• 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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const int8_t** restrict a,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const int8_t* zero,
31     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 2);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (2 * sizeof(void*)) == 0);
39   assert(a_offset % 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 
51   do {
52     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60     int32x4_t vacc1x0 = vacc0x0;
61     int32x4_t vacc1x1 = vacc0x1;
62     int32x4_t vacc1x2 = vacc0x2;
63     int32x4_t vacc1x3 = vacc0x3;
64     int32x4_t vacc1x4 = vacc0x4;
65     int32x4_t vacc1x5 = vacc0x5;
66     int32x4_t vacc1x6 = vacc0x6;
67     int32x4_t vacc1x7 = vacc0x7;
68 
69     size_t p = ks;
70     do {
71       const int8_t* restrict a0 = a[0];
72       if XNN_UNPREDICTABLE(a0 != zero) {
73         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74       }
75       const int8_t* restrict a1 = a[1];
76       if XNN_UNPREDICTABLE(a1 != zero) {
77         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78       }
79       a += 2;
80 
81       size_t k = kc;
82       // 2x partial unrolled loop to load 16 bytes at a time using MLA.
83       while (k >= 16 * sizeof(int8_t)) {
84         const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
85         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
86         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
87         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
88 
89         const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
90         const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
91         const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
92         const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
93         const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
94         const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
95         const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
96         const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
97 
98         const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
99         int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
100         int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
101         vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
102         vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
103         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
104         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
105         const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
106         int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
107         int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
108         vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
109         vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
110         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
111         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
112         const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
113         int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
114         int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
115         vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
116         vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
117         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
118         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
119         const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
120         int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
121         int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
122         vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
123         vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
124         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
125         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
126         const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
127         int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
128         int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
129         vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
130         vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
131         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
132         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
133         const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
134         int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
135         int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
136         vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
137         vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
138         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
139         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
140         const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
141         int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
142         int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
143         vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
144         vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
145         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
146         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
147         const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
148         int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
149         int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
150         vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
151         vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
152         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
153         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
154 
155         k -= 16 * sizeof(int8_t);
156       }
157 
158       // Handle 8 bytes at a time using MUL.
159       if (k != 0) {
160         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
161         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
162 
163         const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164         const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
165         const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
166         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
167         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
168         const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169         const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
170         const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
171         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
172         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
173         const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174         const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
175         const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
176         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
177         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
178         const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179         const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
180         const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
181         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
182         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
183         const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184         const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
185         const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
186         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
187         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
188         const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
189         const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
190         const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
191         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
192         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
193         const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
194         const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
195         const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
196         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
197         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
198         const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
199         const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
200         const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
201         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
202         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
203 
204         k -= 8 * sizeof(int8_t);
205       }
206 
207       p -= 2 * sizeof(void*);
208     } while (p != 0);
209 
210 #if XNN_ARCH_ARM64
211     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
212     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
213     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
214     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
215     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
216     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
217     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
218     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
219 
220     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
221     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
222     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
223     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
224 #else
225     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
226     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
227     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
228     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
229     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
230     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
231     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
232     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
233     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
234     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
235     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
236     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
237     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
238     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
239     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
240     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
241     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
242     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
243     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
244     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
245     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
246     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
247     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
248     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
249     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
250     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
251     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
252     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
253 #endif
254 
255     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
256     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
257     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
258     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
259 
260     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
261     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
262     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
263     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
264     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
265     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
266 
267     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
268     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
269     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
270     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
271 
272     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
273 #if XNN_ARCH_ARM64
274     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
275     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
276 
277     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
278     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
279 
280     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
281 #else
282     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
283     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
284 
285     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
286     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
287 
288     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
289 #endif
290 
291     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
292     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
293 
294     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
295     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
296 
297     if (nc >= 8) {
298       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
299       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
300 
301       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
302       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
303 
304       a = (const int8_t**restrict) ((uintptr_t) a - ks);
305 
306       nc -= 8;
307     } else {
308       if (nc & 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         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
312       }
313       if (nc & 2) {
314         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
315         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
316         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
317       }
318       if (nc & 1) {
319         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
320         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
321       }
322 
323       nc = 0;
324     }
325   } while (nc != 0);
326 }
327