xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/3x16c4s2-minmax-rndnu-neon-mull.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_3x16c4s2__neon_mull(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_3x16c4s2__neon_mull(
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 <= 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   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   kc = round_up_po2(kc, 8 * sizeof(int8_t));
54   do {
55     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
56     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
59     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
60     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
61     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
62     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
63     int32x4_t vacc1x01 = vacc0x01;
64     int32x4_t vacc1x23 = vacc0x23;
65     int32x4_t vacc1x45 = vacc0x45;
66     int32x4_t vacc1x67 = vacc0x67;
67     int32x4_t vacc1x89 = vacc0x89;
68     int32x4_t vacc1xAB = vacc0xAB;
69     int32x4_t vacc1xCD = vacc0xCD;
70     int32x4_t vacc1xEF = vacc0xEF;
71     int32x4_t vacc2x01 = vacc0x01;
72     int32x4_t vacc2x23 = vacc0x23;
73     int32x4_t vacc2x45 = vacc0x45;
74     int32x4_t vacc2x67 = vacc0x67;
75     int32x4_t vacc2x89 = vacc0x89;
76     int32x4_t vacc2xAB = vacc0xAB;
77     int32x4_t vacc2xCD = vacc0xCD;
78     int32x4_t vacc2xEF = vacc0xEF;
79 
80     size_t p = ks;
81     do {
82       const int8_t* restrict a0 = a[0];
83       if XNN_UNPREDICTABLE(a0 != zero) {
84         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
85       }
86       const int8_t* restrict a1 = a[1];
87       if XNN_UNPREDICTABLE(a1 != zero) {
88         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
89       }
90       const int8_t* restrict a2 = a[2];
91       if XNN_UNPREDICTABLE(a2 != zero) {
92         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93       }
94       a += 3;
95 
96       size_t k = kc;
97       do {
98         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
99         int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
100         int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
101 
102         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
103         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
104         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
105         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
106         const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
107         const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108         const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109         const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
112         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
113         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
114         const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
115         const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
116         const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
117         const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
118 
119         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
120         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
121         int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
122         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
123         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
124         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
125         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
126         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
127         int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
128         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
129         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
130         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
131         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
132         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
133         int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
134         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
135         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
136         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
137         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
138         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
139         int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
140         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
141         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
142         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
143         int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
144         int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
145         int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
146         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
147         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
148         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
149         int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
150         int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
151         int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
152         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
153         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
154         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
155         int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
156         int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
157         int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
158         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
159         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
160         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
161         int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
162         int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
163         int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
164         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
165         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
166         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
167         va0x0 = vext_s8(va0x0, va0x0, 4);
168         va1x0 = vext_s8(va1x0, va1x0, 4);
169         va2x0 = vext_s8(va2x0, va2x0, 4);
170         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
171         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
172         int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
173         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
174         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
175         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
176         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
177         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
178         int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
179         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
180         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
181         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
182         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
183         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
184         int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
185         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
186         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
187         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
188         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
189         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
190         int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
191         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
192         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
193         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
194         int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
195         int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
196         int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
197         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
198         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
199         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
200         int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
201         int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
202         int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
203         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
204         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
205         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
206         int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
207         int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
208         int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
209         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
210         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
211         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
212         int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
213         int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
214         int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
215         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
216         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
217         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
218 
219         k -= 8 * sizeof(int8_t);
220       } while (k != 0);
221 
222       p -= 3 * sizeof(void*);
223     } while (p != 0);
224 
225 #if XNN_ARCH_ARM64
226     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
227     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
228     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
229     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
230     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
231     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
232     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
233     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
234     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
235     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
236     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
237     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
238 #else
239     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
240     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
241     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
242     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
243     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
244     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
245     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
246     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
247     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
248     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
249     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
250     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
251     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
252     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
253     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
254     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
255     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
256     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
257     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
258     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
259     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
260     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
261     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
262     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
263     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
264     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
265     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
266     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
267     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
268     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
269     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
270     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
271     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
272     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
273     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
274     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
275 #endif
276 
277     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
278     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
279     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
280 
281     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
282     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
283     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
284     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
285     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
286     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
287     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
288     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
289     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
290     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
291     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
292     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
293 
294     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
295     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
296     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
297     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
298     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
299     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
300     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
301     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
302     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
303     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
304     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
305     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
306 
307     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
308     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
309     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
310     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
311     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
312     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
313     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
314     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
315     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
316     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
317     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
318     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
319 
320     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
321 #if XNN_ARCH_ARM64
322     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
323     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
324     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
325     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
326     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
327     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
328 
329     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
330     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
331     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
332     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
333     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
334     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
335 
336     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
337     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
338     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
339 #else
340     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
341     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
342     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
343     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
344     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
345     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
346 
347     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
348     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
349     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
350     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
351     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
352     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
353 
354     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
355     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
356     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
357 #endif
358 
359     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
360     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
361     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
362     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
363 
364     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
365     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
366     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
367     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
368 
369     if (nc >= 16) {
370       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
371       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
372       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
373 
374       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
375       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
376       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
377 
378       a = (const int8_t**restrict) ((uintptr_t) a - ks);
379 
380       nc -= 16;
381     } else {
382       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
383       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
384       if (nc & 8) {
385         vst1_s8(c2, vout2x01234567); c2 += 8;
386         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
387         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
388         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
389         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
390       }
391       if (nc & 4) {
392         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
393         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
394         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
395         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
396         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
397       }
398       if (nc & 2) {
399         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
400         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
401         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
402         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
403         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
404       }
405       if (nc & 1) {
406         vst1_lane_s8(c2, vout2x01234567, 0);
407         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
408         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
409       }
410 
411       nc = 0;
412     }
413   } while (nc != 0);
414 }
415