Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-16x16c4-minmax-avx512amx.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
19
20 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x16c4__avx512amx(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)])
35 {
36 assert(mr != 0);
37 assert(mr <= 16);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 // TODO: amxintrin.h only provide intrinsics for __x86_64__
46 // Update if amxintrin changes
47 #if defined(__x86_64__)
48 __attribute__((aligned(64))) int32_t vintile[16 * 16];
49 __attribute__((aligned(64))) int32_t res0[16 * 16];
50
51 kc = round_up_po2(kc, 4 * sizeof(int8_t));
52 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
53 const __mmask16 kremainder_mask = _cvtu32_mask16((UINT32_C(1) << (kremainder >> 2)) - 1);
54
55 // Define tile config data structure
56 struct __tile_config {
57 uint8_t palette_id;
58 uint8_t start_row;
59 uint8_t reserved_0[14];
60 uint16_t colsb[8];
61 uint16_t reserved_1[8];
62 uint8_t rows[8];
63 uint8_t reserved_2[8];
64 };
65
66 // Load tile configuration
67 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
68 tile_data.palette_id = 1;
69 tile_data.rows[0] = mr; // tmm0 = res 0
70 tile_data.rows[1] = mr; // tmm1 = res 1
71 tile_data.rows[2] = mr; // tmm2 = res 2
72 tile_data.rows[3] = mr; // tmm3 = res 3
73 tile_data.rows[4] = mr; // tmm4 = input
74 tile_data.rows[5] = 16; // tmm5 = weights
75 tile_data.rows[6] = mr; // tmm6 = input remainder
76 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
77
78 tile_data.colsb[0] = 64; // tmm0 = res 0
79 tile_data.colsb[1] = 64; // tmm1 = res 1
80 tile_data.colsb[2] = 64; // tmm2 = res 1
81 tile_data.colsb[3] = 64; // tmm3 = res 1
82 tile_data.colsb[4] = 64; // tmm4 = input
83 tile_data.colsb[5] = 64; // tmm5 = weights
84 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
85 tile_data.colsb[7] = 64; // tmm7 = weights remainder
86
87 //_tile_loadconfig(&tile_data);
88 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
89
90 float* c0 = c;
91 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
92 if XNN_UNPREDICTABLE(mr < 2) {
93 c1 = c0;
94 }
95 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
96 if XNN_UNPREDICTABLE(mr <= 2) {
97 c2 = c1;
98 }
99 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
100 if XNN_UNPREDICTABLE(mr < 4) {
101 c3 = c2;
102 }
103 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
104 if XNN_UNPREDICTABLE(mr <= 4) {
105 c4 = c3;
106 }
107 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
108 if XNN_UNPREDICTABLE(mr < 6) {
109 c5 = c4;
110 }
111 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
112 if XNN_UNPREDICTABLE(mr <= 6) {
113 c6 = c5;
114 }
115 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
116 if XNN_UNPREDICTABLE(mr < 8) {
117 c7 = c6;
118 }
119 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
120 if XNN_UNPREDICTABLE(mr <= 8) {
121 c8 = c7;
122 }
123 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
124 if XNN_UNPREDICTABLE(mr < 10) {
125 c9 = c8;
126 }
127 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
128 if XNN_UNPREDICTABLE(mr <= 10) {
129 c10 = c9;
130 }
131 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
132 if XNN_UNPREDICTABLE(mr < 12) {
133 c11 = c10;
134 }
135 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
136 if XNN_UNPREDICTABLE(mr <= 12) {
137 c12 = c11;
138 }
139 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
140 if XNN_UNPREDICTABLE(mr < 14) {
141 c13 = c12;
142 }
143 float* c14 = (float*) ((uintptr_t) c13 + cm_stride);
144 if XNN_UNPREDICTABLE(mr <= 14) {
145 c14 = c13;
146 }
147 float* c15 = (float*) ((uintptr_t) c14 + cm_stride);
148 if XNN_UNPREDICTABLE(mr != 16) {
149 c15 = c14;
150 }
151
152 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
153 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
154 // XNN_FORCE_REALIZATION(voutput_min);
155 // XNN_FORCE_REALIZATION(voutput_max);
156
157 do {
158 const __m512i vksum0123456789ABCDEF = _mm512_loadu_epi32((const int32_t*) w + 0);
159 w = (const int32_t*) w + 16;
160
161 // Zero tile accumulator
162 __asm__ volatile (
163 "tilezero %%tmm0\n"
164 ::);
165
166 size_t p = ks;
167 do {
168 const int8_t* restrict a0 = a[0];
169 if XNN_UNPREDICTABLE(a0 != zero) {
170 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
171 } else {
172 a0 = zero_data;
173 }
174 const int8_t* restrict a1 = a[1];
175 if XNN_UNPREDICTABLE(a1 != zero) {
176 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
177 } else {
178 a1 = zero_data;
179 }
180 const int8_t* restrict a2 = a[2];
181 if XNN_UNPREDICTABLE(a2 != zero) {
182 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
183 } else {
184 a2 = zero_data;
185 }
186 const int8_t* restrict a3 = a[3];
187 if XNN_UNPREDICTABLE(a3 != zero) {
188 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
189 } else {
190 a3 = zero_data;
191 }
192 const int8_t* restrict a4 = a[4];
193 if XNN_UNPREDICTABLE(a4 != zero) {
194 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
195 } else {
196 a4 = zero_data;
197 }
198 const int8_t* restrict a5 = a[5];
199 if XNN_UNPREDICTABLE(a5 != zero) {
200 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
201 } else {
202 a5 = zero_data;
203 }
204 const int8_t* restrict a6 = a[6];
205 if XNN_UNPREDICTABLE(a6 != zero) {
206 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
207 } else {
208 a6 = zero_data;
209 }
210 const int8_t* restrict a7 = a[7];
211 if XNN_UNPREDICTABLE(a7 != zero) {
212 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
213 } else {
214 a7 = zero_data;
215 }
216 const int8_t* restrict a8 = a[8];
217 if XNN_UNPREDICTABLE(a8 != zero) {
218 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
219 } else {
220 a8 = zero_data;
221 }
222 const int8_t* restrict a9 = a[9];
223 if XNN_UNPREDICTABLE(a9 != zero) {
224 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
225 } else {
226 a9 = zero_data;
227 }
228 const int8_t* restrict a10 = a[10];
229 if XNN_UNPREDICTABLE(a10 != zero) {
230 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
231 } else {
232 a10 = zero_data;
233 }
234 const int8_t* restrict a11 = a[11];
235 if XNN_UNPREDICTABLE(a11 != zero) {
236 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
237 } else {
238 a11 = zero_data;
239 }
240 const int8_t* restrict a12 = a[12];
241 if XNN_UNPREDICTABLE(a12 != zero) {
242 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
243 } else {
244 a12 = zero_data;
245 }
246 const int8_t* restrict a13 = a[13];
247 if XNN_UNPREDICTABLE(a13 != zero) {
248 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
249 } else {
250 a13 = zero_data;
251 }
252 const int8_t* restrict a14 = a[14];
253 if XNN_UNPREDICTABLE(a14 != zero) {
254 a14 = (const int8_t*) ((uintptr_t) a14 + a_offset);
255 } else {
256 a14 = zero_data;
257 }
258 const int8_t* restrict a15 = a[15];
259 if XNN_UNPREDICTABLE(a15 != zero) {
260 a15 = (const int8_t*) ((uintptr_t) a15 + a_offset);
261 } else {
262 a15 = zero_data;
263 }
264 a += 16;
265
266 size_t k = kc;
267 if (mr == 1)
268 {
269 while (k >= 64 * sizeof(int8_t)) {
270 _tile_loadd(4, a0, 64); // Directly load input for mr=1
271 a15 += 64;
272 _tile_loadd(5, (const int8_t*) w + 0, 64);
273 _tile_dpbssd(0, 4, 5);
274
275 w = (const int8_t*) w + 1024;
276 k -= 64 * sizeof(int8_t);
277 }
278 }
279 else {
280 while (k >= 64 * sizeof(int8_t)) {
281 const __m512i vin0 = _mm512_loadu_epi32(a0);
282 a0 += 64;
283 _mm512_store_epi32(vintile + 0, vin0);
284 const __m512i vin1 = _mm512_loadu_epi32(a1);
285 a1 += 64;
286 _mm512_store_epi32(vintile + 16, vin1);
287 const __m512i vin2 = _mm512_loadu_epi32(a2);
288 a2 += 64;
289 _mm512_store_epi32(vintile + 32, vin2);
290 const __m512i vin3 = _mm512_loadu_epi32(a3);
291 a3 += 64;
292 _mm512_store_epi32(vintile + 48, vin3);
293 const __m512i vin4 = _mm512_loadu_epi32(a4);
294 a4 += 64;
295 _mm512_store_epi32(vintile + 64, vin4);
296 const __m512i vin5 = _mm512_loadu_epi32(a5);
297 a5 += 64;
298 _mm512_store_epi32(vintile + 80, vin5);
299 const __m512i vin6 = _mm512_loadu_epi32(a6);
300 a6 += 64;
301 _mm512_store_epi32(vintile + 96, vin6);
302 const __m512i vin7 = _mm512_loadu_epi32(a7);
303 a7 += 64;
304 _mm512_store_epi32(vintile + 112, vin7);
305 const __m512i vin8 = _mm512_loadu_epi32(a8);
306 a8 += 64;
307 _mm512_store_epi32(vintile + 128, vin8);
308 const __m512i vin9 = _mm512_loadu_epi32(a9);
309 a9 += 64;
310 _mm512_store_epi32(vintile + 144, vin9);
311 const __m512i vin10 = _mm512_loadu_epi32(a10);
312 a10 += 64;
313 _mm512_store_epi32(vintile + 160, vin10);
314 const __m512i vin11 = _mm512_loadu_epi32(a11);
315 a11 += 64;
316 _mm512_store_epi32(vintile + 176, vin11);
317 const __m512i vin12 = _mm512_loadu_epi32(a12);
318 a12 += 64;
319 _mm512_store_epi32(vintile + 192, vin12);
320 const __m512i vin13 = _mm512_loadu_epi32(a13);
321 a13 += 64;
322 _mm512_store_epi32(vintile + 208, vin13);
323 const __m512i vin14 = _mm512_loadu_epi32(a14);
324 a14 += 64;
325 _mm512_store_epi32(vintile + 224, vin14);
326 const __m512i vin15 = _mm512_loadu_epi32(a15);
327 a15 += 64;
328 _mm512_store_epi32(vintile + 240, vin15);
329 _tile_loadd(4, vintile, 64);
330 _tile_loadd(5, (const int8_t*) w + 0, 64);
331 _tile_dpbssd(0, 4, 5);
332
333 w = (const int8_t*) w + 1024;
334 k -= 64 * sizeof(int8_t);
335 }
336 }
337
338 if XNN_UNLIKELY(k != 0) {
339 const __m512i vin0 = _mm512_maskz_loadu_epi32(kremainder_mask, a0);
340 a0 += kremainder;
341 _mm512_store_epi32(vintile + 0, vin0);
342 const __m512i vin1 = _mm512_maskz_loadu_epi32(kremainder_mask, a1);
343 a1 += kremainder;
344 _mm512_store_epi32(vintile + 16, vin1);
345 const __m512i vin2 = _mm512_maskz_loadu_epi32(kremainder_mask, a2);
346 a2 += kremainder;
347 _mm512_store_epi32(vintile + 32, vin2);
348 const __m512i vin3 = _mm512_maskz_loadu_epi32(kremainder_mask, a3);
349 a3 += kremainder;
350 _mm512_store_epi32(vintile + 48, vin3);
351 const __m512i vin4 = _mm512_maskz_loadu_epi32(kremainder_mask, a4);
352 a4 += kremainder;
353 _mm512_store_epi32(vintile + 64, vin4);
354 const __m512i vin5 = _mm512_maskz_loadu_epi32(kremainder_mask, a5);
355 a5 += kremainder;
356 _mm512_store_epi32(vintile + 80, vin5);
357 const __m512i vin6 = _mm512_maskz_loadu_epi32(kremainder_mask, a6);
358 a6 += kremainder;
359 _mm512_store_epi32(vintile + 96, vin6);
360 const __m512i vin7 = _mm512_maskz_loadu_epi32(kremainder_mask, a7);
361 a7 += kremainder;
362 _mm512_store_epi32(vintile + 112, vin7);
363 const __m512i vin8 = _mm512_maskz_loadu_epi32(kremainder_mask, a8);
364 a8 += kremainder;
365 _mm512_store_epi32(vintile + 128, vin8);
366 const __m512i vin9 = _mm512_maskz_loadu_epi32(kremainder_mask, a9);
367 a9 += kremainder;
368 _mm512_store_epi32(vintile + 144, vin9);
369 const __m512i vin10 = _mm512_maskz_loadu_epi32(kremainder_mask, a10);
370 a10 += kremainder;
371 _mm512_store_epi32(vintile + 160, vin10);
372 const __m512i vin11 = _mm512_maskz_loadu_epi32(kremainder_mask, a11);
373 a11 += kremainder;
374 _mm512_store_epi32(vintile + 176, vin11);
375 const __m512i vin12 = _mm512_maskz_loadu_epi32(kremainder_mask, a12);
376 a12 += kremainder;
377 _mm512_store_epi32(vintile + 192, vin12);
378 const __m512i vin13 = _mm512_maskz_loadu_epi32(kremainder_mask, a13);
379 a13 += kremainder;
380 _mm512_store_epi32(vintile + 208, vin13);
381 const __m512i vin14 = _mm512_maskz_loadu_epi32(kremainder_mask, a14);
382 a14 += kremainder;
383 _mm512_store_epi32(vintile + 224, vin14);
384 const __m512i vin15 = _mm512_maskz_loadu_epi32(kremainder_mask, a15);
385 a15 += kremainder;
386 _mm512_store_epi32(vintile + 240, vin15);
387 _tile_loadd(6, vintile, 64);
388 _tile_loadd(7, (const int8_t*) w + 0, 64);
389 _tile_dpbssd(0, 6, 7);
390
391 w = (const int8_t*) w + kremainder * 16;
392 k -= kremainder * sizeof(int8_t);
393 }
394
395 p -= 16 * sizeof(void*);
396 } while (p != 0);
397
398 // Add tile to bias
399 _tile_stored(0, res0, 64);
400
401 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
402 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
403 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
404 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
405 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
406 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
407 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
408 __m512i vacc7x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
409 __m512i vacc8x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
410 __m512i vacc9x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
411 __m512i vacc10x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
412 __m512i vacc11x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
413 __m512i vacc12x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
414 __m512i vacc13x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
415 __m512i vacc14x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
416 __m512i vacc15x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, _mm512_set1_epi32((int) quantization_params->zero_point));
417 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
418 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
419 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
420 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
421 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
422 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
423 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
424 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, _mm512_load_epi32(res0 + 112));
425 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, _mm512_load_epi32(res0 + 128));
426 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, _mm512_load_epi32(res0 + 144));
427 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, _mm512_load_epi32(res0 + 160));
428 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, _mm512_load_epi32(res0 + 176));
429 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, _mm512_load_epi32(res0 + 192));
430 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, _mm512_load_epi32(res0 + 208));
431 vacc14x0123456789ABCDEF = _mm512_add_epi32(vacc14x0123456789ABCDEF, _mm512_load_epi32(res0 + 224));
432 vacc15x0123456789ABCDEF = _mm512_add_epi32(vacc15x0123456789ABCDEF, _mm512_load_epi32(res0 + 240));
433
434 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
435 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
436 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
437 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
438 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
439 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
440 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
441 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
442 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
443 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
444 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
445 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
446 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
447 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
448 __m512 vscaled14x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc14x0123456789ABCDEF);
449 __m512 vscaled15x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc15x0123456789ABCDEF);
450
451 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
452 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
453 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
454 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
455 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
456 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
457 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
458 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
459 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
460 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
461 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
462 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
463 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
464 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
465 vscaled14x0123456789ABCDEF = _mm512_mul_ps(vscaled14x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
466 vscaled15x0123456789ABCDEF = _mm512_mul_ps(vscaled15x0123456789ABCDEF, _mm512_set1_ps(quantization_params->inv_scale));
467
468 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
469 w = (const int32_t*) w + 16;
470 const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
471 w = (const int32_t*) w + 16;
472
473 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
474 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
475 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
476 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
477 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
478 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
479 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
480 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
481 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
482 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
483 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
484 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
485 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
486 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
487 vscaled14x0123456789ABCDEF = _mm512_fmadd_ps(vscaled14x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
488 vscaled15x0123456789ABCDEF = _mm512_fmadd_ps(vscaled15x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
489
490 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
491 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
492 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
493 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
494 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
495 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
496 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
497 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
498 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
499 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
500 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
501 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
502 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
503 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
504 vscaled14x0123456789ABCDEF = _mm512_max_ps(vscaled14x0123456789ABCDEF, voutput_min);
505 vscaled15x0123456789ABCDEF = _mm512_max_ps(vscaled15x0123456789ABCDEF, voutput_min);
506
507 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
508 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
509 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
510 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
511 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
512 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
513 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
514 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
515 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
516 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
517 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
518 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
519 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
520 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
521 vscaled14x0123456789ABCDEF = _mm512_min_ps(vscaled14x0123456789ABCDEF, voutput_max);
522 vscaled15x0123456789ABCDEF = _mm512_min_ps(vscaled15x0123456789ABCDEF, voutput_max);
523
524 if XNN_LIKELY(nc >= 16) {
525 _mm512_storeu_ps(c15 + 0, vscaled15x0123456789ABCDEF);
526 c15 = (float*) ((uintptr_t) c15 + cn_stride);
527 _mm512_storeu_ps(c14 + 0, vscaled14x0123456789ABCDEF);
528 c14 = (float*) ((uintptr_t) c14 + cn_stride);
529 _mm512_storeu_ps(c13 + 0, vscaled13x0123456789ABCDEF);
530 c13 = (float*) ((uintptr_t) c13 + cn_stride);
531 _mm512_storeu_ps(c12 + 0, vscaled12x0123456789ABCDEF);
532 c12 = (float*) ((uintptr_t) c12 + cn_stride);
533 _mm512_storeu_ps(c11 + 0, vscaled11x0123456789ABCDEF);
534 c11 = (float*) ((uintptr_t) c11 + cn_stride);
535 _mm512_storeu_ps(c10 + 0, vscaled10x0123456789ABCDEF);
536 c10 = (float*) ((uintptr_t) c10 + cn_stride);
537 _mm512_storeu_ps(c9 + 0, vscaled9x0123456789ABCDEF);
538 c9 = (float*) ((uintptr_t) c9 + cn_stride);
539 _mm512_storeu_ps(c8 + 0, vscaled8x0123456789ABCDEF);
540 c8 = (float*) ((uintptr_t) c8 + cn_stride);
541 _mm512_storeu_ps(c7 + 0, vscaled7x0123456789ABCDEF);
542 c7 = (float*) ((uintptr_t) c7 + cn_stride);
543 _mm512_storeu_ps(c6 + 0, vscaled6x0123456789ABCDEF);
544 c6 = (float*) ((uintptr_t) c6 + cn_stride);
545 _mm512_storeu_ps(c5 + 0, vscaled5x0123456789ABCDEF);
546 c5 = (float*) ((uintptr_t) c5 + cn_stride);
547 _mm512_storeu_ps(c4 + 0, vscaled4x0123456789ABCDEF);
548 c4 = (float*) ((uintptr_t) c4 + cn_stride);
549 _mm512_storeu_ps(c3 + 0, vscaled3x0123456789ABCDEF);
550 c3 = (float*) ((uintptr_t) c3 + cn_stride);
551 _mm512_storeu_ps(c2 + 0, vscaled2x0123456789ABCDEF);
552 c2 = (float*) ((uintptr_t) c2 + cn_stride);
553 _mm512_storeu_ps(c1 + 0, vscaled1x0123456789ABCDEF);
554 c1 = (float*) ((uintptr_t) c1 + cn_stride);
555 _mm512_storeu_ps(c0 + 0, vscaled0x0123456789ABCDEF);
556 c0 = (float*) ((uintptr_t) c0 + cn_stride);
557
558 a = (const int8_t**restrict) ((uintptr_t) a - ks);
559 nc -= 16;
560 } else {
561 // Prepare mask for valid 32-bit elements (depends on nc).
562 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
563 _mm512_mask_storeu_ps(c15 + 0, vmask0, vscaled15x0123456789ABCDEF);
564 _mm512_mask_storeu_ps(c14 + 0, vmask0, vscaled14x0123456789ABCDEF);
565 _mm512_mask_storeu_ps(c13 + 0, vmask0, vscaled13x0123456789ABCDEF);
566 _mm512_mask_storeu_ps(c12 + 0, vmask0, vscaled12x0123456789ABCDEF);
567 _mm512_mask_storeu_ps(c11 + 0, vmask0, vscaled11x0123456789ABCDEF);
568 _mm512_mask_storeu_ps(c10 + 0, vmask0, vscaled10x0123456789ABCDEF);
569 _mm512_mask_storeu_ps(c9 + 0, vmask0, vscaled9x0123456789ABCDEF);
570 _mm512_mask_storeu_ps(c8 + 0, vmask0, vscaled8x0123456789ABCDEF);
571 _mm512_mask_storeu_ps(c7 + 0, vmask0, vscaled7x0123456789ABCDEF);
572 _mm512_mask_storeu_ps(c6 + 0, vmask0, vscaled6x0123456789ABCDEF);
573 _mm512_mask_storeu_ps(c5 + 0, vmask0, vscaled5x0123456789ABCDEF);
574 _mm512_mask_storeu_ps(c4 + 0, vmask0, vscaled4x0123456789ABCDEF);
575 _mm512_mask_storeu_ps(c3 + 0, vmask0, vscaled3x0123456789ABCDEF);
576 _mm512_mask_storeu_ps(c2 + 0, vmask0, vscaled2x0123456789ABCDEF);
577 _mm512_mask_storeu_ps(c1 + 0, vmask0, vscaled1x0123456789ABCDEF);
578 _mm512_mask_storeu_ps(c0 + 0, vmask0, vscaled0x0123456789ABCDEF);
579 nc = 0;
580 }
581 } while (nc != 0);
582
583 // Release tile config
584 // _tile_release();
585 __asm__ volatile ("tilerelease" ::);
586 #endif // defined(__x86_64__)
587 }