xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x16c16-minmax-rndnu-neon-mlal.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c16-neon-mlal.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_3x16c16__neon_mlal(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_3x16c16__neon_mlal(
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 <= 3);
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, 16 * 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 
55   do {
56     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64     int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65     int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66     int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67     int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
68     int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
69     int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
70     int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
71     int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
72     int32x4_t vacc1x0 = vacc0x0;
73     int32x4_t vacc1x1 = vacc0x1;
74     int32x4_t vacc1x2 = vacc0x2;
75     int32x4_t vacc1x3 = vacc0x3;
76     int32x4_t vacc1x4 = vacc0x4;
77     int32x4_t vacc1x5 = vacc0x5;
78     int32x4_t vacc1x6 = vacc0x6;
79     int32x4_t vacc1x7 = vacc0x7;
80     int32x4_t vacc1x8 = vacc0x8;
81     int32x4_t vacc1x9 = vacc0x9;
82     int32x4_t vacc1x10 = vacc0x10;
83     int32x4_t vacc1x11 = vacc0x11;
84     int32x4_t vacc1x12 = vacc0x12;
85     int32x4_t vacc1x13 = vacc0x13;
86     int32x4_t vacc1x14 = vacc0x14;
87     int32x4_t vacc1x15 = vacc0x15;
88     int32x4_t vacc2x0 = vacc0x0;
89     int32x4_t vacc2x1 = vacc0x1;
90     int32x4_t vacc2x2 = vacc0x2;
91     int32x4_t vacc2x3 = vacc0x3;
92     int32x4_t vacc2x4 = vacc0x4;
93     int32x4_t vacc2x5 = vacc0x5;
94     int32x4_t vacc2x6 = vacc0x6;
95     int32x4_t vacc2x7 = vacc0x7;
96     int32x4_t vacc2x8 = vacc0x8;
97     int32x4_t vacc2x9 = vacc0x9;
98     int32x4_t vacc2x10 = vacc0x10;
99     int32x4_t vacc2x11 = vacc0x11;
100     int32x4_t vacc2x12 = vacc0x12;
101     int32x4_t vacc2x13 = vacc0x13;
102     int32x4_t vacc2x14 = vacc0x14;
103     int32x4_t vacc2x15 = vacc0x15;
104 
105     // KC loop of 16
106     size_t k = kc;
107     while (k != 0) {
108       const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
109       const int8x16_t va1 = vld1q_s8(a1); a1 += 16;
110       const int8x16_t va2 = vld1q_s8(a2); a2 += 16;
111 
112       const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
113       const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
114       const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
115       const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
116       const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
117       const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
118       const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
119       const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
120       const int8x16_t vb8 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
121       const int8x16_t vb9 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
122       const int8x16_t vb10 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
123       const int8x16_t vb11 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
124       const int8x16_t vb12 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
125       const int8x16_t vb13 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
126       const int8x16_t vb14 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
127       const int8x16_t vb15 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
128 
129       int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
130       int16x8_t vprod1x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va1));
131       int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2));
132       vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
133       vprod1x0 = vmlal_s8(vprod1x0, vget_high_s8(vb0), vget_high_s8(va1));
134       vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2));
135       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
136       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
137       vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
138       int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
139       int16x8_t vprod1x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va1));
140       int16x8_t vprod2x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va2));
141       vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
142       vprod1x1 = vmlal_s8(vprod1x1, vget_high_s8(vb1), vget_high_s8(va1));
143       vprod2x1 = vmlal_s8(vprod2x1, vget_high_s8(vb1), vget_high_s8(va2));
144       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
145       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
146       vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
147       int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
148       int16x8_t vprod1x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va1));
149       int16x8_t vprod2x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va2));
150       vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
151       vprod1x2 = vmlal_s8(vprod1x2, vget_high_s8(vb2), vget_high_s8(va1));
152       vprod2x2 = vmlal_s8(vprod2x2, vget_high_s8(vb2), vget_high_s8(va2));
153       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
154       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
155       vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
156       int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
157       int16x8_t vprod1x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va1));
158       int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2));
159       vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
160       vprod1x3 = vmlal_s8(vprod1x3, vget_high_s8(vb3), vget_high_s8(va1));
161       vprod2x3 = vmlal_s8(vprod2x3, vget_high_s8(vb3), vget_high_s8(va2));
162       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
163       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
164       vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
165       int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
166       int16x8_t vprod1x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va1));
167       int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2));
168       vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
169       vprod1x4 = vmlal_s8(vprod1x4, vget_high_s8(vb4), vget_high_s8(va1));
170       vprod2x4 = vmlal_s8(vprod2x4, vget_high_s8(vb4), vget_high_s8(va2));
171       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
172       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
173       vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
174       int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
175       int16x8_t vprod1x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va1));
176       int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2));
177       vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
178       vprod1x5 = vmlal_s8(vprod1x5, vget_high_s8(vb5), vget_high_s8(va1));
179       vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2));
180       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
181       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
182       vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
183       int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
184       int16x8_t vprod1x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va1));
185       int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2));
186       vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
187       vprod1x6 = vmlal_s8(vprod1x6, vget_high_s8(vb6), vget_high_s8(va1));
188       vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2));
189       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
190       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
191       vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
192       int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
193       int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1));
194       int16x8_t vprod2x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va2));
195       vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
196       vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1));
197       vprod2x7 = vmlal_s8(vprod2x7, vget_high_s8(vb7), vget_high_s8(va2));
198       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
199       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
200       vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
201       int16x8_t vprod0x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va0));
202       int16x8_t vprod1x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va1));
203       int16x8_t vprod2x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va2));
204       vprod0x8 = vmlal_s8(vprod0x8, vget_high_s8(vb8), vget_high_s8(va0));
205       vprod1x8 = vmlal_s8(vprod1x8, vget_high_s8(vb8), vget_high_s8(va1));
206       vprod2x8 = vmlal_s8(vprod2x8, vget_high_s8(vb8), vget_high_s8(va2));
207       vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
208       vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
209       vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
210       int16x8_t vprod0x9 = vmull_s8(vget_low_s8(vb9), vget_low_s8(va0));
211       int16x8_t vprod1x9 = vmull_s8(vget_low_s8(vb9), vget_low_s8(va1));
212       int16x8_t vprod2x9 = vmull_s8(vget_low_s8(vb9), vget_low_s8(va2));
213       vprod0x9 = vmlal_s8(vprod0x9, vget_high_s8(vb9), vget_high_s8(va0));
214       vprod1x9 = vmlal_s8(vprod1x9, vget_high_s8(vb9), vget_high_s8(va1));
215       vprod2x9 = vmlal_s8(vprod2x9, vget_high_s8(vb9), vget_high_s8(va2));
216       vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
217       vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
218       vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
219       int16x8_t vprod0x10 = vmull_s8(vget_low_s8(vb10), vget_low_s8(va0));
220       int16x8_t vprod1x10 = vmull_s8(vget_low_s8(vb10), vget_low_s8(va1));
221       int16x8_t vprod2x10 = vmull_s8(vget_low_s8(vb10), vget_low_s8(va2));
222       vprod0x10 = vmlal_s8(vprod0x10, vget_high_s8(vb10), vget_high_s8(va0));
223       vprod1x10 = vmlal_s8(vprod1x10, vget_high_s8(vb10), vget_high_s8(va1));
224       vprod2x10 = vmlal_s8(vprod2x10, vget_high_s8(vb10), vget_high_s8(va2));
225       vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
226       vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
227       vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
228       int16x8_t vprod0x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va0));
229       int16x8_t vprod1x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va1));
230       int16x8_t vprod2x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va2));
231       vprod0x11 = vmlal_s8(vprod0x11, vget_high_s8(vb11), vget_high_s8(va0));
232       vprod1x11 = vmlal_s8(vprod1x11, vget_high_s8(vb11), vget_high_s8(va1));
233       vprod2x11 = vmlal_s8(vprod2x11, vget_high_s8(vb11), vget_high_s8(va2));
234       vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
235       vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
236       vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
237       int16x8_t vprod0x12 = vmull_s8(vget_low_s8(vb12), vget_low_s8(va0));
238       int16x8_t vprod1x12 = vmull_s8(vget_low_s8(vb12), vget_low_s8(va1));
239       int16x8_t vprod2x12 = vmull_s8(vget_low_s8(vb12), vget_low_s8(va2));
240       vprod0x12 = vmlal_s8(vprod0x12, vget_high_s8(vb12), vget_high_s8(va0));
241       vprod1x12 = vmlal_s8(vprod1x12, vget_high_s8(vb12), vget_high_s8(va1));
242       vprod2x12 = vmlal_s8(vprod2x12, vget_high_s8(vb12), vget_high_s8(va2));
243       vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
244       vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
245       vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
246       int16x8_t vprod0x13 = vmull_s8(vget_low_s8(vb13), vget_low_s8(va0));
247       int16x8_t vprod1x13 = vmull_s8(vget_low_s8(vb13), vget_low_s8(va1));
248       int16x8_t vprod2x13 = vmull_s8(vget_low_s8(vb13), vget_low_s8(va2));
249       vprod0x13 = vmlal_s8(vprod0x13, vget_high_s8(vb13), vget_high_s8(va0));
250       vprod1x13 = vmlal_s8(vprod1x13, vget_high_s8(vb13), vget_high_s8(va1));
251       vprod2x13 = vmlal_s8(vprod2x13, vget_high_s8(vb13), vget_high_s8(va2));
252       vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
253       vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
254       vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
255       int16x8_t vprod0x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va0));
256       int16x8_t vprod1x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va1));
257       int16x8_t vprod2x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va2));
258       vprod0x14 = vmlal_s8(vprod0x14, vget_high_s8(vb14), vget_high_s8(va0));
259       vprod1x14 = vmlal_s8(vprod1x14, vget_high_s8(vb14), vget_high_s8(va1));
260       vprod2x14 = vmlal_s8(vprod2x14, vget_high_s8(vb14), vget_high_s8(va2));
261       vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
262       vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
263       vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
264       int16x8_t vprod0x15 = vmull_s8(vget_low_s8(vb15), vget_low_s8(va0));
265       int16x8_t vprod1x15 = vmull_s8(vget_low_s8(vb15), vget_low_s8(va1));
266       int16x8_t vprod2x15 = vmull_s8(vget_low_s8(vb15), vget_low_s8(va2));
267       vprod0x15 = vmlal_s8(vprod0x15, vget_high_s8(vb15), vget_high_s8(va0));
268       vprod1x15 = vmlal_s8(vprod1x15, vget_high_s8(vb15), vget_high_s8(va1));
269       vprod2x15 = vmlal_s8(vprod2x15, vget_high_s8(vb15), vget_high_s8(va2));
270       vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
271       vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
272       vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
273 
274       k -= 16 * sizeof(int8_t);
275     }
276 
277 #if XNN_ARCH_ARM64
278     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
279     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
280     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
281     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
282     const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
283     const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
284     const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
285     const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
286     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
287     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
288     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
289     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
290     const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9);
291     const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11);
292     const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13);
293     const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15);
294     const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
295     const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
296     const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
297     const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
298     const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9);
299     const int32x4_t vsum2xAB = vpaddq_s32(vacc2x10, vacc2x11);
300     const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13);
301     const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15);
302     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
303     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
304     int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
305     int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
306     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
307     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
308     int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB);
309     int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF);
310     int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
311     int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
312     int32x4_t vacc2x89AB = vpaddq_s32(vsum2x89, vsum2xAB);
313     int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF);
314 #else
315     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
316     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
317     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
318     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
319     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
320     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
321     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
322     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
323     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
324     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
325     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
326     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
327     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
328     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
329     const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
330     const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
331     const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
332     const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
333     const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
334     const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
335     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
336     const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
337     const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
338     const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
339     const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
340     const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
341     const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
342     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
343     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
344     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
345     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
346     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
347     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
348     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
349     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
350     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
351     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
352     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
353     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
354     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
355     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
356     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
357     const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));
358     const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));
359     const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));
360     const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));
361     const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9);
362     const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB);
363     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB );
364     const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));
365     const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));
366     const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));
367     const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));
368     const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD);
369     const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF);
370     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF );
371     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
372     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
373     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
374     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
375     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
376     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
377     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
378     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
379     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
380     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
381     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
382     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
383     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
384     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
385     const int32x2_t vpsum2x8 = vadd_s32(vget_low_s32(vacc2x8), vget_high_s32(vacc2x8));
386     const int32x2_t vpsum2x9 = vadd_s32(vget_low_s32(vacc2x9), vget_high_s32(vacc2x9));
387     const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10));
388     const int32x2_t vpsum2xB = vadd_s32(vget_low_s32(vacc2x11), vget_high_s32(vacc2x11));
389     const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9);
390     const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB);
391     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB );
392     const int32x2_t vpsum2xC = vadd_s32(vget_low_s32(vacc2x12), vget_high_s32(vacc2x12));
393     const int32x2_t vpsum2xD = vadd_s32(vget_low_s32(vacc2x13), vget_high_s32(vacc2x13));
394     const int32x2_t vpsum2xE = vadd_s32(vget_low_s32(vacc2x14), vget_high_s32(vacc2x14));
395     const int32x2_t vpsum2xF = vadd_s32(vget_low_s32(vacc2x15), vget_high_s32(vacc2x15));
396     const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD);
397     const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF);
398     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF );
399 #endif
400 
401     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
402     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
403     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
404 
405     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
406     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
407     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
408     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
409     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
410     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
411     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
412     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
413     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
414     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
415     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
416     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
417 
418     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
419     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
420     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
421     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
422     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
423     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
424     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
425     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
426     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
427     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
428     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
429     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
430 
431     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
432     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
433     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
434     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
435     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
436     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
437     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
438     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
439     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
440     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
441     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
442     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
443 
444     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
445 #if XNN_ARCH_ARM64
446     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
447     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
448     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
449     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
450     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
451     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
452     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
453     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
454     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
455 #else
456     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
457     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
458     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
459     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
460     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
461     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
462 
463     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
464     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
465     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
466 #endif
467     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
468     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
469 
470     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
471     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
472     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
473 
474     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
475     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
476     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
477 
478     if (nc >= 16) {
479       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
480       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
481       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
482 
483       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
484       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
485       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
486 
487       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
488       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
489       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
490 
491       nc -= 16;
492     } else {
493       // Final case where not all of the 16 columns fit in the destination.
494       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
495       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
496       if (nc & 8) {
497         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
498         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
499         vst1_s8(c2, vout2x01234567); c2 += 8;
500         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
501         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
502       }
503       if (nc & 4) {
504         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
505         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
506         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
507         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
508         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
509       }
510       if (nc & 2) {
511         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
512         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
513         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
514         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
515         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
516       }
517       if (nc & 1) {
518         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
519         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
520         vst1_lane_s8(c2, vout2x01234567, 0);
521       }
522 
523       nc = 0;
524     }
525   } while (nc != 0);
526 }
527