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