• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c8-neon-mull.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/igemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal(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_2x8c8__neon_mlal(
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 <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (2 * 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   kc = round_up_po2(kc, 8 * sizeof(int8_t));
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr != 2) {
47     c1 = c0;
48   }
49 
50   do {
51     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59     int32x4_t vacc1x0 = vacc0x0;
60     int32x4_t vacc1x1 = vacc0x1;
61     int32x4_t vacc1x2 = vacc0x2;
62     int32x4_t vacc1x3 = vacc0x3;
63     int32x4_t vacc1x4 = vacc0x4;
64     int32x4_t vacc1x5 = vacc0x5;
65     int32x4_t vacc1x6 = vacc0x6;
66     int32x4_t vacc1x7 = vacc0x7;
67 
68     size_t p = ks;
69     do {
70       const int8_t* restrict a0 = a[0];
71       if XNN_UNPREDICTABLE(a0 != zero) {
72         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73       }
74       const int8_t* restrict a1 = a[1];
75       if XNN_UNPREDICTABLE(a1 != zero) {
76         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77       }
78       a += 2;
79 
80       size_t k = kc;
81       // 2x partial unrolled loop to load 16 bytes at a time using MLA.
82       while (k >= 16 * sizeof(int8_t)) {
83         const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
84         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
85         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
86         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
87 
88         const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
89         const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
90         const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
91         const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
92         const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
93         const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
94         const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
95         const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
96 
97         const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
98         int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
99         int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
100         vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
101         vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
102         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
103         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
104         const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
105         int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
106         int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
107         vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
108         vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
109         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
110         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
111         const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
112         int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
113         int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
114         vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
115         vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
116         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
117         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
118         const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
119         int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
120         int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
121         vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
122         vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
123         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
124         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
125         const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
126         int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
127         int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
128         vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
129         vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
130         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
131         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
132         const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
133         int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
134         int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
135         vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
136         vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
137         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
138         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
139         const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
140         int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
141         int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
142         vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
143         vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
144         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
145         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
146         const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
147         int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
148         int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
149         vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
150         vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
151         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
152         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
153 
154         k -= 16 * sizeof(int8_t);
155       }
156 
157       // Handle 8 bytes at a time using MUL.
158       if (k != 0) {
159         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
160         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
161 
162         const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163         const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
164         const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
165         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
166         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
167         const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168         const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
169         const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
170         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
171         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
172         const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173         const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
174         const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
175         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
176         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
177         const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178         const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
179         const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
180         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
181         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
182         const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183         const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
184         const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
185         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
186         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
187         const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188         const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
189         const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
190         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
191         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
192         const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
193         const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
194         const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
195         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
196         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
197         const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
198         const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
199         const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
200         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
201         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
202 
203         k -= 8 * sizeof(int8_t);
204       }
205 
206       p -= 2 * sizeof(void*);
207     } while (p != 0);
208 
209 #if XNN_ARCH_ARM64
210     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
211     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
212     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
213     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
214     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
215     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
216     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
217     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
218 
219     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
220     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
221     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
222     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
223 #else
224     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
225     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
226     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
227     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
228     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
229     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
230     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
231     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
232     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
233     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
234     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
235     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
236     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
237     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
238     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
239     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
240     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
241     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
242     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
243     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
244     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
245     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
246     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
247     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
248     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
249     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
250     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
251     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
252 #endif
253 
254     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
255     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
256     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
257 
258     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
259     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
260     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
261     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
262 
263     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
264     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
265     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
266     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
267 
268     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
269     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
270     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
271     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
272 
273     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
274 #if XNN_ARCH_ARM64
275     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
276     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
277 
278     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
279     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
280 
281     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
282 #else
283     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
284     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
285 
286     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
287     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
288 
289     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
290 #endif
291 
292     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
293     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
294 
295     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
296     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
297 
298     if (nc >= 8) {
299       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
300       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
301 
302       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
303       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
304 
305       a = (const int8_t**restrict) ((uintptr_t) a - ks);
306 
307       nc -= 8;
308     } else {
309       if (nc & 4) {
310         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
311         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
312         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
313       }
314       if (nc & 2) {
315         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
316         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
317         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
318       }
319       if (nc & 1) {
320         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
321         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
322       }
323 
324       nc = 0;
325     }
326   } while (nc != 0);
327 }
328