xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/4x16c2-minmax-rndnu-neon-mull-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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_ld1r(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_ld1r(
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 int16x4_t va00 = vld1_dup_s16((const void*)a0);
100         const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
101         const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
102         const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
103         const int16x4_t va10 = vld1_dup_s16((const void*)a1);
104         const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
105         const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
106         const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
107         const int16x4_t va20 = vld1_dup_s16((const void*)a2);
108         const int16x4_t va21 = vld1_dup_s16((const void*)(a2 + 2));
109         const int16x4_t va22 = vld1_dup_s16((const void*)(a2 + 4));
110         const int16x4_t va23 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
111         const int16x4_t va30 = vld1_dup_s16((const void*)a3);
112         const int16x4_t va31 = vld1_dup_s16((const void*)(a3 + 2));
113         const int16x4_t va32 = vld1_dup_s16((const void*)(a3 + 4));
114         const int16x4_t va33 = vld1_dup_s16((const void*)(a3 + 6)); a3 += 8;
115 
116         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132 
133         const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
134         const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
135         const int8x8_t va2c0 = vreinterpret_s8_s16(va20);
136         const int8x8_t va3c0 = vreinterpret_s8_s16(va30);
137 
138         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
139         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
140         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
141         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
142         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
143         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
144         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
145         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
146         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
147         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
148         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
149         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
150         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
151         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
152         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
153         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
154         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
155         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
156         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
157         const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
158         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
159         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
160         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
161         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
162         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
163         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
164         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
165         const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
166         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
167         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
168         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
169         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
170         const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
171         const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
172         const int8x8_t va2c1 = vreinterpret_s8_s16(va21);
173         const int8x8_t va3c1 = vreinterpret_s8_s16(va31);
174 
175         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
176         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
177         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
178         const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
179         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
180         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
181         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
182         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
183         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
184         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
185         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
186         const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
187         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
188         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
189         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
190         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
191         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
192         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
193         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
194         const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
195         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
196         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
197         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
198         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
199         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
200         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
201         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
202         const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
203         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
204         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
205         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
206         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
207         const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
208         const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
209         const int8x8_t va2c2 = vreinterpret_s8_s16(va22);
210         const int8x8_t va3c2 = vreinterpret_s8_s16(va32);
211 
212         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
213         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
214         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
215         const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
216         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
217         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
218         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
219         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
220         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
221         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
222         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
223         const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
224         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
225         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
226         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
227         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
228         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
229         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
230         const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
231         const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
232         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
233         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
234         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
235         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
236         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
237         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
238         const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
239         const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
240         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
241         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
242         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
243         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
244         const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
245         const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
246         const int8x8_t va2c3 = vreinterpret_s8_s16(va23);
247         const int8x8_t va3c3 = vreinterpret_s8_s16(va33);
248 
249         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
250         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
251         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
252         const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
253         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
254         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
255         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
256         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
257         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
258         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
259         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
260         const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
261         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
262         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
263         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
264         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
265         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
266         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
267         const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
268         const int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3, va3c3);
269         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
270         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
271         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
272         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
273         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
274         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
275         const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
276         const int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3, va3c3);
277         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
278         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
279         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
280         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
281 
282         k -= 8 * sizeof(int8_t);
283       }
284 
285       if XNN_UNLIKELY(k != 0) {
286         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
287         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
288         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
289         const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
290 
291         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
292         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
293         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
294         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
295 
296         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
297         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
298         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
299         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
300         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
301         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
302         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
303         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
304         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
305         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
306         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
307         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
308         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
309         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
310         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
311         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
312         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
313         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
314         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
315         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
316         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
317         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
318         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
319         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
320         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
321         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
322         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
323         const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
324         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
325         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
326         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
327         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
328         const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
329         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
330         const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
331         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
332 
333         if (k > 2 * sizeof(int8_t)) {
334           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
335           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
336           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
337           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
338 
339           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
340           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
341           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
342           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
343           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
344           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
345           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
346           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
347           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
348           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
349           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
350           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
351           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
352           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
353           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
354           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
355           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
356           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
357           const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
358           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
359           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
360           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
361           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
362           const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
363           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
364           const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
365           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
366           const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
367           const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
368           vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
369           const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
370           vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
371           const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
372           vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
373           const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
374           vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
375 
376           if (k > 4 * sizeof(int8_t)) {
377             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
378             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
379             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
380             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
381 
382             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
383             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
384             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
385             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
386             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
387             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
388             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
389             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
390             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
391             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
392             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
393             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
394             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
395             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
396             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
397             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
398             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
399             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
400             const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
401             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
402             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
403             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
404             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
405             const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
406             vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
407             const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
408             vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
409             const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
410             const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
411             vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
412             const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
413             vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
414             const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
415             vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
416             const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
417             vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
418           }
419         }
420       }
421       p -= 4 * sizeof(void*);
422     } while (p != 0);
423 
424     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
425     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
426     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
427 
428     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
429     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
430     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
431     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
432     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
433     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
434     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
435     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
436     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
437     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
438     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
439     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
440     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
441     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
442     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
443     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
444 
445     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
446     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
447     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
448     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
449     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
450     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
451     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
452     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
453     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
454     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
455     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
456     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
457     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
458     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
459     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
460     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
461 
462     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
463     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
464     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
465     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
466     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
467     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
468     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
469     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
470     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
471     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
472     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
473     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
474     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
475     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
476     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
477     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
478 
479     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
480 #if XNN_ARCH_ARM64
481     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
482     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
483     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
484     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
485     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
486     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
487     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
488     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
489 
490     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
491     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
492     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
493     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
494     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
495     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
496     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
497     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
498 
499     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
500     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
501     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
502     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
503 #else
504     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
505     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
506     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
507     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
508     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
509     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
510     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
511     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
512 
513     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
514     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
515     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
516     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
517     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
518     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
519     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
520     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
521 
522     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
523     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
524     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
525     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
526 #endif
527 
528     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
529     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
530     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
531     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
532     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
533 
534     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
535     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
536     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
537     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
538     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
539 
540     if (nc >= 16) {
541       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
542       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
543       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
544       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
545 
546       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
547       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
548       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
549       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
550 
551       a = (const int8_t**restrict) ((uintptr_t) a - ks);
552 
553       nc -= 16;
554     } else {
555       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
556       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
557       if (nc & 8) {
558         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
559         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
560         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
561         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
562         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
563         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
564       }
565       if (nc & 4) {
566         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
567         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
568         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
569         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
570         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
571         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
572       }
573       if (nc & 2) {
574         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
575         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
576         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
577         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
578         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
579         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
580       }
581       if (nc & 1) {
582         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
583         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
584         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
585         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
586       }
587 
588       nc = 0;
589     }
590   } while (nc != 0);
591 }
592