Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-14x16c8-minmax-avx512vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c8-avx512vnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_14x16c8__avx512vnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const int8_t* zero_data,
35 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
36 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 14);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(kc % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 kc = round_up_po2(kc, 8 * sizeof(int8_t));
48 float* c0 = c;
49 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 c1 = c0;
52 }
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 c2 = c1;
56 }
57 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr < 4) {
59 c3 = c2;
60 }
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 c4 = c3;
64 }
65 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr < 6) {
67 c5 = c4;
68 }
69 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
70 if XNN_UNPREDICTABLE(mr <= 6) {
71 c6 = c5;
72 }
73 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
74 if XNN_UNPREDICTABLE(mr < 8) {
75 c7 = c6;
76 }
77 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
78 if XNN_UNPREDICTABLE(mr <= 8) {
79 c8 = c7;
80 }
81 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
82 if XNN_UNPREDICTABLE(mr < 10) {
83 c9 = c8;
84 }
85 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
86 if XNN_UNPREDICTABLE(mr <= 10) {
87 c10 = c9;
88 }
89 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
90 if XNN_UNPREDICTABLE(mr < 12) {
91 c11 = c10;
92 }
93 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
94 if XNN_UNPREDICTABLE(mr <= 12) {
95 c12 = c11;
96 }
97 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
98 if XNN_UNPREDICTABLE(mr != 14) {
99 c13 = c12;
100 }
101
102 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
103 XNN_FORCE_REALIZATION(vsign_mask);
104 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
105 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
106 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
107 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
108 // XNN_FORCE_REALIZATION(vinput_zero_point);
109 // XNN_FORCE_REALIZATION(vinput_inv_scale);
110 // XNN_FORCE_REALIZATION(voutput_min);
111 // XNN_FORCE_REALIZATION(voutput_max);
112 do {
113 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
114 const __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
115 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
116 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
117 __m512i vacc1x01234567 = vacc0x01234567;
118 __m512i vacc1x89ABCDEF = vacc0x89ABCDEF;
119 __m512i vacc2x01234567 = vacc0x01234567;
120 __m512i vacc2x89ABCDEF = vacc0x89ABCDEF;
121 __m512i vacc3x01234567 = vacc0x01234567;
122 __m512i vacc3x89ABCDEF = vacc0x89ABCDEF;
123 __m512i vacc4x01234567 = vacc0x01234567;
124 __m512i vacc4x89ABCDEF = vacc0x89ABCDEF;
125 __m512i vacc5x01234567 = vacc0x01234567;
126 __m512i vacc5x89ABCDEF = vacc0x89ABCDEF;
127 __m512i vacc6x01234567 = vacc0x01234567;
128 __m512i vacc6x89ABCDEF = vacc0x89ABCDEF;
129 __m512i vacc7x01234567 = vacc0x01234567;
130 __m512i vacc7x89ABCDEF = vacc0x89ABCDEF;
131 __m512i vacc8x01234567 = vacc0x01234567;
132 __m512i vacc8x89ABCDEF = vacc0x89ABCDEF;
133 __m512i vacc9x01234567 = vacc0x01234567;
134 __m512i vacc9x89ABCDEF = vacc0x89ABCDEF;
135 __m512i vacc10x01234567 = vacc0x01234567;
136 __m512i vacc10x89ABCDEF = vacc0x89ABCDEF;
137 __m512i vacc11x01234567 = vacc0x01234567;
138 __m512i vacc11x89ABCDEF = vacc0x89ABCDEF;
139 __m512i vacc12x01234567 = vacc0x01234567;
140 __m512i vacc12x89ABCDEF = vacc0x89ABCDEF;
141 __m512i vacc13x01234567 = vacc0x01234567;
142 __m512i vacc13x89ABCDEF = vacc0x89ABCDEF;
143 w = (const int32_t*) w + 16;
144
145 size_t p = ks;
146 do {
147 const int8_t* restrict a0 = a[0];
148 if XNN_UNPREDICTABLE(a0 != zero) {
149 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
150 } else {
151 a0 = zero_data;
152 }
153 const int8_t* restrict a1 = a[1];
154 if XNN_UNPREDICTABLE(a1 != zero) {
155 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
156 } else {
157 a1 = zero_data;
158 }
159 const int8_t* restrict a2 = a[2];
160 if XNN_UNPREDICTABLE(a2 != zero) {
161 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
162 } else {
163 a2 = zero_data;
164 }
165 const int8_t* restrict a3 = a[3];
166 if XNN_UNPREDICTABLE(a3 != zero) {
167 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
168 } else {
169 a3 = zero_data;
170 }
171 const int8_t* restrict a4 = a[4];
172 if XNN_UNPREDICTABLE(a4 != zero) {
173 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
174 } else {
175 a4 = zero_data;
176 }
177 const int8_t* restrict a5 = a[5];
178 if XNN_UNPREDICTABLE(a5 != zero) {
179 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
180 } else {
181 a5 = zero_data;
182 }
183 const int8_t* restrict a6 = a[6];
184 if XNN_UNPREDICTABLE(a6 != zero) {
185 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
186 } else {
187 a6 = zero_data;
188 }
189 const int8_t* restrict a7 = a[7];
190 if XNN_UNPREDICTABLE(a7 != zero) {
191 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
192 } else {
193 a7 = zero_data;
194 }
195 const int8_t* restrict a8 = a[8];
196 if XNN_UNPREDICTABLE(a8 != zero) {
197 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
198 } else {
199 a8 = zero_data;
200 }
201 const int8_t* restrict a9 = a[9];
202 if XNN_UNPREDICTABLE(a9 != zero) {
203 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
204 } else {
205 a9 = zero_data;
206 }
207 const int8_t* restrict a10 = a[10];
208 if XNN_UNPREDICTABLE(a10 != zero) {
209 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
210 } else {
211 a10 = zero_data;
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 } else {
217 a11 = zero_data;
218 }
219 const int8_t* restrict a12 = a[12];
220 if XNN_UNPREDICTABLE(a12 != zero) {
221 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
222 } else {
223 a12 = zero_data;
224 }
225 const int8_t* restrict a13 = a[13];
226 if XNN_UNPREDICTABLE(a13 != zero) {
227 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
228 } else {
229 a13 = zero_data;
230 }
231 a += 14;
232
233 size_t k = kc;
234 while (k >= 16 * sizeof(int8_t)) {
235 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
236 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
237 a0 += 16;
238 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
239 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
240 a1 += 16;
241 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
242 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
243 a2 += 16;
244 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
245 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
246 a3 += 16;
247 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
248 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
249 a4 += 16;
250 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
251 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
252 a5 += 16;
253 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
254 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
255 a6 += 16;
256 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
257 const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
258 a7 += 16;
259 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
260 const __m512i va8x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
261 a8 += 16;
262 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
263 const __m512i va9x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
264 a9 += 16;
265 const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask);
266 const __m512i va10x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
267 a10 += 16;
268 const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask);
269 const __m512i va11x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
270 a11 += 16;
271 const __m512i va12x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a12)), vsign_mask);
272 const __m512i va12x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a12 + 8)), vsign_mask);
273 a12 += 16;
274 const __m512i va13x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a13)), vsign_mask);
275 const __m512i va13x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a13 + 8)), vsign_mask);
276 a13 += 16;
277
278 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
279 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
280 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
281 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
282 xnn_prefetch_to_l1((const int8_t*) w + 768);
283 xnn_prefetch_to_l1((const int8_t*) w + 832);
284
285 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
286 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
287 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
288 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
289 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
290 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
291 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
292 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
293 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
294 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
295 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
296 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
297 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
298 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
299 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
300 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
301 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
302 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
303 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
304 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
305 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567);
306 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567);
307 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567);
308 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567);
309 vacc12x01234567 = _mm512_dpbusd_epi32(vacc12x01234567, va12x01234567, vb01234567x01234567);
310 vacc12x89ABCDEF = _mm512_dpbusd_epi32(vacc12x89ABCDEF, va12x01234567, vb89ABCDEFx01234567);
311 vacc13x01234567 = _mm512_dpbusd_epi32(vacc13x01234567, va13x01234567, vb01234567x01234567);
312 vacc13x89ABCDEF = _mm512_dpbusd_epi32(vacc13x89ABCDEF, va13x01234567, vb89ABCDEFx01234567);
313 xnn_prefetch_to_l1((const int8_t*) w + 896);
314 xnn_prefetch_to_l1((const int8_t*) w + 960);
315 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
316 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
317 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
318 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
319 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
320 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
321 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
322 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
323 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
324 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
325 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
326 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
327 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
328 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
329 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF);
330 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF);
331 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x89ABCDEF, vb01234567x89ABCDEF);
332 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x89ABCDEF, vb89ABCDEFx89ABCDEF);
333 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x89ABCDEF, vb01234567x89ABCDEF);
334 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x89ABCDEF, vb89ABCDEFx89ABCDEF);
335 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x89ABCDEF, vb01234567x89ABCDEF);
336 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x89ABCDEF, vb89ABCDEFx89ABCDEF);
337 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x89ABCDEF, vb01234567x89ABCDEF);
338 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x89ABCDEF, vb89ABCDEFx89ABCDEF);
339 vacc12x01234567 = _mm512_dpbusd_epi32(vacc12x01234567, va12x89ABCDEF, vb01234567x89ABCDEF);
340 vacc12x89ABCDEF = _mm512_dpbusd_epi32(vacc12x89ABCDEF, va12x89ABCDEF, vb89ABCDEFx89ABCDEF);
341 vacc13x01234567 = _mm512_dpbusd_epi32(vacc13x01234567, va13x89ABCDEF, vb01234567x89ABCDEF);
342 vacc13x89ABCDEF = _mm512_dpbusd_epi32(vacc13x89ABCDEF, va13x89ABCDEF, vb89ABCDEFx89ABCDEF);
343
344 w = (const int8_t*) w + 256;
345 k -= 16 * sizeof(int8_t);
346 }
347
348 if (k != 0) {
349 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
350 a0 += 8;
351 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
352 a1 += 8;
353 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
354 a2 += 8;
355 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
356 a3 += 8;
357 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
358 a4 += 8;
359 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
360 a5 += 8;
361 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
362 a6 += 8;
363 const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask);
364 a7 += 8;
365 const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask);
366 a8 += 8;
367 const __m512i va9x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a9)), vsign_mask);
368 a9 += 8;
369 const __m512i va10x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a10)), vsign_mask);
370 a10 += 8;
371 const __m512i va11x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a11)), vsign_mask);
372 a11 += 8;
373 const __m512i va12x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a12)), vsign_mask);
374 a12 += 8;
375 const __m512i va13x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a13)), vsign_mask);
376 a13 += 8;
377
378 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
379 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
380
381 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
382 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
383 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
384 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
385 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
386 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
387 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
388 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
389 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
390 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
391 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
392 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
393 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
394 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
395 vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567);
396 vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567);
397 vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567);
398 vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567);
399 vacc9x01234567 = _mm512_dpbusd_epi32(vacc9x01234567, va9x01234567, vb01234567x01234567);
400 vacc9x89ABCDEF = _mm512_dpbusd_epi32(vacc9x89ABCDEF, va9x01234567, vb89ABCDEFx01234567);
401 vacc10x01234567 = _mm512_dpbusd_epi32(vacc10x01234567, va10x01234567, vb01234567x01234567);
402 vacc10x89ABCDEF = _mm512_dpbusd_epi32(vacc10x89ABCDEF, va10x01234567, vb89ABCDEFx01234567);
403 vacc11x01234567 = _mm512_dpbusd_epi32(vacc11x01234567, va11x01234567, vb01234567x01234567);
404 vacc11x89ABCDEF = _mm512_dpbusd_epi32(vacc11x89ABCDEF, va11x01234567, vb89ABCDEFx01234567);
405 vacc12x01234567 = _mm512_dpbusd_epi32(vacc12x01234567, va12x01234567, vb01234567x01234567);
406 vacc12x89ABCDEF = _mm512_dpbusd_epi32(vacc12x89ABCDEF, va12x01234567, vb89ABCDEFx01234567);
407 vacc13x01234567 = _mm512_dpbusd_epi32(vacc13x01234567, va13x01234567, vb01234567x01234567);
408 vacc13x89ABCDEF = _mm512_dpbusd_epi32(vacc13x89ABCDEF, va13x01234567, vb89ABCDEFx01234567);
409 xnn_prefetch_to_l1((const int8_t*) w + 896);
410 xnn_prefetch_to_l1((const int8_t*) w + 960);
411
412 w = (const int8_t*) w + 128;
413 k -= 8 * sizeof(int8_t);
414 }
415
416 p -= 14 * sizeof(void*);
417 } while (p != 0);
418
419
420 // Add adjacent pairs
421 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
422 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
423 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
424 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
425 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
426 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
427 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
428 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
429 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
430 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
431 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
432 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
433 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
434 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
435 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
436 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
437 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
438 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
439 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
440 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
441 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
442 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
443 const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32));
444 const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32));
445 __m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF);
446 const __m512i vsum8x01234567 = _mm512_add_epi32(vacc8x01234567, _mm512_srli_epi64(vacc8x01234567, 32));
447 const __m512i vsum8x89ABCDEF = _mm512_add_epi32(vacc8x89ABCDEF, _mm512_srli_epi64(vacc8x89ABCDEF, 32));
448 __m512i vacc8x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum8x01234567, vidx, vsum8x89ABCDEF);
449 const __m512i vsum9x01234567 = _mm512_add_epi32(vacc9x01234567, _mm512_srli_epi64(vacc9x01234567, 32));
450 const __m512i vsum9x89ABCDEF = _mm512_add_epi32(vacc9x89ABCDEF, _mm512_srli_epi64(vacc9x89ABCDEF, 32));
451 __m512i vacc9x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum9x01234567, vidx, vsum9x89ABCDEF);
452 const __m512i vsum10x01234567 = _mm512_add_epi32(vacc10x01234567, _mm512_srli_epi64(vacc10x01234567, 32));
453 const __m512i vsum10x89ABCDEF = _mm512_add_epi32(vacc10x89ABCDEF, _mm512_srli_epi64(vacc10x89ABCDEF, 32));
454 __m512i vacc10x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum10x01234567, vidx, vsum10x89ABCDEF);
455 const __m512i vsum11x01234567 = _mm512_add_epi32(vacc11x01234567, _mm512_srli_epi64(vacc11x01234567, 32));
456 const __m512i vsum11x89ABCDEF = _mm512_add_epi32(vacc11x89ABCDEF, _mm512_srli_epi64(vacc11x89ABCDEF, 32));
457 __m512i vacc11x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum11x01234567, vidx, vsum11x89ABCDEF);
458 const __m512i vsum12x01234567 = _mm512_add_epi32(vacc12x01234567, _mm512_srli_epi64(vacc12x01234567, 32));
459 const __m512i vsum12x89ABCDEF = _mm512_add_epi32(vacc12x89ABCDEF, _mm512_srli_epi64(vacc12x89ABCDEF, 32));
460 __m512i vacc12x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum12x01234567, vidx, vsum12x89ABCDEF);
461 const __m512i vsum13x01234567 = _mm512_add_epi32(vacc13x01234567, _mm512_srli_epi64(vacc13x01234567, 32));
462 const __m512i vsum13x89ABCDEF = _mm512_add_epi32(vacc13x89ABCDEF, _mm512_srli_epi64(vacc13x89ABCDEF, 32));
463 __m512i vacc13x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum13x01234567, vidx, vsum13x89ABCDEF);
464
465 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
466 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
467 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
468 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
469 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
470 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
471 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
472 __m512 vscaled7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF);
473 __m512 vscaled8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF);
474 __m512 vscaled9x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc9x0123456789ABCDEF);
475 __m512 vscaled10x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc10x0123456789ABCDEF);
476 __m512 vscaled11x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc11x0123456789ABCDEF);
477 __m512 vscaled12x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc12x0123456789ABCDEF);
478 __m512 vscaled13x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc13x0123456789ABCDEF);
479
480 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
481 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vinput_inv_scale);
482 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vinput_inv_scale);
483 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vinput_inv_scale);
484 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vinput_inv_scale);
485 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vinput_inv_scale);
486 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vinput_inv_scale);
487 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, vinput_inv_scale);
488 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, vinput_inv_scale);
489 vscaled9x0123456789ABCDEF = _mm512_mul_ps(vscaled9x0123456789ABCDEF, vinput_inv_scale);
490 vscaled10x0123456789ABCDEF = _mm512_mul_ps(vscaled10x0123456789ABCDEF, vinput_inv_scale);
491 vscaled11x0123456789ABCDEF = _mm512_mul_ps(vscaled11x0123456789ABCDEF, vinput_inv_scale);
492 vscaled12x0123456789ABCDEF = _mm512_mul_ps(vscaled12x0123456789ABCDEF, vinput_inv_scale);
493 vscaled13x0123456789ABCDEF = _mm512_mul_ps(vscaled13x0123456789ABCDEF, vinput_inv_scale);
494
495 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
496 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
497 w = (const float*) w + 32;
498
499 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
500 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
501 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
502 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
503 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
504 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
505 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
506 vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vscaled7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
507 vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vscaled8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
508 vscaled9x0123456789ABCDEF = _mm512_fmadd_ps(vscaled9x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
509 vscaled10x0123456789ABCDEF = _mm512_fmadd_ps(vscaled10x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
510 vscaled11x0123456789ABCDEF = _mm512_fmadd_ps(vscaled11x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
511 vscaled12x0123456789ABCDEF = _mm512_fmadd_ps(vscaled12x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
512 vscaled13x0123456789ABCDEF = _mm512_fmadd_ps(vscaled13x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
513
514 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
515 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
516 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
517 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
518 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
519 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
520 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
521 vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min);
522 vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min);
523 vscaled9x0123456789ABCDEF = _mm512_max_ps(vscaled9x0123456789ABCDEF, voutput_min);
524 vscaled10x0123456789ABCDEF = _mm512_max_ps(vscaled10x0123456789ABCDEF, voutput_min);
525 vscaled11x0123456789ABCDEF = _mm512_max_ps(vscaled11x0123456789ABCDEF, voutput_min);
526 vscaled12x0123456789ABCDEF = _mm512_max_ps(vscaled12x0123456789ABCDEF, voutput_min);
527 vscaled13x0123456789ABCDEF = _mm512_max_ps(vscaled13x0123456789ABCDEF, voutput_min);
528
529 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
530 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
531 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
532 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
533 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
534 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
535 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
536 vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max);
537 vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max);
538 vscaled9x0123456789ABCDEF = _mm512_min_ps(vscaled9x0123456789ABCDEF, voutput_max);
539 vscaled10x0123456789ABCDEF = _mm512_min_ps(vscaled10x0123456789ABCDEF, voutput_max);
540 vscaled11x0123456789ABCDEF = _mm512_min_ps(vscaled11x0123456789ABCDEF, voutput_max);
541 vscaled12x0123456789ABCDEF = _mm512_min_ps(vscaled12x0123456789ABCDEF, voutput_max);
542 vscaled13x0123456789ABCDEF = _mm512_min_ps(vscaled13x0123456789ABCDEF, voutput_max);
543
544 if XNN_LIKELY(nc >= 16) {
545 _mm512_storeu_ps(c13, vscaled13x0123456789ABCDEF);
546 c13 = (float*) ((uintptr_t) c13 + cn_stride);
547 _mm512_storeu_ps(c12, vscaled12x0123456789ABCDEF);
548 c12 = (float*) ((uintptr_t) c12 + cn_stride);
549 _mm512_storeu_ps(c11, vscaled11x0123456789ABCDEF);
550 c11 = (float*) ((uintptr_t) c11 + cn_stride);
551 _mm512_storeu_ps(c10, vscaled10x0123456789ABCDEF);
552 c10 = (float*) ((uintptr_t) c10 + cn_stride);
553 _mm512_storeu_ps(c9, vscaled9x0123456789ABCDEF);
554 c9 = (float*) ((uintptr_t) c9 + cn_stride);
555 _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF);
556 c8 = (float*) ((uintptr_t) c8 + cn_stride);
557 _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF);
558 c7 = (float*) ((uintptr_t) c7 + cn_stride);
559 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
560 c6 = (float*) ((uintptr_t) c6 + cn_stride);
561 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
562 c5 = (float*) ((uintptr_t) c5 + cn_stride);
563 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
564 c4 = (float*) ((uintptr_t) c4 + cn_stride);
565 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
566 c3 = (float*) ((uintptr_t) c3 + cn_stride);
567 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
568 c2 = (float*) ((uintptr_t) c2 + cn_stride);
569 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
570 c1 = (float*) ((uintptr_t) c1 + cn_stride);
571 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
572 c0 = (float*) ((uintptr_t) c0 + cn_stride);
573 a = (const int8_t**restrict) ((uintptr_t) a - ks);
574 nc -= 16;
575 } else {
576 // Prepare mask for valid 32-bit elements (depends on nc).
577 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
578 _mm512_mask_storeu_ps(c13, vmask, vscaled13x0123456789ABCDEF);
579 _mm512_mask_storeu_ps(c12, vmask, vscaled12x0123456789ABCDEF);
580 _mm512_mask_storeu_ps(c11, vmask, vscaled11x0123456789ABCDEF);
581 _mm512_mask_storeu_ps(c10, vmask, vscaled10x0123456789ABCDEF);
582 _mm512_mask_storeu_ps(c9, vmask, vscaled9x0123456789ABCDEF);
583 _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF);
584 _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF);
585 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
586 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
587 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
588 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
589 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
590 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
591 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
592 nc = 0;
593 }
594 } while (nc != 0);
595 }