Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-6x8c8-minmax-fp32-avxvnni-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_6x8c8__avxvnni_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 <= 6);
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
68 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
69 XNN_FORCE_REALIZATION(vsign_mask);
70 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);
71 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
72 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
73 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
74 // XNN_FORCE_REALIZATION(voutput_zero_point);
75 // XNN_FORCE_REALIZATION(voutput_min);
76 do {
77 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
78 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
79 __m256i vacc1x0123 = vacc0x0123;
80 __m256i vacc1x4567 = vacc0x4567;
81 __m256i vacc2x0123 = vacc0x0123;
82 __m256i vacc2x4567 = vacc0x4567;
83 __m256i vacc3x0123 = vacc0x0123;
84 __m256i vacc3x4567 = vacc0x4567;
85 __m256i vacc4x0123 = vacc0x0123;
86 __m256i vacc4x4567 = vacc0x4567;
87 __m256i vacc5x0123 = vacc0x0123;
88 __m256i vacc5x4567 = vacc0x4567;
89 w = (const int32_t*) w + 8;
90
91 size_t p = ks;
92 do {
93 const int8_t* restrict a0 = a[0];
94 if XNN_UNPREDICTABLE(a0 != zero) {
95 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
96 }
97 const int8_t* restrict a1 = a[1];
98 if XNN_UNPREDICTABLE(a1 != zero) {
99 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
100 }
101 const int8_t* restrict a2 = a[2];
102 if XNN_UNPREDICTABLE(a2 != zero) {
103 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
104 }
105 const int8_t* restrict a3 = a[3];
106 if XNN_UNPREDICTABLE(a3 != zero) {
107 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
108 }
109 const int8_t* restrict a4 = a[4];
110 if XNN_UNPREDICTABLE(a4 != zero) {
111 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
112 }
113 const int8_t* restrict a5 = a[5];
114 if XNN_UNPREDICTABLE(a5 != zero) {
115 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
116 }
117 a += 6;
118
119 size_t k = kc;
120 while (k >= 16 * sizeof(int8_t)) {
121 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
122 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
123 a0 += 16;
124 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
125 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
126 a1 += 16;
127 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
128 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
129 a2 += 16;
130 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
131 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
132 a3 += 16;
133 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
134 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
135 a4 += 16;
136 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
137 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
138 a5 += 16;
139
140 const __m256i vb01234567x0123 = _mm256_load_si256(w);
141 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
142 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
143 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
144 xnn_prefetch_to_l1((const int8_t*) w + 896);
145
146 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
147 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
148 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
149 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
150 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
151 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
152 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
153 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
154 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
155 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
156 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
157 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
158 xnn_prefetch_to_l1((const int8_t*) w + 960);
159 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
160 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
161 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
162 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
163 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
164 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
165 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
166 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
167 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
168 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
169 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
170 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
171
172 w = (const int8_t*) w + 128;
173 k -= 16 * sizeof(int8_t);
174 }
175
176 if (k != 0) {
177 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
178 a0 += 8;
179 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
180 a1 += 8;
181 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
182 a2 += 8;
183 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
184 a3 += 8;
185 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
186 a4 += 8;
187 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
188 a5 += 8;
189
190 const __m256i vb01234567x0123 = _mm256_load_si256(w);
191 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
192
193 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
194 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
195 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
196 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
197 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
198 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
199 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
200 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
201 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
202 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
203 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
204 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
205 xnn_prefetch_to_l1((const int8_t*) w + 960);
206
207 w = (const int8_t*) w + 64;
208 k -= 8 * sizeof(int8_t);
209 }
210
211 p -= 6 * sizeof(void*);
212 } while (p != 0);
213
214
215 // Add adjacent pairs
216 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
217 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
218 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
219 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
220 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
221 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
222 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
223 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
224 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
225 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
226 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
227 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
228
229 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
230 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
231 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
232 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
233 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
234 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
235
236 const __m256 vscale01234567 = _mm256_load_ps(w);
237 w = (const float*) w + 8;
238 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
239 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
240 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
241 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
242 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
243 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
244
245 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
246 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
247 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
248 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
249 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
250 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
251
252 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
253 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
254 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
255 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
256 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
257 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
258
259 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
260 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
261 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
262 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
263 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
264 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
265
266 vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));
267 __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));
268 vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));
269 __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));
270 vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1)));
271 __m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567));
272 vacc3x01234567 = _mm256_packs_epi32(vacc3x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc3x01234567, 1)));
273 __m128i voutb3x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc3x01234567, vacc3x01234567));
274 vacc4x01234567 = _mm256_packs_epi32(vacc4x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc4x01234567, 1)));
275 __m128i voutb4x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc4x01234567, vacc4x01234567));
276 vacc5x01234567 = _mm256_packs_epi32(vacc5x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc5x01234567, 1)));
277 __m128i voutb5x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc5x01234567, vacc5x01234567));
278
279 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
280 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
281 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
282 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
283 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
284 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
285
286 if (nc >= 8) {
287 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
288 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
289 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
290 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
291 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
292 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
293 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
294 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
295 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
296 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
297 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
298 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
299 a = (const int8_t**restrict) ((uintptr_t) a - ks);
300 nc -= 8;
301 } else {
302 if (nc & 4) {
303 _mm_storeu_si32(c5, voutb5x01234567);
304 c5 += 4;
305 _mm_storeu_si32(c4, voutb4x01234567);
306 c4 += 4;
307 _mm_storeu_si32(c3, voutb3x01234567);
308 c3 += 4;
309 _mm_storeu_si32(c2, voutb2x01234567);
310 c2 += 4;
311 _mm_storeu_si32(c1, voutb1x01234567);
312 c1 += 4;
313 _mm_storeu_si32(c0, voutb0x01234567);
314 c0 += 4;
315 voutb5x01234567 = _mm_srli_epi64(voutb5x01234567, 32);
316 voutb4x01234567 = _mm_srli_epi64(voutb4x01234567, 32);
317 voutb3x01234567 = _mm_srli_epi64(voutb3x01234567, 32);
318 voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32);
319 voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);
320 voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);
321 }
322 if (nc & 2) {
323 unaligned_store_u16(c5, (uint16_t) _mm_extract_epi16(voutb5x01234567, 0));
324 c5 += 2;
325 unaligned_store_u16(c4, (uint16_t) _mm_extract_epi16(voutb4x01234567, 0));
326 c4 += 2;
327 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(voutb3x01234567, 0));
328 c3 += 2;
329 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0));
330 c2 += 2;
331 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));
332 c1 += 2;
333 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));
334 c0 += 2;
335 voutb5x01234567 = _mm_srli_epi32(voutb5x01234567, 16);
336 voutb4x01234567 = _mm_srli_epi32(voutb4x01234567, 16);
337 voutb3x01234567 = _mm_srli_epi32(voutb3x01234567, 16);
338 voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16);
339 voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);
340 voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);
341 }
342 if (nc & 1) {
343 *c5 = (int8_t) _mm_extract_epi8(voutb5x01234567, 0);
344 *c4 = (int8_t) _mm_extract_epi8(voutb4x01234567, 0);
345 *c3 = (int8_t) _mm_extract_epi8(voutb3x01234567, 0);
346 *c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0);
347 *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);
348 *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);
349 }
350 nc = 0;
351 }
352 } while (nc != 0);
353 }