Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-7x16c8-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_qs8_qc8w_igemm_minmax_fp32_ukernel_7x16c8__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 int8_t* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 7);
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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr < 6) {
65 c5 = c4;
66 }
67 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 6) {
69 c6 = c5;
70 }
71
72 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
73 XNN_FORCE_REALIZATION(vsign_mask);
74 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);
75 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
76 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
77 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
78 // XNN_FORCE_REALIZATION(voutput_zero_point);
79 XNN_FORCE_REALIZATION(voutput_min);
80 do {
81 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm256_load_si256((const __m256i*) w));
82 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm256_load_si256((const __m256i*) ((const int32_t*) w + 8)));
83 __m512i vacc1x01234567 = vacc0x01234567;
84 __m512i vacc1x89ABCDEF = vacc0x89ABCDEF;
85 __m512i vacc2x01234567 = vacc0x01234567;
86 __m512i vacc2x89ABCDEF = vacc0x89ABCDEF;
87 __m512i vacc3x01234567 = vacc0x01234567;
88 __m512i vacc3x89ABCDEF = vacc0x89ABCDEF;
89 __m512i vacc4x01234567 = vacc0x01234567;
90 __m512i vacc4x89ABCDEF = vacc0x89ABCDEF;
91 __m512i vacc5x01234567 = vacc0x01234567;
92 __m512i vacc5x89ABCDEF = vacc0x89ABCDEF;
93 __m512i vacc6x01234567 = vacc0x01234567;
94 __m512i vacc6x89ABCDEF = vacc0x89ABCDEF;
95 w = (const int32_t*) w + 16;
96
97 size_t p = ks;
98 do {
99 const int8_t* restrict a0 = a[0];
100 if XNN_UNPREDICTABLE(a0 != zero) {
101 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
102 }
103 const int8_t* restrict a1 = a[1];
104 if XNN_UNPREDICTABLE(a1 != zero) {
105 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
106 }
107 const int8_t* restrict a2 = a[2];
108 if XNN_UNPREDICTABLE(a2 != zero) {
109 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
110 }
111 const int8_t* restrict a3 = a[3];
112 if XNN_UNPREDICTABLE(a3 != zero) {
113 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
114 }
115 const int8_t* restrict a4 = a[4];
116 if XNN_UNPREDICTABLE(a4 != zero) {
117 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
118 }
119 const int8_t* restrict a5 = a[5];
120 if XNN_UNPREDICTABLE(a5 != zero) {
121 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
122 }
123 const int8_t* restrict a6 = a[6];
124 if XNN_UNPREDICTABLE(a6 != zero) {
125 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
126 }
127 a += 7;
128
129 size_t k = kc;
130 while (k >= 16 * sizeof(int8_t)) {
131 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
132 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
133 a0 += 16;
134 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
135 const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
136 a1 += 16;
137 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
138 const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
139 a2 += 16;
140 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
141 const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
142 a3 += 16;
143 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
144 const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
145 a4 += 16;
146 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
147 const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
148 a5 += 16;
149 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
150 const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
151 a6 += 16;
152
153 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
154 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
155 const __m512i vb01234567x89ABCDEF = _mm512_load_si512((const int8_t*) w + 128);
156 const __m512i vb89ABCDEFx89ABCDEF = _mm512_load_si512((const int8_t*) w + 192);
157 xnn_prefetch_to_l1((const int8_t*) w + 768);
158 xnn_prefetch_to_l1((const int8_t*) w + 832);
159
160 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
161 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
162 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
163 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
164 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
165 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
166 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
167 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
168 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
169 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
170 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
171 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
172 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
173 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
174 xnn_prefetch_to_l1((const int8_t*) w + 896);
175 xnn_prefetch_to_l1((const int8_t*) w + 960);
176 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
177 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
178 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF);
179 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF);
180 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF);
181 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF);
182 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF);
183 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF);
184 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF);
185 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF);
186 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF);
187 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF);
188 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF);
189 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF);
190
191 w = (const int8_t*) w + 256;
192 k -= 16 * sizeof(int8_t);
193 }
194
195 if (k != 0) {
196 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
197 a0 += 8;
198 const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask);
199 a1 += 8;
200 const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask);
201 a2 += 8;
202 const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask);
203 a3 += 8;
204 const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask);
205 a4 += 8;
206 const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask);
207 a5 += 8;
208 const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask);
209 a6 += 8;
210
211 const __m512i vb01234567x01234567 = _mm512_load_si512(w);
212 const __m512i vb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
213
214 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
215 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
216 vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567);
217 vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567);
218 vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567);
219 vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567);
220 vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567);
221 vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567);
222 vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567);
223 vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567);
224 vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567);
225 vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567);
226 vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567);
227 vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567);
228 xnn_prefetch_to_l1((const int8_t*) w + 896);
229 xnn_prefetch_to_l1((const int8_t*) w + 960);
230
231 w = (const int8_t*) w + 128;
232 k -= 8 * sizeof(int8_t);
233 }
234
235 p -= 7 * sizeof(void*);
236 } while (p != 0);
237
238
239 // Add adjacent pairs
240 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
241 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
242 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
243 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
244 const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32));
245 const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32));
246 __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF);
247 const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32));
248 const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32));
249 __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF);
250 const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32));
251 const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32));
252 __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF);
253 const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32));
254 const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32));
255 __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF);
256 const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32));
257 const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32));
258 __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF);
259 const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32));
260 const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32));
261 __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF);
262
263 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
264 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
265 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
266 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
267 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
268 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
269 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
270
271 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
272 w = (const float*) w + 16;
273 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
274 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, vscale012345678ABCDEF);
275 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, vscale012345678ABCDEF);
276 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, vscale012345678ABCDEF);
277 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, vscale012345678ABCDEF);
278 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, vscale012345678ABCDEF);
279 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, vscale012345678ABCDEF);
280
281 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
282 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max_less_zero_point);
283 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max_less_zero_point);
284 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max_less_zero_point);
285 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max_less_zero_point);
286 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max_less_zero_point);
287 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max_less_zero_point);
288
289 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
290 vacc1x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled1x0123456789ABCDEF);
291 vacc2x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled2x0123456789ABCDEF);
292 vacc3x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled3x0123456789ABCDEF);
293 vacc4x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled4x0123456789ABCDEF);
294 vacc5x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled5x0123456789ABCDEF);
295 vacc6x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled6x0123456789ABCDEF);
296
297 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
298 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, voutput_zero_point);
299 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, voutput_zero_point);
300 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, voutput_zero_point);
301 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, voutput_zero_point);
302 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, voutput_zero_point);
303 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, voutput_zero_point);
304
305 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
306 __m128i vout1x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc1x0123456789ABCDEF);
307 __m128i vout2x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc2x0123456789ABCDEF);
308 __m128i vout3x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc3x0123456789ABCDEF);
309 __m128i vout4x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc4x0123456789ABCDEF);
310 __m128i vout5x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc5x0123456789ABCDEF);
311 __m128i vout6x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc6x0123456789ABCDEF);
312
313 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
314 vout1x0123456789ABCDEF = _mm_max_epi8(vout1x0123456789ABCDEF, voutput_min);
315 vout2x0123456789ABCDEF = _mm_max_epi8(vout2x0123456789ABCDEF, voutput_min);
316 vout3x0123456789ABCDEF = _mm_max_epi8(vout3x0123456789ABCDEF, voutput_min);
317 vout4x0123456789ABCDEF = _mm_max_epi8(vout4x0123456789ABCDEF, voutput_min);
318 vout5x0123456789ABCDEF = _mm_max_epi8(vout5x0123456789ABCDEF, voutput_min);
319 vout6x0123456789ABCDEF = _mm_max_epi8(vout6x0123456789ABCDEF, voutput_min);
320
321 if (nc >= 16) {
322 _mm_storeu_si128((__m128i*) c6, vout6x0123456789ABCDEF);
323 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
324 _mm_storeu_si128((__m128i*) c5, vout5x0123456789ABCDEF);
325 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
326 _mm_storeu_si128((__m128i*) c4, vout4x0123456789ABCDEF);
327 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
328 _mm_storeu_si128((__m128i*) c3, vout3x0123456789ABCDEF);
329 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
330 _mm_storeu_si128((__m128i*) c2, vout2x0123456789ABCDEF);
331 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
332 _mm_storeu_si128((__m128i*) c1, vout1x0123456789ABCDEF);
333 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
334 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
335 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
336 a = (const int8_t**restrict) ((uintptr_t) a - ks);
337 nc -= 16;
338 } else {
339 // Prepare mask for valid 8-bit elements (depends on nc).
340 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
341 _mm_mask_storeu_epi8(c6, vmask, vout6x0123456789ABCDEF);
342 _mm_mask_storeu_epi8(c5, vmask, vout5x0123456789ABCDEF);
343 _mm_mask_storeu_epi8(c4, vmask, vout4x0123456789ABCDEF);
344 _mm_mask_storeu_epi8(c3, vmask, vout3x0123456789ABCDEF);
345 _mm_mask_storeu_epi8(c2, vmask, vout2x0123456789ABCDEF);
346 _mm_mask_storeu_epi8(c1, vmask, vout1x0123456789ABCDEF);
347 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
348 nc = 0;
349 }
350 } while (nc != 0);
351 }