• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 4);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 4 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56   if XNN_UNPREDICTABLE(mr != 4) {
57     a3 = a2;
58     c3 = c2;
59   }
60 
61   do {
62     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
63     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
64     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
65     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
66     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
67     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
68     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
69     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
70     int32x4_t vacc1x01 = vacc0x01;
71     int32x4_t vacc1x23 = vacc0x23;
72     int32x4_t vacc1x45 = vacc0x45;
73     int32x4_t vacc1x67 = vacc0x67;
74     int32x4_t vacc1x89 = vacc0x89;
75     int32x4_t vacc1xAB = vacc0xAB;
76     int32x4_t vacc1xCD = vacc0xCD;
77     int32x4_t vacc1xEF = vacc0xEF;
78     int32x4_t vacc2x01 = vacc0x01;
79     int32x4_t vacc2x23 = vacc0x23;
80     int32x4_t vacc2x45 = vacc0x45;
81     int32x4_t vacc2x67 = vacc0x67;
82     int32x4_t vacc2x89 = vacc0x89;
83     int32x4_t vacc2xAB = vacc0xAB;
84     int32x4_t vacc2xCD = vacc0xCD;
85     int32x4_t vacc2xEF = vacc0xEF;
86     int32x4_t vacc3x01 = vacc0x01;
87     int32x4_t vacc3x23 = vacc0x23;
88     int32x4_t vacc3x45 = vacc0x45;
89     int32x4_t vacc3x67 = vacc0x67;
90     int32x4_t vacc3x89 = vacc0x89;
91     int32x4_t vacc3xAB = vacc0xAB;
92     int32x4_t vacc3xCD = vacc0xCD;
93     int32x4_t vacc3xEF = vacc0xEF;
94 
95     size_t k = kc;
96 
97     while (k >= 16 * sizeof(int8_t)) {
98       const int32x2x2_t va0x0 = vld2_dup_s32((const void*)a0); a0 += 8;
99       const int32x2x2_t va0x1 = vld2_dup_s32((const void*)a0); a0 += 8;
100       const int32x2x2_t va1x0 = vld2_dup_s32((const void*)a1); a1 += 8;
101       const int32x2x2_t va1x1 = vld2_dup_s32((const void*)a1); a1 += 8;
102       const int32x2x2_t va2x0 = vld2_dup_s32((const void*)a2); a2 += 8;
103       const int32x2x2_t va2x1 = vld2_dup_s32((const void*)a2); a2 += 8;
104       const int32x2x2_t va3x0 = vld2_dup_s32((const void*)a3); a3 += 8;
105       const int32x2x2_t va3x1 = vld2_dup_s32((const void*)a3); a3 += 8;
106 
107       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 
124       const int8x8_t va0c0x0 = vreinterpret_s8_s32(va0x0.val[0]);
125       const int8x8_t va0c0x1 = vreinterpret_s8_s32(va0x1.val[0]);
126       const int8x8_t va1c0x0 = vreinterpret_s8_s32(va1x0.val[0]);
127       const int8x8_t va1c0x1 = vreinterpret_s8_s32(va1x1.val[0]);
128       const int8x8_t va2c0x0 = vreinterpret_s8_s32(va2x0.val[0]);
129       const int8x8_t va2c0x1 = vreinterpret_s8_s32(va2x1.val[0]);
130       const int8x8_t va3c0x0 = vreinterpret_s8_s32(va3x0.val[0]);
131       const int8x8_t va3c0x1 = vreinterpret_s8_s32(va3x1.val[0]);
132 
133       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
134       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
135       int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
136       int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3c0x0);
137       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
139       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
140       vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
141       vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3c0x1);
142       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
143       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
144       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
145       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
146       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
147       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
148       int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
149       int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3c0x0);
150       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
152       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
153       vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
154       vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3c0x1);
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       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
160       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
161       int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
162       int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0);
163       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
165       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
166       vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
167       vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1);
168       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
169       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
170       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
171       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
172       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
173       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
174       int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
175       int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3c0x0);
176       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
178       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
179       vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
180       vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3c0x1);
181       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
182       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
183       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
184       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
185       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
186       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0);
187       int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0);
188       int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3c0x0);
189       const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
190       vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
191       vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1);
192       vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1);
193       vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1);
194       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
195       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
196       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
197       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
198       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
199       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1c0x0);
200       int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0);
201       int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3c0x0);
202       const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
203       vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
204       vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1c0x1);
205       vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1);
206       vprod3xABc0 = vmlal_s8(vprod3xABc0, vbABc0x1, va3c0x1);
207       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
208       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
209       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
210       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
211       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
212       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1c0x0);
213       int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0);
214       int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3c0x0);
215       const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216       vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
217       vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1c0x1);
218       vprod2xCDc0 = vmlal_s8(vprod2xCDc0, vbCDc0x1, va2c0x1);
219       vprod3xCDc0 = vmlal_s8(vprod3xCDc0, vbCDc0x1, va3c0x1);
220       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
221       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
222       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
223       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
224       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
225       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1c0x0);
226       int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0);
227       int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0);
228       const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229       vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
230       vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1c0x1);
231       vprod2xEFc0 = vmlal_s8(vprod2xEFc0, vbEFc0x1, va2c0x1);
232       vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1);
233       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
234       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
235       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
236       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
237       const int8x8_t va0c1x0 = vreinterpret_s8_s32(va0x0.val[1]);
238       const int8x8_t va0c1x1 = vreinterpret_s8_s32(va0x1.val[1]);
239       const int8x8_t va1c1x0 = vreinterpret_s8_s32(va1x0.val[1]);
240       const int8x8_t va1c1x1 = vreinterpret_s8_s32(va1x1.val[1]);
241       const int8x8_t va2c1x0 = vreinterpret_s8_s32(va2x0.val[1]);
242       const int8x8_t va2c1x1 = vreinterpret_s8_s32(va2x1.val[1]);
243       const int8x8_t va3c1x0 = vreinterpret_s8_s32(va3x0.val[1]);
244       const int8x8_t va3c1x1 = vreinterpret_s8_s32(va3x1.val[1]);
245 
246       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
247       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
248       int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
249       int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3c1x0);
250       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
252       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
253       vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
254       vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3c1x1);
255       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
256       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
257       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
258       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
259       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
260       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
261       int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
262       int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3c1x0);
263       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
264       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
265       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
266       vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
267       vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3c1x1);
268       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
269       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
270       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
271       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
272       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
273       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
274       int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
275       int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3c1x0);
276       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
277       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
278       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
279       vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
280       vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3c1x1);
281       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
282       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
283       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
284       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
285       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
286       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
287       int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
288       int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3c1x0);
289       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
290       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
291       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
292       vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
293       vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3c1x1);
294       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
295       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
296       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
297       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
298       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
299       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0);
300       int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2c1x0);
301       int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3c1x0);
302       const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
303       vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
304       vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1);
305       vprod2x89c1 = vmlal_s8(vprod2x89c1, vb89c1x1, va2c1x1);
306       vprod3x89c1 = vmlal_s8(vprod3x89c1, vb89c1x1, va3c1x1);
307       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
308       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
309       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
310       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
311       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
312       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0);
313       int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2c1x0);
314       int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3c1x0);
315       const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
316       vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
317       vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1);
318       vprod2xABc1 = vmlal_s8(vprod2xABc1, vbABc1x1, va2c1x1);
319       vprod3xABc1 = vmlal_s8(vprod3xABc1, vbABc1x1, va3c1x1);
320       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
321       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
322       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
323       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
324       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
325       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0);
326       int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2c1x0);
327       int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0);
328       const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
329       vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
330       vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1c1x1);
331       vprod2xCDc1 = vmlal_s8(vprod2xCDc1, vbCDc1x1, va2c1x1);
332       vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1);
333       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
334       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
335       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
336       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
337       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
338       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1c1x0);
339       int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2c1x0);
340       int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3c1x0);
341       const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
342       vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
343       vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1c1x1);
344       vprod2xEFc1 = vmlal_s8(vprod2xEFc1, vbEFc1x1, va2c1x1);
345       vprod3xEFc1 = vmlal_s8(vprod3xEFc1, vbEFc1x1, va3c1x1);
346       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
347       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
348       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
349       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
350 
351       k -= 16 * sizeof(int8_t);
352     }
353 
354     if (k >= 8 * sizeof(int8_t)) {
355       const int32x2x2_t va0 = vld2_dup_s32((const void*)a0); a0 += 8;
356       const int32x2x2_t va1 = vld2_dup_s32((const void*)a1); a1 += 8;
357       const int32x2x2_t va2 = vld2_dup_s32((const void*)a2); a2 += 8;
358       const int32x2x2_t va3 = vld2_dup_s32((const void*)a3); a3 += 8;
359 
360       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
361       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
362       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
363       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
364       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
365       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
366       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
367       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
368       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
369       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
370       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
371       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
372       const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
373       const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
374       const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
375       const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
376 
377       const int8x8_t va0c0 = vreinterpret_s8_s32(va0.val[0]);
378       const int8x8_t va1c0 = vreinterpret_s8_s32(va1.val[0]);
379       const int8x8_t va2c0 = vreinterpret_s8_s32(va2.val[0]);
380       const int8x8_t va3c0 = vreinterpret_s8_s32(va3.val[0]);
381 
382       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
383       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
384       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
385       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
386       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
387       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
388       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
389       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
390       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
391       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
392       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
393       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
394       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
395       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
396       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
397       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
398       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
399       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
400       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
401       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
402       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
403       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
404       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
405       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
406       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
407       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
408       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
409       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
410       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
411       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
412       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
413       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
414       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
415       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
416       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
417       const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
418       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
419       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
420       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
421       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
422       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
423       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
424       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
425       const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
426       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
427       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
428       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
429       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
430       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
431       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
432       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
433       const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
434       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
435       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
436       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
437       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
438       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
439       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
440       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
441       const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
442       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
443       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
444       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
445       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
446       const int8x8_t va0c1 = vreinterpret_s8_s32(va0.val[1]);
447       const int8x8_t va1c1 = vreinterpret_s8_s32(va1.val[1]);
448       const int8x8_t va2c1 = vreinterpret_s8_s32(va2.val[1]);
449       const int8x8_t va3c1 = vreinterpret_s8_s32(va3.val[1]);
450 
451       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
452       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
453       const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
454       const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
455       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
456       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
457       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
458       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
459       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
460       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
461       const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
462       const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
463       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
464       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
465       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
466       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
467       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
468       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
469       const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
470       const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
471       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
472       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
473       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
474       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
475       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
476       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
477       const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
478       const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
479       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
480       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
481       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
482       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
483       const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
484       const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
485       const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1);
486       const int16x8_t vprod3x89c1 = vmull_s8(vb89c1, va3c1);
487       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
488       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
489       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
490       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
491       const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
492       const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
493       const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1);
494       const int16x8_t vprod3xABc1 = vmull_s8(vbABc1, va3c1);
495       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
496       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
497       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
498       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
499       const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
500       const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
501       const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1);
502       const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1);
503       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
504       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
505       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
506       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
507       const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
508       const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
509       const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1);
510       const int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1, va3c1);
511       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
512       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
513       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
514       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
515 
516       k -= 8 * sizeof(int8_t);
517     }
518 
519     if XNN_UNLIKELY(k != 0) {
520       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
521       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
522       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
523       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
524 
525       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
526       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
527       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
528       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
529       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
530       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
531       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
532       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
533 
534       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
535       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
536       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
537       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
538       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
539       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
540       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
541       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
542       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
543       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
544       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
545       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
546       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
547       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
548       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
549       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
550       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
551       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
552       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
553       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
554       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
555       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
556       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
557       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
558       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
559       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
560       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
561       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
562       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
563       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
564       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
565       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
566       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
567       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
568       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
569       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
570       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
571       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
572       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
573       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
574       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
575       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
576       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
577       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
578       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
579       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
580       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
581       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
582       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
583       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
584       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
585       const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
586       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
587       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
588       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
589       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
590       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
591       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
592       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
593       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
594       const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
595       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
596       const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
597       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
598       const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
599       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
600       const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
601       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
602     }
603 
604 #if XNN_ARCH_ARM64
605     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
606     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
607     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
608     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
609     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
610     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
611     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
612     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
613     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
614     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
615     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
616     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
617     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
618     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
619     int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB);
620     int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF);
621 #else
622     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
623     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
624     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
625     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
626     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
627     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
628     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
629     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
630     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
631     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
632     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
633     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
634     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
635     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
636     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
637     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
638     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
639     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
640     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
641     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
642     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
643     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
644     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
645     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
646     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
647     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
648     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
649     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
650     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
651     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
652     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
653     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
654     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
655     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
656     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
657     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
658     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
659     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
660     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
661     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
662     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
663     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
664     const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89));
665     const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB));
666     int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB);
667     const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD));
668     const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF));
669     int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF);
670 #endif
671 
672     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
673     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
674     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
675 
676     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
677     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
678     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
679     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
680     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
681     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
682     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
683     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
684     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
685     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
686     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
687     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
688     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
689     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
690     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
691     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
692 
693     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
694     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
695     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
696     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
697     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
698     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
699     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
700     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
701     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
702     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
703     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
704     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
705     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
706     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
707     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
708     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
709 
710     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
711     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
712     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
713     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
714     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
715     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
716     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
717     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
718     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
719     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
720     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
721     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
722     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
723     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
724     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
725     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
726 
727     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
728 #if XNN_ARCH_ARM64
729     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
730     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
731     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
732     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
733     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
734     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
735     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
736     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
737 
738     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
739     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
740     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
741     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
742     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
743     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
744     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
745     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
746 
747     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
748     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
749     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
750     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
751 #else
752     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
753     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
754     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
755     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
756     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
757     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
758     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
759     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
760 
761     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
762     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
763     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
764     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
765     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
766     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
767     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
768     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
769 
770     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
771     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
772     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
773     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
774 #endif
775 
776     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
777     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
778     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
779     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
780     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
781 
782     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
783     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
784     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
785     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
786     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
787 
788     if (nc >= 16) {
789       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
790       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
791       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
792       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
793 
794       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
795       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
796       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
797       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
798 
799       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
800       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
801       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
802       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
803 
804       nc -= 16;
805     } else {
806       // Final case where not all of the 16 columns fit in the destination.
807       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
808       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
809       if (nc & 8) {
810         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
811         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
812         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
813         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
814         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
815         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
816       }
817       if (nc & 4) {
818         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
819         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
820         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
821         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
822         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
823         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
824       }
825       if (nc & 2) {
826         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
827         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
828         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
829         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
830         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
831         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
832       }
833       if (nc & 1) {
834         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
835         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
836         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
837         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
838       }
839 
840       nc = 0;
841     }
842   } while (nc != 0);
843 }
844