xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/4x16c2-minmax-rndnu-neon-mull-ld2r.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_ld2r(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_ld2r(
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 int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
100         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
101         const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
102         const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
103         const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
104         const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
105         const int16x4x2_t va30 = vld2_dup_s16((const void*)a3);
106         const int16x4x2_t va31 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
107 
108         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 
125         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
126         const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
127         const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
128         const int8x8_t va3c0 = vreinterpret_s8_s16(va30.val[0]);
129 
130         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
131         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
132         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
133         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
134         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
135         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
136         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
137         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
138         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
139         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
140         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
141         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
142         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
143         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
144         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
145         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
146         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
147         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
148         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
149         const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
150         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
151         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
152         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
153         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
154         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
155         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
156         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
157         const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
158         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
159         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
160         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
161         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
162         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
163         const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
164         const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
165         const int8x8_t va3c1 = vreinterpret_s8_s16(va30.val[1]);
166 
167         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
168         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
169         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
170         const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
171         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
172         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
173         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
174         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
175         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
176         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
177         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
178         const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
179         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
180         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
181         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
182         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
183         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
184         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
185         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
186         const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
187         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
188         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
189         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
190         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
191         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
192         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
193         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
194         const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
195         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
196         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
197         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
198         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
199         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
200         const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
201         const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
202         const int8x8_t va3c2 = vreinterpret_s8_s16(va31.val[0]);
203 
204         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
205         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
206         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
207         const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
208         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
209         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
210         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
211         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
212         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
213         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
214         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
215         const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
216         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
217         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
218         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
219         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
220         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
221         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
222         const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
223         const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
224         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
225         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
226         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
227         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
228         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
229         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
230         const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
231         const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
232         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
233         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
234         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
235         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
236         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
237         const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
238         const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
239         const int8x8_t va3c3 = vreinterpret_s8_s16(va31.val[1]);
240 
241         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
242         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
243         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
244         const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
245         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
246         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
247         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
248         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
249         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
250         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
251         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
252         const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
253         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
254         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
255         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
256         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
257         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
258         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
259         const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
260         const int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3, va3c3);
261         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
262         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
263         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
264         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
265         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
266         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
267         const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
268         const int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3, va3c3);
269         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
270         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
271         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
272         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
273 
274         k -= 8 * sizeof(int8_t);
275       }
276 
277       if XNN_UNLIKELY(k != 0) {
278         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
279         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
280         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
281         const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
282 
283         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
284         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
285         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
286         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
287 
288         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
289         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
290         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
291         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
292         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
293         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
294         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
295         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
296         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
297         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
298         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
299         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
300         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
301         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
302         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
303         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
304         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
305         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
306         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
307         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
308         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
309         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
310         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
311         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
312         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
313         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
314         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
315         const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
316         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
317         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
318         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
319         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
320         const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
321         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
322         const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
323         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
324 
325         if (k > 2 * sizeof(int8_t)) {
326           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
327           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
328           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
329           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
330 
331           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
332           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
333           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
334           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
335           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
336           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
337           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
338           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
339           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
340           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
341           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
342           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
343           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
344           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
345           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
346           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
347           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
348           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
349           const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
350           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
351           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
352           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
353           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
354           const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
355           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
356           const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
357           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
358           const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
359           const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
360           vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
361           const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
362           vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
363           const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
364           vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
365           const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
366           vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
367 
368           if (k > 4 * sizeof(int8_t)) {
369             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
370             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
371             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
372             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
373 
374             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
375             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
376             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
377             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
378             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
379             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
380             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
381             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
382             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
383             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
384             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
385             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
386             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
387             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
388             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
389             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
390             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
391             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
392             const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
393             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
394             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
395             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
396             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
397             const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
398             vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
399             const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
400             vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
401             const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
402             const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
403             vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
404             const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
405             vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
406             const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
407             vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
408             const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
409             vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
410           }
411         }
412       }
413       p -= 4 * sizeof(void*);
414     } while (p != 0);
415 
416     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
417     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
418     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
419 
420     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
421     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
422     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
423     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
424     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
425     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
426     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
427     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
428     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
429     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
430     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
431     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
432     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
433     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
434     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
435     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
436 
437     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
438     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
439     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
440     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
441     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
442     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
443     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
444     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
445     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
446     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
447     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
448     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
449     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
450     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
451     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
452     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
453 
454     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
455     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
456     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
457     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
458     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
459     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
460     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
461     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
462     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
463     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
464     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
465     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
466     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
467     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
468     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
469     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
470 
471     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
472 #if XNN_ARCH_ARM64
473     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
474     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
475     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
476     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
477     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
478     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
479     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
480     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
481 
482     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
483     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
484     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
485     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
486     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
487     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
488     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
489     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
490 
491     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
492     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
493     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
494     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
495 #else
496     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
497     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
498     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
499     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
500     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
501     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
502     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
503     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
504 
505     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
506     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
507     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
508     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
509     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
510     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
511     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
512     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
513 
514     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
515     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
516     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
517     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
518 #endif
519 
520     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
521     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
522     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
523     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
524     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
525 
526     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
527     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
528     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
529     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
530     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
531 
532     if (nc >= 16) {
533       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
534       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
535       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
536       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
537 
538       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
539       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
540       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
541       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
542 
543       a = (const int8_t**restrict) ((uintptr_t) a - ks);
544 
545       nc -= 16;
546     } else {
547       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
548       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
549       if (nc & 8) {
550         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
551         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
552         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
553         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
554         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
555         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
556       }
557       if (nc & 4) {
558         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
559         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
560         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
561         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
562         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
563         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
564       }
565       if (nc & 2) {
566         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
567         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
568         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
569         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
570         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
571         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
572       }
573       if (nc & 1) {
574         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
575         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
576         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
577         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
578       }
579 
580       nc = 0;
581     }
582   } while (nc != 0);
583 }
584