• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-padal-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_ukernel_3x16c2__neon_mlal_padal_dup(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
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, 2);
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 vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
58     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
59     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
60     int32x4_t vacc1x0123 = vacc0x0123;
61     int32x4_t vacc1x4567 = vacc0x4567;
62     int32x4_t vacc1x89AB = vacc0x89AB;
63     int32x4_t vacc1xCDEF = vacc0xCDEF;
64     int32x4_t vacc2x0123 = vacc0x0123;
65     int32x4_t vacc2x4567 = vacc0x4567;
66     int32x4_t vacc2x89AB = vacc0x89AB;
67     int32x4_t vacc2xCDEF = vacc0xCDEF;
68 
69     size_t k = kc;
70 
71     while (k >= 16 * sizeof(int8_t)) {
72       const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
73       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
74       const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
75       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
76       const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
77       const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
78 
79       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 
96       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
97       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
98       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
99       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
101       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
102       vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
103       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
104       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
105       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
106       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
107       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
108       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
109       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
111       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
112       vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
113       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
114       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
115       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
116       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
117       int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
118       int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
119       const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120       vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
121       vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
122       vprod2x89ABc0 = vmlal_s8(vprod2x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
123       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
124       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
125       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
126       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
127       int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
128       int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
129       const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130       vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
131       vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
132       vprod2xCDEFc0 = vmlal_s8(vprod2xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
133       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
134       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
135       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
136       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
137       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
138       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
139       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
141       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
142       vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
143       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
144       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
145       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
146       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
147       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
148       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
149       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
151       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
152       vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
153       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
154       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
155       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
156       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
157       int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
158       int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
159       const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160       vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
161       vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
162       vprod2x89ABc1 = vmlal_s8(vprod2x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
163       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
164       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
165       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
166       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
167       int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
168       int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
169       const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170       vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
171       vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
172       vprod2xCDEFc1 = vmlal_s8(vprod2xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
173       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
174       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
175       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
176       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
177       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
178       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
179       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
181       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
182       vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
183       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
184       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
185       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
186       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
187       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
188       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
189       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
190       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
191       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
192       vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
193       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
194       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
195       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
196       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
197       int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
198       int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
199       const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
200       vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
201       vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
202       vprod2x89ABc2 = vmlal_s8(vprod2x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
203       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
204       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
205       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
206       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
207       int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
208       int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
209       const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
210       vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
211       vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
212       vprod2xCDEFc2 = vmlal_s8(vprod2xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
213       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
214       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
215       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
216       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
217       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
218       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
219       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
221       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
222       vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
223       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
224       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
225       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
226       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
227       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
228       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
229       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
231       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
232       vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
233       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
234       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
235       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
236       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
237       int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
238       int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
239       const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
240       vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
241       vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
242       vprod2x89ABc3 = vmlal_s8(vprod2x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
243       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
244       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
245       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
246       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
247       int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
248       int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
249       const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250       vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
251       vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
252       vprod2xCDEFc3 = vmlal_s8(vprod2xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
253       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
254       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
255       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
256 
257       k -= 16 * sizeof(int8_t);
258     }
259 
260     if (k >= 8 * sizeof(int8_t)) {
261       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
262       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
263       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
264 
265       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
266       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
267       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
268       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
269       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
270       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
271       const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
272       const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
273       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
274       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
275       const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
276       const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
277       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
278       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
279       const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
280       const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
281 
282       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
283       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
284       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
285       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
286       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
287       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
288       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
289       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
290       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
291       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
292       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
293       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
294       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
295       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
296       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
297       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
298       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
299       const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
300       const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
301       const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
302       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
303       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
304       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
305       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
306       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
307       const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
308       const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
309       const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
310       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
311       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
312       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
313       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
314       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
315       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
316       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
317       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
318       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
319       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
320       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
321       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
322       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
323       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
324       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
325       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
326       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
327       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
328       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
329       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
330       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
331       const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
332       const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
333       const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
334       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
335       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
336       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
337       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
338       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
339       const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
340       const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
341       const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
342       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
343       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
344       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
345       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
346       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
347       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
348       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
349       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
350       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
351       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
352       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
353       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
354       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
355       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
356       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
357       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
358       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
359       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
360       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
361       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
362       const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
363       const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
364       const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
365       const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
366       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
367       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
368       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
369       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
370       const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
371       const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
372       const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
373       const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
374       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
375       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
376       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
377       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
378 
379       k -= 8 * sizeof(int8_t);
380     }
381 
382     if XNN_UNLIKELY(k != 0) {
383       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
384       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
385       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
386 
387       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
388       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
389       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
390       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
391 
392       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
393       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
394       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
395       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
396       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
397       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
398       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
399       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
400       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
401       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
402       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
403       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
404       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
405       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
406       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
407       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
408       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
409       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
410       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
411       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
412       const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
413       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
414       const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
415       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
416 
417       if (k > 2 * sizeof(int8_t)) {
418         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
419         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
420         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
421         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
422 
423         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
424         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
425         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
426         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
427         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
428         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
429         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
430         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
431         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
432         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
433         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
434         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
435         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
436         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
437         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
438         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
439         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
440         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
441         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
442         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
443         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
444         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
445         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
446         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
447 
448         if (k > 4 * sizeof(int8_t)) {
449           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
450           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
451           const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
452           const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
453 
454           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
455           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
456           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
457           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
458           const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
459           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
460           const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
461           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
462           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
463           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
464           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
465           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
466           const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
467           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
468           const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
469           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
470           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
471           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
472           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
473           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
474           const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
475           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
476           const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
477           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
478         }
479       }
480     }
481     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
482     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
483     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
484     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);
485     vacc0xCDEF = vqrdmulhq_s32(vacc0xCDEF, vmultiplier);
486     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
487     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
488     vacc1x89AB = vqrdmulhq_s32(vacc1x89AB, vmultiplier);
489     vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);
490     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
491     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
492     vacc2x89AB = vqrdmulhq_s32(vacc2x89AB, vmultiplier);
493     vacc2xCDEF = vqrdmulhq_s32(vacc2xCDEF, vmultiplier);
494 
495     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
496     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
497     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
498     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
499     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);
500     vacc0xCDEF = vsraq_n_s32(vacc0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31);
501     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
502     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
503     vacc1x89AB = vsraq_n_s32(vacc1x89AB, vbicq_s32(vacc1x89AB, vzero_shift_mask), 31);
504     vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);
505     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
506     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
507     vacc2x89AB = vsraq_n_s32(vacc2x89AB, vbicq_s32(vacc2x89AB, vzero_shift_mask), 31);
508     vacc2xCDEF = vsraq_n_s32(vacc2xCDEF, vbicq_s32(vacc2xCDEF, vzero_shift_mask), 31);
509 
510     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
511     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
512     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);
513     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_shift);
514     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
515     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
516     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_shift);
517     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);
518     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
519     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
520     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_shift);
521     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_shift);
522 
523     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
524 #if XNN_ARCH_ARM64
525     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
526     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
527     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
528     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
529     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
530     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
531 
532     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
533     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
534     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
535 #else
536     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
537     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
538     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
539     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
540     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
541     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
542 
543     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
544     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
545     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
546 #endif
547     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
548     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
549 
550     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
551     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
552     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
553 
554     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
555     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
556     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
557 
558     if (nc >= 16) {
559       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
560       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
561       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
562 
563       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
564       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
565       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
566 
567       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
568       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
569       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
570 
571       nc -= 16;
572     } else {
573       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
574       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
575       if (nc & 8) {
576         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
577         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
578         vst1_s8(c2, vout2x01234567); c2 += 8;
579         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
580         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
581       }
582       if (nc & 4) {
583         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
584         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
585         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
586         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
587         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
588       }
589       if (nc & 2) {
590         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
591         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
592         vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
593         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
594         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
595       }
596       if (nc & 1) {
597         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
598         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
599         vst1_lane_s8(c2, vout2x01234567, 0);
600       }
601 
602       nc = 0;
603     }
604   } while (nc != 0);
605 }
606