Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-16x16c4-minmax-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_qd8_f32_qc8w_igemm_minmax_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 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)])
36 {
37 assert(mr != 0);
38 assert(mr <= 16);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 // TODO: amxintrin.h only provide intrinsics for __x86_64__
47 // Update if amxintrin changes
48 #if defined(__x86_64__)
49 __attribute__((aligned(64))) int32_t vintile[16 * 16];
50 __attribute__((aligned(64))) int32_t res0[16 * 16];
51
52 kc = round_up_po2(kc, 4 * sizeof(int8_t));
53 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
54 const __mmask16 kremainder_mask = _cvtu32_mask16((UINT32_C(1) << (kremainder >> 2)) - 1);
55
56 // Define tile config data structure
57 struct __tile_config {
58 uint8_t palette_id;
59 uint8_t start_row;
60 uint8_t reserved_0[14];
61 uint16_t colsb[8];
62 uint16_t reserved_1[8];
63 uint8_t rows[8];
64 uint8_t reserved_2[8];
65 };
66
67 // Load tile configuration
68 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
69 tile_data.palette_id = 1;
70 tile_data.rows[0] = mr; // tmm0 = res 0
71 tile_data.rows[1] = mr; // tmm1 = res 1
72 tile_data.rows[2] = mr; // tmm2 = res 2
73 tile_data.rows[3] = mr; // tmm3 = res 3
74 tile_data.rows[4] = mr; // tmm4 = input
75 tile_data.rows[5] = 16; // tmm5 = weights
76 tile_data.rows[6] = mr; // tmm6 = input remainder
77 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
78
79 tile_data.colsb[0] = 64; // tmm0 = res 0
80 tile_data.colsb[1] = 64; // tmm1 = res 1
81 tile_data.colsb[2] = 64; // tmm2 = res 1
82 tile_data.colsb[3] = 64; // tmm3 = res 1
83 tile_data.colsb[4] = 64; // tmm4 = input
84 tile_data.colsb[5] = 64; // tmm5 = weights
85 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
86 tile_data.colsb[7] = 64; // tmm7 = weights remainder
87
88 //_tile_loadconfig(&tile_data);
89 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
90
91 float* c0 = c;
92 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
93 if XNN_UNPREDICTABLE(mr < 2) {
94 c1 = c0;
95 }
96 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
97 if XNN_UNPREDICTABLE(mr <= 2) {
98 c2 = c1;
99 }
100 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
101 if XNN_UNPREDICTABLE(mr < 4) {
102 c3 = c2;
103 }
104 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
105 if XNN_UNPREDICTABLE(mr <= 4) {
106 c4 = c3;
107 }
108 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
109 if XNN_UNPREDICTABLE(mr < 6) {
110 c5 = c4;
111 }
112 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
113 if XNN_UNPREDICTABLE(mr <= 6) {
114 c6 = c5;
115 }
116 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
117 if XNN_UNPREDICTABLE(mr < 8) {
118 c7 = c6;
119 }
120 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
121 if XNN_UNPREDICTABLE(mr <= 8) {
122 c8 = c7;
123 }
124 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
125 if XNN_UNPREDICTABLE(mr < 10) {
126 c9 = c8;
127 }
128 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
129 if XNN_UNPREDICTABLE(mr <= 10) {
130 c10 = c9;
131 }
132 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
133 if XNN_UNPREDICTABLE(mr < 12) {
134 c11 = c10;
135 }
136 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
137 if XNN_UNPREDICTABLE(mr <= 12) {
138 c12 = c11;
139 }
140 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
141 if XNN_UNPREDICTABLE(mr < 14) {
142 c13 = c12;
143 }
144 float* c14 = (float*) ((uintptr_t) c13 + cm_stride);
145 if XNN_UNPREDICTABLE(mr <= 14) {
146 c14 = c13;
147 }
148 float* c15 = (float*) ((uintptr_t) c14 + cm_stride);
149 if XNN_UNPREDICTABLE(mr != 16) {
150 c15 = c14;
151 }
152
153 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
154 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
155 // XNN_FORCE_REALIZATION(voutput_min);
156 // XNN_FORCE_REALIZATION(voutput_max);
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 } else {
173 a0 = zero_data;
174 }
175 const int8_t* restrict a1 = a[1];
176 if XNN_UNPREDICTABLE(a1 != zero) {
177 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
178 } else {
179 a1 = zero_data;
180 }
181 const int8_t* restrict a2 = a[2];
182 if XNN_UNPREDICTABLE(a2 != zero) {
183 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
184 } else {
185 a2 = zero_data;
186 }
187 const int8_t* restrict a3 = a[3];
188 if XNN_UNPREDICTABLE(a3 != zero) {
189 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
190 } else {
191 a3 = zero_data;
192 }
193 const int8_t* restrict a4 = a[4];
194 if XNN_UNPREDICTABLE(a4 != zero) {
195 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
196 } else {
197 a4 = zero_data;
198 }
199 const int8_t* restrict a5 = a[5];
200 if XNN_UNPREDICTABLE(a5 != zero) {
201 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
202 } else {
203 a5 = zero_data;
204 }
205 const int8_t* restrict a6 = a[6];
206 if XNN_UNPREDICTABLE(a6 != zero) {
207 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
208 } else {
209 a6 = zero_data;
210 }
211 const int8_t* restrict a7 = a[7];
212 if XNN_UNPREDICTABLE(a7 != zero) {
213 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
214 } else {
215 a7 = zero_data;
216 }
217 const int8_t* restrict a8 = a[8];
218 if XNN_UNPREDICTABLE(a8 != zero) {
219 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
220 } else {
221 a8 = zero_data;
222 }
223 const int8_t* restrict a9 = a[9];
224 if XNN_UNPREDICTABLE(a9 != zero) {
225 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
226 } else {
227 a9 = zero_data;
228 }
229 const int8_t* restrict a10 = a[10];
230 if XNN_UNPREDICTABLE(a10 != zero) {
231 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
232 } else {
233 a10 = zero_data;
234 }
235 const int8_t* restrict a11 = a[11];
236 if XNN_UNPREDICTABLE(a11 != zero) {
237 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
238 } else {
239 a11 = zero_data;
240 }
241 const int8_t* restrict a12 = a[12];
242 if XNN_UNPREDICTABLE(a12 != zero) {
243 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
244 } else {
245 a12 = zero_data;
246 }
247 const int8_t* restrict a13 = a[13];
248 if XNN_UNPREDICTABLE(a13 != zero) {
249 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
250 } else {
251 a13 = zero_data;
252 }
253 const int8_t* restrict a14 = a[14];
254 if XNN_UNPREDICTABLE(a14 != zero) {
255 a14 = (const int8_t*) ((uintptr_t) a14 + a_offset);
256 } else {
257 a14 = zero_data;
258 }
259 const int8_t* restrict a15 = a[15];
260 if XNN_UNPREDICTABLE(a15 != zero) {
261 a15 = (const int8_t*) ((uintptr_t) a15 + a_offset);
262 } else {
263 a15 = zero_data;
264 }
265 a += 16;
266
267 size_t k = kc;
268 if (mr == 1)
269 {
270 while (k >= 64 * sizeof(int8_t)) {
271 _tile_loadd(4, a0, 64); // Directly load input for mr=1
272 a15 += 64;
273 _tile_loadd(5, (const int8_t*) w + 0, 64);
274 _tile_dpbssd(0, 4, 5);
275 xnn_prefetch_to_l1((const int8_t*) w + 4096);
276 xnn_prefetch_to_l1((const int8_t*) w + 4160);
277 xnn_prefetch_to_l1((const int8_t*) w + 4224);
278 xnn_prefetch_to_l1((const int8_t*) w + 4288);
279 xnn_prefetch_to_l1((const int8_t*) w + 4352);
280 xnn_prefetch_to_l1((const int8_t*) w + 4416);
281 xnn_prefetch_to_l1((const int8_t*) w + 4480);
282 xnn_prefetch_to_l1((const int8_t*) w + 4544);
283 xnn_prefetch_to_l1((const int8_t*) w + 4608);
284 xnn_prefetch_to_l1((const int8_t*) w + 4672);
285 xnn_prefetch_to_l1((const int8_t*) w + 4736);
286 xnn_prefetch_to_l1((const int8_t*) w + 4800);
287 xnn_prefetch_to_l1((const int8_t*) w + 4864);
288 xnn_prefetch_to_l1((const int8_t*) w + 4928);
289 xnn_prefetch_to_l1((const int8_t*) w + 4992);
290 xnn_prefetch_to_l1((const int8_t*) w + 5056);
291 xnn_prefetch_to_l1((const int8_t*) w + 5120);
292 xnn_prefetch_to_l1((const int8_t*) w + 5184);
293 xnn_prefetch_to_l1((const int8_t*) w + 5248);
294 xnn_prefetch_to_l1((const int8_t*) w + 5312);
295 xnn_prefetch_to_l1((const int8_t*) w + 5376);
296 xnn_prefetch_to_l1((const int8_t*) w + 5440);
297 xnn_prefetch_to_l1((const int8_t*) w + 5504);
298 xnn_prefetch_to_l1((const int8_t*) w + 5568);
299 xnn_prefetch_to_l1((const int8_t*) w + 5632);
300 xnn_prefetch_to_l1((const int8_t*) w + 5696);
301 xnn_prefetch_to_l1((const int8_t*) w + 5760);
302 xnn_prefetch_to_l1((const int8_t*) w + 5824);
303 xnn_prefetch_to_l1((const int8_t*) w + 5888);
304 xnn_prefetch_to_l1((const int8_t*) w + 5952);
305 xnn_prefetch_to_l1((const int8_t*) w + 6016);
306 xnn_prefetch_to_l1((const int8_t*) w + 6080);
307
308 w = (const int8_t*) w + 1024;
309 k -= 64 * sizeof(int8_t);
310 }
311 }
312 else {
313 while (k >= 64 * sizeof(int8_t)) {
314 const __m512i vin0 = _mm512_loadu_epi32(a0);
315 a0 += 64;
316 _mm512_store_epi32(vintile + 0, vin0);
317 const __m512i vin1 = _mm512_loadu_epi32(a1);
318 a1 += 64;
319 _mm512_store_epi32(vintile + 16, vin1);
320 const __m512i vin2 = _mm512_loadu_epi32(a2);
321 a2 += 64;
322 _mm512_store_epi32(vintile + 32, vin2);
323 const __m512i vin3 = _mm512_loadu_epi32(a3);
324 a3 += 64;
325 _mm512_store_epi32(vintile + 48, vin3);
326 const __m512i vin4 = _mm512_loadu_epi32(a4);
327 a4 += 64;
328 _mm512_store_epi32(vintile + 64, vin4);
329 const __m512i vin5 = _mm512_loadu_epi32(a5);
330 a5 += 64;
331 _mm512_store_epi32(vintile + 80, vin5);
332 const __m512i vin6 = _mm512_loadu_epi32(a6);
333 a6 += 64;
334 _mm512_store_epi32(vintile + 96, vin6);
335 const __m512i vin7 = _mm512_loadu_epi32(a7);
336 a7 += 64;
337 _mm512_store_epi32(vintile + 112, vin7);
338 const __m512i vin8 = _mm512_loadu_epi32(a8);
339 a8 += 64;
340 _mm512_store_epi32(vintile + 128, vin8);
341 const __m512i vin9 = _mm512_loadu_epi32(a9);
342 a9 += 64;
343 _mm512_store_epi32(vintile + 144, vin9);
344 const __m512i vin10 = _mm512_loadu_epi32(a10);
345 a10 += 64;
346 _mm512_store_epi32(vintile + 160, vin10);
347 const __m512i vin11 = _mm512_loadu_epi32(a11);
348 a11 += 64;
349 _mm512_store_epi32(vintile + 176, vin11);
350 const __m512i vin12 = _mm512_loadu_epi32(a12);
351 a12 += 64;
352 _mm512_store_epi32(vintile + 192, vin12);
353 const __m512i vin13 = _mm512_loadu_epi32(a13);
354 a13 += 64;
355 _mm512_store_epi32(vintile + 208, vin13);
356 const __m512i vin14 = _mm512_loadu_epi32(a14);
357 a14 += 64;
358 _mm512_store_epi32(vintile + 224, vin14);
359 const __m512i vin15 = _mm512_loadu_epi32(a15);
360 a15 += 64;
361 _mm512_store_epi32(vintile + 240, vin15);
362 _tile_loadd(4, vintile, 64);
363 _tile_loadd(5, (const int8_t*) w + 0, 64);
364 _tile_dpbssd(0, 4, 5);
365 xnn_prefetch_to_l1((const int8_t*) w + 4096);
366 xnn_prefetch_to_l1((const int8_t*) w + 4160);
367 xnn_prefetch_to_l1((const int8_t*) w + 4224);
368 xnn_prefetch_to_l1((const int8_t*) w + 4288);
369 xnn_prefetch_to_l1((const int8_t*) w + 4352);
370 xnn_prefetch_to_l1((const int8_t*) w + 4416);
371 xnn_prefetch_to_l1((const int8_t*) w + 4480);
372 xnn_prefetch_to_l1((const int8_t*) w + 4544);
373 xnn_prefetch_to_l1((const int8_t*) w + 4608);
374 xnn_prefetch_to_l1((const int8_t*) w + 4672);
375 xnn_prefetch_to_l1((const int8_t*) w + 4736);
376 xnn_prefetch_to_l1((const int8_t*) w + 4800);
377 xnn_prefetch_to_l1((const int8_t*) w + 4864);
378 xnn_prefetch_to_l1((const int8_t*) w + 4928);
379 xnn_prefetch_to_l1((const int8_t*) w + 4992);
380 xnn_prefetch_to_l1((const int8_t*) w + 5056);
381 xnn_prefetch_to_l1((const int8_t*) w + 5120);
382 xnn_prefetch_to_l1((const int8_t*) w + 5184);
383 xnn_prefetch_to_l1((const int8_t*) w + 5248);
384 xnn_prefetch_to_l1((const int8_t*) w + 5312);
385 xnn_prefetch_to_l1((const int8_t*) w + 5376);
386 xnn_prefetch_to_l1((const int8_t*) w + 5440);
387 xnn_prefetch_to_l1((const int8_t*) w + 5504);
388 xnn_prefetch_to_l1((const int8_t*) w + 5568);
389 xnn_prefetch_to_l1((const int8_t*) w + 5632);
390 xnn_prefetch_to_l1((const int8_t*) w + 5696);
391 xnn_prefetch_to_l1((const int8_t*) w + 5760);
392 xnn_prefetch_to_l1((const int8_t*) w + 5824);
393 xnn_prefetch_to_l1((const int8_t*) w + 5888);
394 xnn_prefetch_to_l1((const int8_t*) w + 5952);
395 xnn_prefetch_to_l1((const int8_t*) w + 6016);
396 xnn_prefetch_to_l1((const int8_t*) w + 6080);
397
398 w = (const int8_t*) w + 1024;
399 k -= 64 * sizeof(int8_t);
400 }
401 }
402
403 if XNN_UNLIKELY(k != 0) {
404 const __m512i vin0 = _mm512_maskz_loadu_epi32(kremainder_mask, a0);
405 a0 += kremainder;
406 _mm512_store_epi32(vintile + 0, vin0);
407 const __m512i vin1 = _mm512_maskz_loadu_epi32(kremainder_mask, a1);
408 a1 += kremainder;
409 _mm512_store_epi32(vintile + 16, vin1);
410 const __m512i vin2 = _mm512_maskz_loadu_epi32(kremainder_mask, a2);
411 a2 += kremainder;
412 _mm512_store_epi32(vintile + 32, vin2);
413 const __m512i vin3 = _mm512_maskz_loadu_epi32(kremainder_mask, a3);
414 a3 += kremainder;
415 _mm512_store_epi32(vintile + 48, vin3);
416 const __m512i vin4 = _mm512_maskz_loadu_epi32(kremainder_mask, a4);
417 a4 += kremainder;
418 _mm512_store_epi32(vintile + 64, vin4);
419 const __m512i vin5 = _mm512_maskz_loadu_epi32(kremainder_mask, a5);
420 a5 += kremainder;
421 _mm512_store_epi32(vintile + 80, vin5);
422 const __m512i vin6 = _mm512_maskz_loadu_epi32(kremainder_mask, a6);
423 a6 += kremainder;
424 _mm512_store_epi32(vintile + 96, vin6);
425 const __m512i vin7 = _mm512_maskz_loadu_epi32(kremainder_mask, a7);
426 a7 += kremainder;
427 _mm512_store_epi32(vintile + 112, vin7);
428 const __m512i vin8 = _mm512_maskz_loadu_epi32(kremainder_mask, a8);
429 a8 += kremainder;
430 _mm512_store_epi32(vintile + 128, vin8);
431 const __m512i vin9 = _mm512_maskz_loadu_epi32(kremainder_mask, a9);
432 a9 += kremainder;
433 _mm512_store_epi32(vintile + 144, vin9);
434 const __m512i vin10 = _mm512_maskz_loadu_epi32(kremainder_mask, a10);
435 a10 += kremainder;
436 _mm512_store_epi32(vintile + 160, vin10);
437 const __m512i vin11 = _mm512_maskz_loadu_epi32(kremainder_mask, a11);
438 a11 += kremainder;
439 _mm512_store_epi32(vintile + 176, vin11);
440 const __m512i vin12 = _mm512_maskz_loadu_epi32(kremainder_mask, a12);
441 a12 += kremainder;
442 _mm512_store_epi32(vintile + 192, vin12);
443 const __m512i vin13 = _mm512_maskz_loadu_epi32(kremainder_mask, a13);
444 a13 += kremainder;
445 _mm512_store_epi32(vintile + 208, vin13);
446 const __m512i vin14 = _mm512_maskz_loadu_epi32(kremainder_mask, a14);
447 a14 += kremainder;
448 _mm512_store_epi32(vintile + 224, vin14);
449 const __m512i vin15 = _mm512_maskz_loadu_epi32(kremainder_mask, a15);
450 a15 += kremainder;
451 _mm512_store_epi32(vintile + 240, vin15);
452 _tile_loadd(6, vintile, 64);
453 _tile_loadd(7, (const int8_t*) w + 0, 64);
454 _tile_dpbssd(0, 6, 7);
455
456 w = (const int8_t*) w + kremainder * 16;
457 k -= kremainder * sizeof(int8_t);
458 }
459
460 p -= 16 * sizeof(void*);
461 } while (p != 0);
462
463 // Add tile to bias
464 _tile_stored(0, res0, 64);
465
466 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
467 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
468 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
469 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
470 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
471 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
472 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
473 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
474 __m512i vacc8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
475 __m512i vacc9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
476 __m512i vacc10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
477 __m512i vacc11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
478 __m512i vacc12x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
479 __m512i vacc13x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
480 __m512i vacc14x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
481 __m512i vacc15x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
482 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
483 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
484 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
485 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
486 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
487 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
488 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
489 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, _mm512_load_epi32(res0 + 112));
490 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, _mm512_load_epi32(res0 + 128));
491 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, _mm512_load_epi32(res0 + 144));
492 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, _mm512_load_epi32(res0 + 160));
493 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, _mm512_load_epi32(res0 + 176));
494 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, _mm512_load_epi32(res0 + 192));
495 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, _mm512_load_epi32(res0 + 208));
496 vacc14x0123456789ABCDEF = _mm512_add_epi32(vacc14x0123456789ABCDEF, _mm512_load_epi32(res0 + 224));
497 vacc15x0123456789ABCDEF = _mm512_add_epi32(vacc15x0123456789ABCDEF, _mm512_load_epi32(res0 + 240));
498
499 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
500 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
501 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
502 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
503 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
504 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
505 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
506 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
507 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
508 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
509 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
510 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
511 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
512 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
513 __m512 vscaled14x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc14x0123456789ABCDEF);
514 __m512 vscaled15x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc15x0123456789ABCDEF);
515
516 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
517 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
518 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
519 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
520 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
521 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
522 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
523 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
524 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
525 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
526 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
527 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
528 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
529 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
530 vscaled14x0123456789ABCDEF = _mm512_mul_ps(vscaled14x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
531 vscaled15x0123456789ABCDEF = _mm512_mul_ps(vscaled15x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
532
533 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
534 w = (const int32_t*) w + 16;
535 const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
536 w = (const int32_t*) w + 16;
537
538 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
539 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
540 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
541 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
542 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
543 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
544 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
545 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
546 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
547 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
548 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
549 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
550 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
551 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
552 vscaled14x0123456789ABCDEF = _mm512_fmadd_ps(vscaled14x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
553 vscaled15x0123456789ABCDEF = _mm512_fmadd_ps(vscaled15x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
554
555 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
556 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
557 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
558 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
559 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
560 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
561 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
562 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
563 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
564 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
565 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
566 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
567 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
568 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
569 vscaled14x0123456789ABCDEF = _mm512_max_ps(vscaled14x0123456789ABCDEF, voutput_min);
570 vscaled15x0123456789ABCDEF = _mm512_max_ps(vscaled15x0123456789ABCDEF, voutput_min);
571
572 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
573 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
574 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
575 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
576 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
577 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
578 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
579 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
580 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
581 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
582 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
583 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
584 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
585 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
586 vscaled14x0123456789ABCDEF = _mm512_min_ps(vscaled14x0123456789ABCDEF, voutput_max);
587 vscaled15x0123456789ABCDEF = _mm512_min_ps(vscaled15x0123456789ABCDEF, voutput_max);
588
589 if XNN_LIKELY(nc >= 16) {
590 _mm512_storeu_ps(c15 + 0, vscaled15x0123456789ABCDEF);
591 c15 = (float*) ((uintptr_t) c15 + cn_stride);
592 _mm512_storeu_ps(c14 + 0, vscaled14x0123456789ABCDEF);
593 c14 = (float*) ((uintptr_t) c14 + cn_stride);
594 _mm512_storeu_ps(c13 + 0, vscaled13x0123456789ABCDEF);
595 c13 = (float*) ((uintptr_t) c13 + cn_stride);
596 _mm512_storeu_ps(c12 + 0, vscaled12x0123456789ABCDEF);
597 c12 = (float*) ((uintptr_t) c12 + cn_stride);
598 _mm512_storeu_ps(c11 + 0, vscaled11x0123456789ABCDEF);
599 c11 = (float*) ((uintptr_t) c11 + cn_stride);
600 _mm512_storeu_ps(c10 + 0, vscaled10x0123456789ABCDEF);
601 c10 = (float*) ((uintptr_t) c10 + cn_stride);
602 _mm512_storeu_ps(c9 + 0, vscaled9x0123456789ABCDEF);
603 c9 = (float*) ((uintptr_t) c9 + cn_stride);
604 _mm512_storeu_ps(c8 + 0, vscaled8x0123456789ABCDEF);
605 c8 = (float*) ((uintptr_t) c8 + cn_stride);
606 _mm512_storeu_ps(c7 + 0, vscaled7x0123456789ABCDEF);
607 c7 = (float*) ((uintptr_t) c7 + cn_stride);
608 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
609 c6 = (float*) ((uintptr_t) c6 + cn_stride);
610 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
611 c5 = (float*) ((uintptr_t) c5 + cn_stride);
612 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
613 c4 = (float*) ((uintptr_t) c4 + cn_stride);
614 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
615 c3 = (float*) ((uintptr_t) c3 + cn_stride);
616 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
617 c2 = (float*) ((uintptr_t) c2 + cn_stride);
618 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
619 c1 = (float*) ((uintptr_t) c1 + cn_stride);
620 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
621 c0 = (float*) ((uintptr_t) c0 + cn_stride);
622
623 a = (const int8_t**restrict) ((uintptr_t) a - ks);
624 nc -= 16;
625 } else {
626 // Prepare mask for valid 32-bit elements (depends on nc).
627 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
628 _mm512_mask_storeu_ps(c15 + 0, vmask0, vscaled15x0123456789ABCDEF);
629 _mm512_mask_storeu_ps(c14 + 0, vmask0, vscaled14x0123456789ABCDEF);
630 _mm512_mask_storeu_ps(c13 + 0, vmask0, vscaled13x0123456789ABCDEF);
631 _mm512_mask_storeu_ps(c12 + 0, vmask0, vscaled12x0123456789ABCDEF);
632 _mm512_mask_storeu_ps(c11 + 0, vmask0, vscaled11x0123456789ABCDEF);
633 _mm512_mask_storeu_ps(c10 + 0, vmask0, vscaled10x0123456789ABCDEF);
634 _mm512_mask_storeu_ps(c9 + 0, vmask0, vscaled9x0123456789ABCDEF);
635 _mm512_mask_storeu_ps(c8 + 0, vmask0, vscaled8x0123456789ABCDEF);
636 _mm512_mask_storeu_ps(c7 + 0, vmask0, vscaled7x0123456789ABCDEF);
637 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
638 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
639 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
640 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
641 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
642 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
643 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
644 nc = 0;
645 }
646 } while (nc != 0);
647
648 // Release tile config
649 // _tile_release();
650 __asm__ volatile ("tilerelease" ::);
651 #endif // defined(__x86_64__)
652 }