xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x16c4s2-minmax-rndnu-neon-mlal.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c4-neon-mull-shuffle.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_rndnu_ukernel_2x16c4s2__neon_mlal(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)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal(
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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (2 * 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   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 
49   kc = round_up_po2(kc, 8 * sizeof(int8_t));
50   do {
51     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
52     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
53     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
54     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
55     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
56     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
59     int32x4_t vacc1x01 = vacc0x01;
60     int32x4_t vacc1x23 = vacc0x23;
61     int32x4_t vacc1x45 = vacc0x45;
62     int32x4_t vacc1x67 = vacc0x67;
63     int32x4_t vacc1x89 = vacc0x89;
64     int32x4_t vacc1xAB = vacc0xAB;
65     int32x4_t vacc1xCD = vacc0xCD;
66     int32x4_t vacc1xEF = vacc0xEF;
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       a += 2;
79 
80       size_t k = kc;
81       while (k >= 16 * sizeof(int8_t)) {
82         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
83         int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
84         int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
85         int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
86 
87         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91         const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92         const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93         const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94         const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
96         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
97         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
98         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
99         const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
100         const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
101         const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
102         const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
103 
104         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
105         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
106         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
107         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
108         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
109         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
110         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
111         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
112         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
113         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
114         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
115         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
116         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
117         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
118         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
119         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
120         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
121         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
122         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
123         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
124         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
125         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
126         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
127         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
128         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
129         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
130         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
131         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
132         int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
133         int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
134         const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
135         vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1);
136         vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1);
137         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
138         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
139         int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
140         int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
141         const int8x8_t vbABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
142         vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0x1);
143         vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1x1);
144         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
145         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
146         int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
147         int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
148         const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
149         vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0x1);
150         vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1x1);
151         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
152         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
153         int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
154         int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
155         const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
156         vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0x1);
157         vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1x1);
158         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
159         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
160         va0x0 = vext_s8(va0x0, va0x0, 4);
161         va0x1 = vext_s8(va0x1, va0x1, 4);
162         va1x0 = vext_s8(va1x0, va1x0, 4);
163         va1x1 = vext_s8(va1x1, va1x1, 4);
164         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
165         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
166         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
167         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
168         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
169         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
170         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
171         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
172         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
173         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
174         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
175         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
176         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
177         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
178         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
179         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
180         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
181         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
182         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
183         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
184         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
185         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
186         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
187         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
188         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
189         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
190         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
191         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
192         int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
193         int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
194         const int8x8_t vb89c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
195         vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0x1);
196         vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1);
197         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
198         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
199         int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
200         int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
201         const int8x8_t vbABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
202         vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0x1);
203         vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1);
204         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
205         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
206         int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
207         int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
208         const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
209         vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0x1);
210         vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1x1);
211         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
212         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
213         int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
214         int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
215         const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
216         vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0x1);
217         vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1x1);
218         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
219         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
220 
221         k -= 16 * sizeof(int8_t);
222       }
223       if (k != 0) {
224         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
225         int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
226 
227         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
228         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
229         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
230         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
231         const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
232         const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
233         const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
234         const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
235         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
236         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
237         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
238         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
239         const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
240         const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
241         const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
242         const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
243 
244         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
245         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
246         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
247         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
248         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
249         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
250         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
251         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
252         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
253         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
254         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
255         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
256         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
257         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
258         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
259         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
260         int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
261         int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
262         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
263         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
264         int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
265         int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
266         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
267         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
268         int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
269         int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
270         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
271         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
272         int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
273         int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
274         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
275         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
276         va0x0 = vext_s8(va0x0, va0x0, 4);
277         va1x0 = vext_s8(va1x0, va1x0, 4);
278         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
279         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
280         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
281         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
282         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
283         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
284         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
285         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
286         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
287         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
288         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
289         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
290         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
291         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
292         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
293         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
294         int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
295         int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
296         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
297         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
298         int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
299         int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
300         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
301         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
302         int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
303         int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
304         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
305         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
306         int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
307         int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
308         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
309         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
310 
311       }
312 
313       p -= 2 * sizeof(void*);
314     } while (p != 0);
315 
316 #if XNN_ARCH_ARM64
317     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
318     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
319     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
320     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
321     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
322     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
323     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
324     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
325 #else
326     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
327     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
328     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
329     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
330     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
331     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
332     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
333     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
334     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
335     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
336     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
337     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
338     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
339     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
340     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
341     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
342     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
343     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
344     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
345     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
346     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
347     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
348     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
349     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
350 #endif
351 
352     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
353     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
354     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
355 
356     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
357     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
358     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
359     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
360     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
361     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
362     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
363     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
364 
365     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
366     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
367     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
368     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
369     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
370     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
371     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
372     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
373 
374     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
375     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
376     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
377     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
378     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
379     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
380     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
381     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
382 
383     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
384 #if XNN_ARCH_ARM64
385     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
386     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
387     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
388     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
389 
390     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
391     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
392     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
393     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
394 
395     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
396     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
397 #else
398     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
399     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
400     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
401     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
402 
403     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
404     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
405     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
406     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
407 
408     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
409     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
410 #endif
411 
412     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
413     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
414     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
415 
416     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
417     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
418     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
419 
420     if (nc >= 16) {
421       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
422       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
423 
424       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
425       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
426 
427       a = (const int8_t**restrict) ((uintptr_t) a - ks);
428 
429       nc -= 16;
430     } else {
431       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
432       if (nc & 8) {
433         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
434         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
435         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
436       }
437       if (nc & 4) {
438         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
439         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
440         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
441       }
442       if (nc & 2) {
443         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
444         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
445         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
446       }
447       if (nc & 1) {
448         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
449         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
450       }
451 
452       nc = 0;
453     }
454   } while (nc != 0);
455 }
456