xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/3x16c2-minmax-rndnu-neon-mlal-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c2-neon-mull-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
29     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 3);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (3 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
43   int8_t* c0 = c;
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52 
53   do {
54     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
55     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
58     int32x4_t vacc1x0123 = vacc0x0123;
59     int32x4_t vacc1x4567 = vacc0x4567;
60     int32x4_t vacc1x89AB = vacc0x89AB;
61     int32x4_t vacc1xCDEF = vacc0xCDEF;
62     int32x4_t vacc2x0123 = vacc0x0123;
63     int32x4_t vacc2x4567 = vacc0x4567;
64     int32x4_t vacc2x89AB = vacc0x89AB;
65     int32x4_t vacc2xCDEF = vacc0xCDEF;
66 
67     size_t p = ks;
68     do {
69       const int8_t* restrict a0 = a[0];
70       if XNN_UNPREDICTABLE(a0 != zero) {
71         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72       }
73       const int8_t* restrict a1 = a[1];
74       if XNN_UNPREDICTABLE(a1 != zero) {
75         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
76       }
77       const int8_t* restrict a2 = a[2];
78       if XNN_UNPREDICTABLE(a2 != zero) {
79         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
80       }
81       a += 3;
82 
83       size_t k = kc;
84 
85       while (k >= 16 * sizeof(int8_t)) {
86         const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
87         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
88         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
89         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
90         const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
91         const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
92 
93         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95         const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96         const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99         const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100         const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103         const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104         const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107         const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 
110         const int8x8_t va0c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0));
111         const int8x8_t va0c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0));
112         const int8x8_t va1c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0));
113         const int8x8_t va1c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0));
114         const int8x8_t va2c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0));
115         const int8x8_t va2c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0));
116 
117         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
118         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
119         int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
120         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
122         vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
123         vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
124         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
125         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
126         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
127         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
128         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
129         int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
130         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
132         vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
133         vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
134         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
135         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
136         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
137         int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
138         int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0);
139         int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2c0x0);
140         const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141         vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
142         vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1);
143         vprod2x89ABc0 = vmlal_s8(vprod2x89ABc0, vb89ABc0x1, va2c0x1);
144         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
145         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
146         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
147         int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
148         int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1c0x0);
149         int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2c0x0);
150         const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151         vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
152         vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1c0x1);
153         vprod2xCDEFc0 = vmlal_s8(vprod2xCDEFc0, vbCDEFc0x1, va2c0x1);
154         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
155         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
156         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
157         const int8x8_t va0c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1));
158         const int8x8_t va0c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1));
159         const int8x8_t va1c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1));
160         const int8x8_t va1c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1));
161         const int8x8_t va2c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1));
162         const int8x8_t va2c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1));
163 
164         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
165         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
166         int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
167         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
169         vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
170         vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
171         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
172         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
173         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
174         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
175         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
176         int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
177         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
179         vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
180         vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
181         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
182         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
183         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
184         int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
185         int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0);
186         int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2c1x0);
187         const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188         vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
189         vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1c1x1);
190         vprod2x89ABc1 = vmlal_s8(vprod2x89ABc1, vb89ABc1x1, va2c1x1);
191         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
192         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
193         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
194         int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
195         int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0);
196         int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2c1x0);
197         const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
198         vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
199         vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1);
200         vprod2xCDEFc1 = vmlal_s8(vprod2xCDEFc1, vbCDEFc1x1, va2c1x1);
201         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
202         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
203         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
204         const int8x8_t va0c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2));
205         const int8x8_t va0c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2));
206         const int8x8_t va1c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2));
207         const int8x8_t va1c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2));
208         const int8x8_t va2c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2));
209         const int8x8_t va2c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2));
210 
211         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
212         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
213         int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
214         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
215         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
216         vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
217         vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
218         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
219         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
220         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
221         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
222         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
223         int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
224         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
226         vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
227         vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
228         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
229         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
230         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
231         int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
232         int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1c2x0);
233         int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2c2x0);
234         const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235         vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
236         vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1c2x1);
237         vprod2x89ABc2 = vmlal_s8(vprod2x89ABc2, vb89ABc2x1, va2c2x1);
238         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
239         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
240         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
241         int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
242         int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1c2x0);
243         int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2c2x0);
244         const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245         vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
246         vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1c2x1);
247         vprod2xCDEFc2 = vmlal_s8(vprod2xCDEFc2, vbCDEFc2x1, va2c2x1);
248         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
249         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
250         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
251         const int8x8_t va0c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3));
252         const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3));
253         const int8x8_t va1c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3));
254         const int8x8_t va1c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3));
255         const int8x8_t va2c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3));
256         const int8x8_t va2c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3));
257 
258         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
259         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
260         int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
261         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
262         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
263         vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
264         vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
265         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
266         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
267         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
268         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
269         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
270         int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
271         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
272         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
273         vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
274         vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
275         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
276         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
277         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
278         int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
279         int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1c3x0);
280         int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0);
281         const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
282         vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
283         vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1);
284         vprod2x89ABc3 = vmlal_s8(vprod2x89ABc3, vb89ABc3x1, va2c3x1);
285         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
286         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
287         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
288         int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
289         int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1c3x0);
290         int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2c3x0);
291         const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
292         vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
293         vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1);
294         vprod2xCDEFc3 = vmlal_s8(vprod2xCDEFc3, vbCDEFc3x1, va2c3x1);
295         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
296         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
297         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
298 
299         k -= 16 * sizeof(int8_t);
300       }
301 
302       if (k >= 8 * sizeof(int8_t)) {
303         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
304         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
305         const int8x8_t va2 = vld1_s8(a2); a2 += 8;
306 
307         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
308         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
309         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
310         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
311         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
312         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
313         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
314         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
315         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
316         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
317         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
318         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
319         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
320         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
321         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
322         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
323 
324         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
325         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
326         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
327 
328         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
329         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
330         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
331         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
332         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
333         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
334         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
335         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
336         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
337         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
338         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
339         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
340         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
341         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
342         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
343         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
344         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
345         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
346         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
347         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
348         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
349         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
350         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
351         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
352         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
353         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
354         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
355 
356         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
357         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
358         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
359         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
360         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
361         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
362         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
363         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
364         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
365         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
366         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
367         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
368         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
369         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
370         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
371         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
372         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
373         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
374         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
375         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
376         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
377         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
378         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
379         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
380         const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
381         const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
382         const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
383 
384         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
385         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
386         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
387         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
388         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
389         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
390         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
391         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
392         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
393         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
394         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
395         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
396         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
397         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
398         const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
399         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
400         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
401         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
402         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
403         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
404         const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
405         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
406         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
407         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
408         const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
409         const int8x8_t va1c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3));
410         const int8x8_t va2c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3));
411 
412         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
413         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
414         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
415         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
416         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
417         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
418         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
419         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
420         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
421         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
422         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
423         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
424         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
425         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
426         const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
427         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
428         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
429         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
430         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
431         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
432         const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
433         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
434         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
435         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
436 
437         k -= 8 * sizeof(int8_t);
438       }
439 
440       if XNN_UNLIKELY(k != 0) {
441         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
442         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
443         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
444 
445         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
446         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
447         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
448         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
449 
450         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
451         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
452         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
453         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
454         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
455         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
456         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
457         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
458         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
459         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
460         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
461         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
462         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
463         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
464         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
465         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
466         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
467         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
468         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
469         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
470         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
471         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
472         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
473         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
474         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
475         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
476         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
477 
478         if (k > 2 * sizeof(int8_t)) {
479           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
480           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
481           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
482           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
483 
484           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
485           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
486           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
487           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
488           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
489           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
490           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
491           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
492           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
493           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
494           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
495           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
496           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
497           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
498           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
499           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
500           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
501           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
502           const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
503           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
504           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
505           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
506           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
507           const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
508           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
509           const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
510           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
511 
512           if (k > 4 * sizeof(int8_t)) {
513             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
514             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
515             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
516             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
517 
518             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
519             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
520             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
521             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
522             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
523             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
524             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
525             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
526             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
527             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
528             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
529             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
530             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
531             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
532             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
533             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
534             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
535             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
536             const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
537             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
538             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
539             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
540             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
541             const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
542             vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
543             const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
544             vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
545           }
546         }
547       }
548       p -= 3 * sizeof(void*);
549     } while (p != 0);
550 
551     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
552     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
553     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
554 
555     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
556     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
557     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
558     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
559     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
560     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
561     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
562     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
563     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
564     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
565     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
566     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
567 
568     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
569     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
570     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
571     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
572     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
573     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
574     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
575     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
576     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
577     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
578     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
579     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
580 
581     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
582     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
583     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
584     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
585     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
586     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
587     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
588     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
589     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
590     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
591     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
592     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
593 
594     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
595 #if XNN_ARCH_ARM64
596     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
597     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
598     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
599     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
600     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
601     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
602 
603     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
604     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
605     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
606     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
607     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
608     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
609 
610     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
611     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
612     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
613 #else
614     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
615     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
616     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
617     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
618     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
619     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
620 
621     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
622     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
623     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
624     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
625     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
626     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
627 
628     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
629     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
630     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
631 #endif
632 
633     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
634     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
635     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
636     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
637 
638     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
639     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
640     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
641     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
642 
643     if (nc >= 16) {
644       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
645       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
646       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
647 
648       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
649       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
650       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
651 
652       a = (const int8_t**restrict) ((uintptr_t) a - ks);
653 
654       nc -= 16;
655     } else {
656       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
657       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
658       if (nc & 8) {
659         vst1_s8(c2, vout2x01234567); c2 += 8;
660         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
661         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
662         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
663         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
664       }
665       if (nc & 4) {
666         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
667         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
668         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
669         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
670         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
671       }
672       if (nc & 2) {
673         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
674         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
675         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
676         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
677         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
678       }
679       if (nc & 1) {
680         vst1_lane_s8(c2, vout2x01234567, 0);
681         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
682         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
683       }
684 
685       nc = 0;
686     }
687   } while (nc != 0);
688 }
689