xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x16c4s2-minmax-rndnu-neon-mull.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_3x16c4s2__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_3x16c4s2__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 <= 3);
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 
54   kc = round_up_po2(kc, 8 * sizeof(int8_t));
55   do {
56     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
59     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
60     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
61     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
62     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
63     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
64     int32x4_t vacc1x01 = vacc0x01;
65     int32x4_t vacc1x23 = vacc0x23;
66     int32x4_t vacc1x45 = vacc0x45;
67     int32x4_t vacc1x67 = vacc0x67;
68     int32x4_t vacc1x89 = vacc0x89;
69     int32x4_t vacc1xAB = vacc0xAB;
70     int32x4_t vacc1xCD = vacc0xCD;
71     int32x4_t vacc1xEF = vacc0xEF;
72     int32x4_t vacc2x01 = vacc0x01;
73     int32x4_t vacc2x23 = vacc0x23;
74     int32x4_t vacc2x45 = vacc0x45;
75     int32x4_t vacc2x67 = vacc0x67;
76     int32x4_t vacc2x89 = vacc0x89;
77     int32x4_t vacc2xAB = vacc0xAB;
78     int32x4_t vacc2xCD = vacc0xCD;
79     int32x4_t vacc2xEF = vacc0xEF;
80 
81     size_t k = kc;
82     do {
83       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
84       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
85       int8x8_t va2x0 = vld1_s8(a2); a2 += 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       int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
107       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
108       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
109       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
110       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
111       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
112       int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
113       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
114       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
115       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
116       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
117       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
118       int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
119       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
120       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
121       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
122       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
123       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
124       int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
125       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
126       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
127       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
128       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
129       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
130       int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
131       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
132       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
133       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
134       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
135       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
136       int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
137       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
138       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
139       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
140       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
141       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
142       int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
143       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
144       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
145       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
146       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
147       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
148       int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
149       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
150       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
151       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
152       va0x0 = vext_s8(va0x0, va0x0, 4);
153       va1x0 = vext_s8(va1x0, va1x0, 4);
154       va2x0 = vext_s8(va2x0, va2x0, 4);
155       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
156       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
157       int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
158       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
159       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
160       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
161       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
162       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
163       int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
164       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
165       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
166       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
167       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
168       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
169       int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
170       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
171       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
172       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
173       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
174       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
175       int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
176       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
177       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
178       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
179       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
180       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
181       int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
182       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
183       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
184       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
185       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
186       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
187       int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
188       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
189       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
190       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
191       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
192       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
193       int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
194       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
195       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
196       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
197       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
198       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
199       int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
200       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
201       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
202       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
203 
204       k -= 8 * sizeof(int8_t);
205     } while (k != 0);
206 
207 #if XNN_ARCH_ARM64
208     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
209     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
210     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
211     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
212     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
213     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
214     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
215     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
216     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
217     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
218     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
219     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
220 #else
221     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
222     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
223     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
224     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
225     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
226     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
227     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
228     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
229     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
230     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
231     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
232     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
233     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
234     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
235     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
236     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
237     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
238     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
239     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
240     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
241     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
242     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
243     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
244     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
245     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
246     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
247     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
248     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
249     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
250     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
251     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
252     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
253     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
254     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
255     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
256     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
257 #endif
258 
259     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
260     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
261     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
262 
263     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
264     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
265     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
266     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
267     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
268     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
269     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
270     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
271     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
272     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
273     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
274     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
275 
276     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
277     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
278     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
279     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
280     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
281     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
282     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
283     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
284     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
285     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
286     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
287     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
288 
289     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
290     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
291     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
292     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
293     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
294     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
295     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
296     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
297     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
298     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
299     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
300     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, 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 
311     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
312     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
313     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
314     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
315     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
316     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
317 
318     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
319     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
320     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
321 #else
322     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
323     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
324     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
325     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
326     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
327     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(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 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
337     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
338     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
339 #endif
340 
341     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
342     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
343     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
344     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
345 
346     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
347     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
348     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
349     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
350 
351     if (nc >= 16) {
352       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
353       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
354       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
355 
356       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
357       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
358       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
359 
360       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
361       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
362       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
363 
364       nc -= 16;
365     } else {
366       // Final case where not all of the 16 columns fit in the destination.
367       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
368       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
369       if (nc & 8) {
370         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
371         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
372         vst1_s8(c2, vout2x01234567); c2 += 8;
373         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
374         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
375       }
376       if (nc & 4) {
377         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
378         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
379         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
380         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
381         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
382       }
383       if (nc & 2) {
384         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
385         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
386         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
387         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
388         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
389       }
390       if (nc & 1) {
391         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
392         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
393         vst1_lane_s8(c2, vout2x01234567, 0);
394       }
395 
396       nc = 0;
397     }
398   } while (nc != 0);
399 }
400