• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c2-neon-mull-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 
xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup(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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
29     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 4);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (4 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
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   do {
58     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
59     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
60     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
61     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
62     int32x4_t vacc1x0123 = vacc0x0123;
63     int32x4_t vacc1x4567 = vacc0x4567;
64     int32x4_t vacc1x89AB = vacc0x89AB;
65     int32x4_t vacc1xCDEF = vacc0xCDEF;
66     int32x4_t vacc2x0123 = vacc0x0123;
67     int32x4_t vacc2x4567 = vacc0x4567;
68     int32x4_t vacc2x89AB = vacc0x89AB;
69     int32x4_t vacc2xCDEF = vacc0xCDEF;
70     int32x4_t vacc3x0123 = vacc0x0123;
71     int32x4_t vacc3x4567 = vacc0x4567;
72     int32x4_t vacc3x89AB = vacc0x89AB;
73     int32x4_t vacc3xCDEF = vacc0xCDEF;
74 
75     size_t p = ks;
76     do {
77       const int8_t* restrict a0 = a[0];
78       if XNN_UNPREDICTABLE(a0 != zero) {
79         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
80       }
81       const int8_t* restrict a1 = a[1];
82       if XNN_UNPREDICTABLE(a1 != zero) {
83         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
84       }
85       const int8_t* restrict a2 = a[2];
86       if XNN_UNPREDICTABLE(a2 != zero) {
87         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
88       }
89       const int8_t* restrict a3 = a[3];
90       if XNN_UNPREDICTABLE(a3 != zero) {
91         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
92       }
93       a += 4;
94 
95       size_t k = kc;
96 
97 
98       while (k >= 8 * sizeof(int8_t)) {
99         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
100         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
101         const int8x8_t va2 = vld1_s8(a2); a2 += 8;
102         const int8x8_t va3 = vld1_s8(a3); a3 += 8;
103 
104         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 
121         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
122         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
123         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
124         const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
125 
126         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
127         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
128         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
129         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
130         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
131         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
132         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
133         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
134         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
135         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
136         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
137         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
138         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
139         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
140         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
141         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
142         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
143         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
144         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
145         const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
146         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
147         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
148         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
149         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
150         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
151         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
152         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
153         const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
154         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
155         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
156         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
157         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
158         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
159         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
160         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
161         const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
162 
163         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
164         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
165         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
166         const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
167         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
168         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
169         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
170         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
171         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
172         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
173         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
174         const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
175         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
176         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
177         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
178         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
179         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
180         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
181         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
182         const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
183         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
184         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
185         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
186         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
187         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
188         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
189         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
190         const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
191         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
192         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
193         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
194         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
195         const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
196         const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
197         const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
198         const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
199 
200         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
201         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
202         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
203         const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
204         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
205         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
206         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
207         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
208         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
209         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
210         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
211         const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
212         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
213         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
214         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
215         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
216         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
217         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
218         const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
219         const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
220         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
221         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
222         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
223         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
224         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
225         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
226         const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
227         const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
228         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
229         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
230         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
231         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
232         const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
233         const int8x8_t va1c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3));
234         const int8x8_t va2c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3));
235         const int8x8_t va3c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 3));
236 
237         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
238         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
239         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
240         const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
241         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
242         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
243         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
244         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
245         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
246         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
247         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
248         const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
249         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
250         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
251         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
252         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
253         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
254         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
255         const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
256         const int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3, va3c3);
257         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
258         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
259         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
260         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
261         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
262         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
263         const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
264         const int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3, va3c3);
265         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
266         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
267         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
268         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
269 
270         k -= 8 * sizeof(int8_t);
271       }
272 
273       if XNN_UNLIKELY(k != 0) {
274         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
275         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
276         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
277         const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
278 
279         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
280         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
281         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
282         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
283 
284         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
285         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
286         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
287         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
288         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
289         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
290         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
291         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
292         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
293         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
294         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
295         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
296         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
297         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
298         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
299         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
300         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
301         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
302         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
303         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
304         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
305         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
306         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
307         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
308         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
309         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
310         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
311         const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
312         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
313         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
314         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
315         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
316         const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
317         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
318         const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
319         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
320 
321         if (k > 2 * sizeof(int8_t)) {
322           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
323           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
324           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
325           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
326 
327           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
328           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
329           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
330           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
331           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
332           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
333           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
334           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
335           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
336           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
337           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
338           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
339           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
340           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
341           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
342           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
343           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
344           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
345           const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
346           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
347           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
348           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
349           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
350           const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
351           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
352           const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
353           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
354           const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
355           const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
356           vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
357           const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
358           vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
359           const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
360           vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
361           const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
362           vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
363 
364           if (k > 4 * sizeof(int8_t)) {
365             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
366             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
367             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
368             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
369 
370             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
371             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
372             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
373             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
374             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
375             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
376             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
377             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
378             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
379             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
380             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
381             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
382             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
383             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
384             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
385             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
386             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
387             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
388             const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
389             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
390             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
391             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
392             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
393             const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
394             vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
395             const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
396             vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
397             const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
398             const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
399             vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
400             const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
401             vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
402             const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
403             vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
404             const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
405             vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
406           }
407         }
408       }
409       p -= 4 * sizeof(void*);
410     } while (p != 0);
411 
412     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
413     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
414     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
415 
416     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
417     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
418     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
419     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
420     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
421     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
422     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
423     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
424     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
425     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
426     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
427     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
428     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
429     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
430     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
431     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
432 
433     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
434     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
435     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
436     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
437     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
438     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
439     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
440     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
441     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
442     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
443     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
444     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
445     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
446     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
447     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
448     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
449 
450     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
451     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
452     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
453     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
454     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
455     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
456     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
457     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
458     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
459     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
460     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
461     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
462     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
463     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
464     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
465     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
466 
467     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
468 #if XNN_ARCH_ARM64
469     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
470     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
471     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
472     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
473     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
474     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
475     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
476     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
477 
478     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
479     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
480     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
481     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
482     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
483     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
484     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
485     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
486 
487     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
488     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
489     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
490     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
491 #else
492     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
493     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
494     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
495     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
496     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
497     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
498     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
499     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
500 
501     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
502     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
503     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
504     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
505     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
506     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
507     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
508     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
509 
510     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
511     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
512     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
513     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
514 #endif
515 
516     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
517     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
518     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
519     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
520     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
521 
522     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
523     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
524     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
525     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
526     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
527 
528     if (nc >= 16) {
529       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
530       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
531       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
532       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
533 
534       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
535       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
536       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
537       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
538 
539       a = (const int8_t**restrict) ((uintptr_t) a - ks);
540 
541       nc -= 16;
542     } else {
543       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
544       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
545       if (nc & 8) {
546         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
547         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
548         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
549         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
550         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
551         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
552       }
553       if (nc & 4) {
554         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
555         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
556         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
557         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
558         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
559         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
560       }
561       if (nc & 2) {
562         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
563         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
564         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
565         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
566         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
567         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
568       }
569       if (nc & 1) {
570         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
571         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
572         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
573         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
574       }
575 
576       nc = 0;
577     }
578   } while (nc != 0);
579 }
580