1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t** restrict a,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
31 {
32   assert(mr != 0);
33   assert(mr <= 3);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (3 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(int8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   kc = round_up_po2(kc, 2);
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr < 2) {
47     c1 = c0;
48   }
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     c2 = c1;
52   }
53 
54   do {
55     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
58     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
59     int32x4_t vacc1x0123 = vacc0x0123;
60     int32x4_t vacc1x4567 = vacc0x4567;
61     int32x4_t vacc1x89AB = vacc0x89AB;
62     int32x4_t vacc1xCDEF = vacc0xCDEF;
63     int32x4_t vacc2x0123 = vacc0x0123;
64     int32x4_t vacc2x4567 = vacc0x4567;
65     int32x4_t vacc2x89AB = vacc0x89AB;
66     int32x4_t vacc2xCDEF = vacc0xCDEF;
67 
68     size_t p = ks;
69     do {
70       const int8_t* restrict a0 = a[0];
71       if XNN_UNPREDICTABLE(a0 != zero) {
72         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73       }
74       const int8_t* restrict a1 = a[1];
75       if XNN_UNPREDICTABLE(a1 != zero) {
76         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77       }
78       const int8_t* restrict a2 = a[2];
79       if XNN_UNPREDICTABLE(a2 != zero) {
80         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
81       }
82       a += 3;
83 
84       size_t k = kc;
85 
86       while (k >= 16 * sizeof(int8_t)) {
87         const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
88         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
89         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
90         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
91         const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
92         const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
93 
94         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96         const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97         const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100         const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104         const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109         const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 
111         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
112         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
113         int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
114         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
116         vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
117         vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
118         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
119         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
120         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
121         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
122         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
123         int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
124         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
126         vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
127         vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
128         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
129         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
130         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
131         int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
132         int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
133         int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
134         const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135         vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
136         vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
137         vprod2x89ABc0 = vmlal_s8(vprod2x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
138         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
139         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
140         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
141         int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
142         int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
143         int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
144         const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
145         vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
146         vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
147         vprod2xCDEFc0 = vmlal_s8(vprod2xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
148         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
149         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
150         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
151         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
152         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
153         int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
154         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
156         vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
157         vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
158         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
159         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
160         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
161         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
162         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
163         int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
164         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
166         vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
167         vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
168         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
169         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
170         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
171         int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
172         int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
173         int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
174         const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175         vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
176         vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
177         vprod2x89ABc1 = vmlal_s8(vprod2x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
178         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
179         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
180         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
181         int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
182         int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
183         int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
184         const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185         vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
186         vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
187         vprod2xCDEFc1 = vmlal_s8(vprod2xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
188         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
189         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
190         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
191         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
192         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
193         int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
194         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
195         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
196         vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
197         vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
198         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
199         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
200         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
201         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
202         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
203         int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
204         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
206         vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
207         vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
208         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
209         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
210         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
211         int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
212         int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
213         int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
214         const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
215         vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
216         vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
217         vprod2x89ABc2 = vmlal_s8(vprod2x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
218         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
219         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
220         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
221         int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
222         int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
223         int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
224         const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225         vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
226         vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
227         vprod2xCDEFc2 = vmlal_s8(vprod2xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
228         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
229         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
230         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
231         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
232         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
233         int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
234         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
236         vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
237         vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
238         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
239         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
240         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
241         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
242         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
243         int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
244         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
246         vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
247         vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
248         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
249         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
250         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
251         int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
252         int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
253         int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
254         const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255         vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
256         vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
257         vprod2x89ABc3 = vmlal_s8(vprod2x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
258         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
259         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
260         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
261         int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
262         int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
263         int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
264         const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
265         vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
266         vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
267         vprod2xCDEFc3 = vmlal_s8(vprod2xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
268         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
269         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
270         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
271 
272         k -= 16 * sizeof(int8_t);
273       }
274 
275       if (k >= 8 * sizeof(int8_t)) {
276         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
277         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
278         const int8x8_t va2 = vld1_s8(a2); a2 += 8;
279 
280         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
281         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
282         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
283         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
284         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
285         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
286         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
287         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
288         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
289         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
290         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
291         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
292         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
293         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
294         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
295         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
296 
297         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
298         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
299         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
300         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
301         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
302         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
303         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
304         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
305         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
306         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
307         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
308         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
309         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
310         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
311         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
312         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
313         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
314         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
315         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
316         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
317         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
318         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
319         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
320         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
321         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
322         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
323         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
324         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
325         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
326         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
327         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
328         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
329         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
330         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
331         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
332         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
333         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
334         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
335         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
336         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
337         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
338         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
339         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
340         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
341         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
342         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
343         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
344         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
345         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
346         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
347         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
348         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
349         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
350         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
351         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
352         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
353         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
354         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
355         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
356         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
357         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
358         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
359         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
360         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
361         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
362         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
363         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
364         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
365         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
366         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
367         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
368         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
369         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
370         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
371         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
372         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
373         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
374         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
375         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
376         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
377         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
378         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
379         const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
380         const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
381         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
382         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
383         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
384         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
385         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
386         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
387         const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
388         const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
389         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
390         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
391         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
392         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
393 
394         k -= 8 * sizeof(int8_t);
395       }
396 
397       if XNN_UNLIKELY(k != 0) {
398         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
399         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
400         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
401 
402         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
403         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
404         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
405         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
406 
407         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
408         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
409         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
410         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
411         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
412         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
413         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
414         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
415         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
416         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
417         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
418         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
419         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
420         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
421         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
422         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
423         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
424         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
425         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
426         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
427         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
428         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
429         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
430         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
431 
432         if (k > 2 * sizeof(int8_t)) {
433           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
434           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
435           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
436           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
437 
438           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
439           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
440           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
441           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
442           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
443           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
444           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
445           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
446           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
447           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
448           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
449           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
450           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
451           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
452           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
453           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
454           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
455           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
456           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
457           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
458           const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
459           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
460           const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
461           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
462 
463           if (k > 4 * sizeof(int8_t)) {
464             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
465             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
466             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
467             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
468 
469             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
470             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
471             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
472             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
473             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
474             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
475             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
476             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
477             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
478             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
479             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
480             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
481             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
482             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
483             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
484             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
485             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
486             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
487             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
488             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
489             const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
490             vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
491             const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
492             vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
493           }
494         }
495       }
496       p -= 3 * sizeof(void*);
497     } while (p != 0);
498 
499     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
500     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
501     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
502     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);
503     vacc0xCDEF = vqrdmulhq_s32(vacc0xCDEF, vmultiplier);
504     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
505     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
506     vacc1x89AB = vqrdmulhq_s32(vacc1x89AB, vmultiplier);
507     vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);
508     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
509     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
510     vacc2x89AB = vqrdmulhq_s32(vacc2x89AB, vmultiplier);
511     vacc2xCDEF = vqrdmulhq_s32(vacc2xCDEF, vmultiplier);
512 
513     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
514     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
515     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
516     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
517     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);
518     vacc0xCDEF = vsraq_n_s32(vacc0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31);
519     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
520     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
521     vacc1x89AB = vsraq_n_s32(vacc1x89AB, vbicq_s32(vacc1x89AB, vzero_shift_mask), 31);
522     vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);
523     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
524     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
525     vacc2x89AB = vsraq_n_s32(vacc2x89AB, vbicq_s32(vacc2x89AB, vzero_shift_mask), 31);
526     vacc2xCDEF = vsraq_n_s32(vacc2xCDEF, vbicq_s32(vacc2xCDEF, vzero_shift_mask), 31);
527 
528     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
529     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
530     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);
531     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_shift);
532     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
533     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
534     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_shift);
535     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);
536     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
537     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
538     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_shift);
539     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_shift);
540 
541     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
542 #if XNN_ARCH_ARM64
543     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
544     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
545     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
546     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
547     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
548     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
549 
550     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
551     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
552     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
553 #else
554     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
555     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
556     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
557     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
558     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
559     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
560 
561     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
562     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
563     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
564 #endif
565     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
566     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
567 
568     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
569     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
570     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
571 
572     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
573     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
574     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
575 
576     if (nc >= 16) {
577       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
578       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
579       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
580 
581       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
582       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
583       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
584 
585       a = (const int8_t**restrict) ((uintptr_t) a - ks);
586 
587       nc -= 16;
588     } else {
589       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
590       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
591       if (nc & 8) {
592         vst1_s8(c2, vout2x01234567); c2 += 8;
593         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
594         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
595         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
596         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
597       }
598       if (nc & 4) {
599         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
600         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
601         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
602         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
603         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
604       }
605       if (nc & 2) {
606         vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
607         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
608         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
609         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
610         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
611       }
612       if (nc & 1) {
613         vst1_lane_s8(c2, vout2x01234567, 0);
614         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
615         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
616       }
617 
618       nc = 0;
619     }
620   } while (nc != 0);
621 }
622