xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x16c4-minmax-rndnu-neon-mull-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_mull_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_mull_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 
98     while (k >= 8 * sizeof(int8_t)) {
99       const int32x2_t va00 = vld1_dup_s32((const void*)a0);
100       const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
101       const int32x2_t va10 = vld1_dup_s32((const void*)a1);
102       const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
103       const int32x2_t va20 = vld1_dup_s32((const void*)a2);
104       const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
105       const int32x2_t va30 = vld1_dup_s32((const void*)a3);
106       const int32x2_t va31 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8;
107 
108       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120       const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121       const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122       const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123       const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 
125       const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
126       const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
127       const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
128       const int8x8_t va3c0 = vreinterpret_s8_s32(va30);
129 
130       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
131       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
132       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
133       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
134       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
135       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
136       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
137       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
138       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
139       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
140       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
141       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
142       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
143       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
144       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
145       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
146       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
147       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
148       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
149       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
150       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
151       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
152       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
153       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
154       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
155       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
156       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
157       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
158       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
159       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
160       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
161       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
162       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
163       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
164       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
165       const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
166       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
167       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
168       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
169       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
170       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
171       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
172       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
173       const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
174       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
175       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
176       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
177       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
178       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
179       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
180       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
181       const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
182       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
183       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
184       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
185       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
186       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
187       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
188       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
189       const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
190       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
191       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
192       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
193       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
194       const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
195       const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
196       const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
197       const int8x8_t va3c1 = vreinterpret_s8_s32(va31);
198 
199       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
200       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
201       const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
202       const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
203       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
204       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
205       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
206       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
207       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
208       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
209       const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
210       const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
211       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
212       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
213       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
214       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
215       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
216       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
217       const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
218       const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
219       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
220       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
221       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
222       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
223       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
224       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
225       const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
226       const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
227       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
228       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
229       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
230       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
231       const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
232       const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
233       const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1);
234       const int16x8_t vprod3x89c1 = vmull_s8(vb89c1, va3c1);
235       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
236       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
237       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
238       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
239       const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
240       const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
241       const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1);
242       const int16x8_t vprod3xABc1 = vmull_s8(vbABc1, va3c1);
243       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
244       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
245       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
246       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
247       const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
248       const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
249       const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1);
250       const int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1, va3c1);
251       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
252       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
253       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
254       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
255       const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
256       const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
257       const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1);
258       const int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1, va3c1);
259       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
260       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
261       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
262       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
263 
264       k -= 8 * sizeof(int8_t);
265     }
266 
267     if XNN_UNLIKELY(k != 0) {
268       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
269       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
270       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
271       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
272 
273       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
274       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
275       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
276       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
277       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
278       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
279       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
280       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
281 
282       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
283       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
284       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
285       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
286       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
287       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
288       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
289       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
290       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
291       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
292       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
293       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
294       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
295       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
296       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
297       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
298       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
299       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
300       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
301       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
302       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
303       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
304       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
305       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
306       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
307       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
308       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
309       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
310       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
311       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
312       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
313       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
314       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
315       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
316       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
317       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
318       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
319       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
320       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
321       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
322       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
323       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
324       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
325       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
326       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
327       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
328       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
329       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
330       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
331       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
332       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
333       const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
334       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
335       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
336       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
337       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
338       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
339       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
340       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
341       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
342       const int16x8_t vprod3x89c0 = vmull_s8(vb89c0, va3c0);
343       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
344       const int16x8_t vprod3xABc0 = vmull_s8(vbABc0, va3c0);
345       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
346       const int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0, va3c0);
347       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
348       const int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0, va3c0);
349       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
350     }
351 
352 #if XNN_ARCH_ARM64
353     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
354     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
355     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
356     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
357     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
358     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
359     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
360     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
361     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
362     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
363     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
364     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
365     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
366     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
367     int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB);
368     int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF);
369 #else
370     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
371     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
372     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
373     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
374     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
375     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
376     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
377     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
378     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
379     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
380     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
381     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
382     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
383     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
384     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
385     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
386     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
387     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
388     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
389     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
390     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
391     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
392     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
393     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
394     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
395     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
396     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
397     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
398     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
399     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
400     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
401     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
402     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
403     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
404     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
405     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
406     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
407     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
408     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
409     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
410     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
411     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
412     const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89));
413     const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB));
414     int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB);
415     const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD));
416     const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF));
417     int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF);
418 #endif
419 
420     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
421     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
422     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
423 
424     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
425     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
426     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
427     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
428     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
429     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
430     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
431     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
432     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
433     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
434     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
435     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
436     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
437     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
438     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
439     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
440 
441     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
442     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
443     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
444     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
445     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
446     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
447     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
448     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
449     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
450     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
451     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
452     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
453     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
454     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
455     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
456     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
457 
458     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
459     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
460     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
461     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
462     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
463     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
464     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
465     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
466     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
467     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
468     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
469     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
470     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
471     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
472     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
473     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
474 
475     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
476 #if XNN_ARCH_ARM64
477     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
478     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
479     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
480     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
481     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
482     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
483     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
484     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
485 
486     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
487     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
488     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
489     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
490     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
491     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
492     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
493     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
494 
495     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
496     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
497     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
498     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
499 #else
500     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
501     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
502     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
503     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
504     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
505     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
506     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
507     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
508 
509     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
510     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
511     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
512     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
513     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
514     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
515     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
516     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
517 
518     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
519     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
520     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
521     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
522 #endif
523 
524     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
525     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
526     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
527     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
528     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
529 
530     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
531     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
532     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
533     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
534     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
535 
536     if (nc >= 16) {
537       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
538       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
539       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
540       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
541 
542       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
543       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
544       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
545       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
546 
547       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
548       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
549       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
550       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
551 
552       nc -= 16;
553     } else {
554       // Final case where not all of the 16 columns fit in the destination.
555       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
556       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
557       if (nc & 8) {
558         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
559         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
560         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
561         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
562         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
563         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
564       }
565       if (nc & 4) {
566         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
567         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
568         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
569         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
570         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
571         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
572       }
573       if (nc & 2) {
574         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
575         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
576         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
577         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
578         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
579         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
580       }
581       if (nc & 1) {
582         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
583         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
584         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
585         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
586       }
587 
588       nc = 0;
589     }
590   } while (nc != 0);
591 }
592