 GemmMicrokernelTester()
31305 .mr(1)
31306 .nr(8)
31307 .kr(1)
31308 .sr(1)
31309 .m(1)
31310 .n(8)
31311 .k(k)
31312 .ks(3)
31313 .a_offset(43)
Marat Dukhan50323b82022-01-11 00:12:01 -080031314 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031315 }
31316 }
31317
31318 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, zero) {
31319 TEST_REQUIRES_ARM_NEON;
31320 for (uint32_t mz = 0; mz < 1; mz++) {
31321 for (size_t k = 1; k <= 40; k += 9) {
31322 GemmMicrokernelTester()
31323 .mr(1)
31324 .nr(8)
31325 .kr(1)
31326 .sr(1)
31327 .m(1)
31328 .n(8)
31329 .k(k)
31330 .ks(3)
31331 .a_offset(43)
31332 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080031333 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031334 }
31335 }
31336 }
31337
31338 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, qmin) {
31339 TEST_REQUIRES_ARM_NEON;
31340 GemmMicrokernelTester()
31341 .mr(1)
31342 .nr(8)
31343 .kr(1)
31344 .sr(1)
31345 .m(1)
31346 .n(8)
31347 .k(8)
31348 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080031349 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031350 }
31351
31352 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, qmax) {
31353 TEST_REQUIRES_ARM_NEON;
31354 GemmMicrokernelTester()
31355 .mr(1)
31356 .nr(8)
31357 .kr(1)
31358 .sr(1)
31359 .m(1)
31360 .n(8)
31361 .k(8)
31362 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080031363 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031364 }
31365
31366 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MLAL_LANE, strided_cm) {
31367 TEST_REQUIRES_ARM_NEON;
31368 GemmMicrokernelTester()
31369 .mr(1)
31370 .nr(8)
31371 .kr(1)
31372 .sr(1)
31373 .m(1)
31374 .n(8)
31375 .k(8)
31376 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080031377 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031378 }
31379#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
31380
31381
31382#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barchard22fbe772021-07-20 15:56:32 -070031383 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8) {
31384 TEST_REQUIRES_ARM_NEON;
31385 GemmMicrokernelTester()
31386 .mr(1)
31387 .nr(16)
31388 .kr(1)
31389 .sr(1)
31390 .m(1)
31391 .n(16)
31392 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080031393 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031394 }
31395
31396 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, strided_cn) {
31397 TEST_REQUIRES_ARM_NEON;
31398 GemmMicrokernelTester()
31399 .mr(1)
31400 .nr(16)
31401 .kr(1)
31402 .sr(1)
31403 .m(1)
31404 .n(16)
31405 .k(8)
31406 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080031407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031408 }
31409
31410 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_subtile) {
31411 TEST_REQUIRES_ARM_NEON;
31412 for (uint32_t m = 1; m <= 1; m++) {
31413 for (uint32_t n = 1; n <= 16; n++) {
31414 GemmMicrokernelTester()
31415 .mr(1)
31416 .nr(16)
31417 .kr(1)
31418 .sr(1)
31419 .m(m)
31420 .n(n)
31421 .k(8)
31422 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031423 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031424 }
31425 }
31426 }
31427
31428 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_subtile_m) {
31429 TEST_REQUIRES_ARM_NEON;
31430 for (uint32_t m = 1; m <= 1; m++) {
31431 GemmMicrokernelTester()
31432 .mr(1)
31433 .nr(16)
31434 .kr(1)
31435 .sr(1)
31436 .m(m)
31437 .n(16)
31438 .k(8)
31439 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031440 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031441 }
31442 }
31443
31444 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_subtile_n) {
31445 TEST_REQUIRES_ARM_NEON;
31446 for (uint32_t n = 1; n <= 16; n++) {
31447 GemmMicrokernelTester()
31448 .mr(1)
31449 .nr(16)
31450 .kr(1)
31451 .sr(1)
31452 .m(1)
31453 .n(n)
31454 .k(8)
31455 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031456 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031457 }
31458 }
31459
31460 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_lt_8) {
31461 TEST_REQUIRES_ARM_NEON;
31462 for (size_t k = 1; k < 8; k++) {
31463 GemmMicrokernelTester()
31464 .mr(1)
31465 .nr(16)
31466 .kr(1)
31467 .sr(1)
31468 .m(1)
31469 .n(16)
31470 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031471 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031472 }
31473 }
31474
31475 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_lt_8_subtile) {
31476 TEST_REQUIRES_ARM_NEON;
31477 for (size_t k = 1; k < 8; k++) {
31478 for (uint32_t m = 1; m <= 1; m++) {
31479 for (uint32_t n = 1; n <= 16; n++) {
31480 GemmMicrokernelTester()
31481 .mr(1)
31482 .nr(16)
31483 .kr(1)
31484 .sr(1)
31485 .m(m)
31486 .n(n)
31487 .k(k)
31488 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031489 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031490 }
31491 }
31492 }
31493 }
31494
31495 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_gt_8) {
31496 TEST_REQUIRES_ARM_NEON;
31497 for (size_t k = 9; k < 16; k++) {
31498 GemmMicrokernelTester()
31499 .mr(1)
31500 .nr(16)
31501 .kr(1)
31502 .sr(1)
31503 .m(1)
31504 .n(16)
31505 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031506 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031507 }
31508 }
31509
31510 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_gt_8_subtile) {
31511 TEST_REQUIRES_ARM_NEON;
31512 for (size_t k = 9; k < 16; k++) {
31513 for (uint32_t m = 1; m <= 1; m++) {
31514 for (uint32_t n = 1; n <= 16; n++) {
31515 GemmMicrokernelTester()
31516 .mr(1)
31517 .nr(16)
31518 .kr(1)
31519 .sr(1)
31520 .m(m)
31521 .n(n)
31522 .k(k)
31523 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031524 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031525 }
31526 }
31527 }
31528 }
31529
31530 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_div_8) {
31531 TEST_REQUIRES_ARM_NEON;
31532 for (size_t k = 16; k <= 80; k += 8) {
31533 GemmMicrokernelTester()
31534 .mr(1)
31535 .nr(16)
31536 .kr(1)
31537 .sr(1)
31538 .m(1)
31539 .n(16)
31540 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031541 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031542 }
31543 }
31544
31545 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_div_8_subtile) {
31546 TEST_REQUIRES_ARM_NEON;
31547 for (size_t k = 16; k <= 80; k += 8) {
31548 for (uint32_t m = 1; m <= 1; m++) {
31549 for (uint32_t n = 1; n <= 16; n++) {
31550 GemmMicrokernelTester()
31551 .mr(1)
31552 .nr(16)
31553 .kr(1)
31554 .sr(1)
31555 .m(m)
31556 .n(n)
31557 .k(k)
31558 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031559 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031560 }
31561 }
31562 }
31563 }
31564
31565 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16) {
31566 TEST_REQUIRES_ARM_NEON;
31567 for (uint32_t n = 17; n < 32; n++) {
31568 for (size_t k = 1; k <= 40; k += 9) {
31569 GemmMicrokernelTester()
31570 .mr(1)
31571 .nr(16)
31572 .kr(1)
31573 .sr(1)
31574 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031575 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070031576 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031577 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031578 }
31579 }
31580 }
31581
31582 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16_strided_cn) {
31583 TEST_REQUIRES_ARM_NEON;
31584 for (uint32_t n = 17; n < 32; n++) {
31585 for (size_t k = 1; k <= 40; k += 9) {
31586 GemmMicrokernelTester()
31587 .mr(1)
31588 .nr(16)
31589 .kr(1)
31590 .sr(1)
31591 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031592 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070031593 .k(k)
31594 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080031595 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031596 }
31597 }
31598 }
31599
31600 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16_subtile) {
31601 TEST_REQUIRES_ARM_NEON;
31602 for (uint32_t n = 17; n < 32; n++) {
31603 for (size_t k = 1; k <= 40; k += 9) {
31604 for (uint32_t m = 1; m <= 1; m++) {
31605 GemmMicrokernelTester()
31606 .mr(1)
31607 .nr(16)
31608 .kr(1)
31609 .sr(1)
31610 .m(m)
31611 .n(n)
31612 .k(k)
31613 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031614 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031615 }
31616 }
31617 }
31618 }
31619
31620 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16) {
31621 TEST_REQUIRES_ARM_NEON;
31622 for (uint32_t n = 32; n <= 48; n += 16) {
31623 for (size_t k = 1; k <= 40; k += 9) {
31624 GemmMicrokernelTester()
31625 .mr(1)
31626 .nr(16)
31627 .kr(1)
31628 .sr(1)
31629 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031630 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070031631 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080031632 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031633 }
31634 }
31635 }
31636
31637 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16_strided_cn) {
31638 TEST_REQUIRES_ARM_NEON;
31639 for (uint32_t n = 32; n <= 48; n += 16) {
31640 for (size_t k = 1; k <= 40; k += 9) {
31641 GemmMicrokernelTester()
31642 .mr(1)
31643 .nr(16)
31644 .kr(1)
31645 .sr(1)
31646 .m(1)
31647 .n(n)
31648 .k(k)
31649 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080031650 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031651 }
31652 }
31653 }
31654
31655 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16_subtile) {
31656 TEST_REQUIRES_ARM_NEON;
31657 for (uint32_t n = 32; n <= 48; n += 16) {
31658 for (size_t k = 1; k <= 40; k += 9) {
31659 for (uint32_t m = 1; m <= 1; m++) {
31660 GemmMicrokernelTester()
31661 .mr(1)
31662 .nr(16)
31663 .kr(1)
31664 .sr(1)
31665 .m(m)
31666 .n(n)
31667 .k(k)
31668 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031669 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031670 }
31671 }
31672 }
31673 }
31674
31675 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, small_kernel) {
31676 TEST_REQUIRES_ARM_NEON;
31677 for (size_t k = 1; k <= 40; k += 9) {
31678 GemmMicrokernelTester()
31679 .mr(1)
31680 .nr(16)
31681 .kr(1)
31682 .sr(1)
31683 .m(1)
31684 .n(16)
31685 .k(k)
31686 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080031687 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031688 }
31689 }
31690
31691 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, small_kernel_subtile) {
31692 TEST_REQUIRES_ARM_NEON;
31693 for (size_t k = 1; k <= 40; k += 9) {
31694 for (uint32_t m = 1; m <= 1; m++) {
31695 for (uint32_t n = 1; n <= 16; n++) {
31696 GemmMicrokernelTester()
31697 .mr(1)
31698 .nr(16)
31699 .kr(1)
31700 .sr(1)
31701 .m(m)
31702 .n(n)
31703 .k(k)
31704 .ks(3)
31705 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031706 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031707 }
31708 }
31709 }
31710 }
31711
31712 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16_small_kernel) {
31713 TEST_REQUIRES_ARM_NEON;
31714 for (uint32_t n = 17; n < 32; n++) {
31715 for (size_t k = 1; k <= 40; k += 9) {
31716 GemmMicrokernelTester()
31717 .mr(1)
31718 .nr(16)
31719 .kr(1)
31720 .sr(1)
31721 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031722 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070031723 .k(k)
31724 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080031725 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031726 }
31727 }
31728 }
31729
31730 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16_small_kernel) {
31731 TEST_REQUIRES_ARM_NEON;
31732 for (uint32_t n = 32; n <= 48; n += 16) {
31733 for (size_t k = 1; k <= 40; k += 9) {
31734 GemmMicrokernelTester()
31735 .mr(1)
31736 .nr(16)
31737 .kr(1)
31738 .sr(1)
31739 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080031740 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070031741 .k(k)
31742 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080031743 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031744 }
31745 }
31746 }
31747
31748 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, strided_cm_subtile) {
31749 TEST_REQUIRES_ARM_NEON;
31750 for (size_t k = 1; k <= 40; k += 9) {
31751 for (uint32_t m = 1; m <= 1; m++) {
31752 for (uint32_t n = 1; n <= 16; n++) {
31753 GemmMicrokernelTester()
31754 .mr(1)
31755 .nr(16)
31756 .kr(1)
31757 .sr(1)
31758 .m(m)
31759 .n(n)
31760 .k(k)
31761 .cm_stride(19)
31762 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080031763 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031764 }
31765 }
31766 }
31767 }
31768
31769 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, a_offset) {
31770 TEST_REQUIRES_ARM_NEON;
31771 for (size_t k = 1; k <= 40; k += 9) {
31772 GemmMicrokernelTester()
31773 .mr(1)
31774 .nr(16)
31775 .kr(1)
31776 .sr(1)
31777 .m(1)
31778 .n(16)
31779 .k(k)
31780 .ks(3)
31781 .a_offset(43)
Marat Dukhan50323b82022-01-11 00:12:01 -080031782 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031783 }
31784 }
31785
31786 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, zero) {
31787 TEST_REQUIRES_ARM_NEON;
31788 for (uint32_t mz = 0; mz < 1; mz++) {
31789 for (size_t k = 1; k <= 40; k += 9) {
31790 GemmMicrokernelTester()
31791 .mr(1)
31792 .nr(16)
31793 .kr(1)
31794 .sr(1)
31795 .m(1)
31796 .n(16)
31797 .k(k)
31798 .ks(3)
31799 .a_offset(43)
31800 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080031801 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031802 }
31803 }
31804 }
31805
31806 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, qmin) {
31807 TEST_REQUIRES_ARM_NEON;
31808 GemmMicrokernelTester()
31809 .mr(1)
31810 .nr(16)
31811 .kr(1)
31812 .sr(1)
31813 .m(1)
31814 .n(16)
31815 .k(8)
31816 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080031817 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031818 }
31819
31820 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, qmax) {
31821 TEST_REQUIRES_ARM_NEON;
31822 GemmMicrokernelTester()
31823 .mr(1)
31824 .nr(16)
31825 .kr(1)
31826 .sr(1)
31827 .m(1)
31828 .n(16)
31829 .k(8)
31830 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080031831 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031832 }
31833
31834 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, strided_cm) {
31835 TEST_REQUIRES_ARM_NEON;
31836 GemmMicrokernelTester()
31837 .mr(1)
31838 .nr(16)
31839 .kr(1)
31840 .sr(1)
31841 .m(1)
31842 .n(16)
31843 .k(8)
31844 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080031845 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070031846 }
31847#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
31848
31849
31850#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031851 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_eq_8) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031852 TEST_REQUIRES_ARM_NEON;
31853 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031854 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031855 .nr(16)
31856 .kr(1)
31857 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031858 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031859 .n(16)
31860 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031861 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031862 }
31863
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031864 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, strided_cn) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031865 TEST_REQUIRES_ARM_NEON;
31866 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031867 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031868 .nr(16)
31869 .kr(1)
31870 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031871 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031872 .n(16)
31873 .k(8)
31874 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031876 }
31877
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031878 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_eq_8_subtile) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031879 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031880 for (uint32_t m = 1; m <= 6; m++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031881 for (uint32_t n = 1; n <= 16; n++) {
31882 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031883 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031884 .nr(16)
31885 .kr(1)
31886 .sr(1)
31887 .m(m)
31888 .n(n)
31889 .k(8)
31890 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031891 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031892 }
31893 }
31894 }
31895
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031896 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_eq_8_subtile_m) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031897 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031898 for (uint32_t m = 1; m <= 6; m++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031899 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031900 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031901 .nr(16)
31902 .kr(1)
31903 .sr(1)
31904 .m(m)
31905 .n(16)
31906 .k(8)
31907 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031908 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031909 }
31910 }
31911
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031912 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_eq_8_subtile_n) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031913 TEST_REQUIRES_ARM_NEON;
31914 for (uint32_t n = 1; n <= 16; n++) {
31915 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031916 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031917 .nr(16)
31918 .kr(1)
31919 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031920 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031921 .n(n)
31922 .k(8)
31923 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031924 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031925 }
31926 }
31927
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031928 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_lt_8) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031929 TEST_REQUIRES_ARM_NEON;
31930 for (size_t k = 1; k < 8; k++) {
31931 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031932 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031933 .nr(16)
31934 .kr(1)
31935 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031936 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031937 .n(16)
31938 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031939 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031940 }
31941 }
31942
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031943 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_lt_8_subtile) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031944 TEST_REQUIRES_ARM_NEON;
31945 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031946 for (uint32_t m = 1; m <= 6; m++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031947 for (uint32_t n = 1; n <= 16; n++) {
31948 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031949 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031950 .nr(16)
31951 .kr(1)
31952 .sr(1)
31953 .m(m)
31954 .n(n)
31955 .k(k)
31956 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031957 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031958 }
31959 }
31960 }
31961 }
31962
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031963 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_gt_8) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031964 TEST_REQUIRES_ARM_NEON;
31965 for (size_t k = 9; k < 16; k++) {
31966 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031967 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031968 .nr(16)
31969 .kr(1)
31970 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031971 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031972 .n(16)
31973 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031974 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031975 }
31976 }
31977
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031978 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_gt_8_subtile) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031979 TEST_REQUIRES_ARM_NEON;
31980 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031981 for (uint32_t m = 1; m <= 6; m++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031982 for (uint32_t n = 1; n <= 16; n++) {
31983 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031984 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080031985 .nr(16)
31986 .kr(1)
31987 .sr(1)
31988 .m(m)
31989 .n(n)
31990 .k(k)
31991 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031992 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080031993 }
31994 }
31995 }
31996 }
31997
Zhi An Ngc27f04b2022-01-11 09:34:07 -080031998 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_div_8) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080031999 TEST_REQUIRES_ARM_NEON;
32000 for (size_t k = 16; k <= 80; k += 8) {
32001 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032002 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032003 .nr(16)
32004 .kr(1)
32005 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032006 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032007 .n(16)
32008 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032009 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032010 }
32011 }
32012
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032013 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_div_8_subtile) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032014 TEST_REQUIRES_ARM_NEON;
32015 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032016 for (uint32_t m = 1; m <= 6; m++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032017 for (uint32_t n = 1; n <= 16; n++) {
32018 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032019 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032020 .nr(16)
32021 .kr(1)
32022 .sr(1)
32023 .m(m)
32024 .n(n)
32025 .k(k)
32026 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032027 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032028 }
32029 }
32030 }
32031 }
32032
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032033 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_gt_16) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032034 TEST_REQUIRES_ARM_NEON;
32035 for (uint32_t n = 17; n < 32; n++) {
32036 for (size_t k = 1; k <= 40; k += 9) {
32037 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032038 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032039 .nr(16)
32040 .kr(1)
32041 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032042 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032043 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032044 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032045 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032046 }
32047 }
32048 }
32049
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032050 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_gt_16_strided_cn) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032051 TEST_REQUIRES_ARM_NEON;
32052 for (uint32_t n = 17; n < 32; n++) {
32053 for (size_t k = 1; k <= 40; k += 9) {
32054 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032055 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032056 .nr(16)
32057 .kr(1)
32058 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032059 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032060 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032061 .k(k)
32062 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032063 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032064 }
32065 }
32066 }
32067
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032068 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_gt_16_subtile) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032069 TEST_REQUIRES_ARM_NEON;
32070 for (uint32_t n = 17; n < 32; n++) {
32071 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032072 for (uint32_t m = 1; m <= 6; m++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032073 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032074 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032075 .nr(16)
32076 .kr(1)
32077 .sr(1)
32078 .m(m)
32079 .n(n)
32080 .k(k)
32081 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032082 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032083 }
32084 }
32085 }
32086 }
32087
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032088 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_div_16) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032089 TEST_REQUIRES_ARM_NEON;
32090 for (uint32_t n = 32; n <= 48; n += 16) {
32091 for (size_t k = 1; k <= 40; k += 9) {
32092 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032093 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032094 .nr(16)
32095 .kr(1)
32096 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032097 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032098 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032099 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032100 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032101 }
32102 }
32103 }
32104
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032105 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_div_16_strided_cn) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032106 TEST_REQUIRES_ARM_NEON;
32107 for (uint32_t n = 32; n <= 48; n += 16) {
32108 for (size_t k = 1; k <= 40; k += 9) {
32109 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032110 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032111 .nr(16)
32112 .kr(1)
32113 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032114 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032115 .n(n)
32116 .k(k)
32117 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032118 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032119 }
32120 }
32121 }
32122
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032123 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_div_16_subtile) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032124 TEST_REQUIRES_ARM_NEON;
32125 for (uint32_t n = 32; n <= 48; n += 16) {
32126 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032127 for (uint32_t m = 1; m <= 6; m++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032128 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032129 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032130 .nr(16)
32131 .kr(1)
32132 .sr(1)
32133 .m(m)
32134 .n(n)
32135 .k(k)
32136 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032137 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032138 }
32139 }
32140 }
32141 }
32142
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032143 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, small_kernel) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032144 TEST_REQUIRES_ARM_NEON;
32145 for (size_t k = 1; k <= 40; k += 9) {
32146 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032147 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032148 .nr(16)
32149 .kr(1)
32150 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032151 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032152 .n(16)
32153 .k(k)
32154 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032155 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032156 }
32157 }
32158
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032159 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, small_kernel_subtile) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032160 TEST_REQUIRES_ARM_NEON;
32161 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032162 for (uint32_t m = 1; m <= 6; m++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032163 for (uint32_t n = 1; n <= 16; n++) {
32164 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032165 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032166 .nr(16)
32167 .kr(1)
32168 .sr(1)
32169 .m(m)
32170 .n(n)
32171 .k(k)
32172 .ks(3)
32173 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032174 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032175 }
32176 }
32177 }
32178 }
32179
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032180 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_gt_16_small_kernel) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032181 TEST_REQUIRES_ARM_NEON;
32182 for (uint32_t n = 17; n < 32; n++) {
32183 for (size_t k = 1; k <= 40; k += 9) {
32184 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032185 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032186 .nr(16)
32187 .kr(1)
32188 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032189 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032190 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032191 .k(k)
32192 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032193 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032194 }
32195 }
32196 }
32197
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032198 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_div_16_small_kernel) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032199 TEST_REQUIRES_ARM_NEON;
32200 for (uint32_t n = 32; n <= 48; n += 16) {
32201 for (size_t k = 1; k <= 40; k += 9) {
32202 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032203 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032204 .nr(16)
32205 .kr(1)
32206 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032207 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032208 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032209 .k(k)
32210 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032211 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032212 }
32213 }
32214 }
32215
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032216 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, strided_cm_subtile) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032217 TEST_REQUIRES_ARM_NEON;
32218 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032219 for (uint32_t m = 1; m <= 6; m++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032220 for (uint32_t n = 1; n <= 16; n++) {
32221 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032222 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032223 .nr(16)
32224 .kr(1)
32225 .sr(1)
32226 .m(m)
32227 .n(n)
32228 .k(k)
32229 .cm_stride(19)
32230 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032231 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032232 }
32233 }
32234 }
32235 }
32236
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032237 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, a_offset) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032238 TEST_REQUIRES_ARM_NEON;
32239 for (size_t k = 1; k <= 40; k += 9) {
32240 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032241 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032242 .nr(16)
32243 .kr(1)
32244 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032245 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032246 .n(16)
32247 .k(k)
32248 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032249 .a_offset(251)
32250 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032251 }
32252 }
32253
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032254 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, zero) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032255 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032256 for (uint32_t mz = 0; mz < 6; mz++) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032257 for (size_t k = 1; k <= 40; k += 9) {
32258 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032259 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032260 .nr(16)
32261 .kr(1)
32262 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032263 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032264 .n(16)
32265 .k(k)
32266 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032267 .a_offset(251)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032268 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032269 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032270 }
32271 }
32272 }
32273
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032274 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, qmin) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032275 TEST_REQUIRES_ARM_NEON;
32276 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032277 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032278 .nr(16)
32279 .kr(1)
32280 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032281 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032282 .n(16)
32283 .k(8)
32284 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032285 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032286 }
32287
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032288 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, qmax) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032289 TEST_REQUIRES_ARM_NEON;
32290 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032291 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032292 .nr(16)
32293 .kr(1)
32294 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032295 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032296 .n(16)
32297 .k(8)
32298 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032299 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032300 }
32301
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032302 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, strided_cm) {
Frank Barchard27bf92c2021-11-24 15:47:52 -080032303 TEST_REQUIRES_ARM_NEON;
32304 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032305 .mr(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032306 .nr(16)
32307 .kr(1)
32308 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032309 .m(6)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032310 .n(16)
32311 .k(8)
32312 .cm_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032313 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032314 }
32315#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
32316
32317
32318#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barchard27bf92c2021-11-24 15:47:52 -080032319 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_eq_8) {
32320 TEST_REQUIRES_ARM_NEON;
32321 GemmMicrokernelTester()
32322 .mr(6)
32323 .nr(8)
32324 .kr(1)
32325 .sr(1)
32326 .m(6)
32327 .n(8)
32328 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080032329 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032330 }
32331
32332 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, strided_cn) {
32333 TEST_REQUIRES_ARM_NEON;
32334 GemmMicrokernelTester()
32335 .mr(6)
32336 .nr(8)
32337 .kr(1)
32338 .sr(1)
32339 .m(6)
32340 .n(8)
32341 .k(8)
32342 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080032343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032344 }
32345
32346 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_eq_8_subtile) {
32347 TEST_REQUIRES_ARM_NEON;
32348 for (uint32_t m = 1; m <= 6; m++) {
32349 for (uint32_t n = 1; n <= 8; n++) {
32350 GemmMicrokernelTester()
32351 .mr(6)
32352 .nr(8)
32353 .kr(1)
32354 .sr(1)
32355 .m(m)
32356 .n(n)
32357 .k(8)
32358 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032359 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032360 }
32361 }
32362 }
32363
32364 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_eq_8_subtile_m) {
32365 TEST_REQUIRES_ARM_NEON;
32366 for (uint32_t m = 1; m <= 6; m++) {
32367 GemmMicrokernelTester()
32368 .mr(6)
32369 .nr(8)
32370 .kr(1)
32371 .sr(1)
32372 .m(m)
32373 .n(8)
32374 .k(8)
32375 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032376 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032377 }
32378 }
32379
32380 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_eq_8_subtile_n) {
32381 TEST_REQUIRES_ARM_NEON;
32382 for (uint32_t n = 1; n <= 8; n++) {
32383 GemmMicrokernelTester()
32384 .mr(6)
32385 .nr(8)
32386 .kr(1)
32387 .sr(1)
32388 .m(6)
32389 .n(n)
32390 .k(8)
32391 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032392 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032393 }
32394 }
32395
32396 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_lt_8) {
32397 TEST_REQUIRES_ARM_NEON;
32398 for (size_t k = 1; k < 8; k++) {
32399 GemmMicrokernelTester()
32400 .mr(6)
32401 .nr(8)
32402 .kr(1)
32403 .sr(1)
32404 .m(6)
32405 .n(8)
32406 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080032407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032408 }
32409 }
32410
32411 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_lt_8_subtile) {
32412 TEST_REQUIRES_ARM_NEON;
32413 for (size_t k = 1; k < 8; k++) {
32414 for (uint32_t m = 1; m <= 6; m++) {
32415 for (uint32_t n = 1; n <= 8; n++) {
32416 GemmMicrokernelTester()
32417 .mr(6)
32418 .nr(8)
32419 .kr(1)
32420 .sr(1)
32421 .m(m)
32422 .n(n)
32423 .k(k)
32424 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032425 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032426 }
32427 }
32428 }
32429 }
32430
32431 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_gt_8) {
32432 TEST_REQUIRES_ARM_NEON;
32433 for (size_t k = 9; k < 16; k++) {
32434 GemmMicrokernelTester()
32435 .mr(6)
32436 .nr(8)
32437 .kr(1)
32438 .sr(1)
32439 .m(6)
32440 .n(8)
32441 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080032442 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032443 }
32444 }
32445
32446 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_gt_8_subtile) {
32447 TEST_REQUIRES_ARM_NEON;
32448 for (size_t k = 9; k < 16; k++) {
32449 for (uint32_t m = 1; m <= 6; m++) {
32450 for (uint32_t n = 1; n <= 8; n++) {
32451 GemmMicrokernelTester()
32452 .mr(6)
32453 .nr(8)
32454 .kr(1)
32455 .sr(1)
32456 .m(m)
32457 .n(n)
32458 .k(k)
32459 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032460 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032461 }
32462 }
32463 }
32464 }
32465
32466 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_div_8) {
32467 TEST_REQUIRES_ARM_NEON;
32468 for (size_t k = 16; k <= 80; k += 8) {
32469 GemmMicrokernelTester()
32470 .mr(6)
32471 .nr(8)
32472 .kr(1)
32473 .sr(1)
32474 .m(6)
32475 .n(8)
32476 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080032477 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032478 }
32479 }
32480
32481 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, k_div_8_subtile) {
32482 TEST_REQUIRES_ARM_NEON;
32483 for (size_t k = 16; k <= 80; k += 8) {
32484 for (uint32_t m = 1; m <= 6; m++) {
32485 for (uint32_t n = 1; n <= 8; n++) {
32486 GemmMicrokernelTester()
32487 .mr(6)
32488 .nr(8)
32489 .kr(1)
32490 .sr(1)
32491 .m(m)
32492 .n(n)
32493 .k(k)
32494 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032495 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032496 }
32497 }
32498 }
32499 }
32500
32501 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, n_gt_8) {
32502 TEST_REQUIRES_ARM_NEON;
32503 for (uint32_t n = 9; n < 16; n++) {
32504 for (size_t k = 1; k <= 40; k += 9) {
32505 GemmMicrokernelTester()
32506 .mr(6)
32507 .nr(8)
32508 .kr(1)
32509 .sr(1)
32510 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032511 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032512 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080032513 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032514 }
32515 }
32516 }
32517
32518 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, n_gt_8_strided_cn) {
32519 TEST_REQUIRES_ARM_NEON;
32520 for (uint32_t n = 9; n < 16; n++) {
32521 for (size_t k = 1; k <= 40; k += 9) {
32522 GemmMicrokernelTester()
32523 .mr(6)
32524 .nr(8)
32525 .kr(1)
32526 .sr(1)
32527 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032528 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032529 .k(k)
32530 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080032531 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032532 }
32533 }
32534 }
32535
32536 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, n_gt_8_subtile) {
32537 TEST_REQUIRES_ARM_NEON;
32538 for (uint32_t n = 9; n < 16; n++) {
32539 for (size_t k = 1; k <= 40; k += 9) {
32540 for (uint32_t m = 1; m <= 6; m++) {
32541 GemmMicrokernelTester()
32542 .mr(6)
32543 .nr(8)
32544 .kr(1)
32545 .sr(1)
32546 .m(m)
32547 .n(n)
32548 .k(k)
32549 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032550 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032551 }
32552 }
32553 }
32554 }
32555
32556 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, n_div_8) {
32557 TEST_REQUIRES_ARM_NEON;
32558 for (uint32_t n = 16; n <= 24; n += 8) {
32559 for (size_t k = 1; k <= 40; k += 9) {
32560 GemmMicrokernelTester()
32561 .mr(6)
32562 .nr(8)
32563 .kr(1)
32564 .sr(1)
32565 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032566 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032567 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080032568 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032569 }
32570 }
32571 }
32572
32573 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, n_div_8_strided_cn) {
32574 TEST_REQUIRES_ARM_NEON;
32575 for (uint32_t n = 16; n <= 24; n += 8) {
32576 for (size_t k = 1; k <= 40; k += 9) {
32577 GemmMicrokernelTester()
32578 .mr(6)
32579 .nr(8)
32580 .kr(1)
32581 .sr(1)
32582 .m(6)
32583 .n(n)
32584 .k(k)
32585 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080032586 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032587 }
32588 }
32589 }
32590
32591 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, n_div_8_subtile) {
32592 TEST_REQUIRES_ARM_NEON;
32593 for (uint32_t n = 16; n <= 24; n += 8) {
32594 for (size_t k = 1; k <= 40; k += 9) {
32595 for (uint32_t m = 1; m <= 6; m++) {
32596 GemmMicrokernelTester()
32597 .mr(6)
32598 .nr(8)
32599 .kr(1)
32600 .sr(1)
32601 .m(m)
32602 .n(n)
32603 .k(k)
32604 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032605 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032606 }
32607 }
32608 }
32609 }
32610
32611 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, small_kernel) {
32612 TEST_REQUIRES_ARM_NEON;
32613 for (size_t k = 1; k <= 40; k += 9) {
32614 GemmMicrokernelTester()
32615 .mr(6)
32616 .nr(8)
32617 .kr(1)
32618 .sr(1)
32619 .m(6)
32620 .n(8)
32621 .k(k)
32622 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080032623 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032624 }
32625 }
32626
32627 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, small_kernel_subtile) {
32628 TEST_REQUIRES_ARM_NEON;
32629 for (size_t k = 1; k <= 40; k += 9) {
32630 for (uint32_t m = 1; m <= 6; m++) {
32631 for (uint32_t n = 1; n <= 8; n++) {
32632 GemmMicrokernelTester()
32633 .mr(6)
32634 .nr(8)
32635 .kr(1)
32636 .sr(1)
32637 .m(m)
32638 .n(n)
32639 .k(k)
32640 .ks(3)
32641 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032642 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032643 }
32644 }
32645 }
32646 }
32647
32648 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, n_gt_8_small_kernel) {
32649 TEST_REQUIRES_ARM_NEON;
32650 for (uint32_t n = 9; n < 16; n++) {
32651 for (size_t k = 1; k <= 40; k += 9) {
32652 GemmMicrokernelTester()
32653 .mr(6)
32654 .nr(8)
32655 .kr(1)
32656 .sr(1)
32657 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032658 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032659 .k(k)
32660 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080032661 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032662 }
32663 }
32664 }
32665
32666 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, n_div_8_small_kernel) {
32667 TEST_REQUIRES_ARM_NEON;
32668 for (uint32_t n = 16; n <= 24; n += 8) {
32669 for (size_t k = 1; k <= 40; k += 9) {
32670 GemmMicrokernelTester()
32671 .mr(6)
32672 .nr(8)
32673 .kr(1)
32674 .sr(1)
32675 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032676 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080032677 .k(k)
32678 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080032679 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032680 }
32681 }
32682 }
32683
32684 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, strided_cm_subtile) {
32685 TEST_REQUIRES_ARM_NEON;
32686 for (size_t k = 1; k <= 40; k += 9) {
32687 for (uint32_t m = 1; m <= 6; m++) {
32688 for (uint32_t n = 1; n <= 8; n++) {
32689 GemmMicrokernelTester()
32690 .mr(6)
32691 .nr(8)
32692 .kr(1)
32693 .sr(1)
32694 .m(m)
32695 .n(n)
32696 .k(k)
32697 .cm_stride(11)
32698 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080032699 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032700 }
32701 }
32702 }
32703 }
32704
32705 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, a_offset) {
32706 TEST_REQUIRES_ARM_NEON;
32707 for (size_t k = 1; k <= 40; k += 9) {
32708 GemmMicrokernelTester()
32709 .mr(6)
32710 .nr(8)
32711 .kr(1)
32712 .sr(1)
32713 .m(6)
32714 .n(8)
32715 .k(k)
32716 .ks(3)
32717 .a_offset(251)
Marat Dukhan50323b82022-01-11 00:12:01 -080032718 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032719 }
32720 }
32721
32722 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, zero) {
32723 TEST_REQUIRES_ARM_NEON;
32724 for (uint32_t mz = 0; mz < 6; mz++) {
32725 for (size_t k = 1; k <= 40; k += 9) {
32726 GemmMicrokernelTester()
32727 .mr(6)
32728 .nr(8)
32729 .kr(1)
32730 .sr(1)
32731 .m(6)
32732 .n(8)
32733 .k(k)
32734 .ks(3)
32735 .a_offset(251)
32736 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080032737 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032738 }
32739 }
32740 }
32741
32742 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, qmin) {
32743 TEST_REQUIRES_ARM_NEON;
32744 GemmMicrokernelTester()
32745 .mr(6)
32746 .nr(8)
32747 .kr(1)
32748 .sr(1)
32749 .m(6)
32750 .n(8)
32751 .k(8)
32752 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080032753 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032754 }
32755
32756 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, qmax) {
32757 TEST_REQUIRES_ARM_NEON;
32758 GemmMicrokernelTester()
32759 .mr(6)
32760 .nr(8)
32761 .kr(1)
32762 .sr(1)
32763 .m(6)
32764 .n(8)
32765 .k(8)
32766 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080032767 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032768 }
32769
32770 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE_PRFM, strided_cm) {
32771 TEST_REQUIRES_ARM_NEON;
32772 GemmMicrokernelTester()
32773 .mr(6)
32774 .nr(8)
32775 .kr(1)
32776 .sr(1)
32777 .m(6)
32778 .n(8)
32779 .k(8)
32780 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080032781 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080032782 }
32783#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
32784
32785
32786#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032787 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_eq_8) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032788 TEST_REQUIRES_ARM_NEON;
32789 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032790 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032791 .nr(16)
32792 .kr(1)
32793 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032794 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032795 .n(16)
32796 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032797 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032798 }
32799
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032800 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, strided_cn) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032801 TEST_REQUIRES_ARM_NEON;
32802 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032803 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032804 .nr(16)
32805 .kr(1)
32806 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032807 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032808 .n(16)
32809 .k(8)
32810 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032812 }
32813
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032814 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_eq_8_subtile) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032815 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032816 for (uint32_t m = 1; m <= 2; m++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032817 for (uint32_t n = 1; n <= 16; n++) {
32818 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032819 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032820 .nr(16)
32821 .kr(1)
32822 .sr(1)
32823 .m(m)
32824 .n(n)
32825 .k(8)
32826 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032827 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032828 }
32829 }
32830 }
32831
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032832 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_eq_8_subtile_m) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032833 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032834 for (uint32_t m = 1; m <= 2; m++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032835 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032836 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032837 .nr(16)
32838 .kr(1)
32839 .sr(1)
32840 .m(m)
32841 .n(16)
32842 .k(8)
32843 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032844 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032845 }
32846 }
32847
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032848 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_eq_8_subtile_n) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032849 TEST_REQUIRES_ARM_NEON;
32850 for (uint32_t n = 1; n <= 16; n++) {
32851 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032852 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032853 .nr(16)
32854 .kr(1)
32855 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032856 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032857 .n(n)
32858 .k(8)
32859 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032860 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032861 }
32862 }
32863
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032864 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_lt_8) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032865 TEST_REQUIRES_ARM_NEON;
32866 for (size_t k = 1; k < 8; k++) {
32867 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032868 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032869 .nr(16)
32870 .kr(1)
32871 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032872 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032873 .n(16)
32874 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032876 }
32877 }
32878
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032879 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_lt_8_subtile) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032880 TEST_REQUIRES_ARM_NEON;
32881 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032882 for (uint32_t m = 1; m <= 2; m++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032883 for (uint32_t n = 1; n <= 16; n++) {
32884 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032885 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032886 .nr(16)
32887 .kr(1)
32888 .sr(1)
32889 .m(m)
32890 .n(n)
32891 .k(k)
32892 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032893 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032894 }
32895 }
32896 }
32897 }
32898
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032899 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_gt_8) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032900 TEST_REQUIRES_ARM_NEON;
32901 for (size_t k = 9; k < 16; k++) {
32902 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032903 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032904 .nr(16)
32905 .kr(1)
32906 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032907 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032908 .n(16)
32909 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032910 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032911 }
32912 }
32913
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032914 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_gt_8_subtile) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032915 TEST_REQUIRES_ARM_NEON;
32916 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032917 for (uint32_t m = 1; m <= 2; m++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032918 for (uint32_t n = 1; n <= 16; n++) {
32919 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032920 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032921 .nr(16)
32922 .kr(1)
32923 .sr(1)
32924 .m(m)
32925 .n(n)
32926 .k(k)
32927 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032928 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032929 }
32930 }
32931 }
32932 }
32933
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032934 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_div_8) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032935 TEST_REQUIRES_ARM_NEON;
32936 for (size_t k = 16; k <= 80; k += 8) {
32937 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032938 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032939 .nr(16)
32940 .kr(1)
32941 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032942 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032943 .n(16)
32944 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032945 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032946 }
32947 }
32948
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032949 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_div_8_subtile) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032950 TEST_REQUIRES_ARM_NEON;
32951 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032952 for (uint32_t m = 1; m <= 2; m++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032953 for (uint32_t n = 1; n <= 16; n++) {
32954 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032955 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032956 .nr(16)
32957 .kr(1)
32958 .sr(1)
32959 .m(m)
32960 .n(n)
32961 .k(k)
32962 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032963 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032964 }
32965 }
32966 }
32967 }
32968
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032969 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_gt_16) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032970 TEST_REQUIRES_ARM_NEON;
32971 for (uint32_t n = 17; n < 32; n++) {
32972 for (size_t k = 1; k <= 40; k += 9) {
32973 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032974 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032975 .nr(16)
32976 .kr(1)
32977 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032978 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032979 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070032980 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032981 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070032982 }
32983 }
32984 }
32985
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032986 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_gt_16_strided_cn) {
Frank Barchard22fbe772021-07-20 15:56:32 -070032987 TEST_REQUIRES_ARM_NEON;
32988 for (uint32_t n = 17; n < 32; n++) {
32989 for (size_t k = 1; k <= 40; k += 9) {
32990 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032991 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070032992 .nr(16)
32993 .kr(1)
32994 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032995 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080032996 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070032997 .k(k)
32998 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080032999 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033000 }
33001 }
33002 }
33003
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033004 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_gt_16_subtile) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033005 TEST_REQUIRES_ARM_NEON;
33006 for (uint32_t n = 17; n < 32; n++) {
33007 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033008 for (uint32_t m = 1; m <= 2; m++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033009 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033010 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033011 .nr(16)
33012 .kr(1)
33013 .sr(1)
33014 .m(m)
33015 .n(n)
33016 .k(k)
33017 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033018 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033019 }
33020 }
33021 }
33022 }
33023
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033024 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_div_16) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033025 TEST_REQUIRES_ARM_NEON;
33026 for (uint32_t n = 32; n <= 48; n += 16) {
33027 for (size_t k = 1; k <= 40; k += 9) {
33028 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033029 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033030 .nr(16)
33031 .kr(1)
33032 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033033 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033034 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070033035 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033036 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033037 }
33038 }
33039 }
33040
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033041 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_div_16_strided_cn) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033042 TEST_REQUIRES_ARM_NEON;
33043 for (uint32_t n = 32; n <= 48; n += 16) {
33044 for (size_t k = 1; k <= 40; k += 9) {
33045 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033046 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033047 .nr(16)
33048 .kr(1)
33049 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033050 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033051 .n(n)
33052 .k(k)
33053 .cn_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033054 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033055 }
33056 }
33057 }
33058
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033059 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_div_16_subtile) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033060 TEST_REQUIRES_ARM_NEON;
33061 for (uint32_t n = 32; n <= 48; n += 16) {
33062 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033063 for (uint32_t m = 1; m <= 2; m++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033064 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033065 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033066 .nr(16)
33067 .kr(1)
33068 .sr(1)
33069 .m(m)
33070 .n(n)
33071 .k(k)
33072 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033073 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033074 }
33075 }
33076 }
33077 }
33078
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033079 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, small_kernel) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033080 TEST_REQUIRES_ARM_NEON;
33081 for (size_t k = 1; k <= 40; k += 9) {
33082 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033083 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033084 .nr(16)
33085 .kr(1)
33086 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033087 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033088 .n(16)
33089 .k(k)
33090 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033091 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033092 }
33093 }
33094
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033095 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, small_kernel_subtile) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033096 TEST_REQUIRES_ARM_NEON;
33097 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033098 for (uint32_t m = 1; m <= 2; m++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033099 for (uint32_t n = 1; n <= 16; n++) {
33100 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033101 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033102 .nr(16)
33103 .kr(1)
33104 .sr(1)
33105 .m(m)
33106 .n(n)
33107 .k(k)
33108 .ks(3)
33109 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033110 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033111 }
33112 }
33113 }
33114 }
33115
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033116 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_gt_16_small_kernel) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033117 TEST_REQUIRES_ARM_NEON;
33118 for (uint32_t n = 17; n < 32; n++) {
33119 for (size_t k = 1; k <= 40; k += 9) {
33120 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033121 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033122 .nr(16)
33123 .kr(1)
33124 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033125 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033126 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070033127 .k(k)
33128 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033129 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033130 }
33131 }
33132 }
33133
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033134 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_div_16_small_kernel) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033135 TEST_REQUIRES_ARM_NEON;
33136 for (uint32_t n = 32; n <= 48; n += 16) {
33137 for (size_t k = 1; k <= 40; k += 9) {
33138 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033139 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033140 .nr(16)
33141 .kr(1)
33142 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033143 .m(2)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033144 .n(n)
Frank Barchard22fbe772021-07-20 15:56:32 -070033145 .k(k)
33146 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033147 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033148 }
33149 }
33150 }
33151
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033152 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, strided_cm_subtile) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033153 TEST_REQUIRES_ARM_NEON;
33154 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033155 for (uint32_t m = 1; m <= 2; m++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033156 for (uint32_t n = 1; n <= 16; n++) {
33157 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033158 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033159 .nr(16)
33160 .kr(1)
33161 .sr(1)
33162 .m(m)
33163 .n(n)
33164 .k(k)
33165 .cm_stride(19)
33166 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033167 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033168 }
33169 }
33170 }
33171 }
33172
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033173 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, a_offset) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033174 TEST_REQUIRES_ARM_NEON;
33175 for (size_t k = 1; k <= 40; k += 9) {
33176 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033177 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033178 .nr(16)
33179 .kr(1)
33180 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033181 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033182 .n(16)
33183 .k(k)
33184 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033185 .a_offset(83)
33186 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033187 }
33188 }
33189
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033190 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, zero) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033191 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033192 for (uint32_t mz = 0; mz < 2; mz++) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033193 for (size_t k = 1; k <= 40; k += 9) {
33194 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033195 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033196 .nr(16)
33197 .kr(1)
33198 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033199 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033200 .n(16)
33201 .k(k)
33202 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033203 .a_offset(83)
Frank Barchard22fbe772021-07-20 15:56:32 -070033204 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033205 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033206 }
33207 }
33208 }
33209
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033210 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, qmin) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033211 TEST_REQUIRES_ARM_NEON;
33212 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033213 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033214 .nr(16)
33215 .kr(1)
33216 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033217 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033218 .n(16)
33219 .k(8)
33220 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033221 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033222 }
33223
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033224 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, qmax) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033225 TEST_REQUIRES_ARM_NEON;
33226 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033227 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033228 .nr(16)
33229 .kr(1)
33230 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033231 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033232 .n(16)
33233 .k(8)
33234 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033235 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033236 }
33237
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033238 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, strided_cm) {
Frank Barchard22fbe772021-07-20 15:56:32 -070033239 TEST_REQUIRES_ARM_NEON;
33240 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033241 .mr(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033242 .nr(16)
33243 .kr(1)
33244 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033245 .m(2)
Frank Barchard22fbe772021-07-20 15:56:32 -070033246 .n(16)
33247 .k(8)
33248 .cm_stride(19)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033249 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard22fbe772021-07-20 15:56:32 -070033250 }
33251#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
33252
33253
33254#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Frank Barchard27bf92c2021-11-24 15:47:52 -080033255 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_eq_8) {
33256 TEST_REQUIRES_ARM_NEON;
33257 GemmMicrokernelTester()
33258 .mr(6)
33259 .nr(16)
33260 .kr(1)
33261 .sr(1)
33262 .m(6)
33263 .n(16)
33264 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080033265 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033266 }
33267
33268 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, strided_cn) {
33269 TEST_REQUIRES_ARM_NEON;
33270 GemmMicrokernelTester()
33271 .mr(6)
33272 .nr(16)
33273 .kr(1)
33274 .sr(1)
33275 .m(6)
33276 .n(16)
33277 .k(8)
33278 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080033279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033280 }
33281
33282 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_eq_8_subtile) {
33283 TEST_REQUIRES_ARM_NEON;
33284 for (uint32_t m = 1; m <= 6; m++) {
33285 for (uint32_t n = 1; n <= 16; n++) {
33286 GemmMicrokernelTester()
33287 .mr(6)
33288 .nr(16)
33289 .kr(1)
33290 .sr(1)
33291 .m(m)
33292 .n(n)
33293 .k(8)
33294 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033295 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033296 }
33297 }
33298 }
33299
33300 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_eq_8_subtile_m) {
33301 TEST_REQUIRES_ARM_NEON;
33302 for (uint32_t m = 1; m <= 6; m++) {
33303 GemmMicrokernelTester()
33304 .mr(6)
33305 .nr(16)
33306 .kr(1)
33307 .sr(1)
33308 .m(m)
33309 .n(16)
33310 .k(8)
33311 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033312 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033313 }
33314 }
33315
33316 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_eq_8_subtile_n) {
33317 TEST_REQUIRES_ARM_NEON;
33318 for (uint32_t n = 1; n <= 16; n++) {
33319 GemmMicrokernelTester()
33320 .mr(6)
33321 .nr(16)
33322 .kr(1)
33323 .sr(1)
33324 .m(6)
33325 .n(n)
33326 .k(8)
33327 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033328 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033329 }
33330 }
33331
33332 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_lt_8) {
33333 TEST_REQUIRES_ARM_NEON;
33334 for (size_t k = 1; k < 8; k++) {
33335 GemmMicrokernelTester()
33336 .mr(6)
33337 .nr(16)
33338 .kr(1)
33339 .sr(1)
33340 .m(6)
33341 .n(16)
33342 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080033343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033344 }
33345 }
33346
33347 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_lt_8_subtile) {
33348 TEST_REQUIRES_ARM_NEON;
33349 for (size_t k = 1; k < 8; k++) {
33350 for (uint32_t m = 1; m <= 6; m++) {
33351 for (uint32_t n = 1; n <= 16; n++) {
33352 GemmMicrokernelTester()
33353 .mr(6)
33354 .nr(16)
33355 .kr(1)
33356 .sr(1)
33357 .m(m)
33358 .n(n)
33359 .k(k)
33360 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033361 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033362 }
33363 }
33364 }
33365 }
33366
33367 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_gt_8) {
33368 TEST_REQUIRES_ARM_NEON;
33369 for (size_t k = 9; k < 16; k++) {
33370 GemmMicrokernelTester()
33371 .mr(6)
33372 .nr(16)
33373 .kr(1)
33374 .sr(1)
33375 .m(6)
33376 .n(16)
33377 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080033378 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033379 }
33380 }
33381
33382 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_gt_8_subtile) {
33383 TEST_REQUIRES_ARM_NEON;
33384 for (size_t k = 9; k < 16; k++) {
33385 for (uint32_t m = 1; m <= 6; m++) {
33386 for (uint32_t n = 1; n <= 16; n++) {
33387 GemmMicrokernelTester()
33388 .mr(6)
33389 .nr(16)
33390 .kr(1)
33391 .sr(1)
33392 .m(m)
33393 .n(n)
33394 .k(k)
33395 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033396 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033397 }
33398 }
33399 }
33400 }
33401
33402 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_div_8) {
33403 TEST_REQUIRES_ARM_NEON;
33404 for (size_t k = 16; k <= 80; k += 8) {
33405 GemmMicrokernelTester()
33406 .mr(6)
33407 .nr(16)
33408 .kr(1)
33409 .sr(1)
33410 .m(6)
33411 .n(16)
33412 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080033413 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033414 }
33415 }
33416
33417 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, k_div_8_subtile) {
33418 TEST_REQUIRES_ARM_NEON;
33419 for (size_t k = 16; k <= 80; k += 8) {
33420 for (uint32_t m = 1; m <= 6; m++) {
33421 for (uint32_t n = 1; n <= 16; n++) {
33422 GemmMicrokernelTester()
33423 .mr(6)
33424 .nr(16)
33425 .kr(1)
33426 .sr(1)
33427 .m(m)
33428 .n(n)
33429 .k(k)
33430 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033431 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033432 }
33433 }
33434 }
33435 }
33436
33437 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, n_gt_16) {
33438 TEST_REQUIRES_ARM_NEON;
33439 for (uint32_t n = 17; n < 32; n++) {
33440 for (size_t k = 1; k <= 40; k += 9) {
33441 GemmMicrokernelTester()
33442 .mr(6)
33443 .nr(16)
33444 .kr(1)
33445 .sr(1)
33446 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033447 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080033448 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080033449 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033450 }
33451 }
33452 }
33453
33454 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, n_gt_16_strided_cn) {
33455 TEST_REQUIRES_ARM_NEON;
33456 for (uint32_t n = 17; n < 32; n++) {
33457 for (size_t k = 1; k <= 40; k += 9) {
33458 GemmMicrokernelTester()
33459 .mr(6)
33460 .nr(16)
33461 .kr(1)
33462 .sr(1)
33463 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033464 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080033465 .k(k)
33466 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080033467 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033468 }
33469 }
33470 }
33471
33472 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, n_gt_16_subtile) {
33473 TEST_REQUIRES_ARM_NEON;
33474 for (uint32_t n = 17; n < 32; n++) {
33475 for (size_t k = 1; k <= 40; k += 9) {
33476 for (uint32_t m = 1; m <= 6; m++) {
33477 GemmMicrokernelTester()
33478 .mr(6)
33479 .nr(16)
33480 .kr(1)
33481 .sr(1)
33482 .m(m)
33483 .n(n)
33484 .k(k)
33485 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033486 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033487 }
33488 }
33489 }
33490 }
33491
33492 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, n_div_16) {
33493 TEST_REQUIRES_ARM_NEON;
33494 for (uint32_t n = 32; n <= 48; n += 16) {
33495 for (size_t k = 1; k <= 40; k += 9) {
33496 GemmMicrokernelTester()
33497 .mr(6)
33498 .nr(16)
33499 .kr(1)
33500 .sr(1)
33501 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033502 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080033503 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080033504 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033505 }
33506 }
33507 }
33508
33509 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, n_div_16_strided_cn) {
33510 TEST_REQUIRES_ARM_NEON;
33511 for (uint32_t n = 32; n <= 48; n += 16) {
33512 for (size_t k = 1; k <= 40; k += 9) {
33513 GemmMicrokernelTester()
33514 .mr(6)
33515 .nr(16)
33516 .kr(1)
33517 .sr(1)
33518 .m(6)
33519 .n(n)
33520 .k(k)
33521 .cn_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080033522 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033523 }
33524 }
33525 }
33526
33527 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, n_div_16_subtile) {
33528 TEST_REQUIRES_ARM_NEON;
33529 for (uint32_t n = 32; n <= 48; n += 16) {
33530 for (size_t k = 1; k <= 40; k += 9) {
33531 for (uint32_t m = 1; m <= 6; m++) {
33532 GemmMicrokernelTester()
33533 .mr(6)
33534 .nr(16)
33535 .kr(1)
33536 .sr(1)
33537 .m(m)
33538 .n(n)
33539 .k(k)
33540 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033541 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033542 }
33543 }
33544 }
33545 }
33546
33547 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, small_kernel) {
33548 TEST_REQUIRES_ARM_NEON;
33549 for (size_t k = 1; k <= 40; k += 9) {
33550 GemmMicrokernelTester()
33551 .mr(6)
33552 .nr(16)
33553 .kr(1)
33554 .sr(1)
33555 .m(6)
33556 .n(16)
33557 .k(k)
33558 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080033559 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033560 }
33561 }
33562
33563 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, small_kernel_subtile) {
33564 TEST_REQUIRES_ARM_NEON;
33565 for (size_t k = 1; k <= 40; k += 9) {
33566 for (uint32_t m = 1; m <= 6; m++) {
33567 for (uint32_t n = 1; n <= 16; n++) {
33568 GemmMicrokernelTester()
33569 .mr(6)
33570 .nr(16)
33571 .kr(1)
33572 .sr(1)
33573 .m(m)
33574 .n(n)
33575 .k(k)
33576 .ks(3)
33577 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033578 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033579 }
33580 }
33581 }
33582 }
33583
33584 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, n_gt_16_small_kernel) {
33585 TEST_REQUIRES_ARM_NEON;
33586 for (uint32_t n = 17; n < 32; n++) {
33587 for (size_t k = 1; k <= 40; k += 9) {
33588 GemmMicrokernelTester()
33589 .mr(6)
33590 .nr(16)
33591 .kr(1)
33592 .sr(1)
33593 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033594 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080033595 .k(k)
33596 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080033597 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033598 }
33599 }
33600 }
33601
33602 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, n_div_16_small_kernel) {
33603 TEST_REQUIRES_ARM_NEON;
33604 for (uint32_t n = 32; n <= 48; n += 16) {
33605 for (size_t k = 1; k <= 40; k += 9) {
33606 GemmMicrokernelTester()
33607 .mr(6)
33608 .nr(16)
33609 .kr(1)
33610 .sr(1)
33611 .m(6)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033612 .n(n)
Frank Barchard27bf92c2021-11-24 15:47:52 -080033613 .k(k)
33614 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080033615 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033616 }
33617 }
33618 }
33619
33620 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, strided_cm_subtile) {
33621 TEST_REQUIRES_ARM_NEON;
33622 for (size_t k = 1; k <= 40; k += 9) {
33623 for (uint32_t m = 1; m <= 6; m++) {
33624 for (uint32_t n = 1; n <= 16; n++) {
33625 GemmMicrokernelTester()
33626 .mr(6)
33627 .nr(16)
33628 .kr(1)
33629 .sr(1)
33630 .m(m)
33631 .n(n)
33632 .k(k)
33633 .cm_stride(19)
33634 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080033635 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033636 }
33637 }
33638 }
33639 }
33640
33641 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, a_offset) {
33642 TEST_REQUIRES_ARM_NEON;
33643 for (size_t k = 1; k <= 40; k += 9) {
33644 GemmMicrokernelTester()
33645 .mr(6)
33646 .nr(16)
33647 .kr(1)
33648 .sr(1)
33649 .m(6)
33650 .n(16)
33651 .k(k)
33652 .ks(3)
33653 .a_offset(251)
Marat Dukhan50323b82022-01-11 00:12:01 -080033654 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033655 }
33656 }
33657
33658 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, zero) {
33659 TEST_REQUIRES_ARM_NEON;
33660 for (uint32_t mz = 0; mz < 6; mz++) {
33661 for (size_t k = 1; k <= 40; k += 9) {
33662 GemmMicrokernelTester()
33663 .mr(6)
33664 .nr(16)
33665 .kr(1)
33666 .sr(1)
33667 .m(6)
33668 .n(16)
33669 .k(k)
33670 .ks(3)
33671 .a_offset(251)
33672 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080033673 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033674 }
33675 }
33676 }
33677
33678 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, qmin) {
33679 TEST_REQUIRES_ARM_NEON;
33680 GemmMicrokernelTester()
33681 .mr(6)
33682 .nr(16)
33683 .kr(1)
33684 .sr(1)
33685 .m(6)
33686 .n(16)
33687 .k(8)
33688 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080033689 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033690 }
33691
33692 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, qmax) {
33693 TEST_REQUIRES_ARM_NEON;
33694 GemmMicrokernelTester()
33695 .mr(6)
33696 .nr(16)
33697 .kr(1)
33698 .sr(1)
33699 .m(6)
33700 .n(16)
33701 .k(8)
33702 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080033703 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033704 }
33705
33706 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE_PRFM, strided_cm) {
33707 TEST_REQUIRES_ARM_NEON;
33708 GemmMicrokernelTester()
33709 .mr(6)
33710 .nr(16)
33711 .kr(1)
33712 .sr(1)
33713 .m(6)
33714 .n(16)
33715 .k(8)
33716 .cm_stride(19)
Marat Dukhan50323b82022-01-11 00:12:01 -080033717 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard27bf92c2021-11-24 15:47:52 -080033718 }
33719#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
33720
33721
33722#if XNN_ARCH_ARM || XNN_ARCH_ARM64
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033723 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_eq_8) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033724 TEST_REQUIRES_ARM_NEON;
33725 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033726 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033727 .nr(8)
33728 .kr(1)
33729 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033730 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033731 .n(8)
33732 .k(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033733 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033734 }
33735
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033736 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, strided_cn) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033737 TEST_REQUIRES_ARM_NEON;
33738 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033739 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033740 .nr(8)
33741 .kr(1)
33742 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033743 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033744 .n(8)
33745 .k(8)
33746 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033748 }
33749
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033750 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033751 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033752 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033753 for (uint32_t n = 1; n <= 8; n++) {
33754 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033755 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033756 .nr(8)
33757 .kr(1)
33758 .sr(1)
33759 .m(m)
33760 .n(n)
33761 .k(8)
33762 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033763 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033764 }
33765 }
33766 }
33767
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033768 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile_m) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033769 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033770 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033771 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033772 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033773 .nr(8)
33774 .kr(1)
33775 .sr(1)
33776 .m(m)
33777 .n(8)
33778 .k(8)
33779 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033780 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033781 }
33782 }
33783
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033784 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile_n) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033785 TEST_REQUIRES_ARM_NEON;
33786 for (uint32_t n = 1; n <= 8; n++) {
33787 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033788 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033789 .nr(8)
33790 .kr(1)
33791 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033792 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033793 .n(n)
33794 .k(8)
33795 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033796 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033797 }
33798 }
33799
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033800 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_lt_8) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033801 TEST_REQUIRES_ARM_NEON;
33802 for (size_t k = 1; k < 8; k++) {
33803 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033804 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033805 .nr(8)
33806 .kr(1)
33807 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033808 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033809 .n(8)
33810 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033812 }
33813 }
33814
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033815 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_lt_8_subtile) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033816 TEST_REQUIRES_ARM_NEON;
33817 for (size_t k = 1; k < 8; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033818 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033819 for (uint32_t n = 1; n <= 8; n++) {
33820 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033821 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033822 .nr(8)
33823 .kr(1)
33824 .sr(1)
33825 .m(m)
33826 .n(n)
33827 .k(k)
33828 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033829 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033830 }
33831 }
33832 }
33833 }
33834
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033835 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_gt_8) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033836 TEST_REQUIRES_ARM_NEON;
33837 for (size_t k = 9; k < 16; k++) {
33838 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033839 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033840 .nr(8)
33841 .kr(1)
33842 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033843 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033844 .n(8)
33845 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033846 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033847 }
33848 }
33849
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033850 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_gt_8_subtile) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033851 TEST_REQUIRES_ARM_NEON;
33852 for (size_t k = 9; k < 16; k++) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033853 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033854 for (uint32_t n = 1; n <= 8; n++) {
33855 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033856 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033857 .nr(8)
33858 .kr(1)
33859 .sr(1)
33860 .m(m)
33861 .n(n)
33862 .k(k)
33863 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033864 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033865 }
33866 }
33867 }
33868 }
33869
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033870 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_div_8) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033871 TEST_REQUIRES_ARM_NEON;
33872 for (size_t k = 16; k <= 80; k += 8) {
33873 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033874 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033875 .nr(8)
33876 .kr(1)
33877 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033878 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033879 .n(8)
33880 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033881 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033882 }
33883 }
33884
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033885 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, k_div_8_subtile) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033886 TEST_REQUIRES_ARM_NEON;
33887 for (size_t k = 16; k <= 80; k += 8) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033888 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033889 for (uint32_t n = 1; n <= 8; n++) {
33890 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033891 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033892 .nr(8)
33893 .kr(1)
33894 .sr(1)
33895 .m(m)
33896 .n(n)
33897 .k(k)
33898 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033899 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033900 }
33901 }
33902 }
33903 }
33904
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033905 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, n_gt_8) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033906 TEST_REQUIRES_ARM_NEON;
33907 for (uint32_t n = 9; n < 16; n++) {
33908 for (size_t k = 1; k <= 40; k += 9) {
33909 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033910 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033911 .nr(8)
33912 .kr(1)
33913 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033914 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033915 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070033916 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033917 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033918 }
33919 }
33920 }
33921
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033922 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, n_gt_8_strided_cn) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033923 TEST_REQUIRES_ARM_NEON;
33924 for (uint32_t n = 9; n < 16; n++) {
33925 for (size_t k = 1; k <= 40; k += 9) {
33926 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033927 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033928 .nr(8)
33929 .kr(1)
33930 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033931 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033932 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070033933 .k(k)
33934 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033935 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033936 }
33937 }
33938 }
33939
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033940 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, n_gt_8_subtile) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033941 TEST_REQUIRES_ARM_NEON;
33942 for (uint32_t n = 9; n < 16; n++) {
33943 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033944 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033945 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033946 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033947 .nr(8)
33948 .kr(1)
33949 .sr(1)
33950 .m(m)
33951 .n(n)
33952 .k(k)
33953 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033954 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033955 }
33956 }
33957 }
33958 }
33959
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033960 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, n_div_8) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033961 TEST_REQUIRES_ARM_NEON;
33962 for (uint32_t n = 16; n <= 24; n += 8) {
33963 for (size_t k = 1; k <= 40; k += 9) {
33964 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033965 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033966 .nr(8)
33967 .kr(1)
33968 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033969 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080033970 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070033971 .k(k)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033972 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033973 }
33974 }
33975 }
33976
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033977 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, n_div_8_strided_cn) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033978 TEST_REQUIRES_ARM_NEON;
33979 for (uint32_t n = 16; n <= 24; n += 8) {
33980 for (size_t k = 1; k <= 40; k += 9) {
33981 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033982 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033983 .nr(8)
33984 .kr(1)
33985 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033986 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070033987 .n(n)
33988 .k(k)
33989 .cn_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033990 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070033991 }
33992 }
33993 }
33994
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033995 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, n_div_8_subtile) {
Frank Barchard510b8e02021-07-26 17:25:18 -070033996 TEST_REQUIRES_ARM_NEON;
33997 for (uint32_t n = 16; n <= 24; n += 8) {
33998 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080033999 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034000 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034001 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034002 .nr(8)
34003 .kr(1)
34004 .sr(1)
34005 .m(m)
34006 .n(n)
34007 .k(k)
34008 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034009 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034010 }
34011 }
34012 }
34013 }
34014
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034015 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, small_kernel) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034016 TEST_REQUIRES_ARM_NEON;
34017 for (size_t k = 1; k <= 40; k += 9) {
34018 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034019 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034020 .nr(8)
34021 .kr(1)
34022 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034023 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034024 .n(8)
34025 .k(k)
34026 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034027 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034028 }
34029 }
34030
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034031 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, small_kernel_subtile) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034032 TEST_REQUIRES_ARM_NEON;
34033 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034034 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034035 for (uint32_t n = 1; n <= 8; n++) {
34036 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034037 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034038 .nr(8)
34039 .kr(1)
34040 .sr(1)
34041 .m(m)
34042 .n(n)
34043 .k(k)
34044 .ks(3)
34045 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034046 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034047 }
34048 }
34049 }
34050 }
34051
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034052 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, n_gt_8_small_kernel) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034053 TEST_REQUIRES_ARM_NEON;
34054 for (uint32_t n = 9; n < 16; n++) {
34055 for (size_t k = 1; k <= 40; k += 9) {
34056 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034057 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034058 .nr(8)
34059 .kr(1)
34060 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034061 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034062 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070034063 .k(k)
34064 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034065 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034066 }
34067 }
34068 }
34069
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034070 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, n_div_8_small_kernel) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034071 TEST_REQUIRES_ARM_NEON;
34072 for (uint32_t n = 16; n <= 24; n += 8) {
34073 for (size_t k = 1; k <= 40; k += 9) {
34074 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034075 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034076 .nr(8)
34077 .kr(1)
34078 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034079 .m(1)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034080 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070034081 .k(k)
34082 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034083 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034084 }
34085 }
34086 }
34087
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034088 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, strided_cm_subtile) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034089 TEST_REQUIRES_ARM_NEON;
34090 for (size_t k = 1; k <= 40; k += 9) {
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034091 for (uint32_t m = 1; m <= 1; m++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034092 for (uint32_t n = 1; n <= 8; n++) {
34093 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034094 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034095 .nr(8)
34096 .kr(1)
34097 .sr(1)
34098 .m(m)
34099 .n(n)
34100 .k(k)
34101 .cm_stride(11)
34102 .iterations(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034103 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034104 }
34105 }
34106 }
34107 }
34108
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034109 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, a_offset) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034110 TEST_REQUIRES_ARM_NEON;
34111 for (size_t k = 1; k <= 40; k += 9) {
34112 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034113 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034114 .nr(8)
34115 .kr(1)
34116 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034117 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034118 .n(8)
34119 .k(k)
34120 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034121 .a_offset(43)
34122 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034123 }
34124 }
34125
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034126 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, zero) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034127 TEST_REQUIRES_ARM_NEON;
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034128 for (uint32_t mz = 0; mz < 1; mz++) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034129 for (size_t k = 1; k <= 40; k += 9) {
34130 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034131 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034132 .nr(8)
34133 .kr(1)
34134 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034135 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034136 .n(8)
34137 .k(k)
34138 .ks(3)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034139 .a_offset(43)
Frank Barchard510b8e02021-07-26 17:25:18 -070034140 .zero_index(mz)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034141 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034142 }
34143 }
34144 }
34145
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034146 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, qmin) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034147 TEST_REQUIRES_ARM_NEON;
34148 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034149 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034150 .nr(8)
34151 .kr(1)
34152 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034153 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034154 .n(8)
34155 .k(8)
34156 .qmin(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034157 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034158 }
34159
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034160 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, qmax) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034161 TEST_REQUIRES_ARM_NEON;
34162 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034163 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034164 .nr(8)
34165 .kr(1)
34166 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034167 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034168 .n(8)
34169 .k(8)
34170 .qmax(128)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034171 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034172 }
34173
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034174 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8__NEON_MULL_ADDW_DUP, strided_cm) {
Frank Barchard510b8e02021-07-26 17:25:18 -070034175 TEST_REQUIRES_ARM_NEON;
34176 GemmMicrokernelTester()
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034177 .mr(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034178 .nr(8)
34179 .kr(1)
34180 .sr(1)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034181 .m(1)
Frank Barchard510b8e02021-07-26 17:25:18 -070034182 .n(8)
34183 .k(8)
34184 .cm_stride(11)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034185 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034186 }
34187#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
34188
34189
34190#if XNN_ARCH_ARM || XNN_ARCH_ARM64
34191 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_eq_8) {
34192 TEST_REQUIRES_ARM_NEON;
34193 GemmMicrokernelTester()
34194 .mr(4)
34195 .nr(8)
34196 .kr(1)
34197 .sr(1)
34198 .m(4)
34199 .n(8)
34200 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080034201 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034202 }
34203
34204 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, strided_cn) {
34205 TEST_REQUIRES_ARM_NEON;
34206 GemmMicrokernelTester()
34207 .mr(4)
34208 .nr(8)
34209 .kr(1)
34210 .sr(1)
34211 .m(4)
34212 .n(8)
34213 .k(8)
34214 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080034215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034216 }
34217
34218 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile) {
34219 TEST_REQUIRES_ARM_NEON;
34220 for (uint32_t m = 1; m <= 4; m++) {
34221 for (uint32_t n = 1; n <= 8; n++) {
34222 GemmMicrokernelTester()
34223 .mr(4)
34224 .nr(8)
34225 .kr(1)
34226 .sr(1)
34227 .m(m)
34228 .n(n)
34229 .k(8)
34230 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034231 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034232 }
34233 }
34234 }
34235
34236 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile_m) {
34237 TEST_REQUIRES_ARM_NEON;
34238 for (uint32_t m = 1; m <= 4; m++) {
34239 GemmMicrokernelTester()
34240 .mr(4)
34241 .nr(8)
34242 .kr(1)
34243 .sr(1)
34244 .m(m)
34245 .n(8)
34246 .k(8)
34247 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034248 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034249 }
34250 }
34251
34252 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile_n) {
34253 TEST_REQUIRES_ARM_NEON;
34254 for (uint32_t n = 1; n <= 8; n++) {
34255 GemmMicrokernelTester()
34256 .mr(4)
34257 .nr(8)
34258 .kr(1)
34259 .sr(1)
34260 .m(4)
34261 .n(n)
34262 .k(8)
34263 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034264 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034265 }
34266 }
34267
34268 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_lt_8) {
34269 TEST_REQUIRES_ARM_NEON;
34270 for (size_t k = 1; k < 8; k++) {
34271 GemmMicrokernelTester()
34272 .mr(4)
34273 .nr(8)
34274 .kr(1)
34275 .sr(1)
34276 .m(4)
34277 .n(8)
34278 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034280 }
34281 }
34282
34283 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_lt_8_subtile) {
34284 TEST_REQUIRES_ARM_NEON;
34285 for (size_t k = 1; k < 8; k++) {
34286 for (uint32_t m = 1; m <= 4; m++) {
34287 for (uint32_t n = 1; n <= 8; n++) {
34288 GemmMicrokernelTester()
34289 .mr(4)
34290 .nr(8)
34291 .kr(1)
34292 .sr(1)
34293 .m(m)
34294 .n(n)
34295 .k(k)
34296 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034297 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034298 }
34299 }
34300 }
34301 }
34302
34303 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_gt_8) {
34304 TEST_REQUIRES_ARM_NEON;
34305 for (size_t k = 9; k < 16; k++) {
34306 GemmMicrokernelTester()
34307 .mr(4)
34308 .nr(8)
34309 .kr(1)
34310 .sr(1)
34311 .m(4)
34312 .n(8)
34313 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034314 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034315 }
34316 }
34317
34318 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_gt_8_subtile) {
34319 TEST_REQUIRES_ARM_NEON;
34320 for (size_t k = 9; k < 16; k++) {
34321 for (uint32_t m = 1; m <= 4; m++) {
34322 for (uint32_t n = 1; n <= 8; n++) {
34323 GemmMicrokernelTester()
34324 .mr(4)
34325 .nr(8)
34326 .kr(1)
34327 .sr(1)
34328 .m(m)
34329 .n(n)
34330 .k(k)
34331 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034332 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034333 }
34334 }
34335 }
34336 }
34337
34338 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_div_8) {
34339 TEST_REQUIRES_ARM_NEON;
34340 for (size_t k = 16; k <= 80; k += 8) {
34341 GemmMicrokernelTester()
34342 .mr(4)
34343 .nr(8)
34344 .kr(1)
34345 .sr(1)
34346 .m(4)
34347 .n(8)
34348 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034349 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034350 }
34351 }
34352
34353 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_div_8_subtile) {
34354 TEST_REQUIRES_ARM_NEON;
34355 for (size_t k = 16; k <= 80; k += 8) {
34356 for (uint32_t m = 1; m <= 4; m++) {
34357 for (uint32_t n = 1; n <= 8; n++) {
34358 GemmMicrokernelTester()
34359 .mr(4)
34360 .nr(8)
34361 .kr(1)
34362 .sr(1)
34363 .m(m)
34364 .n(n)
34365 .k(k)
34366 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034367 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034368 }
34369 }
34370 }
34371 }
34372
34373 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_gt_8) {
34374 TEST_REQUIRES_ARM_NEON;
34375 for (uint32_t n = 9; n < 16; n++) {
34376 for (size_t k = 1; k <= 40; k += 9) {
34377 GemmMicrokernelTester()
34378 .mr(4)
34379 .nr(8)
34380 .kr(1)
34381 .sr(1)
34382 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034383 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070034384 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034385 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034386 }
34387 }
34388 }
34389
34390 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_gt_8_strided_cn) {
34391 TEST_REQUIRES_ARM_NEON;
34392 for (uint32_t n = 9; n < 16; n++) {
34393 for (size_t k = 1; k <= 40; k += 9) {
34394 GemmMicrokernelTester()
34395 .mr(4)
34396 .nr(8)
34397 .kr(1)
34398 .sr(1)
34399 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034400 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070034401 .k(k)
34402 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080034403 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034404 }
34405 }
34406 }
34407
34408 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_gt_8_subtile) {
34409 TEST_REQUIRES_ARM_NEON;
34410 for (uint32_t n = 9; n < 16; n++) {
34411 for (size_t k = 1; k <= 40; k += 9) {
34412 for (uint32_t m = 1; m <= 4; m++) {
34413 GemmMicrokernelTester()
34414 .mr(4)
34415 .nr(8)
34416 .kr(1)
34417 .sr(1)
34418 .m(m)
34419 .n(n)
34420 .k(k)
34421 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034422 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034423 }
34424 }
34425 }
34426 }
34427
34428 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_div_8) {
34429 TEST_REQUIRES_ARM_NEON;
34430 for (uint32_t n = 16; n <= 24; n += 8) {
34431 for (size_t k = 1; k <= 40; k += 9) {
34432 GemmMicrokernelTester()
34433 .mr(4)
34434 .nr(8)
34435 .kr(1)
34436 .sr(1)
34437 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034438 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070034439 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034440 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034441 }
34442 }
34443 }
34444
34445 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_div_8_strided_cn) {
34446 TEST_REQUIRES_ARM_NEON;
34447 for (uint32_t n = 16; n <= 24; n += 8) {
34448 for (size_t k = 1; k <= 40; k += 9) {
34449 GemmMicrokernelTester()
34450 .mr(4)
34451 .nr(8)
34452 .kr(1)
34453 .sr(1)
34454 .m(4)
34455 .n(n)
34456 .k(k)
34457 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080034458 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034459 }
34460 }
34461 }
34462
34463 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_div_8_subtile) {
34464 TEST_REQUIRES_ARM_NEON;
34465 for (uint32_t n = 16; n <= 24; n += 8) {
34466 for (size_t k = 1; k <= 40; k += 9) {
34467 for (uint32_t m = 1; m <= 4; m++) {
34468 GemmMicrokernelTester()
34469 .mr(4)
34470 .nr(8)
34471 .kr(1)
34472 .sr(1)
34473 .m(m)
34474 .n(n)
34475 .k(k)
34476 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034477 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034478 }
34479 }
34480 }
34481 }
34482
34483 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, small_kernel) {
34484 TEST_REQUIRES_ARM_NEON;
34485 for (size_t k = 1; k <= 40; k += 9) {
34486 GemmMicrokernelTester()
34487 .mr(4)
34488 .nr(8)
34489 .kr(1)
34490 .sr(1)
34491 .m(4)
34492 .n(8)
34493 .k(k)
34494 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080034495 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034496 }
34497 }
34498
34499 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, small_kernel_subtile) {
34500 TEST_REQUIRES_ARM_NEON;
34501 for (size_t k = 1; k <= 40; k += 9) {
34502 for (uint32_t m = 1; m <= 4; m++) {
34503 for (uint32_t n = 1; n <= 8; n++) {
34504 GemmMicrokernelTester()
34505 .mr(4)
34506 .nr(8)
34507 .kr(1)
34508 .sr(1)
34509 .m(m)
34510 .n(n)
34511 .k(k)
34512 .ks(3)
34513 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034514 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034515 }
34516 }
34517 }
34518 }
34519
34520 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_gt_8_small_kernel) {
34521 TEST_REQUIRES_ARM_NEON;
34522 for (uint32_t n = 9; n < 16; n++) {
34523 for (size_t k = 1; k <= 40; k += 9) {
34524 GemmMicrokernelTester()
34525 .mr(4)
34526 .nr(8)
34527 .kr(1)
34528 .sr(1)
34529 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034530 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070034531 .k(k)
34532 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080034533 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034534 }
34535 }
34536 }
34537
34538 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_div_8_small_kernel) {
34539 TEST_REQUIRES_ARM_NEON;
34540 for (uint32_t n = 16; n <= 24; n += 8) {
34541 for (size_t k = 1; k <= 40; k += 9) {
34542 GemmMicrokernelTester()
34543 .mr(4)
34544 .nr(8)
34545 .kr(1)
34546 .sr(1)
34547 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034548 .n(n)
Frank Barchard510b8e02021-07-26 17:25:18 -070034549 .k(k)
34550 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080034551 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034552 }
34553 }
34554 }
34555
34556 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, strided_cm_subtile) {
34557 TEST_REQUIRES_ARM_NEON;
34558 for (size_t k = 1; k <= 40; k += 9) {
34559 for (uint32_t m = 1; m <= 4; m++) {
34560 for (uint32_t n = 1; n <= 8; n++) {
34561 GemmMicrokernelTester()
34562 .mr(4)
34563 .nr(8)
34564 .kr(1)
34565 .sr(1)
34566 .m(m)
34567 .n(n)
34568 .k(k)
34569 .cm_stride(11)
34570 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034571 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034572 }
34573 }
34574 }
34575 }
34576
34577 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, a_offset) {
34578 TEST_REQUIRES_ARM_NEON;
34579 for (size_t k = 1; k <= 40; k += 9) {
34580 GemmMicrokernelTester()
34581 .mr(4)
34582 .nr(8)
34583 .kr(1)
34584 .sr(1)
34585 .m(4)
34586 .n(8)
34587 .k(k)
34588 .ks(3)
34589 .a_offset(163)
Marat Dukhan50323b82022-01-11 00:12:01 -080034590 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034591 }
34592 }
34593
34594 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, zero) {
34595 TEST_REQUIRES_ARM_NEON;
34596 for (uint32_t mz = 0; mz < 4; mz++) {
34597 for (size_t k = 1; k <= 40; k += 9) {
34598 GemmMicrokernelTester()
34599 .mr(4)
34600 .nr(8)
34601 .kr(1)
34602 .sr(1)
34603 .m(4)
34604 .n(8)
34605 .k(k)
34606 .ks(3)
34607 .a_offset(163)
34608 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080034609 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034610 }
34611 }
34612 }
34613
34614 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, qmin) {
34615 TEST_REQUIRES_ARM_NEON;
34616 GemmMicrokernelTester()
34617 .mr(4)
34618 .nr(8)
34619 .kr(1)
34620 .sr(1)
34621 .m(4)
34622 .n(8)
34623 .k(8)
34624 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080034625 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034626 }
34627
34628 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, qmax) {
34629 TEST_REQUIRES_ARM_NEON;
34630 GemmMicrokernelTester()
34631 .mr(4)
34632 .nr(8)
34633 .kr(1)
34634 .sr(1)
34635 .m(4)
34636 .n(8)
34637 .k(8)
34638 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080034639 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034640 }
34641
34642 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, strided_cm) {
34643 TEST_REQUIRES_ARM_NEON;
34644 GemmMicrokernelTester()
34645 .mr(4)
34646 .nr(8)
34647 .kr(1)
34648 .sr(1)
34649 .m(4)
34650 .n(8)
34651 .k(8)
34652 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080034653 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Frank Barchard510b8e02021-07-26 17:25:18 -070034654 }
34655#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
34656
34657
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034658#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS && XNN_ENABLE_ASSEMBLY && XNN_PLATFORM_JIT
34659 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_eq_8) {
34660 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034661 struct xnn_code_buffer code_buffer;
34662 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034663 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034664 GemmMicrokernelTester()
34665 .mr(4)
34666 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034667 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034668 .sr(1)
34669 .m(4)
34670 .n(8)
34671 .k(8)
Marat Dukhan50323b82022-01-11 00:12:01 -080034672 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034673 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34674 }
34675
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034676 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, strided_cn) {
34677 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034678 struct xnn_code_buffer code_buffer;
34679 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034680 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034681 GemmMicrokernelTester()
34682 .mr(4)
34683 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034684 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034685 .sr(1)
34686 .m(4)
34687 .n(8)
34688 .k(8)
34689 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080034690 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034691 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34692 }
34693
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034694 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_eq_8_subtile) {
34695 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034696 struct xnn_code_buffer code_buffer;
34697 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034698 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034699 for (uint32_t m = 1; m <= 4; m++) {
34700 for (uint32_t n = 1; n <= 8; n++) {
34701 GemmMicrokernelTester()
34702 .mr(4)
34703 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034704 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034705 .sr(1)
34706 .m(m)
34707 .n(n)
34708 .k(8)
34709 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034710 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034711 }
34712 }
34713 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34714 }
34715
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034716 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_eq_8_subtile_m) {
34717 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034718 struct xnn_code_buffer code_buffer;
34719 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034720 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034721 for (uint32_t m = 1; m <= 4; m++) {
34722 GemmMicrokernelTester()
34723 .mr(4)
34724 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034725 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034726 .sr(1)
34727 .m(m)
34728 .n(8)
34729 .k(8)
34730 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034731 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034732 }
34733 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34734 }
34735
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034736 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_eq_8_subtile_n) {
34737 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034738 struct xnn_code_buffer code_buffer;
34739 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034740 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034741 for (uint32_t n = 1; n <= 8; n++) {
34742 GemmMicrokernelTester()
34743 .mr(4)
34744 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034745 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034746 .sr(1)
34747 .m(4)
34748 .n(n)
34749 .k(8)
34750 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034751 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034752 }
34753 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34754 }
34755
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034756 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_lt_8) {
34757 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034758 struct xnn_code_buffer code_buffer;
34759 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034760 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034761 for (size_t k = 1; k < 8; k++) {
34762 GemmMicrokernelTester()
34763 .mr(4)
34764 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034765 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034766 .sr(1)
34767 .m(4)
34768 .n(8)
34769 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034770 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034771 }
34772 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34773 }
34774
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034775 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_lt_8_subtile) {
34776 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034777 struct xnn_code_buffer code_buffer;
34778 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034779 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034780 for (size_t k = 1; k < 8; k++) {
34781 for (uint32_t m = 1; m <= 4; m++) {
34782 for (uint32_t n = 1; n <= 8; n++) {
34783 GemmMicrokernelTester()
34784 .mr(4)
34785 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034786 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034787 .sr(1)
34788 .m(m)
34789 .n(n)
34790 .k(k)
34791 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034792 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034793 }
34794 }
34795 }
34796 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34797 }
34798
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034799 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_gt_8) {
34800 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034801 struct xnn_code_buffer code_buffer;
34802 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034803 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034804 for (size_t k = 9; k < 16; k++) {
34805 GemmMicrokernelTester()
34806 .mr(4)
34807 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034808 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034809 .sr(1)
34810 .m(4)
34811 .n(8)
34812 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034813 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034814 }
34815 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34816 }
34817
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034818 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_gt_8_subtile) {
34819 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034820 struct xnn_code_buffer code_buffer;
34821 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034822 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034823 for (size_t k = 9; k < 16; k++) {
34824 for (uint32_t m = 1; m <= 4; m++) {
34825 for (uint32_t n = 1; n <= 8; n++) {
34826 GemmMicrokernelTester()
34827 .mr(4)
34828 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034829 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034830 .sr(1)
34831 .m(m)
34832 .n(n)
34833 .k(k)
34834 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034835 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034836 }
34837 }
34838 }
34839 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34840 }
34841
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034842 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_div_8) {
34843 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034844 struct xnn_code_buffer code_buffer;
34845 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034846 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034847 for (size_t k = 16; k <= 80; k += 8) {
34848 GemmMicrokernelTester()
34849 .mr(4)
34850 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034851 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034852 .sr(1)
34853 .m(4)
34854 .n(8)
34855 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034856 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034857 }
34858 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34859 }
34860
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034861 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_div_8_subtile) {
34862 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034863 struct xnn_code_buffer code_buffer;
34864 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034865 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034866 for (size_t k = 16; k <= 80; k += 8) {
34867 for (uint32_t m = 1; m <= 4; m++) {
34868 for (uint32_t n = 1; n <= 8; n++) {
34869 GemmMicrokernelTester()
34870 .mr(4)
34871 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034872 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034873 .sr(1)
34874 .m(m)
34875 .n(n)
34876 .k(k)
34877 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034878 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034879 }
34880 }
34881 }
34882 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34883 }
34884
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034885 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_gt_8) {
34886 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034887 struct xnn_code_buffer code_buffer;
34888 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034889 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034890 for (uint32_t n = 9; n < 16; n++) {
34891 for (size_t k = 1; k <= 40; k += 9) {
34892 GemmMicrokernelTester()
34893 .mr(4)
34894 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034895 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034896 .sr(1)
34897 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034898 .n(n)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034899 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034900 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034901 }
34902 }
34903 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34904 }
34905
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034906 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_gt_8_strided_cn) {
34907 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034908 struct xnn_code_buffer code_buffer;
34909 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034910 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034911 for (uint32_t n = 9; n < 16; n++) {
34912 for (size_t k = 1; k <= 40; k += 9) {
34913 GemmMicrokernelTester()
34914 .mr(4)
34915 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034916 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034917 .sr(1)
34918 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034919 .n(n)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034920 .k(k)
34921 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080034922 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034923 }
34924 }
34925 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34926 }
34927
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034928 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_gt_8_subtile) {
34929 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034930 struct xnn_code_buffer code_buffer;
34931 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034932 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034933 for (uint32_t n = 9; n < 16; n++) {
34934 for (size_t k = 1; k <= 40; k += 9) {
34935 for (uint32_t m = 1; m <= 4; m++) {
34936 GemmMicrokernelTester()
34937 .mr(4)
34938 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034939 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034940 .sr(1)
34941 .m(m)
34942 .n(n)
34943 .k(k)
34944 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080034945 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034946 }
34947 }
34948 }
34949 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34950 }
34951
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034952 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_div_8) {
34953 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034954 struct xnn_code_buffer code_buffer;
34955 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034956 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034957 for (uint32_t n = 16; n <= 24; n += 8) {
34958 for (size_t k = 1; k <= 40; k += 9) {
34959 GemmMicrokernelTester()
34960 .mr(4)
34961 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034962 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034963 .sr(1)
34964 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080034965 .n(n)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034966 .k(k)
Marat Dukhan50323b82022-01-11 00:12:01 -080034967 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034968 }
34969 }
34970 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34971 }
34972
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034973 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_div_8_strided_cn) {
34974 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034975 struct xnn_code_buffer code_buffer;
34976 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034977 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034978 for (uint32_t n = 16; n <= 24; n += 8) {
34979 for (size_t k = 1; k <= 40; k += 9) {
34980 GemmMicrokernelTester()
34981 .mr(4)
34982 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034983 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034984 .sr(1)
34985 .m(4)
34986 .n(n)
34987 .k(k)
34988 .cn_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080034989 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034990 }
34991 }
34992 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
34993 }
34994
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034995 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_div_8_subtile) {
34996 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080034997 struct xnn_code_buffer code_buffer;
34998 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080034999 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035000 for (uint32_t n = 16; n <= 24; n += 8) {
35001 for (size_t k = 1; k <= 40; k += 9) {
35002 for (uint32_t m = 1; m <= 4; m++) {
35003 GemmMicrokernelTester()
35004 .mr(4)
35005 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035006 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035007 .sr(1)
35008 .m(m)
35009 .n(n)
35010 .k(k)
35011 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080035012 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035013 }
35014 }
35015 }
35016 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35017 }
35018
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035019 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, small_kernel) {
35020 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035021 struct xnn_code_buffer code_buffer;
35022 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035023 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035024 for (size_t k = 1; k <= 40; k += 9) {
35025 GemmMicrokernelTester()
35026 .mr(4)
35027 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035028 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035029 .sr(1)
35030 .m(4)
35031 .n(8)
35032 .k(k)
35033 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080035034 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035035 }
35036 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35037 }
35038
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035039 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, small_kernel_subtile) {
35040 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035041 struct xnn_code_buffer code_buffer;
35042 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035043 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035044 for (size_t k = 1; k <= 40; k += 9) {
35045 for (uint32_t m = 1; m <= 4; m++) {
35046 for (uint32_t n = 1; n <= 8; n++) {
35047 GemmMicrokernelTester()
35048 .mr(4)
35049 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035050 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035051 .sr(1)
35052 .m(m)
35053 .n(n)
35054 .k(k)
35055 .ks(3)
35056 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080035057 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035058 }
35059 }
35060 }
35061 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35062 }
35063
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035064 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_gt_8_small_kernel) {
35065 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035066 struct xnn_code_buffer code_buffer;
35067 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035068 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035069 for (uint32_t n = 9; n < 16; n++) {
35070 for (size_t k = 1; k <= 40; k += 9) {
35071 GemmMicrokernelTester()
35072 .mr(4)
35073 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035074 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035075 .sr(1)
35076 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080035077 .n(n)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035078 .k(k)
35079 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080035080 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035081 }
35082 }
35083 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35084 }
35085
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035086 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_div_8_small_kernel) {
35087 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035088 struct xnn_code_buffer code_buffer;
35089 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035090 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035091 for (uint32_t n = 16; n <= 24; n += 8) {
35092 for (size_t k = 1; k <= 40; k += 9) {
35093 GemmMicrokernelTester()
35094 .mr(4)
35095 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035096 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035097 .sr(1)
35098 .m(4)
Zhi An Ngaf9ff852022-01-13 10:48:37 -080035099 .n(n)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035100 .k(k)
35101 .ks(3)
Marat Dukhan50323b82022-01-11 00:12:01 -080035102 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035103 }
35104 }
35105 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35106 }
35107
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035108 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, strided_cm_subtile) {
35109 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035110 struct xnn_code_buffer code_buffer;
35111 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035112 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035113 for (size_t k = 1; k <= 40; k += 9) {
35114 for (uint32_t m = 1; m <= 4; m++) {
35115 for (uint32_t n = 1; n <= 8; n++) {
35116 GemmMicrokernelTester()
35117 .mr(4)
35118 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035119 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035120 .sr(1)
35121 .m(m)
35122 .n(n)
35123 .k(k)
35124 .cm_stride(11)
35125 .iterations(1)
Marat Dukhan50323b82022-01-11 00:12:01 -080035126 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035127 }
35128 }
35129 }
35130 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35131 }
35132
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035133 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, a_offset) {
35134 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035135 struct xnn_code_buffer code_buffer;
35136 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035137 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035138 for (size_t k = 1; k <= 40; k += 9) {
35139 GemmMicrokernelTester()
35140 .mr(4)
35141 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035142 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035143 .sr(1)
35144 .m(4)
35145 .n(8)
35146 .k(k)
35147 .ks(3)
35148 .a_offset(163)
Marat Dukhan50323b82022-01-11 00:12:01 -080035149 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035150 }
35151 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35152 }
35153
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035154 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, zero) {
35155 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035156 struct xnn_code_buffer code_buffer;
35157 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035158 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035159 for (uint32_t mz = 0; mz < 4; mz++) {
35160 for (size_t k = 1; k <= 40; k += 9) {
35161 GemmMicrokernelTester()
35162 .mr(4)
35163 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035164 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035165 .sr(1)
35166 .m(4)
35167 .n(8)
35168 .k(k)
35169 .ks(3)
35170 .a_offset(163)
35171 .zero_index(mz)
Marat Dukhan50323b82022-01-11 00:12:01 -080035172 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035173 }
35174 }
35175 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35176 }
35177
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035178 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, qmin) {
35179 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035180 struct xnn_code_buffer code_buffer;
35181 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035182 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035183 GemmMicrokernelTester()
35184 .mr(4)
35185 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035186 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035187 .sr(1)
35188 .m(4)
35189 .n(8)
35190 .k(8)
35191 .qmin(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080035192 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035193 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35194 }
35195
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035196 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, qmax) {
35197 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035198 struct xnn_code_buffer code_buffer;
35199 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035200 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035201 GemmMicrokernelTester()
35202 .mr(4)
35203 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035204 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035205 .sr(1)
35206 .m(4)
35207 .n(8)
35208 .k(8)
35209 .qmax(128)
Marat Dukhan50323b82022-01-11 00:12:01 -080035210 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035211 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35212 }
35213
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035214 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, strided_cm) {
35215 TEST_REQUIRES_ARM_NEON_DOT;
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035216 struct xnn_code_buffer code_buffer;
35217 ASSERT_EQ(xnn_status_success, xnn_allocate_code_memory(&code_buffer, XNN_DEFAULT_CODE_BUFFER_SIZE));
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035218 ASSERT_EQ(xnn_status_success, xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64(&code_buffer));
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035219 GemmMicrokernelTester()
35220 .mr(4)
35221 .nr(8)
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035222 .kr(4)
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035223 .sr(1)
35224 .m(4)
35225 .n(8)
35226 .k(8)
35227 .cm_stride(11)
Marat Dukhan50323b82022-01-11 00:12:01 -080035228 .Test(reinterpret_cast<xnn_qs8_igemm_minmax_ukernel_function>(code_buffer.code), xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
Zhi An Ng13b57dd2022-01-06 09:33:20 -080035229 ASSERT_EQ(xnn_status_success, xnn_release_code_memory(&code_buffer));
35230 }
Zhi An Ngc27f04b2022-01-11 09:34:07 -080035231#endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS && XNN_ENABLE_ASSEMBLY && XNN_PLATFORM_JIT