Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-16x16c4-minmax-fp32-avx512amx-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-avx512amx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x16c4__avx512amx_prfm(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
34 {
35 assert(mr != 0);
36 assert(mr <= 16);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 // TODO: amxintrin.h only provide intrinsics for __x86_64__
45 // Update if amxintrin changes
46 #if defined(__x86_64__)
47 __attribute__((aligned(64))) int32_t vintile[16 * 16];
48 __attribute__((aligned(64))) int32_t res0[16 * 16];
49
50 kc = round_up_po2(kc, 4 * sizeof(int8_t));
51 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
52 const __mmask16 kremainder_mask = _cvtu32_mask16((UINT32_C(1) << (kremainder >> 2)) - 1);
53
54 // Define tile config data structure
55 struct __tile_config {
56 uint8_t palette_id;
57 uint8_t start_row;
58 uint8_t reserved_0[14];
59 uint16_t colsb[8];
60 uint16_t reserved_1[8];
61 uint8_t rows[8];
62 uint8_t reserved_2[8];
63 };
64
65 // Load tile configuration
66 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
67 tile_data.palette_id = 1;
68 tile_data.rows[0] = mr; // tmm0 = res 0
69 tile_data.rows[1] = mr; // tmm1 = res 1
70 tile_data.rows[2] = mr; // tmm2 = res 2
71 tile_data.rows[3] = mr; // tmm3 = res 3
72 tile_data.rows[4] = mr; // tmm4 = input
73 tile_data.rows[5] = 16; // tmm5 = weights
74 tile_data.rows[6] = mr; // tmm6 = input remainder
75 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
76
77 tile_data.colsb[0] = 64; // tmm0 = res 0
78 tile_data.colsb[1] = 64; // tmm1 = res 1
79 tile_data.colsb[2] = 64; // tmm2 = res 1
80 tile_data.colsb[3] = 64; // tmm3 = res 1
81 tile_data.colsb[4] = 64; // tmm4 = input
82 tile_data.colsb[5] = 64; // tmm5 = weights
83 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
84 tile_data.colsb[7] = 64; // tmm7 = weights remainder
85
86 //_tile_loadconfig(&tile_data);
87 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
88
89 int8_t* c0 = c;
90 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
91 if XNN_UNPREDICTABLE(mr < 2) {
92 c1 = c0;
93 }
94 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
95 if XNN_UNPREDICTABLE(mr <= 2) {
96 c2 = c1;
97 }
98 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
99 if XNN_UNPREDICTABLE(mr < 4) {
100 c3 = c2;
101 }
102 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
103 if XNN_UNPREDICTABLE(mr <= 4) {
104 c4 = c3;
105 }
106 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
107 if XNN_UNPREDICTABLE(mr < 6) {
108 c5 = c4;
109 }
110 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
111 if XNN_UNPREDICTABLE(mr <= 6) {
112 c6 = c5;
113 }
114 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
115 if XNN_UNPREDICTABLE(mr < 8) {
116 c7 = c6;
117 }
118 int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
119 if XNN_UNPREDICTABLE(mr <= 8) {
120 c8 = c7;
121 }
122 int8_t* c9 = (int8_t*) ((uintptr_t) c8 + cm_stride);
123 if XNN_UNPREDICTABLE(mr < 10) {
124 c9 = c8;
125 }
126 int8_t* c10 = (int8_t*) ((uintptr_t) c9 + cm_stride);
127 if XNN_UNPREDICTABLE(mr <= 10) {
128 c10 = c9;
129 }
130 int8_t* c11 = (int8_t*) ((uintptr_t) c10 + cm_stride);
131 if XNN_UNPREDICTABLE(mr < 12) {
132 c11 = c10;
133 }
134 int8_t* c12 = (int8_t*) ((uintptr_t) c11 + cm_stride);
135 if XNN_UNPREDICTABLE(mr <= 12) {
136 c12 = c11;
137 }
138 int8_t* c13 = (int8_t*) ((uintptr_t) c12 + cm_stride);
139 if XNN_UNPREDICTABLE(mr < 14) {
140 c13 = c12;
141 }
142 int8_t* c14 = (int8_t*) ((uintptr_t) c13 + cm_stride);
143 if XNN_UNPREDICTABLE(mr <= 14) {
144 c14 = c13;
145 }
146 int8_t* c15 = (int8_t*) ((uintptr_t) c14 + cm_stride);
147 if XNN_UNPREDICTABLE(mr != 16) {
148 c15 = c14;
149 }
150
151 const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
152 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
153 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
154 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
155 // XNN_FORCE_REALIZATION(voutput_zero_point);
156 // XNN_FORCE_REALIZATION(voutput_min);
157
158 do {
159 const __m512i vksum0123456789ABCDEF = _mm512_loadu_epi32((const int32_t*) w + 0);
160 w = (const int32_t*) w + 16;
161
162 // Zero tile accumulator
163 __asm__ volatile (
164 "tilezero %%tmm0\n"
165 ::);
166
167 size_t p = ks;
168 do {
169 const int8_t* restrict a0 = a[0];
170 if XNN_UNPREDICTABLE(a0 != zero) {
171 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
172 }
173 const int8_t* restrict a1 = a[1];
174 if XNN_UNPREDICTABLE(a1 != zero) {
175 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
176 }
177 const int8_t* restrict a2 = a[2];
178 if XNN_UNPREDICTABLE(a2 != zero) {
179 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
180 }
181 const int8_t* restrict a3 = a[3];
182 if XNN_UNPREDICTABLE(a3 != zero) {
183 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
184 }
185 const int8_t* restrict a4 = a[4];
186 if XNN_UNPREDICTABLE(a4 != zero) {
187 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
188 }
189 const int8_t* restrict a5 = a[5];
190 if XNN_UNPREDICTABLE(a5 != zero) {
191 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
192 }
193 const int8_t* restrict a6 = a[6];
194 if XNN_UNPREDICTABLE(a6 != zero) {
195 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
196 }
197 const int8_t* restrict a7 = a[7];
198 if XNN_UNPREDICTABLE(a7 != zero) {
199 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
200 }
201 const int8_t* restrict a8 = a[8];
202 if XNN_UNPREDICTABLE(a8 != zero) {
203 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
204 }
205 const int8_t* restrict a9 = a[9];
206 if XNN_UNPREDICTABLE(a9 != zero) {
207 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
208 }
209 const int8_t* restrict a10 = a[10];
210 if XNN_UNPREDICTABLE(a10 != zero) {
211 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
212 }
213 const int8_t* restrict a11 = a[11];
214 if XNN_UNPREDICTABLE(a11 != zero) {
215 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
216 }
217 const int8_t* restrict a12 = a[12];
218 if XNN_UNPREDICTABLE(a12 != zero) {
219 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
220 }
221 const int8_t* restrict a13 = a[13];
222 if XNN_UNPREDICTABLE(a13 != zero) {
223 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
224 }
225 const int8_t* restrict a14 = a[14];
226 if XNN_UNPREDICTABLE(a14 != zero) {
227 a14 = (const int8_t*) ((uintptr_t) a14 + a_offset);
228 }
229 const int8_t* restrict a15 = a[15];
230 if XNN_UNPREDICTABLE(a15 != zero) {
231 a15 = (const int8_t*) ((uintptr_t) a15 + a_offset);
232 }
233 a += 16;
234
235 size_t k = kc;
236 if (mr == 1)
237 {
238 while (k >= 64 * sizeof(int8_t)) {
239 _tile_loadd(4, a0, 64); // Directly load input for mr=1
240 a15 += 64;
241 _tile_loadd(5, (const int8_t*) w + 0, 64);
242 _tile_dpbssd(0, 4, 5);
243 xnn_prefetch_to_l1((const int8_t*) w + 4096);
244 xnn_prefetch_to_l1((const int8_t*) w + 4160);
245 xnn_prefetch_to_l1((const int8_t*) w + 4224);
246 xnn_prefetch_to_l1((const int8_t*) w + 4288);
247 xnn_prefetch_to_l1((const int8_t*) w + 4352);
248 xnn_prefetch_to_l1((const int8_t*) w + 4416);
249 xnn_prefetch_to_l1((const int8_t*) w + 4480);
250 xnn_prefetch_to_l1((const int8_t*) w + 4544);
251 xnn_prefetch_to_l1((const int8_t*) w + 4608);
252 xnn_prefetch_to_l1((const int8_t*) w + 4672);
253 xnn_prefetch_to_l1((const int8_t*) w + 4736);
254 xnn_prefetch_to_l1((const int8_t*) w + 4800);
255 xnn_prefetch_to_l1((const int8_t*) w + 4864);
256 xnn_prefetch_to_l1((const int8_t*) w + 4928);
257 xnn_prefetch_to_l1((const int8_t*) w + 4992);
258 xnn_prefetch_to_l1((const int8_t*) w + 5056);
259 xnn_prefetch_to_l1((const int8_t*) w + 5120);
260 xnn_prefetch_to_l1((const int8_t*) w + 5184);
261 xnn_prefetch_to_l1((const int8_t*) w + 5248);
262 xnn_prefetch_to_l1((const int8_t*) w + 5312);
263 xnn_prefetch_to_l1((const int8_t*) w + 5376);
264 xnn_prefetch_to_l1((const int8_t*) w + 5440);
265 xnn_prefetch_to_l1((const int8_t*) w + 5504);
266 xnn_prefetch_to_l1((const int8_t*) w + 5568);
267 xnn_prefetch_to_l1((const int8_t*) w + 5632);
268 xnn_prefetch_to_l1((const int8_t*) w + 5696);
269 xnn_prefetch_to_l1((const int8_t*) w + 5760);
270 xnn_prefetch_to_l1((const int8_t*) w + 5824);
271 xnn_prefetch_to_l1((const int8_t*) w + 5888);
272 xnn_prefetch_to_l1((const int8_t*) w + 5952);
273 xnn_prefetch_to_l1((const int8_t*) w + 6016);
274 xnn_prefetch_to_l1((const int8_t*) w + 6080);
275
276 w = (const int8_t*) w + 1024;
277 k -= 64 * sizeof(int8_t);
278 }
279 }
280 else {
281 while (k >= 64 * sizeof(int8_t)) {
282 const __m512i vin0 = _mm512_loadu_epi32(a0);
283 a0 += 64;
284 _mm512_store_epi32(vintile + 0, vin0);
285 const __m512i vin1 = _mm512_loadu_epi32(a1);
286 a1 += 64;
287 _mm512_store_epi32(vintile + 16, vin1);
288 const __m512i vin2 = _mm512_loadu_epi32(a2);
289 a2 += 64;
290 _mm512_store_epi32(vintile + 32, vin2);
291 const __m512i vin3 = _mm512_loadu_epi32(a3);
292 a3 += 64;
293 _mm512_store_epi32(vintile + 48, vin3);
294 const __m512i vin4 = _mm512_loadu_epi32(a4);
295 a4 += 64;
296 _mm512_store_epi32(vintile + 64, vin4);
297 const __m512i vin5 = _mm512_loadu_epi32(a5);
298 a5 += 64;
299 _mm512_store_epi32(vintile + 80, vin5);
300 const __m512i vin6 = _mm512_loadu_epi32(a6);
301 a6 += 64;
302 _mm512_store_epi32(vintile + 96, vin6);
303 const __m512i vin7 = _mm512_loadu_epi32(a7);
304 a7 += 64;
305 _mm512_store_epi32(vintile + 112, vin7);
306 const __m512i vin8 = _mm512_loadu_epi32(a8);
307 a8 += 64;
308 _mm512_store_epi32(vintile + 128, vin8);
309 const __m512i vin9 = _mm512_loadu_epi32(a9);
310 a9 += 64;
311 _mm512_store_epi32(vintile + 144, vin9);
312 const __m512i vin10 = _mm512_loadu_epi32(a10);
313 a10 += 64;
314 _mm512_store_epi32(vintile + 160, vin10);
315 const __m512i vin11 = _mm512_loadu_epi32(a11);
316 a11 += 64;
317 _mm512_store_epi32(vintile + 176, vin11);
318 const __m512i vin12 = _mm512_loadu_epi32(a12);
319 a12 += 64;
320 _mm512_store_epi32(vintile + 192, vin12);
321 const __m512i vin13 = _mm512_loadu_epi32(a13);
322 a13 += 64;
323 _mm512_store_epi32(vintile + 208, vin13);
324 const __m512i vin14 = _mm512_loadu_epi32(a14);
325 a14 += 64;
326 _mm512_store_epi32(vintile + 224, vin14);
327 const __m512i vin15 = _mm512_loadu_epi32(a15);
328 a15 += 64;
329 _mm512_store_epi32(vintile + 240, vin15);
330 _tile_loadd(4, vintile, 64);
331 _tile_loadd(5, (const int8_t*) w + 0, 64);
332 _tile_dpbssd(0, 4, 5);
333 xnn_prefetch_to_l1((const int8_t*) w + 4096);
334 xnn_prefetch_to_l1((const int8_t*) w + 4160);
335 xnn_prefetch_to_l1((const int8_t*) w + 4224);
336 xnn_prefetch_to_l1((const int8_t*) w + 4288);
337 xnn_prefetch_to_l1((const int8_t*) w + 4352);
338 xnn_prefetch_to_l1((const int8_t*) w + 4416);
339 xnn_prefetch_to_l1((const int8_t*) w + 4480);
340 xnn_prefetch_to_l1((const int8_t*) w + 4544);
341 xnn_prefetch_to_l1((const int8_t*) w + 4608);
342 xnn_prefetch_to_l1((const int8_t*) w + 4672);
343 xnn_prefetch_to_l1((const int8_t*) w + 4736);
344 xnn_prefetch_to_l1((const int8_t*) w + 4800);
345 xnn_prefetch_to_l1((const int8_t*) w + 4864);
346 xnn_prefetch_to_l1((const int8_t*) w + 4928);
347 xnn_prefetch_to_l1((const int8_t*) w + 4992);
348 xnn_prefetch_to_l1((const int8_t*) w + 5056);
349 xnn_prefetch_to_l1((const int8_t*) w + 5120);
350 xnn_prefetch_to_l1((const int8_t*) w + 5184);
351 xnn_prefetch_to_l1((const int8_t*) w + 5248);
352 xnn_prefetch_to_l1((const int8_t*) w + 5312);
353 xnn_prefetch_to_l1((const int8_t*) w + 5376);
354 xnn_prefetch_to_l1((const int8_t*) w + 5440);
355 xnn_prefetch_to_l1((const int8_t*) w + 5504);
356 xnn_prefetch_to_l1((const int8_t*) w + 5568);
357 xnn_prefetch_to_l1((const int8_t*) w + 5632);
358 xnn_prefetch_to_l1((const int8_t*) w + 5696);
359 xnn_prefetch_to_l1((const int8_t*) w + 5760);
360 xnn_prefetch_to_l1((const int8_t*) w + 5824);
361 xnn_prefetch_to_l1((const int8_t*) w + 5888);
362 xnn_prefetch_to_l1((const int8_t*) w + 5952);
363 xnn_prefetch_to_l1((const int8_t*) w + 6016);
364 xnn_prefetch_to_l1((const int8_t*) w + 6080);
365
366 w = (const int8_t*) w + 1024;
367 k -= 64 * sizeof(int8_t);
368 }
369 }
370
371 if XNN_UNLIKELY(k != 0) {
372 const __m512i vin0 = _mm512_maskz_loadu_epi32(kremainder_mask, a0);
373 a0 += kremainder;
374 _mm512_store_epi32(vintile + 0, vin0);
375 const __m512i vin1 = _mm512_maskz_loadu_epi32(kremainder_mask, a1);
376 a1 += kremainder;
377 _mm512_store_epi32(vintile + 16, vin1);
378 const __m512i vin2 = _mm512_maskz_loadu_epi32(kremainder_mask, a2);
379 a2 += kremainder;
380 _mm512_store_epi32(vintile + 32, vin2);
381 const __m512i vin3 = _mm512_maskz_loadu_epi32(kremainder_mask, a3);
382 a3 += kremainder;
383 _mm512_store_epi32(vintile + 48, vin3);
384 const __m512i vin4 = _mm512_maskz_loadu_epi32(kremainder_mask, a4);
385 a4 += kremainder;
386 _mm512_store_epi32(vintile + 64, vin4);
387 const __m512i vin5 = _mm512_maskz_loadu_epi32(kremainder_mask, a5);
388 a5 += kremainder;
389 _mm512_store_epi32(vintile + 80, vin5);
390 const __m512i vin6 = _mm512_maskz_loadu_epi32(kremainder_mask, a6);
391 a6 += kremainder;
392 _mm512_store_epi32(vintile + 96, vin6);
393 const __m512i vin7 = _mm512_maskz_loadu_epi32(kremainder_mask, a7);
394 a7 += kremainder;
395 _mm512_store_epi32(vintile + 112, vin7);
396 const __m512i vin8 = _mm512_maskz_loadu_epi32(kremainder_mask, a8);
397 a8 += kremainder;
398 _mm512_store_epi32(vintile + 128, vin8);
399 const __m512i vin9 = _mm512_maskz_loadu_epi32(kremainder_mask, a9);
400 a9 += kremainder;
401 _mm512_store_epi32(vintile + 144, vin9);
402 const __m512i vin10 = _mm512_maskz_loadu_epi32(kremainder_mask, a10);
403 a10 += kremainder;
404 _mm512_store_epi32(vintile + 160, vin10);
405 const __m512i vin11 = _mm512_maskz_loadu_epi32(kremainder_mask, a11);
406 a11 += kremainder;
407 _mm512_store_epi32(vintile + 176, vin11);
408 const __m512i vin12 = _mm512_maskz_loadu_epi32(kremainder_mask, a12);
409 a12 += kremainder;
410 _mm512_store_epi32(vintile + 192, vin12);
411 const __m512i vin13 = _mm512_maskz_loadu_epi32(kremainder_mask, a13);
412 a13 += kremainder;
413 _mm512_store_epi32(vintile + 208, vin13);
414 const __m512i vin14 = _mm512_maskz_loadu_epi32(kremainder_mask, a14);
415 a14 += kremainder;
416 _mm512_store_epi32(vintile + 224, vin14);
417 const __m512i vin15 = _mm512_maskz_loadu_epi32(kremainder_mask, a15);
418 a15 += kremainder;
419 _mm512_store_epi32(vintile + 240, vin15);
420 _tile_loadd(6, vintile, 64);
421 _tile_loadd(7, (const int8_t*) w + 0, 64);
422 _tile_dpbssd(0, 6, 7);
423
424 w = (const int8_t*) w + kremainder * 16;
425 k -= kremainder * sizeof(int8_t);
426 }
427
428 p -= 16 * sizeof(void*);
429 } while (p != 0);
430
431 // Add tile to bias
432 _tile_stored(0, res0, 64);
433
434 __m512i vacc0x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
435 __m512i vacc1x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
436 __m512i vacc2x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
437 __m512i vacc3x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
438 __m512i vacc4x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
439 __m512i vacc5x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
440 __m512i vacc6x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
441 __m512i vacc7x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 112));
442 __m512i vacc8x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 128));
443 __m512i vacc9x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 144));
444 __m512i vacc10x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 160));
445 __m512i vacc11x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 176));
446 __m512i vacc12x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 192));
447 __m512i vacc13x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 208));
448 __m512i vacc14x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 224));
449 __m512i vacc15x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 240));
450
451 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
452 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
453 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
454 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
455 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
456 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
457 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
458 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
459 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
460 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
461 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
462 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
463 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
464 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
465 __m512 vscaled14x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc14x0123456789ABCDEF);
466 __m512 vscaled15x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc15x0123456789ABCDEF);
467
468 const __m512 vscale0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
469 w = (const int32_t*) w + 16;
470
471 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale0123456789ABCDEF);
472 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale0123456789ABCDEF);
473 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale0123456789ABCDEF);
474 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale0123456789ABCDEF);
475 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale0123456789ABCDEF);
476 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale0123456789ABCDEF);
477 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale0123456789ABCDEF);
478 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vscale0123456789ABCDEF);
479 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vscale0123456789ABCDEF);
480 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vscale0123456789ABCDEF);
481 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, vscale0123456789ABCDEF);
482 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, vscale0123456789ABCDEF);
483 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, vscale0123456789ABCDEF);
484 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, vscale0123456789ABCDEF);
485 vscaled14x0123456789ABCDEF = _mm512_mul_ps(vscaled14x0123456789ABCDEF, vscale0123456789ABCDEF);
486 vscaled15x0123456789ABCDEF = _mm512_mul_ps(vscaled15x0123456789ABCDEF, vscale0123456789ABCDEF);
487
488 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
489 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
490 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
491 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
492 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
493 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
494 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
495 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max_less_zero_point);
496 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max_less_zero_point);
497 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max_less_zero_point);
498 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max_less_zero_point);
499 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max_less_zero_point);
500 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max_less_zero_point);
501 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max_less_zero_point);
502 vscaled14x0123456789ABCDEF = _mm512_min_ps(vscaled14x0123456789ABCDEF, voutput_max_less_zero_point);
503 vscaled15x0123456789ABCDEF = _mm512_min_ps(vscaled15x0123456789ABCDEF, voutput_max_less_zero_point);
504
505 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
506 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
507 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
508 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
509 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
510 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
511 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
512 vacc7x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled7x0123456789ABCDEF);
513 vacc8x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled8x0123456789ABCDEF);
514 vacc9x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled9x0123456789ABCDEF);
515 vacc10x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled10x0123456789ABCDEF);
516 vacc11x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled11x0123456789ABCDEF);
517 vacc12x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled12x0123456789ABCDEF);
518 vacc13x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled13x0123456789ABCDEF);
519 vacc14x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled14x0123456789ABCDEF);
520 vacc15x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled15x0123456789ABCDEF);
521
522 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
523 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
524 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
525 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
526 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
527 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
528 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
529 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, voutput_zero_point);
530 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, voutput_zero_point);
531 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, voutput_zero_point);
532 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, voutput_zero_point);
533 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, voutput_zero_point);
534 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, voutput_zero_point);
535 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, voutput_zero_point);
536 vacc14x0123456789ABCDEF = _mm512_add_epi32(vacc14x0123456789ABCDEF, voutput_zero_point);
537 vacc15x0123456789ABCDEF = _mm512_add_epi32(vacc15x0123456789ABCDEF, voutput_zero_point);
538
539 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
540 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
541 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
542 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
543 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
544 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
545 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
546 __m128i vout7x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc7x0123456789ABCDEF);
547 __m128i vout8x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc8x0123456789ABCDEF);
548 __m128i vout9x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc9x0123456789ABCDEF);
549 __m128i vout10x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc10x0123456789ABCDEF);
550 __m128i vout11x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc11x0123456789ABCDEF);
551 __m128i vout12x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc12x0123456789ABCDEF);
552 __m128i vout13x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc13x0123456789ABCDEF);
553 __m128i vout14x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc14x0123456789ABCDEF);
554 __m128i vout15x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc15x0123456789ABCDEF);
555
556 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
557 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
558 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
559 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
560 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
561 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
562 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
563 vout7x0123456789ABCDEF = _mm_max_epi8(vout7x0123456789ABCDEF, voutput_min);
564 vout8x0123456789ABCDEF = _mm_max_epi8(vout8x0123456789ABCDEF, voutput_min);
565 vout9x0123456789ABCDEF = _mm_max_epi8(vout9x0123456789ABCDEF, voutput_min);
566 vout10x0123456789ABCDEF = _mm_max_epi8(vout10x0123456789ABCDEF, voutput_min);
567 vout11x0123456789ABCDEF = _mm_max_epi8(vout11x0123456789ABCDEF, voutput_min);
568 vout12x0123456789ABCDEF = _mm_max_epi8(vout12x0123456789ABCDEF, voutput_min);
569 vout13x0123456789ABCDEF = _mm_max_epi8(vout13x0123456789ABCDEF, voutput_min);
570 vout14x0123456789ABCDEF = _mm_max_epi8(vout14x0123456789ABCDEF, voutput_min);
571 vout15x0123456789ABCDEF = _mm_max_epi8(vout15x0123456789ABCDEF, voutput_min);
572
573 if XNN_LIKELY(nc >= 16) {
574 _mm_storeu_si128((__m128i*) (c15 + 0), vout15x0123456789ABCDEF);
575 c15 = (int8_t*) ((uintptr_t) c15 + cn_stride);
576 _mm_storeu_si128((__m128i*) (c14 + 0), vout14x0123456789ABCDEF);
577 c14 = (int8_t*) ((uintptr_t) c14 + cn_stride);
578 _mm_storeu_si128((__m128i*) (c13 + 0), vout13x0123456789ABCDEF);
579 c13 = (int8_t*) ((uintptr_t) c13 + cn_stride);
580 _mm_storeu_si128((__m128i*) (c12 + 0), vout12x0123456789ABCDEF);
581 c12 = (int8_t*) ((uintptr_t) c12 + cn_stride);
582 _mm_storeu_si128((__m128i*) (c11 + 0), vout11x0123456789ABCDEF);
583 c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
584 _mm_storeu_si128((__m128i*) (c10 + 0), vout10x0123456789ABCDEF);
585 c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
586 _mm_storeu_si128((__m128i*) (c9 + 0), vout9x0123456789ABCDEF);
587 c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
588 _mm_storeu_si128((__m128i*) (c8 + 0), vout8x0123456789ABCDEF);
589 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
590 _mm_storeu_si128((__m128i*) (c7 + 0), vout7x0123456789ABCDEF);
591 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
592 _mm_storeu_si128((__m128i*) (c6 + 0), vout6x0123456789ABCDEF);
593 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
594 _mm_storeu_si128((__m128i*) (c5 + 0), vout5x0123456789ABCDEF);
595 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
596 _mm_storeu_si128((__m128i*) (c4 + 0), vout4x0123456789ABCDEF);
597 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
598 _mm_storeu_si128((__m128i*) (c3 + 0), vout3x0123456789ABCDEF);
599 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
600 _mm_storeu_si128((__m128i*) (c2 + 0), vout2x0123456789ABCDEF);
601 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
602 _mm_storeu_si128((__m128i*) (c1 + 0), vout1x0123456789ABCDEF);
603 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
604 _mm_storeu_si128((__m128i*) (c0 + 0), vout0x0123456789ABCDEF);
605 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
606 a = (const int8_t**restrict) ((uintptr_t) a - ks);
607 nc -= 16;
608 } else {
609 // Prepare mask for valid 8-bit elements (depends on nc).
610 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
611
612 _mm_mask_storeu_epi8(c15 + 0, vmask0, vout15x0123456789ABCDEF);
613 _mm_mask_storeu_epi8(c14 + 0, vmask0, vout14x0123456789ABCDEF);
614 _mm_mask_storeu_epi8(c13 + 0, vmask0, vout13x0123456789ABCDEF);
615 _mm_mask_storeu_epi8(c12 + 0, vmask0, vout12x0123456789ABCDEF);
616 _mm_mask_storeu_epi8(c11 + 0, vmask0, vout11x0123456789ABCDEF);
617 _mm_mask_storeu_epi8(c10 + 0, vmask0, vout10x0123456789ABCDEF);
618 _mm_mask_storeu_epi8(c9 + 0, vmask0, vout9x0123456789ABCDEF);
619 _mm_mask_storeu_epi8(c8 + 0, vmask0, vout8x0123456789ABCDEF);
620 _mm_mask_storeu_epi8(c7 + 0, vmask0, vout7x0123456789ABCDEF);
621 _mm_mask_storeu_epi8(c6 + 0, vmask0, vout6x0123456789ABCDEF);
622 _mm_mask_storeu_epi8(c5 + 0, vmask0, vout5x0123456789ABCDEF);
623 _mm_mask_storeu_epi8(c4 + 0, vmask0, vout4x0123456789ABCDEF);
624 _mm_mask_storeu_epi8(c3 + 0, vmask0, vout3x0123456789ABCDEF);
625 _mm_mask_storeu_epi8(c2 + 0, vmask0, vout2x0123456789ABCDEF);
626 _mm_mask_storeu_epi8(c1 + 0, vmask0, vout1x0123456789ABCDEF);
627 _mm_mask_storeu_epi8(c0 + 0, vmask0, vout0x0123456789ABCDEF);
628 nc = 0;
629 }
630 } while (nc != 0);
631
632 // Release tile config
633 // _tile_release();
634 __asm__ volatile ("tilerelease" ::);
635 #endif // defined(__x86_64__)
636 }