1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-padal-dup.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_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30   assert(mr != 0);
31   assert(mr <= 3);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 2);
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54 
55   do {
56     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
58     int32x4_t vacc1x0123 = vacc0x0123;
59     int32x4_t vacc1x4567 = vacc0x4567;
60     int32x4_t vacc2x0123 = vacc0x0123;
61     int32x4_t vacc2x4567 = vacc0x4567;
62 
63     size_t k = kc;
64 
65 
66     while (k >= 8 * sizeof(int8_t)) {
67       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
68       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
69       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
70 
71       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 
80       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
81       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
82       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
83       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
84       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
85       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
86       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
87       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
88       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
89       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
90       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
91       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
92       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
93       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
94       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
95       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
96       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
97       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
98       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
99       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
100       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
101       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
102       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
103       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
104       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
105       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
106       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
107       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
108       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
109       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
110       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
111       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
112       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
113       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
114       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
115       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
116       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
117       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
118       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
119       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
120       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
121       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
122       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
123       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
124       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
125       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
126       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
127       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
128 
129       k -= 8 * sizeof(int8_t);
130     }
131 
132     if XNN_UNLIKELY(k != 0) {
133       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
134       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
135       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
136 
137       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 
140       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
141       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
142       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
143       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
144       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
145       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
146       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
147       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
148       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
149       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
150       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
151       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
152 
153       if (k > 2 * sizeof(int8_t)) {
154         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156 
157         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
158         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
159         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
160         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
161         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
162         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
163         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
164         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
165         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
166         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
167         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
168         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
169 
170         if (k > 4 * sizeof(int8_t)) {
171           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 
174           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
175           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
176           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
177           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
178           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
179           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
180           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
181           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
182           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
183           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
184           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
185           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
186         }
187       }
188     }
189     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
190     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
191     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
192     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
193     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
194     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
195     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
196 
197     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
198     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
199     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
200     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
201     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
202     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
203     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
204     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
205 
206     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
207     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
208     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
209     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
210     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
211     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
212 
213     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
214 #if XNN_ARCH_ARM64
215     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
216     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
217     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
218 
219     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
220     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
221 #else
222     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
223     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
224     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
225 
226     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
227     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
228 #endif
229     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
230     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
231 
232     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
233     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
234 
235     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
236     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
237 
238     if (nc >= 8) {
239       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
240       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
241       vst1_s8(c2 + 0, vout2x01234567);
242 
243       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
244       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
245       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
246 
247       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
248       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
249       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
250 
251       nc -= 8;
252     } else {
253       if (nc & 4) {
254         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
255         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
256         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
257         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
258         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
259       }
260       if (nc & 2) {
261         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
262         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
263         vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
264         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
265         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
266       }
267       if (nc & 1) {
268         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
269         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
270         vst1_lane_s8(c2, vout2x01234567, 0);
271       }
272 
273       nc = 0;
274     }
275   } while (nc != 0);
276 }
277