Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-7x8c8-minmax-fp32-avx256vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx8c8-avxvnni.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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_7x8c8__avx256vnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 7);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
63 if XNN_UNPREDICTABLE(mr < 6) {
64 c5 = c4;
65 }
66 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 6) {
68 c6 = c5;
69 }
70
71 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
72 XNN_FORCE_REALIZATION(vsign_mask);
73 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
74 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
75 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
76 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
77 // XNN_FORCE_REALIZATION(voutput_zero_point);
78 // XNN_FORCE_REALIZATION(voutput_min);
79 do {
80 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
81 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
82 __m256i vacc1x0123 = vacc0x0123;
83 __m256i vacc1x4567 = vacc0x4567;
84 __m256i vacc2x0123 = vacc0x0123;
85 __m256i vacc2x4567 = vacc0x4567;
86 __m256i vacc3x0123 = vacc0x0123;
87 __m256i vacc3x4567 = vacc0x4567;
88 __m256i vacc4x0123 = vacc0x0123;
89 __m256i vacc4x4567 = vacc0x4567;
90 __m256i vacc5x0123 = vacc0x0123;
91 __m256i vacc5x4567 = vacc0x4567;
92 __m256i vacc6x0123 = vacc0x0123;
93 __m256i vacc6x4567 = vacc0x4567;
94 w = (const int32_t*) w + 8;
95
96 size_t p = ks;
97 do {
98 const int8_t* restrict a0 = a[0];
99 if XNN_UNPREDICTABLE(a0 != zero) {
100 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
101 }
102 const int8_t* restrict a1 = a[1];
103 if XNN_UNPREDICTABLE(a1 != zero) {
104 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
105 }
106 const int8_t* restrict a2 = a[2];
107 if XNN_UNPREDICTABLE(a2 != zero) {
108 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
109 }
110 const int8_t* restrict a3 = a[3];
111 if XNN_UNPREDICTABLE(a3 != zero) {
112 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
113 }
114 const int8_t* restrict a4 = a[4];
115 if XNN_UNPREDICTABLE(a4 != zero) {
116 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
117 }
118 const int8_t* restrict a5 = a[5];
119 if XNN_UNPREDICTABLE(a5 != zero) {
120 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
121 }
122 const int8_t* restrict a6 = a[6];
123 if XNN_UNPREDICTABLE(a6 != zero) {
124 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
125 }
126 a += 7;
127
128 size_t k = kc;
129 while (k >= 16 * sizeof(int8_t)) {
130 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
131 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
132 a0 += 16;
133 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
134 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
135 a1 += 16;
136 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
137 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
138 a2 += 16;
139 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
140 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
141 a3 += 16;
142 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
143 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
144 a4 += 16;
145 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
146 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
147 a5 += 16;
148 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
149 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
150 a6 += 16;
151
152 const __m256i vb01234567x0123 = _mm256_load_si256(w);
153 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
154 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
155 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
156
157 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
158 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
159 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
160 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
161 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
162 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
163 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
164 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
165 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
166 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
167 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
168 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
169 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
170 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
171 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
172 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
173 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
174 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
175 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
176 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
177 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
178 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
179 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
180 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
181 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
182 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
183 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
184 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
185
186 w = (const int8_t*) w + 128;
187 k -= 16 * sizeof(int8_t);
188 }
189
190 if (k != 0) {
191 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
192 a0 += 8;
193 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
194 a1 += 8;
195 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
196 a2 += 8;
197 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
198 a3 += 8;
199 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
200 a4 += 8;
201 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
202 a5 += 8;
203 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
204 a6 += 8;
205
206 const __m256i vb01234567x0123 = _mm256_load_si256(w);
207 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
208
209 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
210 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
211 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
212 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
213 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
214 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
215 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
216 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
217 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
218 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
219 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
220 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
221 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
222 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
223
224 w = (const int8_t*) w + 64;
225 k -= 8 * sizeof(int8_t);
226 }
227
228 p -= 7 * sizeof(void*);
229 } while (p != 0);
230
231
232 // Add adjacent pairs
233 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
234 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
235 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
236 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
237 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
238 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
239 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
240 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
241 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
242 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
243 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
244 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
245 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
246 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
247
248 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
249 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
250 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
251 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
252 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
253 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
254 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
255
256 const __m256 vscale01234567 = _mm256_load_ps(w);
257 w = (const float*) w + 8;
258 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
259 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
260 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
261 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
262 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
263 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
264 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
265
266 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
267 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
268 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
269 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
270 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
271 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
272 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
273
274 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
275 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
276 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
277 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
278 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
279 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
280 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
281
282 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
283 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
284 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
285 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
286 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
287 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
288 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
289
290 __m128i voutb0x01234567 = _mm256_cvtsepi32_epi8(vacc0x01234567);
291 __m128i voutb1x01234567 = _mm256_cvtsepi32_epi8(vacc1x01234567);
292 __m128i voutb2x01234567 = _mm256_cvtsepi32_epi8(vacc2x01234567);
293 __m128i voutb3x01234567 = _mm256_cvtsepi32_epi8(vacc3x01234567);
294 __m128i voutb4x01234567 = _mm256_cvtsepi32_epi8(vacc4x01234567);
295 __m128i voutb5x01234567 = _mm256_cvtsepi32_epi8(vacc5x01234567);
296 __m128i voutb6x01234567 = _mm256_cvtsepi32_epi8(vacc6x01234567);
297
298 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
299 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
300 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
301 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
302 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
303 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
304 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
305
306 if (nc >= 8) {
307 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
308 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
309 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
310 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
311 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
312 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
313 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
314 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
315 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
316 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
317 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
318 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
319 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
320 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
321 a = (const int8_t**restrict) ((uintptr_t) a - ks);
322 nc -= 8;
323 } else {
324 // Prepare mask for valid 8-bit elements (depends on nc).
325 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
326 _mm_mask_storeu_epi8(c6, vmask, voutb6x01234567);
327 _mm_mask_storeu_epi8(c5, vmask, voutb5x01234567);
328 _mm_mask_storeu_epi8(c4, vmask, voutb4x01234567);
329 _mm_mask_storeu_epi8(c3, vmask, voutb3x01234567);
330 _mm_mask_storeu_epi8(c2, vmask, voutb2x01234567);
331 _mm_mask_storeu_epi8(c1, vmask, voutb1x01234567);
332 _mm_mask_storeu_epi8(c0, vmask, voutb0x01234567);
333 nc = 0;
334 }
335 } while (nc != 0);
336 }