xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x16c2s4-minmax-rndnu-neon-mull.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-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_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mull(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mull(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 4);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr != 4) {
56     a3 = a2;
57     c3 = c2;
58   }
59 
60   kc = round_up_po2(kc, 8 * sizeof(int8_t));
61   do {
62     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
63     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
64     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
65     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
66     int32x4_t vacc1x0123 = vacc0x0123;
67     int32x4_t vacc1x4567 = vacc0x4567;
68     int32x4_t vacc1x89AB = vacc0x89AB;
69     int32x4_t vacc1xCDEF = vacc0xCDEF;
70     int32x4_t vacc2x0123 = vacc0x0123;
71     int32x4_t vacc2x4567 = vacc0x4567;
72     int32x4_t vacc2x89AB = vacc0x89AB;
73     int32x4_t vacc2xCDEF = vacc0xCDEF;
74     int32x4_t vacc3x0123 = vacc0x0123;
75     int32x4_t vacc3x4567 = vacc0x4567;
76     int32x4_t vacc3x89AB = vacc0x89AB;
77     int32x4_t vacc3xCDEF = vacc0xCDEF;
78 
79     size_t k = kc;
80     do {
81       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
82       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
83       int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
84       int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
85 
86       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
96       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
97       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
98       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
99       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
100       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
101       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
102 
103       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
104       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
105       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
106       int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);
107       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
108       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
109       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
110       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
111       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
112       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
113       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
114       int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);
115       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
116       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
117       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
118       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
119       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
120       int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
121       int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2x0);
122       int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0x0, va3x0);
123       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
124       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
125       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
126       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
127       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
128       int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
129       int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2x0);
130       int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0x0, va3x0);
131       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
132       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
133       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
134       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
135       va0x0 = vext_s8(va0x0, va0x0, 2);
136       va1x0 = vext_s8(va1x0, va1x0, 2);
137       va2x0 = vext_s8(va2x0, va2x0, 2);
138       va3x0 = vext_s8(va3x0, va3x0, 2);
139       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
140       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
141       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
142       int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);
143       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
144       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
145       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
146       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
147       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
148       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
149       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
150       int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);
151       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
152       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
153       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
154       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
155       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
156       int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
157       int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2x0);
158       int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3x0);
159       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
160       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
161       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
162       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
163       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
164       int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
165       int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2x0);
166       int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0);
167       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
168       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
169       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
170       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
171       va0x0 = vext_s8(va0x0, va0x0, 2);
172       va1x0 = vext_s8(va1x0, va1x0, 2);
173       va2x0 = vext_s8(va2x0, va2x0, 2);
174       va3x0 = vext_s8(va3x0, va3x0, 2);
175       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
176       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
177       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
178       int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);
179       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
180       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
181       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
182       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
183       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
184       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
185       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
186       int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);
187       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
188       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
189       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
190       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
191       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
192       int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
193       int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2x0);
194       int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2x0, va3x0);
195       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
196       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
197       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
198       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
199       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
200       int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
201       int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2x0);
202       int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2x0, va3x0);
203       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
204       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
205       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
206       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
207       va0x0 = vext_s8(va0x0, va0x0, 2);
208       va1x0 = vext_s8(va1x0, va1x0, 2);
209       va2x0 = vext_s8(va2x0, va2x0, 2);
210       va3x0 = vext_s8(va3x0, va3x0, 2);
211       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
212       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
213       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
214       int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);
215       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
216       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
217       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
218       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
219       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
220       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
221       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
222       int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);
223       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
224       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
225       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
226       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
227       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
228       int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
229       int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0);
230       int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3x0, va3x0);
231       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
232       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
233       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
234       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
235       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
236       int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
237       int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2x0);
238       int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3x0, va3x0);
239       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
240       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
241       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
242       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
243 
244       k -= 8 * sizeof(int8_t);
245     } while (k != 0);
246 
247     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
248     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
249     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
250 
251     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
252     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
253     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
254     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
255     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
256     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
257     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
258     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
259     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
260     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
261     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
262     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
263     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
264     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
265     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
266     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
267 
268     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
269     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
270     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
271     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
272     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
273     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
274     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
275     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
276     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
277     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
278     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
279     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
280     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
281     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
282     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
283     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
284 
285     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
286     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
287     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
288     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
289     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
290     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
291     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
292     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
293     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
294     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
295     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
296     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
297     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
298     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
299     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
300     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
301 
302     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
303 #if XNN_ARCH_ARM64
304     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
305     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
306     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
307     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
308     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
309     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
310     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
311     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
312 
313     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
314     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
315     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
316     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
317     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
318     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
319     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
320     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
321 
322     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
323     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
324     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
325     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
326 #else
327     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
328     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
329     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
330     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
331     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
332     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
333     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
334     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
335 
336     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
337     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
338     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
339     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
340     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
341     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
342     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
343     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
344 
345     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
346     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
347     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
348     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
349 #endif
350 
351     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
352     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
353     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
354     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
355     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
356 
357     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
358     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
359     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
360     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
361     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
362 
363     if (nc >= 16) {
364       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
365       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
366       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
367       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
368 
369       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
370       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
371       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
372       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
373 
374       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
375       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
376       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
377       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
378 
379       nc -= 16;
380     } else {
381       // Final case where not all of the 16 columns fit in the destination.
382       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
383       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
384       if (nc & 8) {
385         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
386         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
387         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
388         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
389         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
390         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
391       }
392       if (nc & 4) {
393         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
394         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
395         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
396         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
397         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
398         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
399       }
400       if (nc & 2) {
401         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
402         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
403         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
404         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
405         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
406         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
407       }
408       if (nc & 1) {
409         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
410         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
411         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
412         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
413       }
414 
415       nc = 0;
416     }
417   } while (nc != 0);
418 }
419