1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neon-mull-shuffle.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <arm_neon.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 int8_t* c0 = c;
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 kc = round_up_po2(kc, 8 * sizeof(int8_t));
58 do {
59 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
60 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
61 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
62 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
63 int32x4_t vacc1x01 = vacc0x01;
64 int32x4_t vacc1x23 = vacc0x23;
65 int32x4_t vacc1x45 = vacc0x45;
66 int32x4_t vacc1x67 = vacc0x67;
67 int32x4_t vacc2x01 = vacc0x01;
68 int32x4_t vacc2x23 = vacc0x23;
69 int32x4_t vacc2x45 = vacc0x45;
70 int32x4_t vacc2x67 = vacc0x67;
71 int32x4_t vacc3x01 = vacc0x01;
72 int32x4_t vacc3x23 = vacc0x23;
73 int32x4_t vacc3x45 = vacc0x45;
74 int32x4_t vacc3x67 = vacc0x67;
75
76 size_t p = ks;
77 do {
78 const int8_t* restrict a0 = a[0];
79 if XNN_UNPREDICTABLE(a0 != zero) {
80 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
81 }
82 const int8_t* restrict a1 = a[1];
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
85 }
86 const int8_t* restrict a2 = a[2];
87 if XNN_UNPREDICTABLE(a2 != zero) {
88 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
89 }
90 const int8_t* restrict a3 = a[3];
91 if XNN_UNPREDICTABLE(a3 != zero) {
92 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
93 }
94 a += 4;
95
96 size_t k = kc;
97 do {
98 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
99 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
100 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
101 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
102
103 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
104 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
105 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
106 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
107 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111
112 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
113 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
114 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
115 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
116 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
117 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
118 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
119 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
120 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
121 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
122 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
123 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
124 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
125 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
126 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
127 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
128 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
129 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
130 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
131 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
132 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
133 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
134 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
135 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
136 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
137 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
138 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
139 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
140 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
141 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
142 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
143 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
144 va0x0 = vext_s8(va0x0, va0x0, 4);
145 va1x0 = vext_s8(va1x0, va1x0, 4);
146 va2x0 = vext_s8(va2x0, va2x0, 4);
147 va3x0 = vext_s8(va3x0, va3x0, 4);
148 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
149 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
150 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
151 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
152 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
153 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
154 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
155 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
156 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
157 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
158 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
159 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
160 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
161 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
162 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
163 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
164 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
165 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
166 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
167 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
168 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
169 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
170 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
171 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
172 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
173 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
174 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
175 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
176 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
177 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
178 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
179 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
180
181 k -= 8 * sizeof(int8_t);
182 } while (k != 0);
183
184 p -= 4 * sizeof(void*);
185 } while (p != 0);
186
187 #if XNN_ARCH_ARM64
188 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
189 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
190 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
191 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
192 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
193 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
194 int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
195 int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
196 #else
197 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
198 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
199 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
200 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
201 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
202 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
203 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
204 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
205 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
206 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
207 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
208 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
209 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
210 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
211 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
212 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
213 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
214 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
215 const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
216 const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
217 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
218 const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
219 const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
220 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
221 #endif
222
223 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
224 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
225 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
226
227 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
228 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
229 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
230 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
231 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
232 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
233 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
234 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
235
236 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
237 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
238 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
239 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
240 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
241 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
242 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
243 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
244
245 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
246 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
247 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
248 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
249 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
250 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
251 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
252 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
253
254 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
255 #if XNN_ARCH_ARM64
256 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
257 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
258 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
259 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
260
261 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
262 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
263 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
264 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
265
266 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
267 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
268 #else
269 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
270 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
271 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
272 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
273
274 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
275 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
276 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
277 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
278
279 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
280 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
281 #endif
282
283 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
284 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
285 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
286
287 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
288 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
289 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
290
291 if (nc >= 8) {
292 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
293 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
294 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
295 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
296
297 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
298 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
299 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
300 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
301
302 a = (const int8_t**restrict) ((uintptr_t) a - ks);
303
304 nc -= 8;
305 } else {
306 if (nc & 4) {
307 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
308 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
309 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
310 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
311 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
312 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
313 }
314 if (nc & 2) {
315 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
316 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
317 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
318 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
319 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
320 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
321 }
322 if (nc & 1) {
323 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
324 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
325 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
326 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
327 }
328
329 nc = 0;
330 }
331 } while (nc != 0);
332 }
333