xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/4x16c4-minmax-rndnu-neon-mull-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c4-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 
17 
xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t** restrict a,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 4);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (4 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(int8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   kc = round_up_po2(kc, 4 * sizeof(int8_t));
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr < 2) {
47     c1 = c0;
48   }
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     c2 = c1;
52   }
53   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54   if XNN_UNPREDICTABLE(mr != 4) {
55     c3 = c2;
56   }
57 
58   do {
59     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
60     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
61     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
62     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
63     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
64     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
65     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
66     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
67     int32x4_t vacc1x01 = vacc0x01;
68     int32x4_t vacc1x23 = vacc0x23;
69     int32x4_t vacc1x45 = vacc0x45;
70     int32x4_t vacc1x67 = vacc0x67;
71     int32x4_t vacc1x89 = vacc0x89;
72     int32x4_t vacc1xAB = vacc0xAB;
73     int32x4_t vacc1xCD = vacc0xCD;
74     int32x4_t vacc1xEF = vacc0xEF;
75     int32x4_t vacc2x01 = vacc0x01;
76     int32x4_t vacc2x23 = vacc0x23;
77     int32x4_t vacc2x45 = vacc0x45;
78     int32x4_t vacc2x67 = vacc0x67;
79     int32x4_t vacc2x89 = vacc0x89;
80     int32x4_t vacc2xAB = vacc0xAB;
81     int32x4_t vacc2xCD = vacc0xCD;
82     int32x4_t vacc2xEF = vacc0xEF;
83     int32x4_t vacc3x01 = vacc0x01;
84     int32x4_t vacc3x23 = vacc0x23;
85     int32x4_t vacc3x45 = vacc0x45;
86     int32x4_t vacc3x67 = vacc0x67;
87     int32x4_t vacc3x89 = vacc0x89;
88     int32x4_t vacc3xAB = vacc0xAB;
89     int32x4_t vacc3xCD = vacc0xCD;
90     int32x4_t vacc3xEF = vacc0xEF;
91 
92     size_t p = ks;
93     do {
94       const int8_t* restrict a0 = a[0];
95       if XNN_UNPREDICTABLE(a0 != zero) {
96         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
97       }
98       const int8_t* restrict a1 = a[1];
99       if XNN_UNPREDICTABLE(a1 != zero) {
100         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
101       }
102       const int8_t* restrict a2 = a[2];
103       if XNN_UNPREDICTABLE(a2 != zero) {
104         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
105       }
106       const int8_t* restrict a3 = a[3];
107       if XNN_UNPREDICTABLE(a3 != zero) {
108         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
109       }
110       a += 4;
111 
112       size_t k = kc;
113 
114 
115       while (k >= 8 * sizeof(int8_t)) {
116         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
117         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
118         const int8x8_t va2 = vld1_s8(a2); a2 += 8;
119         const int8x8_t va3 = vld1_s8(a3); a3 += 8;
120 
121         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133         const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134         const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135         const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136         const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137 
138         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
139         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
140         const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
141         const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
142 
143         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
144         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
145         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
146         const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
147         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
148         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
149         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
150         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
151         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
152         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
153         const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
154         const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
155         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
156         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
157         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
158         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
159         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
160         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
161         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
162         const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
163         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
164         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
165         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
166         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
167         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
168         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
169         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
170         const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
171         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
172         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
173         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
174         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
175         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
176         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
177         const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
178         const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
179         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
180         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
181         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
182         vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
183         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
184         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
185         const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
186         const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
187         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
188         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
189         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
190         vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
191         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
192         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
193         const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
194         const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
195         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
196         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
197         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
198         vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
199         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
200         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
201         const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
202         const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
203         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
204         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
205         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
206         vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
207         const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
208         const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
209         const int8x8_t va2c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 1));
210         const int8x8_t va3c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 1));
211 
212         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
213         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
214         const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
215         const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
216         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
217         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
218         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
219         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
220         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
221         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
222         const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
223         const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
224         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
225         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
226         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
227         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
228         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
229         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
230         const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
231         const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
232         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
233         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
234         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
235         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
236         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
237         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
238         const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
239         const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
240         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
241         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
242         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
243         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
244         const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
245         const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
246         const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1);
247         const int16x8_t vprod3x89c1 = vmull_s8(vb89c1, va3c1);
248         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
249         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
250         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
251         vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
252         const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
253         const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
254         const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1);
255         const int16x8_t vprod3xABc1 = vmull_s8(vbABc1, va3c1);
256         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
257         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
258         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
259         vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
260         const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
261         const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
262         const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1);
263         const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1);
264         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
265         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
266         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
267         vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
268         const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
269         const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
270         const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1);
271         const int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1, va3c1);
272         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
273         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
274         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
275         vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
276 
277         k -= 8 * sizeof(int8_t);
278       }
279 
280       if XNN_UNLIKELY(k != 0) {
281         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
282         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
283         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
284         const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
285 
286         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
287         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
288         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
289         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
290         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
291         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
292         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
293         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
294 
295         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
296         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
297         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
298         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
299         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
300         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
301         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
302         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
303         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
304         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
305         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
306         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
307         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
308         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
309         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
310         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
311         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
312         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
313         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
314         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
315         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
316         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
317         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
318         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
319         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
320         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
321         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
322         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
323         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
324         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
325         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
326         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
327         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
328         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
329         const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
330         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
331         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
332         const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
333         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
334         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
335         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
336         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
337         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
338         const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
339         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
340         const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
341         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
342         const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
343         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
344         const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
345         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
346         const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
347         const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
348         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
349         const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
350         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
351         const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
352         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
353         const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
354         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
355         const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
356         vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
357         const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
358         vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
359         const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
360         vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
361         const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
362         vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
363       }
364       p -= 4 * sizeof(void*);
365     } while (p != 0);
366 
367 #if XNN_ARCH_ARM64
368     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
369     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
370     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
371     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
372     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
373     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
374     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
375     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
376     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
377     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
378     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
379     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
380     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
381     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
382     int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB);
383     int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF);
384 #else
385     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
386     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
387     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
388     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
389     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
390     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
391     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
392     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
393     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
394     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
395     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
396     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
397     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
398     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
399     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
400     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
401     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
402     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
403     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
404     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
405     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
406     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
407     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
408     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
409     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
410     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
411     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
412     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
413     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
414     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
415     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
416     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
417     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
418     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
419     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
420     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
421     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
422     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
423     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
424     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
425     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
426     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
427     const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89));
428     const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB));
429     int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB);
430     const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD));
431     const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF));
432     int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF);
433 #endif
434 
435     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
436     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
437     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
438 
439     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
440     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
441     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
442     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
443     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
444     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
445     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
446     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
447     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
448     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
449     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
450     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
451     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
452     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
453     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
454     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
455 
456     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
457     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
458     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
459     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
460     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
461     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
462     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
463     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
464     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
465     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
466     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
467     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
468     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
469     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
470     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
471     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
472 
473     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
474     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
475     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
476     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
477     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
478     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
479     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
480     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
481     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
482     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
483     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
484     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
485     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
486     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
487     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
488     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
489 
490     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
491 #if XNN_ARCH_ARM64
492     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
493     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
494     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
495     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
496     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
497     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
498     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
499     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
500 
501     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
502     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
503     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
504     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
505     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
506     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
507     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
508     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
509 
510     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
511     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
512     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
513     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
514 #else
515     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
516     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
517     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
518     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
519     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
520     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
521     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
522     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
523 
524     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
525     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
526     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
527     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
528     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
529     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
530     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
531     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
532 
533     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
534     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
535     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
536     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
537 #endif
538 
539     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
540     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
541     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
542     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
543     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
544 
545     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
546     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
547     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
548     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
549     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
550 
551     if (nc >= 16) {
552       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
553       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
554       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
555       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
556 
557       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
558       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
559       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
560       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
561 
562       a = (const int8_t**restrict) ((uintptr_t) a - ks);
563 
564       nc -= 16;
565     } else {
566       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
567       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
568       if (nc & 8) {
569         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
570         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
571         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
572         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
573         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
574         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
575       }
576       if (nc & 4) {
577         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
578         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
579         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
580         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
581         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
582         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
583       }
584       if (nc & 2) {
585         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
586         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
587         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
588         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
589         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
590         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
591       }
592       if (nc & 1) {
593         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
594         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
595         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
596         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
597       }
598 
599       nc = 0;
600     }
601   } while (nc != 0);
602 }
603