xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/4x8c2-minmax-rndnu-neon-mlal-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_4x8c2__neon_mlal_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_4x8c2__neon_mlal_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 vacc1x0123 = vacc0x0123;
61     int32x4_t vacc1x4567 = vacc0x4567;
62     int32x4_t vacc2x0123 = vacc0x0123;
63     int32x4_t vacc2x4567 = vacc0x4567;
64     int32x4_t vacc3x0123 = vacc0x0123;
65     int32x4_t vacc3x4567 = vacc0x4567;
66 
67     size_t p = ks;
68     do {
69       const int8_t* restrict a0 = a[0];
70       if XNN_UNPREDICTABLE(a0 != zero) {
71         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72       }
73       const int8_t* restrict a1 = a[1];
74       if XNN_UNPREDICTABLE(a1 != zero) {
75         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
76       }
77       const int8_t* restrict a2 = a[2];
78       if XNN_UNPREDICTABLE(a2 != zero) {
79         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
80       }
81       const int8_t* restrict a3 = a[3];
82       if XNN_UNPREDICTABLE(a3 != zero) {
83         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
84       }
85       a += 4;
86 
87       size_t k = kc;
88 
89       while (k >= 16 * sizeof(int8_t)) {
90         const int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
91         const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
92         const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
93         const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
94         const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
95         const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
96         const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
97         const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
98         const int16x4_t va10x0 = vld1_dup_s16((const void*)a1);
99         const int16x4_t va11x0 = vld1_dup_s16((const void*)(a1 + 2));
100         const int16x4_t va12x0 = vld1_dup_s16((const void*)(a1 + 4));
101         const int16x4_t va13x0 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
102         const int16x4_t va10x1 = vld1_dup_s16((const void*)a1);
103         const int16x4_t va11x1 = vld1_dup_s16((const void*)(a1 + 2));
104         const int16x4_t va12x1 = vld1_dup_s16((const void*)(a1 + 4));
105         const int16x4_t va13x1 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
106         const int16x4_t va20x0 = vld1_dup_s16((const void*)a2);
107         const int16x4_t va21x0 = vld1_dup_s16((const void*)(a2 + 2));
108         const int16x4_t va22x0 = vld1_dup_s16((const void*)(a2 + 4));
109         const int16x4_t va23x0 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
110         const int16x4_t va20x1 = vld1_dup_s16((const void*)a2);
111         const int16x4_t va21x1 = vld1_dup_s16((const void*)(a2 + 2));
112         const int16x4_t va22x1 = vld1_dup_s16((const void*)(a2 + 4));
113         const int16x4_t va23x1 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
114         const int16x4_t va30x0 = vld1_dup_s16((const void*)a3);
115         const int16x4_t va31x0 = vld1_dup_s16((const void*)(a3 + 2));
116         const int16x4_t va32x0 = vld1_dup_s16((const void*)(a3 + 4));
117         const int16x4_t va33x0 = vld1_dup_s16((const void*)(a3 + 6)); a3 += 8;
118         const int16x4_t va30x1 = vld1_dup_s16((const void*)a3);
119         const int16x4_t va31x1 = vld1_dup_s16((const void*)(a3 + 2));
120         const int16x4_t va32x1 = vld1_dup_s16((const void*)(a3 + 4));
121         const int16x4_t va33x1 = vld1_dup_s16((const void*)(a3 + 6)); a3 += 8;
122 
123         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 
132         const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
133         const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
134         const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0);
135         const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1);
136         const int8x8_t va2c0x0 = vreinterpret_s8_s16(va20x0);
137         const int8x8_t va2c0x1 = vreinterpret_s8_s16(va20x1);
138         const int8x8_t va3c0x0 = vreinterpret_s8_s16(va30x0);
139         const int8x8_t va3c0x1 = vreinterpret_s8_s16(va30x1);
140 
141         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
142         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
143         int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
144         int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3c0x0);
145         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
147         vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
148         vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
149         vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, va3c0x1);
150         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
151         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
152         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
153         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
154         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
155         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
156         int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
157         int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0);
158         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
160         vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
161         vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
162         vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, va3c0x1);
163         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
164         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
165         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
166         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
167         const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
168         const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
169         const int8x8_t va1c1x0 = vreinterpret_s8_s16(va11x0);
170         const int8x8_t va1c1x1 = vreinterpret_s8_s16(va11x1);
171         const int8x8_t va2c1x0 = vreinterpret_s8_s16(va21x0);
172         const int8x8_t va2c1x1 = vreinterpret_s8_s16(va21x1);
173         const int8x8_t va3c1x0 = vreinterpret_s8_s16(va31x0);
174         const int8x8_t va3c1x1 = vreinterpret_s8_s16(va31x1);
175 
176         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
177         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
178         int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
179         int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0);
180         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
182         vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
183         vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
184         vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1);
185         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
186         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
187         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
188         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
189         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
190         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
191         int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
192         int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3c1x0);
193         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
194         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
195         vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
196         vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
197         vprod3x4567c1 = vmlal_s8(vprod3x4567c1, vb4567c1x1, va3c1x1);
198         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
199         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
200         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
201         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
202         const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
203         const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
204         const int8x8_t va1c2x0 = vreinterpret_s8_s16(va12x0);
205         const int8x8_t va1c2x1 = vreinterpret_s8_s16(va12x1);
206         const int8x8_t va2c2x0 = vreinterpret_s8_s16(va22x0);
207         const int8x8_t va2c2x1 = vreinterpret_s8_s16(va22x1);
208         const int8x8_t va3c2x0 = vreinterpret_s8_s16(va32x0);
209         const int8x8_t va3c2x1 = vreinterpret_s8_s16(va32x1);
210 
211         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
212         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
213         int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
214         int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3c2x0);
215         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
217         vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
218         vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
219         vprod3x0123c2 = vmlal_s8(vprod3x0123c2, vb0123c2x1, va3c2x1);
220         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
221         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
222         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
223         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
224         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
225         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
226         int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
227         int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0);
228         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
230         vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
231         vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
232         vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1);
233         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
234         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
235         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
236         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
237         const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
238         const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
239         const int8x8_t va1c3x0 = vreinterpret_s8_s16(va13x0);
240         const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1);
241         const int8x8_t va2c3x0 = vreinterpret_s8_s16(va23x0);
242         const int8x8_t va2c3x1 = vreinterpret_s8_s16(va23x1);
243         const int8x8_t va3c3x0 = vreinterpret_s8_s16(va33x0);
244         const int8x8_t va3c3x1 = vreinterpret_s8_s16(va33x1);
245 
246         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
247         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
248         int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
249         int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3c3x0);
250         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
252         vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
253         vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
254         vprod3x0123c3 = vmlal_s8(vprod3x0123c3, vb0123c3x1, va3c3x1);
255         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
256         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
257         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
258         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
259         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
260         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
261         int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
262         int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0);
263         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
264         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
265         vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
266         vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
267         vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1);
268         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
269         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
270         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
271         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
272 
273         k -= 16 * sizeof(int8_t);
274       }
275 
276       if (k >= 8 * sizeof(int8_t)) {
277         const int16x4_t va00 = vld1_dup_s16((const void*)a0);
278         const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
279         const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
280         const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
281         const int16x4_t va10 = vld1_dup_s16((const void*)a1);
282         const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
283         const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
284         const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
285         const int16x4_t va20 = vld1_dup_s16((const void*)a2);
286         const int16x4_t va21 = vld1_dup_s16((const void*)(a2 + 2));
287         const int16x4_t va22 = vld1_dup_s16((const void*)(a2 + 4));
288         const int16x4_t va23 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
289         const int16x4_t va30 = vld1_dup_s16((const void*)a3);
290         const int16x4_t va31 = vld1_dup_s16((const void*)(a3 + 2));
291         const int16x4_t va32 = vld1_dup_s16((const void*)(a3 + 4));
292         const int16x4_t va33 = vld1_dup_s16((const void*)(a3 + 6)); a3 += 8;
293 
294         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
295         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
296         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
297         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
298         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
299         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
300         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
301         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
302 
303         const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
304         const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
305         const int8x8_t va2c0 = vreinterpret_s8_s16(va20);
306         const int8x8_t va3c0 = vreinterpret_s8_s16(va30);
307 
308         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
309         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
310         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
311         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
312         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
313         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
314         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
315         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
316         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
317         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
318         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
319         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
320         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
321         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
322         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
323         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
324         const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
325         const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
326         const int8x8_t va2c1 = vreinterpret_s8_s16(va21);
327         const int8x8_t va3c1 = vreinterpret_s8_s16(va31);
328 
329         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
330         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
331         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
332         const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
333         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
334         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
335         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
336         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
337         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
338         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
339         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
340         const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
341         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
342         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
343         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
344         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
345         const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
346         const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
347         const int8x8_t va2c2 = vreinterpret_s8_s16(va22);
348         const int8x8_t va3c2 = vreinterpret_s8_s16(va32);
349 
350         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
351         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
352         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
353         const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
354         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
355         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
356         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
357         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
358         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
359         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
360         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
361         const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
362         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
363         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
364         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
365         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
366         const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
367         const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
368         const int8x8_t va2c3 = vreinterpret_s8_s16(va23);
369         const int8x8_t va3c3 = vreinterpret_s8_s16(va33);
370 
371         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
372         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
373         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
374         const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
375         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
376         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
377         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
378         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
379         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
380         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
381         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
382         const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
383         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
384         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
385         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
386         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
387 
388         k -= 8 * sizeof(int8_t);
389       }
390 
391       if XNN_UNLIKELY(k != 0) {
392         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
393         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
394         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
395         const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
396 
397         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
398         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
399 
400         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
401         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
402         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
403         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
404         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
405         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
406         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
407         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
408         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
409         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
410         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
411         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
412         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
413         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
414         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
415         const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
416         const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
417         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
418         const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
419         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
420 
421         if (k > 2 * sizeof(int8_t)) {
422           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
423           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
424 
425           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
426           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
427           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
428           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
429           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
430           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
431           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
432           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
433           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
434           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
435           const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
436           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
437           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
438           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
439           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
440           const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
441           const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
442           vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
443           const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
444           vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
445 
446           if (k > 4 * sizeof(int8_t)) {
447             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
448             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
449 
450             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
451             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
452             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
453             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
454             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
455             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
456             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
457             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
458             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
459             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
460             const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
461             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
462             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
463             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
464             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
465             const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
466             const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
467             vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
468             const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
469             vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
470           }
471         }
472       }
473       p -= 4 * sizeof(void*);
474     } while (p != 0);
475 
476     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
477     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
478     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
479 
480     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
481     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
482     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
483     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
484     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
485     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
486     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
487     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
488 
489     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
490     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
491     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
492     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
493     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
494     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
495     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
496     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
497 
498     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
499     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
500     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
501     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
502     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
503     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
504     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
505     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
506 
507     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
508 #if XNN_ARCH_ARM64
509     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
510     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
511     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
512     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
513 
514     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
515     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
516     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
517     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
518 
519     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
520     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
521 #else
522     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
523     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
524     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
525     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
526 
527     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
528     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
529     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
530     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
531 
532     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
533     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
534 #endif
535 
536     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
537     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
538     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
539 
540     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
541     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
542     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
543 
544     if (nc >= 8) {
545       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
546       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
547       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
548       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
549 
550       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
551       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
552       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
553       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
554 
555       a = (const int8_t**restrict) ((uintptr_t) a - ks);
556 
557       nc -= 8;
558     } else {
559       if (nc & 4) {
560         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
561         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
562         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
563         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
564         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
565         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
566       }
567       if (nc & 2) {
568         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
569         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
570         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
571         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
572         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
573         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
574       }
575       if (nc & 1) {
576         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
577         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
578         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
579         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
580       }
581 
582       nc = 0;
583     }
584   } while (nc != 0);
585 }
586