Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-16x16c4-minmax-fp32-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_qs8_qc8w_igemm_minmax_fp32_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 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
33 {
34 assert(mr != 0);
35 assert(mr <= 16);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 // TODO: amxintrin.h only provide intrinsics for __x86_64__
44 // Update if amxintrin changes
45 #if defined(__x86_64__)
46 __attribute__((aligned(64))) int32_t vintile[16 * 16];
47 __attribute__((aligned(64))) int32_t res0[16 * 16];
48
49 kc = round_up_po2(kc, 4 * sizeof(int8_t));
50 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
51 const __mmask16 kremainder_mask = _cvtu32_mask16((UINT32_C(1) << (kremainder >> 2)) - 1);
52
53 // Define tile config data structure
54 struct __tile_config {
55 uint8_t palette_id;
56 uint8_t start_row;
57 uint8_t reserved_0[14];
58 uint16_t colsb[8];
59 uint16_t reserved_1[8];
60 uint8_t rows[8];
61 uint8_t reserved_2[8];
62 };
63
64 // Load tile configuration
65 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
66 tile_data.palette_id = 1;
67 tile_data.rows[0] = mr; // tmm0 = res 0
68 tile_data.rows[1] = mr; // tmm1 = res 1
69 tile_data.rows[2] = mr; // tmm2 = res 2
70 tile_data.rows[3] = mr; // tmm3 = res 3
71 tile_data.rows[4] = mr; // tmm4 = input
72 tile_data.rows[5] = 16; // tmm5 = weights
73 tile_data.rows[6] = mr; // tmm6 = input remainder
74 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
75
76 tile_data.colsb[0] = 64; // tmm0 = res 0
77 tile_data.colsb[1] = 64; // tmm1 = res 1
78 tile_data.colsb[2] = 64; // tmm2 = res 1
79 tile_data.colsb[3] = 64; // tmm3 = res 1
80 tile_data.colsb[4] = 64; // tmm4 = input
81 tile_data.colsb[5] = 64; // tmm5 = weights
82 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
83 tile_data.colsb[7] = 64; // tmm7 = weights remainder
84
85 //_tile_loadconfig(&tile_data);
86 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
87
88 int8_t* c0 = c;
89 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
90 if XNN_UNPREDICTABLE(mr < 2) {
91 c1 = c0;
92 }
93 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
94 if XNN_UNPREDICTABLE(mr <= 2) {
95 c2 = c1;
96 }
97 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
98 if XNN_UNPREDICTABLE(mr < 4) {
99 c3 = c2;
100 }
101 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
102 if XNN_UNPREDICTABLE(mr <= 4) {
103 c4 = c3;
104 }
105 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
106 if XNN_UNPREDICTABLE(mr < 6) {
107 c5 = c4;
108 }
109 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
110 if XNN_UNPREDICTABLE(mr <= 6) {
111 c6 = c5;
112 }
113 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
114 if XNN_UNPREDICTABLE(mr < 8) {
115 c7 = c6;
116 }
117 int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
118 if XNN_UNPREDICTABLE(mr <= 8) {
119 c8 = c7;
120 }
121 int8_t* c9 = (int8_t*) ((uintptr_t) c8 + cm_stride);
122 if XNN_UNPREDICTABLE(mr < 10) {
123 c9 = c8;
124 }
125 int8_t* c10 = (int8_t*) ((uintptr_t) c9 + cm_stride);
126 if XNN_UNPREDICTABLE(mr <= 10) {
127 c10 = c9;
128 }
129 int8_t* c11 = (int8_t*) ((uintptr_t) c10 + cm_stride);
130 if XNN_UNPREDICTABLE(mr < 12) {
131 c11 = c10;
132 }
133 int8_t* c12 = (int8_t*) ((uintptr_t) c11 + cm_stride);
134 if XNN_UNPREDICTABLE(mr <= 12) {
135 c12 = c11;
136 }
137 int8_t* c13 = (int8_t*) ((uintptr_t) c12 + cm_stride);
138 if XNN_UNPREDICTABLE(mr < 14) {
139 c13 = c12;
140 }
141 int8_t* c14 = (int8_t*) ((uintptr_t) c13 + cm_stride);
142 if XNN_UNPREDICTABLE(mr <= 14) {
143 c14 = c13;
144 }
145 int8_t* c15 = (int8_t*) ((uintptr_t) c14 + cm_stride);
146 if XNN_UNPREDICTABLE(mr != 16) {
147 c15 = c14;
148 }
149
150 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);
151 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
152 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
153 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
154 // XNN_FORCE_REALIZATION(voutput_zero_point);
155 // XNN_FORCE_REALIZATION(voutput_min);
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 }
172 const int8_t* restrict a1 = a[1];
173 if XNN_UNPREDICTABLE(a1 != zero) {
174 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
175 }
176 const int8_t* restrict a2 = a[2];
177 if XNN_UNPREDICTABLE(a2 != zero) {
178 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
179 }
180 const int8_t* restrict a3 = a[3];
181 if XNN_UNPREDICTABLE(a3 != zero) {
182 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
183 }
184 const int8_t* restrict a4 = a[4];
185 if XNN_UNPREDICTABLE(a4 != zero) {
186 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
187 }
188 const int8_t* restrict a5 = a[5];
189 if XNN_UNPREDICTABLE(a5 != zero) {
190 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
191 }
192 const int8_t* restrict a6 = a[6];
193 if XNN_UNPREDICTABLE(a6 != zero) {
194 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
195 }
196 const int8_t* restrict a7 = a[7];
197 if XNN_UNPREDICTABLE(a7 != zero) {
198 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
199 }
200 const int8_t* restrict a8 = a[8];
201 if XNN_UNPREDICTABLE(a8 != zero) {
202 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
203 }
204 const int8_t* restrict a9 = a[9];
205 if XNN_UNPREDICTABLE(a9 != zero) {
206 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
207 }
208 const int8_t* restrict a10 = a[10];
209 if XNN_UNPREDICTABLE(a10 != zero) {
210 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
211 }
212 const int8_t* restrict a11 = a[11];
213 if XNN_UNPREDICTABLE(a11 != zero) {
214 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
215 }
216 const int8_t* restrict a12 = a[12];
217 if XNN_UNPREDICTABLE(a12 != zero) {
218 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
219 }
220 const int8_t* restrict a13 = a[13];
221 if XNN_UNPREDICTABLE(a13 != zero) {
222 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
223 }
224 const int8_t* restrict a14 = a[14];
225 if XNN_UNPREDICTABLE(a14 != zero) {
226 a14 = (const int8_t*) ((uintptr_t) a14 + a_offset);
227 }
228 const int8_t* restrict a15 = a[15];
229 if XNN_UNPREDICTABLE(a15 != zero) {
230 a15 = (const int8_t*) ((uintptr_t) a15 + a_offset);
231 }
232 a += 16;
233
234 size_t k = kc;
235 if (mr == 1)
236 {
237 while (k >= 64 * sizeof(int8_t)) {
238 _tile_loadd(4, a0, 64); // Directly load input for mr=1
239 a15 += 64;
240 _tile_loadd(5, (const int8_t*) w + 0, 64);
241 _tile_dpbssd(0, 4, 5);
242
243 w = (const int8_t*) w + 1024;
244 k -= 64 * sizeof(int8_t);
245 }
246 }
247 else {
248 while (k >= 64 * sizeof(int8_t)) {
249 const __m512i vin0 = _mm512_loadu_epi32(a0);
250 a0 += 64;
251 _mm512_store_epi32(vintile + 0, vin0);
252 const __m512i vin1 = _mm512_loadu_epi32(a1);
253 a1 += 64;
254 _mm512_store_epi32(vintile + 16, vin1);
255 const __m512i vin2 = _mm512_loadu_epi32(a2);
256 a2 += 64;
257 _mm512_store_epi32(vintile + 32, vin2);
258 const __m512i vin3 = _mm512_loadu_epi32(a3);
259 a3 += 64;
260 _mm512_store_epi32(vintile + 48, vin3);
261 const __m512i vin4 = _mm512_loadu_epi32(a4);
262 a4 += 64;
263 _mm512_store_epi32(vintile + 64, vin4);
264 const __m512i vin5 = _mm512_loadu_epi32(a5);
265 a5 += 64;
266 _mm512_store_epi32(vintile + 80, vin5);
267 const __m512i vin6 = _mm512_loadu_epi32(a6);
268 a6 += 64;
269 _mm512_store_epi32(vintile + 96, vin6);
270 const __m512i vin7 = _mm512_loadu_epi32(a7);
271 a7 += 64;
272 _mm512_store_epi32(vintile + 112, vin7);
273 const __m512i vin8 = _mm512_loadu_epi32(a8);
274 a8 += 64;
275 _mm512_store_epi32(vintile + 128, vin8);
276 const __m512i vin9 = _mm512_loadu_epi32(a9);
277 a9 += 64;
278 _mm512_store_epi32(vintile + 144, vin9);
279 const __m512i vin10 = _mm512_loadu_epi32(a10);
280 a10 += 64;
281 _mm512_store_epi32(vintile + 160, vin10);
282 const __m512i vin11 = _mm512_loadu_epi32(a11);
283 a11 += 64;
284 _mm512_store_epi32(vintile + 176, vin11);
285 const __m512i vin12 = _mm512_loadu_epi32(a12);
286 a12 += 64;
287 _mm512_store_epi32(vintile + 192, vin12);
288 const __m512i vin13 = _mm512_loadu_epi32(a13);
289 a13 += 64;
290 _mm512_store_epi32(vintile + 208, vin13);
291 const __m512i vin14 = _mm512_loadu_epi32(a14);
292 a14 += 64;
293 _mm512_store_epi32(vintile + 224, vin14);
294 const __m512i vin15 = _mm512_loadu_epi32(a15);
295 a15 += 64;
296 _mm512_store_epi32(vintile + 240, vin15);
297 _tile_loadd(4, vintile, 64);
298 _tile_loadd(5, (const int8_t*) w + 0, 64);
299 _tile_dpbssd(0, 4, 5);
300
301 w = (const int8_t*) w + 1024;
302 k -= 64 * sizeof(int8_t);
303 }
304 }
305
306 if XNN_UNLIKELY(k != 0) {
307 const __m512i vin0 = _mm512_maskz_loadu_epi32(kremainder_mask, a0);
308 a0 += kremainder;
309 _mm512_store_epi32(vintile + 0, vin0);
310 const __m512i vin1 = _mm512_maskz_loadu_epi32(kremainder_mask, a1);
311 a1 += kremainder;
312 _mm512_store_epi32(vintile + 16, vin1);
313 const __m512i vin2 = _mm512_maskz_loadu_epi32(kremainder_mask, a2);
314 a2 += kremainder;
315 _mm512_store_epi32(vintile + 32, vin2);
316 const __m512i vin3 = _mm512_maskz_loadu_epi32(kremainder_mask, a3);
317 a3 += kremainder;
318 _mm512_store_epi32(vintile + 48, vin3);
319 const __m512i vin4 = _mm512_maskz_loadu_epi32(kremainder_mask, a4);
320 a4 += kremainder;
321 _mm512_store_epi32(vintile + 64, vin4);
322 const __m512i vin5 = _mm512_maskz_loadu_epi32(kremainder_mask, a5);
323 a5 += kremainder;
324 _mm512_store_epi32(vintile + 80, vin5);
325 const __m512i vin6 = _mm512_maskz_loadu_epi32(kremainder_mask, a6);
326 a6 += kremainder;
327 _mm512_store_epi32(vintile + 96, vin6);
328 const __m512i vin7 = _mm512_maskz_loadu_epi32(kremainder_mask, a7);
329 a7 += kremainder;
330 _mm512_store_epi32(vintile + 112, vin7);
331 const __m512i vin8 = _mm512_maskz_loadu_epi32(kremainder_mask, a8);
332 a8 += kremainder;
333 _mm512_store_epi32(vintile + 128, vin8);
334 const __m512i vin9 = _mm512_maskz_loadu_epi32(kremainder_mask, a9);
335 a9 += kremainder;
336 _mm512_store_epi32(vintile + 144, vin9);
337 const __m512i vin10 = _mm512_maskz_loadu_epi32(kremainder_mask, a10);
338 a10 += kremainder;
339 _mm512_store_epi32(vintile + 160, vin10);
340 const __m512i vin11 = _mm512_maskz_loadu_epi32(kremainder_mask, a11);
341 a11 += kremainder;
342 _mm512_store_epi32(vintile + 176, vin11);
343 const __m512i vin12 = _mm512_maskz_loadu_epi32(kremainder_mask, a12);
344 a12 += kremainder;
345 _mm512_store_epi32(vintile + 192, vin12);
346 const __m512i vin13 = _mm512_maskz_loadu_epi32(kremainder_mask, a13);
347 a13 += kremainder;
348 _mm512_store_epi32(vintile + 208, vin13);
349 const __m512i vin14 = _mm512_maskz_loadu_epi32(kremainder_mask, a14);
350 a14 += kremainder;
351 _mm512_store_epi32(vintile + 224, vin14);
352 const __m512i vin15 = _mm512_maskz_loadu_epi32(kremainder_mask, a15);
353 a15 += kremainder;
354 _mm512_store_epi32(vintile + 240, vin15);
355 _tile_loadd(6, vintile, 64);
356 _tile_loadd(7, (const int8_t*) w + 0, 64);
357 _tile_dpbssd(0, 6, 7);
358
359 w = (const int8_t*) w + kremainder * 16;
360 k -= kremainder * sizeof(int8_t);
361 }
362
363 p -= 16 * sizeof(void*);
364 } while (p != 0);
365
366 // Add tile to bias
367 _tile_stored(0, res0, 64);
368
369 __m512i vacc0x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
370 __m512i vacc1x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 16));
371 __m512i vacc2x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 32));
372 __m512i vacc3x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 48));
373 __m512i vacc4x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 64));
374 __m512i vacc5x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 80));
375 __m512i vacc6x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 96));
376 __m512i vacc7x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 112));
377 __m512i vacc8x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 128));
378 __m512i vacc9x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 144));
379 __m512i vacc10x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 160));
380 __m512i vacc11x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 176));
381 __m512i vacc12x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 192));
382 __m512i vacc13x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 208));
383 __m512i vacc14x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 224));
384 __m512i vacc15x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 240));
385
386 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
387 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
388 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
389 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
390 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
391 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
392 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
393 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
394 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
395 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
396 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
397 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
398 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
399 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
400 __m512 vscaled14x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc14x0123456789ABCDEF);
401 __m512 vscaled15x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc15x0123456789ABCDEF);
402
403 const __m512 vscale0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
404 w = (const int32_t*) w + 16;
405
406 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale0123456789ABCDEF);
407 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale0123456789ABCDEF);
408 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale0123456789ABCDEF);
409 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale0123456789ABCDEF);
410 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale0123456789ABCDEF);
411 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale0123456789ABCDEF);
412 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale0123456789ABCDEF);
413 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vscale0123456789ABCDEF);
414 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vscale0123456789ABCDEF);
415 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vscale0123456789ABCDEF);
416 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, vscale0123456789ABCDEF);
417 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, vscale0123456789ABCDEF);
418 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, vscale0123456789ABCDEF);
419 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, vscale0123456789ABCDEF);
420 vscaled14x0123456789ABCDEF = _mm512_mul_ps(vscaled14x0123456789ABCDEF, vscale0123456789ABCDEF);
421 vscaled15x0123456789ABCDEF = _mm512_mul_ps(vscaled15x0123456789ABCDEF, vscale0123456789ABCDEF);
422
423 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
424 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
425 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
426 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
427 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
428 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
429 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
430 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max_less_zero_point);
431 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max_less_zero_point);
432 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max_less_zero_point);
433 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max_less_zero_point);
434 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max_less_zero_point);
435 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max_less_zero_point);
436 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max_less_zero_point);
437 vscaled14x0123456789ABCDEF = _mm512_min_ps(vscaled14x0123456789ABCDEF, voutput_max_less_zero_point);
438 vscaled15x0123456789ABCDEF = _mm512_min_ps(vscaled15x0123456789ABCDEF, voutput_max_less_zero_point);
439
440 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
441 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
442 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
443 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
444 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
445 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
446 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
447 vacc7x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled7x0123456789ABCDEF);
448 vacc8x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled8x0123456789ABCDEF);
449 vacc9x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled9x0123456789ABCDEF);
450 vacc10x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled10x0123456789ABCDEF);
451 vacc11x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled11x0123456789ABCDEF);
452 vacc12x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled12x0123456789ABCDEF);
453 vacc13x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled13x0123456789ABCDEF);
454 vacc14x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled14x0123456789ABCDEF);
455 vacc15x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled15x0123456789ABCDEF);
456
457 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
458 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
459 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
460 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
461 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
462 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
463 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
464 vacc7x0123456789ABCDEF = _mm512_add_epi32(vacc7x0123456789ABCDEF, voutput_zero_point);
465 vacc8x0123456789ABCDEF = _mm512_add_epi32(vacc8x0123456789ABCDEF, voutput_zero_point);
466 vacc9x0123456789ABCDEF = _mm512_add_epi32(vacc9x0123456789ABCDEF, voutput_zero_point);
467 vacc10x0123456789ABCDEF = _mm512_add_epi32(vacc10x0123456789ABCDEF, voutput_zero_point);
468 vacc11x0123456789ABCDEF = _mm512_add_epi32(vacc11x0123456789ABCDEF, voutput_zero_point);
469 vacc12x0123456789ABCDEF = _mm512_add_epi32(vacc12x0123456789ABCDEF, voutput_zero_point);
470 vacc13x0123456789ABCDEF = _mm512_add_epi32(vacc13x0123456789ABCDEF, voutput_zero_point);
471 vacc14x0123456789ABCDEF = _mm512_add_epi32(vacc14x0123456789ABCDEF, voutput_zero_point);
472 vacc15x0123456789ABCDEF = _mm512_add_epi32(vacc15x0123456789ABCDEF, voutput_zero_point);
473
474 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
475 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
476 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
477 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
478 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
479 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
480 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
481 __m128i vout7x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc7x0123456789ABCDEF);
482 __m128i vout8x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc8x0123456789ABCDEF);
483 __m128i vout9x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc9x0123456789ABCDEF);
484 __m128i vout10x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc10x0123456789ABCDEF);
485 __m128i vout11x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc11x0123456789ABCDEF);
486 __m128i vout12x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc12x0123456789ABCDEF);
487 __m128i vout13x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc13x0123456789ABCDEF);
488 __m128i vout14x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc14x0123456789ABCDEF);
489 __m128i vout15x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc15x0123456789ABCDEF);
490
491 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
492 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
493 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
494 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
495 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
496 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
497 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
498 vout7x0123456789ABCDEF = _mm_max_epi8(vout7x0123456789ABCDEF, voutput_min);
499 vout8x0123456789ABCDEF = _mm_max_epi8(vout8x0123456789ABCDEF, voutput_min);
500 vout9x0123456789ABCDEF = _mm_max_epi8(vout9x0123456789ABCDEF, voutput_min);
501 vout10x0123456789ABCDEF = _mm_max_epi8(vout10x0123456789ABCDEF, voutput_min);
502 vout11x0123456789ABCDEF = _mm_max_epi8(vout11x0123456789ABCDEF, voutput_min);
503 vout12x0123456789ABCDEF = _mm_max_epi8(vout12x0123456789ABCDEF, voutput_min);
504 vout13x0123456789ABCDEF = _mm_max_epi8(vout13x0123456789ABCDEF, voutput_min);
505 vout14x0123456789ABCDEF = _mm_max_epi8(vout14x0123456789ABCDEF, voutput_min);
506 vout15x0123456789ABCDEF = _mm_max_epi8(vout15x0123456789ABCDEF, voutput_min);
507
508 if XNN_LIKELY(nc >= 16) {
509 _mm_storeu_si128((__m128i*) (c15 + 0), vout15x0123456789ABCDEF);
510 c15 = (int8_t*) ((uintptr_t) c15 + cn_stride);
511 _mm_storeu_si128((__m128i*) (c14 + 0), vout14x0123456789ABCDEF);
512 c14 = (int8_t*) ((uintptr_t) c14 + cn_stride);
513 _mm_storeu_si128((__m128i*) (c13 + 0), vout13x0123456789ABCDEF);
514 c13 = (int8_t*) ((uintptr_t) c13 + cn_stride);
515 _mm_storeu_si128((__m128i*) (c12 + 0), vout12x0123456789ABCDEF);
516 c12 = (int8_t*) ((uintptr_t) c12 + cn_stride);
517 _mm_storeu_si128((__m128i*) (c11 + 0), vout11x0123456789ABCDEF);
518 c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
519 _mm_storeu_si128((__m128i*) (c10 + 0), vout10x0123456789ABCDEF);
520 c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
521 _mm_storeu_si128((__m128i*) (c9 + 0), vout9x0123456789ABCDEF);
522 c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
523 _mm_storeu_si128((__m128i*) (c8 + 0), vout8x0123456789ABCDEF);
524 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
525 _mm_storeu_si128((__m128i*) (c7 + 0), vout7x0123456789ABCDEF);
526 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
527 _mm_storeu_si128((__m128i*) (c6 + 0), vout6x0123456789ABCDEF);
528 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
529 _mm_storeu_si128((__m128i*) (c5 + 0), vout5x0123456789ABCDEF);
530 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
531 _mm_storeu_si128((__m128i*) (c4 + 0), vout4x0123456789ABCDEF);
532 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
533 _mm_storeu_si128((__m128i*) (c3 + 0), vout3x0123456789ABCDEF);
534 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
535 _mm_storeu_si128((__m128i*) (c2 + 0), vout2x0123456789ABCDEF);
536 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
537 _mm_storeu_si128((__m128i*) (c1 + 0), vout1x0123456789ABCDEF);
538 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
539 _mm_storeu_si128((__m128i*) (c0 + 0), vout0x0123456789ABCDEF);
540 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
541 a = (const int8_t**restrict) ((uintptr_t) a - ks);
542 nc -= 16;
543 } else {
544 // Prepare mask for valid 8-bit elements (depends on nc).
545 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
546
547 _mm_mask_storeu_epi8(c15 + 0, vmask0, vout15x0123456789ABCDEF);
548 _mm_mask_storeu_epi8(c14 + 0, vmask0, vout14x0123456789ABCDEF);
549 _mm_mask_storeu_epi8(c13 + 0, vmask0, vout13x0123456789ABCDEF);
550 _mm_mask_storeu_epi8(c12 + 0, vmask0, vout12x0123456789ABCDEF);
551 _mm_mask_storeu_epi8(c11 + 0, vmask0, vout11x0123456789ABCDEF);
552 _mm_mask_storeu_epi8(c10 + 0, vmask0, vout10x0123456789ABCDEF);
553 _mm_mask_storeu_epi8(c9 + 0, vmask0, vout9x0123456789ABCDEF);
554 _mm_mask_storeu_epi8(c8 + 0, vmask0, vout8x0123456789ABCDEF);
555 _mm_mask_storeu_epi8(c7 + 0, vmask0, vout7x0123456789ABCDEF);
556 _mm_mask_storeu_epi8(c6 + 0, vmask0, vout6x0123456789ABCDEF);
557 _mm_mask_storeu_epi8(c5 + 0, vmask0, vout5x0123456789ABCDEF);
558 _mm_mask_storeu_epi8(c4 + 0, vmask0, vout4x0123456789ABCDEF);
559 _mm_mask_storeu_epi8(c3 + 0, vmask0, vout3x0123456789ABCDEF);
560 _mm_mask_storeu_epi8(c2 + 0, vmask0, vout2x0123456789ABCDEF);
561 _mm_mask_storeu_epi8(c1 + 0, vmask0, vout1x0123456789ABCDEF);
562 _mm_mask_storeu_epi8(c0 + 0, vmask0, vout0x0123456789ABCDEF);
563 nc = 0;
564 }
565 } while (nc != 0);
566
567 // Release tile config
568 // _tile_release();
569 __asm__ volatile ("tilerelease" ::);
570 #endif // defined(__x86_64__)
571 }