xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mlal-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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_ld1r(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_ld1r(
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 int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
99       const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
100       const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
101       const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
102       const int32x2_t va10x0 = vld1_dup_s32((const void*)a1);
103       const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
104       const int32x2_t va10x1 = vld1_dup_s32((const void*)a1);
105       const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
106       const int32x2_t va20x0 = vld1_dup_s32((const void*)a2);
107       const int32x2_t va21x0 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
108       const int32x2_t va20x1 = vld1_dup_s32((const void*)a2);
109       const int32x2_t va21x1 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
110       const int32x2_t va30x0 = vld1_dup_s32((const void*)a3);
111       const int32x2_t va31x0 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8;
112       const int32x2_t va30x1 = vld1_dup_s32((const void*)a3);
113       const int32x2_t va31x1 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8;
114 
115       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 
132       const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
133       const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
134       const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0);
135       const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1);
136       const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0);
137       const int8x8_t va2c0x1 = vreinterpret_s8_s32(va20x1);
138       const int8x8_t va3c0x0 = vreinterpret_s8_s32(va30x0);
139       const int8x8_t va3c0x1 = vreinterpret_s8_s32(va30x1);
140 
141       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
142       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
143       int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
144       int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3c0x0);
145       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
147       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
148       vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
149       vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3c0x1);
150       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
151       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
152       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
153       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
154       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
155       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
156       int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
157       int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3c0x0);
158       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
160       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
161       vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
162       vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3c0x1);
163       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
164       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
165       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
166       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
167       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
168       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
169       int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
170       int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0);
171       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
173       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
174       vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
175       vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1);
176       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
177       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
178       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
179       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
180       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
181       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
182       int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
183       int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3c0x0);
184       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
186       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
187       vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
188       vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3c0x1);
189       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
190       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
191       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
192       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
193       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
194       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0);
195       int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0);
196       int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3c0x0);
197       const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
198       vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
199       vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1);
200       vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1);
201       vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1);
202       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
203       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
204       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
205       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
206       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
207       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1c0x0);
208       int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0);
209       int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3c0x0);
210       const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
211       vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
212       vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1c0x1);
213       vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1);
214       vprod3xABc0 = vmlal_s8(vprod3xABc0, vbABc0x1, va3c0x1);
215       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
216       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
217       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
218       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
219       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
220       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1c0x0);
221       int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0);
222       int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3c0x0);
223       const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
224       vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
225       vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1c0x1);
226       vprod2xCDc0 = vmlal_s8(vprod2xCDc0, vbCDc0x1, va2c0x1);
227       vprod3xCDc0 = vmlal_s8(vprod3xCDc0, vbCDc0x1, va3c0x1);
228       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
229       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
230       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
231       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
232       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
233       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1c0x0);
234       int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0);
235       int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3c0x0);
236       const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237       vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
238       vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1c0x1);
239       vprod2xEFc0 = vmlal_s8(vprod2xEFc0, vbEFc0x1, va2c0x1);
240       vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3c0x1);
241       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
242       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
243       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
244       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
245       const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
246       const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
247       const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0);
248       const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1);
249       const int8x8_t va2c1x0 = vreinterpret_s8_s32(va21x0);
250       const int8x8_t va2c1x1 = vreinterpret_s8_s32(va21x1);
251       const int8x8_t va3c1x0 = vreinterpret_s8_s32(va31x0);
252       const int8x8_t va3c1x1 = vreinterpret_s8_s32(va31x1);
253 
254       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
255       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
256       int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
257       int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3c1x0);
258       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
259       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
260       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
261       vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
262       vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3c1x1);
263       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
264       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
265       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
266       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
267       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
268       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
269       int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
270       int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3c1x0);
271       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
272       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
273       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
274       vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
275       vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3c1x1);
276       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
277       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
278       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
279       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
280       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
281       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
282       int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
283       int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3c1x0);
284       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
285       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
286       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
287       vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
288       vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3c1x1);
289       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
290       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
291       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
292       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
293       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
294       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
295       int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
296       int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3c1x0);
297       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
298       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
299       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
300       vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
301       vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3c1x1);
302       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
303       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
304       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
305       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
306       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
307       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0);
308       int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2c1x0);
309       int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3c1x0);
310       const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
311       vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
312       vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1);
313       vprod2x89c1 = vmlal_s8(vprod2x89c1, vb89c1x1, va2c1x1);
314       vprod3x89c1 = vmlal_s8(vprod3x89c1, vb89c1x1, va3c1x1);
315       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
316       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
317       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
318       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
319       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
320       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0);
321       int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2c1x0);
322       int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3c1x0);
323       const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
324       vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
325       vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1);
326       vprod2xABc1 = vmlal_s8(vprod2xABc1, vbABc1x1, va2c1x1);
327       vprod3xABc1 = vmlal_s8(vprod3xABc1, vbABc1x1, va3c1x1);
328       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
329       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
330       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
331       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
332       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
333       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0);
334       int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2c1x0);
335       int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3c1x0);
336       const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
337       vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
338       vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1c1x1);
339       vprod2xCDc1 = vmlal_s8(vprod2xCDc1, vbCDc1x1, va2c1x1);
340       vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3c1x1);
341       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
342       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
343       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
344       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
345       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
346       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1c1x0);
347       int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2c1x0);
348       int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3c1x0);
349       const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
350       vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
351       vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1c1x1);
352       vprod2xEFc1 = vmlal_s8(vprod2xEFc1, vbEFc1x1, va2c1x1);
353       vprod3xEFc1 = vmlal_s8(vprod3xEFc1, vbEFc1x1, va3c1x1);
354       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
355       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
356       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
357       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
358 
359       k -= 16 * sizeof(int8_t);
360     }
361 
362     if (k >= 8 * sizeof(int8_t)) {
363       const int32x2_t va00 = vld1_dup_s32((const void*)a0);
364       const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
365       const int32x2_t va10 = vld1_dup_s32((const void*)a1);
366       const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
367       const int32x2_t va20 = vld1_dup_s32((const void*)a2);
368       const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
369       const int32x2_t va30 = vld1_dup_s32((const void*)a3);
370       const int32x2_t va31 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8;
371 
372       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
373       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
374       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
375       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
376       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
377       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
378       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
379       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
380       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
381       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
382       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
383       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
384       const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
385       const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
386       const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
387       const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
388 
389       const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
390       const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
391       const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
392       const int8x8_t va3c0 = vreinterpret_s8_s32(va30);
393 
394       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
395       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
396       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
397       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
398       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
399       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
400       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
401       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
402       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
403       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
404       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
405       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
406       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
407       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
408       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
409       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
410       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
411       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
412       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
413       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
414       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
415       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
416       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
417       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
418       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
419       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
420       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
421       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
422       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
423       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
424       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
425       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
426       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
427       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
428       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
429       const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
430       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
431       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
432       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
433       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
434       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
435       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
436       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
437       const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
438       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
439       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
440       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
441       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
442       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
443       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
444       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
445       const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
446       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
447       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
448       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
449       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
450       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
451       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
452       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
453       const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
454       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
455       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
456       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
457       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
458       const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
459       const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
460       const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
461       const int8x8_t va3c1 = vreinterpret_s8_s32(va31);
462 
463       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
464       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
465       const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
466       const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
467       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
468       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
469       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
470       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
471       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
472       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
473       const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
474       const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
475       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
476       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
477       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
478       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
479       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
480       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
481       const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
482       const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
483       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
484       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
485       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
486       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
487       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
488       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
489       const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
490       const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
491       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
492       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
493       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
494       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
495       const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
496       const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
497       const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1);
498       const int16x8_t vprod3x89c1 = vmull_s8(vb89c1, va3c1);
499       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
500       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
501       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
502       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
503       const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
504       const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
505       const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1);
506       const int16x8_t vprod3xABc1 = vmull_s8(vbABc1, va3c1);
507       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
508       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
509       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
510       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
511       const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
512       const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
513       const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1);
514       const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1);
515       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
516       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
517       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
518       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
519       const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
520       const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
521       const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1);
522       const int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1, va3c1);
523       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
524       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
525       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
526       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
527 
528       k -= 8 * sizeof(int8_t);
529     }
530 
531     if XNN_UNLIKELY(k != 0) {
532       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
533       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
534       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
535       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
536 
537       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
538       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
539       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
540       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
541       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
542       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
543       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
544       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
545 
546       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
547       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
548       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
549       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
550       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
551       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
552       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
553       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
554       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
555       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
556       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
557       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
558       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
559       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
560       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
561       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
562       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
563       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
564       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
565       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
566       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
567       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
568       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
569       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
570       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
571       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
572       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
573       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
574       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
575       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
576       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
577       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
578       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
579       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
580       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
581       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
582       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
583       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
584       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
585       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
586       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
587       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
588       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
589       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
590       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
591       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
592       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
593       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
594       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
595       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
596       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
597       const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
598       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
599       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
600       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
601       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
602       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
603       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
604       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
605       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
606       const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
607       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
608       const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
609       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
610       const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
611       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
612       const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
613       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
614     }
615 
616 #if XNN_ARCH_ARM64
617     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
618     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
619     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
620     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
621     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
622     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
623     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
624     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
625     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
626     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
627     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
628     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
629     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
630     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
631     int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB);
632     int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF);
633 #else
634     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
635     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
636     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
637     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
638     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
639     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
640     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
641     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
642     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
643     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
644     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
645     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
646     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
647     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
648     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
649     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
650     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
651     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
652     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
653     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
654     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
655     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
656     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
657     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
658     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
659     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
660     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
661     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
662     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
663     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
664     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
665     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
666     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
667     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
668     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
669     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
670     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
671     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
672     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
673     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
674     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
675     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
676     const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89));
677     const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB));
678     int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB);
679     const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD));
680     const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF));
681     int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF);
682 #endif
683 
684     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
685     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
686     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
687 
688     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
689     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
690     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
691     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
692     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
693     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
694     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
695     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
696     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
697     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
698     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
699     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
700     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
701     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
702     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
703     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
704 
705     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
706     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
707     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
708     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
709     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
710     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
711     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
712     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
713     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
714     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
715     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
716     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
717     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
718     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
719     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
720     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
721 
722     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
723     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
724     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
725     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
726     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
727     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
728     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
729     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
730     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
731     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
732     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
733     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
734     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
735     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
736     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
737     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
738 
739     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
740 #if XNN_ARCH_ARM64
741     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
742     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
743     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
744     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
745     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
746     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
747     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
748     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
749 
750     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
751     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
752     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
753     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
754     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
755     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
756     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
757     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
758 
759     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
760     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
761     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
762     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
763 #else
764     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
765     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
766     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
767     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
768     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
769     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
770     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
771     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
772 
773     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
774     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
775     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
776     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
777     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
778     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
779     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
780     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
781 
782     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
783     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
784     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
785     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
786 #endif
787 
788     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
789     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
790     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
791     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
792     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
793 
794     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
795     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
796     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
797     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
798     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
799 
800     if (nc >= 16) {
801       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
802       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
803       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
804       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
805 
806       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
807       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
808       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
809       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
810 
811       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
812       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
813       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
814       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
815 
816       nc -= 16;
817     } else {
818       // Final case where not all of the 16 columns fit in the destination.
819       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
820       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
821       if (nc & 8) {
822         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
823         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
824         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
825         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
826         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
827         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
828       }
829       if (nc & 4) {
830         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
831         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
832         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
833         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
834         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
835         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
836       }
837       if (nc & 2) {
838         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
839         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
840         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
841         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
842         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
843         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
844       }
845       if (nc & 1) {
846         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
847         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
848         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
849         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
850       }
851 
852       nc = 0;
853     }
854   } while (nc != 0);
855 }
856