• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_mlal_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_mlal_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       while (k >= 16 * sizeof(int8_t)) {
115         const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
116         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
117         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
118         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
119         const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
120         const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
121         const int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
122         const int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
123 
124         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128         const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129         const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130         const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136         const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137         const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138         const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139         const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 
141         const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
142         const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
143         const int8x8_t va1c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 0));
144         const int8x8_t va1c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 0));
145         const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0));
146         const int8x8_t va2c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x1), 0));
147         const int8x8_t va3c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x0), 0));
148         const int8x8_t va3c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x1), 0));
149 
150         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
151         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
152         int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
153         int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3c0x0);
154         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
156         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
157         vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
158         vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3c0x1);
159         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
160         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
161         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
162         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
163         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
164         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
165         int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
166         int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3c0x0);
167         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
169         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
170         vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
171         vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3c0x1);
172         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
173         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
174         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
175         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
176         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
177         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
178         int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
179         int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0);
180         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
182         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
183         vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
184         vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1);
185         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
186         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
187         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
188         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
189         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
190         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
191         int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
192         int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3c0x0);
193         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
194         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
195         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
196         vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
197         vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3c0x1);
198         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
199         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
200         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
201         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
202         int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
203         int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0);
204         int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0);
205         int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3c0x0);
206         const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207         vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
208         vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1);
209         vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1);
210         vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1);
211         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
212         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
213         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
214         vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
215         int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
216         int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1c0x0);
217         int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0);
218         int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3c0x0);
219         const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220         vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
221         vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1c0x1);
222         vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1);
223         vprod3xABc0 = vmlal_s8(vprod3xABc0, vbABc0x1, va3c0x1);
224         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
225         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
226         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
227         vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
228         int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
229         int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1c0x0);
230         int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0);
231         int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3c0x0);
232         const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233         vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
234         vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1c0x1);
235         vprod2xCDc0 = vmlal_s8(vprod2xCDc0, vbCDc0x1, va2c0x1);
236         vprod3xCDc0 = vmlal_s8(vprod3xCDc0, vbCDc0x1, va3c0x1);
237         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
238         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
239         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
240         vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
241         int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
242         int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1c0x0);
243         int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0);
244         int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0);
245         const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246         vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
247         vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1c0x1);
248         vprod2xEFc0 = vmlal_s8(vprod2xEFc0, vbEFc0x1, va2c0x1);
249         vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1);
250         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
251         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
252         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
253         vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
254         const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
255         const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
256         const int8x8_t va1c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 1));
257         const int8x8_t va1c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 1));
258         const int8x8_t va2c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 1));
259         const int8x8_t va2c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x1), 1));
260         const int8x8_t va3c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x0), 1));
261         const int8x8_t va3c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x1), 1));
262 
263         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
264         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
265         int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
266         int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3c1x0);
267         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
268         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
269         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
270         vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
271         vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3c1x1);
272         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
273         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
274         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
275         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
276         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
277         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
278         int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
279         int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3c1x0);
280         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
281         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
282         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
283         vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
284         vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3c1x1);
285         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
286         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
287         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
288         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
289         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
290         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
291         int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
292         int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3c1x0);
293         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
294         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
295         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
296         vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
297         vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3c1x1);
298         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
299         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
300         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
301         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
302         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
303         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
304         int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
305         int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3c1x0);
306         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
307         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
308         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
309         vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
310         vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3c1x1);
311         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
312         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
313         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
314         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
315         int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
316         int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0);
317         int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2c1x0);
318         int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3c1x0);
319         const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
320         vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
321         vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1);
322         vprod2x89c1 = vmlal_s8(vprod2x89c1, vb89c1x1, va2c1x1);
323         vprod3x89c1 = vmlal_s8(vprod3x89c1, vb89c1x1, va3c1x1);
324         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
325         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
326         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
327         vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
328         int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
329         int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0);
330         int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2c1x0);
331         int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3c1x0);
332         const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
333         vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
334         vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1);
335         vprod2xABc1 = vmlal_s8(vprod2xABc1, vbABc1x1, va2c1x1);
336         vprod3xABc1 = vmlal_s8(vprod3xABc1, vbABc1x1, va3c1x1);
337         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
338         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
339         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
340         vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
341         int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
342         int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0);
343         int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2c1x0);
344         int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0);
345         const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
346         vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
347         vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1c1x1);
348         vprod2xCDc1 = vmlal_s8(vprod2xCDc1, vbCDc1x1, va2c1x1);
349         vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1);
350         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
351         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
352         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
353         vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
354         int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
355         int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1c1x0);
356         int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2c1x0);
357         int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3c1x0);
358         const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
359         vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
360         vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1c1x1);
361         vprod2xEFc1 = vmlal_s8(vprod2xEFc1, vbEFc1x1, va2c1x1);
362         vprod3xEFc1 = vmlal_s8(vprod3xEFc1, vbEFc1x1, va3c1x1);
363         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
364         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
365         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
366         vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
367 
368         k -= 16 * sizeof(int8_t);
369       }
370 
371       if (k >= 8 * sizeof(int8_t)) {
372         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
373         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
374         const int8x8_t va2 = vld1_s8(a2); a2 += 8;
375         const int8x8_t va3 = vld1_s8(a3); a3 += 8;
376 
377         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
378         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
379         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
380         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
381         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
382         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
383         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
384         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
385         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
386         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
387         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
388         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
389         const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
390         const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
391         const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
392         const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
393 
394         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
395         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
396         const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
397         const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
398 
399         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
400         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
401         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
402         const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
403         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
404         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
405         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
406         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
407         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
408         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
409         const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
410         const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
411         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
412         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
413         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
414         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
415         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
416         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
417         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
418         const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
419         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
420         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
421         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
422         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
423         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
424         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
425         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
426         const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
427         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
428         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
429         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
430         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
431         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
432         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
433         const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
434         const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
435         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
436         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
437         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
438         vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
439         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
440         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
441         const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
442         const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
443         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
444         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
445         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
446         vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
447         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
448         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
449         const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
450         const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
451         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
452         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
453         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
454         vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
455         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
456         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
457         const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
458         const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
459         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
460         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
461         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
462         vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
463         const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
464         const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
465         const int8x8_t va2c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 1));
466         const int8x8_t va3c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 1));
467 
468         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
469         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
470         const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
471         const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
472         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
473         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
474         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
475         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
476         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
477         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
478         const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
479         const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
480         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
481         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
482         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
483         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
484         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
485         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
486         const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
487         const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
488         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
489         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
490         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
491         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
492         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
493         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
494         const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
495         const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
496         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
497         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
498         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
499         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
500         const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
501         const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
502         const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1);
503         const int16x8_t vprod3x89c1 = vmull_s8(vb89c1, va3c1);
504         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
505         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
506         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
507         vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
508         const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
509         const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
510         const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1);
511         const int16x8_t vprod3xABc1 = vmull_s8(vbABc1, va3c1);
512         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
513         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
514         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
515         vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
516         const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
517         const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
518         const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1);
519         const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1);
520         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
521         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
522         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
523         vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
524         const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
525         const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
526         const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1);
527         const int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1, va3c1);
528         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
529         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
530         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
531         vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
532 
533         k -= 8 * sizeof(int8_t);
534       }
535 
536       if XNN_UNLIKELY(k != 0) {
537         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
538         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
539         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
540         const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
541 
542         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
543         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
544         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
545         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
546         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
547         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
548         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
549         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
550 
551         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
552         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
553         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
554         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
555         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
556         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
557         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
558         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
559         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
560         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
561         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
562         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
563         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
564         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
565         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
566         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
567         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
568         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
569         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
570         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
571         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
572         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
573         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
574         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
575         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
576         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
577         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
578         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
579         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
580         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
581         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
582         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
583         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
584         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
585         const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
586         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
587         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
588         const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
589         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
590         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
591         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
592         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
593         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
594         const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
595         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
596         const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
597         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
598         const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
599         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
600         const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
601         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
602         const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
603         const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
604         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
605         const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
606         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
607         const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
608         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
609         const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
610         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
611         const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
612         vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
613         const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
614         vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
615         const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
616         vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
617         const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
618         vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
619       }
620       p -= 4 * sizeof(void*);
621     } while (p != 0);
622 
623 #if XNN_ARCH_ARM64
624     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
625     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
626     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
627     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
628     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
629     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
630     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
631     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
632     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
633     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
634     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
635     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
636     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
637     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
638     int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB);
639     int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF);
640 #else
641     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
642     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
643     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
644     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
645     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
646     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
647     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
648     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
649     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
650     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
651     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
652     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
653     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
654     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
655     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
656     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
657     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
658     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
659     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
660     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
661     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
662     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
663     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
664     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
665     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
666     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
667     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
668     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
669     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
670     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
671     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
672     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
673     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
674     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
675     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
676     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
677     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
678     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
679     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
680     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
681     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
682     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
683     const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89));
684     const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB));
685     int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB);
686     const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD));
687     const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF));
688     int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF);
689 #endif
690 
691     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
692     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
693     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
694 
695     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
696     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
697     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
698     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
699     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
700     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
701     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
702     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
703     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
704     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
705     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
706     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
707     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
708     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
709     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
710     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
711 
712     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
713     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
714     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
715     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
716     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
717     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
718     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
719     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
720     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
721     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
722     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
723     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
724     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
725     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
726     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
727     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
728 
729     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
730     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
731     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
732     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
733     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
734     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
735     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
736     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
737     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
738     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
739     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
740     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
741     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
742     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
743     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
744     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
745 
746     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
747 #if XNN_ARCH_ARM64
748     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
749     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
750     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
751     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
752     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
753     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
754     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
755     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
756 
757     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
758     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
759     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
760     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
761     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
762     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
763     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
764     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
765 
766     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
767     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
768     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
769     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
770 #else
771     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
772     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
773     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
774     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
775     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
776     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
777     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
778     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
779 
780     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
781     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
782     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
783     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
784     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
785     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
786     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
787     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
788 
789     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
790     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
791     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
792     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
793 #endif
794 
795     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
796     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
797     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
798     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
799     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
800 
801     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
802     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
803     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
804     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
805     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
806 
807     if (nc >= 16) {
808       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
809       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
810       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
811       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
812 
813       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
814       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
815       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
816       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
817 
818       a = (const int8_t**restrict) ((uintptr_t) a - ks);
819 
820       nc -= 16;
821     } else {
822       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
823       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
824       if (nc & 8) {
825         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
826         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
827         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
828         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
829         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
830         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
831       }
832       if (nc & 4) {
833         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
834         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
835         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
836         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
837         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
838         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
839       }
840       if (nc & 2) {
841         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
842         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
843         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
844         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
845         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
846         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
847       }
848       if (nc & 1) {
849         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
850         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
851         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
852         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
853       }
854 
855       nc = 0;
856     }
857   } while (nc != 0);
858 }
859