xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mlal.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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_3x16c4s2__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_3x16c4s2__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 <= 3);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (3 * 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 
53   kc = round_up_po2(kc, 8 * sizeof(int8_t));
54   do {
55     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
56     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
59     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
60     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
61     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
62     int32x4_t vacc0xEF = 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 vacc1x89 = vacc0x89;
68     int32x4_t vacc1xAB = vacc0xAB;
69     int32x4_t vacc1xCD = vacc0xCD;
70     int32x4_t vacc1xEF = vacc0xEF;
71     int32x4_t vacc2x01 = vacc0x01;
72     int32x4_t vacc2x23 = vacc0x23;
73     int32x4_t vacc2x45 = vacc0x45;
74     int32x4_t vacc2x67 = vacc0x67;
75     int32x4_t vacc2x89 = vacc0x89;
76     int32x4_t vacc2xAB = vacc0xAB;
77     int32x4_t vacc2xCD = vacc0xCD;
78     int32x4_t vacc2xEF = vacc0xEF;
79 
80     size_t p = ks;
81     do {
82       const int8_t* restrict a0 = a[0];
83       if XNN_UNPREDICTABLE(a0 != zero) {
84         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
85       }
86       const int8_t* restrict a1 = a[1];
87       if XNN_UNPREDICTABLE(a1 != zero) {
88         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
89       }
90       const int8_t* restrict a2 = a[2];
91       if XNN_UNPREDICTABLE(a2 != zero) {
92         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93       }
94       a += 3;
95 
96       size_t k = kc;
97       while (k >= 16 * sizeof(int8_t)) {
98         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
99         int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
100         int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
101         int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
102         int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
103         int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
104 
105         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
106         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
107         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109         const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110         const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111         const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
112         const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
113         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
114         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
115         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
116         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
117         const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
118         const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
119         const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
120         const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
121 
122         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
123         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
124         int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
125         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
126         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
127         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
128         vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2x1);
129         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
130         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
131         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
132         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
133         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
134         int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
135         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
136         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
137         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
138         vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1);
139         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
140         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
141         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
142         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
143         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
144         int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
145         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
146         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
147         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
148         vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2x1);
149         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
150         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
151         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
152         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
153         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
154         int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
155         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
156         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
157         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
158         vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1);
159         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
160         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
161         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
162         int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
163         int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
164         int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
165         const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
166         vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1);
167         vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1);
168         vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2x1);
169         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
170         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
171         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
172         int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
173         int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
174         int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
175         const int8x8_t vbABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
176         vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0x1);
177         vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1x1);
178         vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2x1);
179         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
180         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
181         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
182         int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
183         int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
184         int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
185         const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
186         vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0x1);
187         vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1x1);
188         vprod2xCDc0 = vmlal_s8(vprod2xCDc0, vbCDc0x1, va2x1);
189         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
190         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
191         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
192         int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
193         int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
194         int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
195         const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
196         vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0x1);
197         vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1x1);
198         vprod2xEFc0 = vmlal_s8(vprod2xEFc0, vbEFc0x1, va2x1);
199         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
200         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
201         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
202         va0x0 = vext_s8(va0x0, va0x0, 4);
203         va0x1 = vext_s8(va0x1, va0x1, 4);
204         va1x0 = vext_s8(va1x0, va1x0, 4);
205         va1x1 = vext_s8(va1x1, va1x1, 4);
206         va2x0 = vext_s8(va2x0, va2x0, 4);
207         va2x1 = vext_s8(va2x1, va2x1, 4);
208         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
209         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
210         int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
211         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
212         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
213         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
214         vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2x1);
215         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
216         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
217         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
218         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
219         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
220         int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
221         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
222         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
223         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
224         vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2x1);
225         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
226         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
227         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
228         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
229         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
230         int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
231         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
232         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
233         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
234         vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2x1);
235         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
236         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
237         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
238         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
239         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
240         int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
241         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
242         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
243         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
244         vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2x1);
245         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
246         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
247         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
248         int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
249         int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
250         int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
251         const int8x8_t vb89c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
252         vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0x1);
253         vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1);
254         vprod2x89c1 = vmlal_s8(vprod2x89c1, vb89c1x1, va2x1);
255         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
256         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
257         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
258         int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
259         int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
260         int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
261         const int8x8_t vbABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
262         vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0x1);
263         vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1);
264         vprod2xABc1 = vmlal_s8(vprod2xABc1, vbABc1x1, va2x1);
265         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
266         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
267         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
268         int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
269         int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
270         int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
271         const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
272         vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0x1);
273         vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1x1);
274         vprod2xCDc1 = vmlal_s8(vprod2xCDc1, vbCDc1x1, va2x1);
275         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
276         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
277         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
278         int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
279         int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
280         int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
281         const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
282         vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0x1);
283         vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1x1);
284         vprod2xEFc1 = vmlal_s8(vprod2xEFc1, vbEFc1x1, va2x1);
285         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
286         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
287         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
288 
289         k -= 16 * sizeof(int8_t);
290       }
291       if (k != 0) {
292         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
293         int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
294         int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
295 
296         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
297         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
298         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
299         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
300         const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
301         const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
302         const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
303         const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
304         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
305         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
306         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
307         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
308         const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
309         const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
310         const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
311         const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
312 
313         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
314         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
315         int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
316         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
317         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
318         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
319         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
320         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
321         int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
322         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
323         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
324         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
325         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
326         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
327         int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
328         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
329         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
330         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
331         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
332         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
333         int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
334         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
335         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
336         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
337         int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
338         int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
339         int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
340         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
341         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
342         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
343         int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
344         int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
345         int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
346         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
347         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
348         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
349         int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
350         int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
351         int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
352         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
353         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
354         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
355         int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
356         int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
357         int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
358         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
359         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
360         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
361         va0x0 = vext_s8(va0x0, va0x0, 4);
362         va1x0 = vext_s8(va1x0, va1x0, 4);
363         va2x0 = vext_s8(va2x0, va2x0, 4);
364         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
365         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
366         int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
367         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
368         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
369         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
370         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
371         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
372         int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
373         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
374         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
375         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
376         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
377         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
378         int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
379         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
380         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
381         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
382         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
383         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
384         int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
385         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
386         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
387         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
388         int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
389         int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
390         int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
391         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
392         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
393         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
394         int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
395         int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
396         int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
397         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
398         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
399         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
400         int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
401         int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
402         int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
403         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
404         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
405         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
406         int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
407         int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
408         int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
409         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
410         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
411         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
412 
413       }
414 
415       p -= 3 * sizeof(void*);
416     } while (p != 0);
417 
418 #if XNN_ARCH_ARM64
419     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
420     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
421     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
422     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
423     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
424     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
425     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
426     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
427     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
428     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
429     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
430     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
431 #else
432     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
433     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
434     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
435     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
436     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
437     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
438     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
439     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
440     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
441     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
442     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
443     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
444     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
445     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
446     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
447     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
448     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
449     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
450     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
451     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
452     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
453     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
454     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
455     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
456     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
457     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
458     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
459     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
460     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
461     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
462     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
463     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
464     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
465     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
466     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
467     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
468 #endif
469 
470     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
471     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
472     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
473 
474     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
475     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
476     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
477     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
478     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
479     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
480     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
481     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
482     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
483     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
484     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
485     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
486 
487     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
488     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
489     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
490     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
491     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
492     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
493     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
494     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
495     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
496     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
497     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
498     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
499 
500     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
501     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
502     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
503     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
504     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
505     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
506     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
507     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
508     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
509     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
510     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
511     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
512 
513     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
514 #if XNN_ARCH_ARM64
515     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
516     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
517     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
518     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
519     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
520     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
521 
522     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
523     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
524     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
525     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
526     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
527     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
528 
529     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
530     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
531     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
532 #else
533     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
534     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
535     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
536     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
537     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
538     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
539 
540     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
541     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
542     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
543     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
544     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
545     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
546 
547     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
548     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
549     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
550 #endif
551 
552     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
553     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
554     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
555     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
556 
557     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
558     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
559     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
560     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
561 
562     if (nc >= 16) {
563       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
564       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
565       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
566 
567       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
568       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
569       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
570 
571       a = (const int8_t**restrict) ((uintptr_t) a - ks);
572 
573       nc -= 16;
574     } else {
575       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
576       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
577       if (nc & 8) {
578         vst1_s8(c2, vout2x01234567); c2 += 8;
579         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
580         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
581         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
582         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
583       }
584       if (nc & 4) {
585         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
586         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
587         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
588         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
589         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
590       }
591       if (nc & 2) {
592         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
593         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
594         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
595         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
596         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
597       }
598       if (nc & 1) {
599         vst1_lane_s8(c2, vout2x01234567, 0);
600         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
601         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
602       }
603 
604       nc = 0;
605     }
606   } while (nc != 0);
607 }
608