xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x16c2-minmax-rndnu-neon-mull-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-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 
xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__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)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(mr != 0);
30   assert(mr <= 4);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(int8_t) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   kc = round_up_po2(kc, 2 * sizeof(int8_t));
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr != 4) {
56     a3 = a2;
57     c3 = c2;
58   }
59 
60   do {
61     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
62     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
63     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
64     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
65     int32x4_t vacc1x0123 = vacc0x0123;
66     int32x4_t vacc1x4567 = vacc0x4567;
67     int32x4_t vacc1x89AB = vacc0x89AB;
68     int32x4_t vacc1xCDEF = vacc0xCDEF;
69     int32x4_t vacc2x0123 = vacc0x0123;
70     int32x4_t vacc2x4567 = vacc0x4567;
71     int32x4_t vacc2x89AB = vacc0x89AB;
72     int32x4_t vacc2xCDEF = vacc0xCDEF;
73     int32x4_t vacc3x0123 = vacc0x0123;
74     int32x4_t vacc3x4567 = vacc0x4567;
75     int32x4_t vacc3x89AB = vacc0x89AB;
76     int32x4_t vacc3xCDEF = vacc0xCDEF;
77 
78     size_t k = kc;
79 
80 
81     while (k >= 8 * sizeof(int8_t)) {
82       const int16x4_t va00 = vld1_dup_s16((const void*)a0);
83       const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
84       const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
85       const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
86       const int16x4_t va10 = vld1_dup_s16((const void*)a1);
87       const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
88       const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
89       const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
90       const int16x4_t va20 = vld1_dup_s16((const void*)a2);
91       const int16x4_t va21 = vld1_dup_s16((const void*)(a2 + 2));
92       const int16x4_t va22 = vld1_dup_s16((const void*)(a2 + 4));
93       const int16x4_t va23 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
94       const int16x4_t va30 = vld1_dup_s16((const void*)a3);
95       const int16x4_t va31 = vld1_dup_s16((const void*)(a3 + 2));
96       const int16x4_t va32 = vld1_dup_s16((const void*)(a3 + 4));
97       const int16x4_t va33 = vld1_dup_s16((const void*)(a3 + 6)); a3 += 8;
98 
99       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105       const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106       const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109       const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110       const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113       const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114       const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 
116       const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
117       const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
118       const int8x8_t va2c0 = vreinterpret_s8_s16(va20);
119       const int8x8_t va3c0 = vreinterpret_s8_s16(va30);
120 
121       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
122       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
123       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
124       const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
125       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
126       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
127       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
128       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
129       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
130       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
131       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
132       const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
133       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
134       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
135       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
136       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
137       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
138       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
139       const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
140       const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
141       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
142       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
143       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
144       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
145       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
146       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
147       const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
148       const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
149       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
150       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
151       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
152       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
153       const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
154       const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
155       const int8x8_t va2c1 = vreinterpret_s8_s16(va21);
156       const int8x8_t va3c1 = vreinterpret_s8_s16(va31);
157 
158       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
159       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
160       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
161       const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
162       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
163       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
164       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
165       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
166       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
167       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
168       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
169       const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
170       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
171       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
172       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
173       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
174       const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
175       const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
176       const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
177       const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
178       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
179       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
180       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
181       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
182       const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
183       const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
184       const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
185       const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
186       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
187       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
188       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
189       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
190       const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
191       const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
192       const int8x8_t va2c2 = vreinterpret_s8_s16(va22);
193       const int8x8_t va3c2 = vreinterpret_s8_s16(va32);
194 
195       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
196       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
197       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
198       const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
199       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
200       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
201       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
202       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
203       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
204       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
205       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
206       const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
207       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
208       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
209       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
210       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
211       const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
212       const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
213       const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
214       const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
215       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
216       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
217       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
218       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
219       const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
220       const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
221       const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
222       const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
223       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
224       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
225       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
226       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
227       const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
228       const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
229       const int8x8_t va2c3 = vreinterpret_s8_s16(va23);
230       const int8x8_t va3c3 = vreinterpret_s8_s16(va33);
231 
232       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
233       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
234       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
235       const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
236       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
237       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
238       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
239       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
240       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
241       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
242       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
243       const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
244       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
245       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
246       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
247       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
248       const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
249       const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
250       const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
251       const int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3, va3c3);
252       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
253       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
254       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
255       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
256       const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
257       const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
258       const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
259       const int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3, va3c3);
260       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
261       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
262       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
263       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
264 
265       k -= 8 * sizeof(int8_t);
266     }
267 
268     if XNN_UNLIKELY(k != 0) {
269       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
270       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
271       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
272       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
273 
274       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
275       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
276       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
277       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
278 
279       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
280       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
281       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
282       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
283       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
284       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
285       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
286       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
287       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
288       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
289       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
290       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
291       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
292       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
293       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
294       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
295       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
296       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
297       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
298       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
299       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
300       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
301       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
302       const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
303       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
304       const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
305       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
306       const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
307       const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
308       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
309       const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
310       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
311       const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
312       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
313       const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
314       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
315 
316       if (k > 2 * sizeof(int8_t)) {
317         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
318         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
319         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
320         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
321 
322         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
323         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
324         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
325         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
326         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
327         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
328         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
329         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
330         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
331         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
332         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
333         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
334         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
335         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
336         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
337         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
338         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
339         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
340         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
341         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
342         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
343         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
344         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
345         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
346         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
347         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
348         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
349         const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
350         const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
351         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
352         const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
353         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
354         const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
355         vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
356         const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
357         vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
358 
359         if (k > 4 * sizeof(int8_t)) {
360           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
361           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
362           const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
363           const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
364 
365           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
366           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
367           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
368           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
369           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
370           const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
371           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
372           const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
373           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
374           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
375           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
376           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
377           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
378           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
379           const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
380           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
381           const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
382           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
383           const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
384           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
385           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
386           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
387           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
388           const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
389           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
390           const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
391           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
392           const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
393           const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
394           vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
395           const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
396           vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
397           const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
398           vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
399           const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
400           vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
401         }
402       }
403     }
404 
405     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
406     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
407     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
408 
409     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
410     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
411     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
412     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
413     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
414     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
415     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
416     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
417     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
418     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
419     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
420     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
421     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
422     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
423     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
424     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
425 
426     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
427     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
428     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
429     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
430     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
431     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
432     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
433     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
434     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
435     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
436     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
437     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
438     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
439     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
440     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
441     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
442 
443     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
444     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
445     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
446     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
447     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
448     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
449     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
450     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
451     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
452     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
453     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
454     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
455     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
456     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
457     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
458     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
459 
460     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
461 #if XNN_ARCH_ARM64
462     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
463     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
464     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
465     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
466     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
467     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
468     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
469     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
470 
471     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
472     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
473     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
474     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
475     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
476     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
477     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
478     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
479 
480     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
481     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
482     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
483     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
484 #else
485     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
486     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
487     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
488     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
489     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
490     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
491     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
492     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
493 
494     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
495     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
496     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
497     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
498     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
499     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
500     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
501     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
502 
503     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
504     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
505     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
506     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
507 #endif
508 
509     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
510     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
511     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
512     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
513     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
514 
515     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
516     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
517     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
518     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
519     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
520 
521     if (nc >= 16) {
522       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
523       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
524       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
525       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
526 
527       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
528       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
529       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
530       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
531 
532       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
533       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
534       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
535       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
536 
537       nc -= 16;
538     } else {
539       // Final case where not all of the 16 columns fit in the destination.
540       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
541       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
542       if (nc & 8) {
543         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
544         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
545         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
546         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
547         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
548         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
549       }
550       if (nc & 4) {
551         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
552         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
553         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
554         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
555         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
556         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
557       }
558       if (nc & 2) {
559         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
560         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
561         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
562         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
563         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
564         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
565       }
566       if (nc & 1) {
567         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
568         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
569         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
570         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
571       }
572 
573       nc = 0;
574     }
575   } while (nc != 0);
576 }
577