xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x16c4s2-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_4x16c4s2__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_4x16c4s2__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 vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
63     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
64     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
65     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
66     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
67     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
68     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
69     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
70     int32x4_t vacc1x01 = vacc0x01;
71     int32x4_t vacc1x23 = vacc0x23;
72     int32x4_t vacc1x45 = vacc0x45;
73     int32x4_t vacc1x67 = vacc0x67;
74     int32x4_t vacc1x89 = vacc0x89;
75     int32x4_t vacc1xAB = vacc0xAB;
76     int32x4_t vacc1xCD = vacc0xCD;
77     int32x4_t vacc1xEF = vacc0xEF;
78     int32x4_t vacc2x01 = vacc0x01;
79     int32x4_t vacc2x23 = vacc0x23;
80     int32x4_t vacc2x45 = vacc0x45;
81     int32x4_t vacc2x67 = vacc0x67;
82     int32x4_t vacc2x89 = vacc0x89;
83     int32x4_t vacc2xAB = vacc0xAB;
84     int32x4_t vacc2xCD = vacc0xCD;
85     int32x4_t vacc2xEF = vacc0xEF;
86     int32x4_t vacc3x01 = vacc0x01;
87     int32x4_t vacc3x23 = vacc0x23;
88     int32x4_t vacc3x45 = vacc0x45;
89     int32x4_t vacc3x67 = vacc0x67;
90     int32x4_t vacc3x89 = vacc0x89;
91     int32x4_t vacc3xAB = vacc0xAB;
92     int32x4_t vacc3xCD = vacc0xCD;
93     int32x4_t vacc3xEF = vacc0xEF;
94 
95     size_t k = kc;
96     do {
97       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
98       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
99       int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
100       int8x8_t va3x0 = vld1_s8(a3); a3 += 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       int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
123       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
124       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
125       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
126       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
127       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
128       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
129       int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
130       int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
131       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
132       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
133       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
134       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
135       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
136       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
137       int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
138       int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
139       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
140       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
141       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
142       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
143       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
144       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
145       int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
146       int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
147       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
148       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
149       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
150       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
151       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
152       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
153       int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
154       int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3x0);
155       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
156       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
157       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
158       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
159       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
160       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
161       int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
162       int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3x0);
163       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
164       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
165       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
166       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
167       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
168       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
169       int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
170       int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3x0);
171       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
172       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
173       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
174       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
175       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
176       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
177       int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
178       int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0);
179       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
180       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
181       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
182       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
183       va0x0 = vext_s8(va0x0, va0x0, 4);
184       va1x0 = vext_s8(va1x0, va1x0, 4);
185       va2x0 = vext_s8(va2x0, va2x0, 4);
186       va3x0 = vext_s8(va3x0, va3x0, 4);
187       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
188       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
189       int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
190       int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
191       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
192       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
193       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
194       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
195       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
196       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
197       int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
198       int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
199       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
200       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
201       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
202       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
203       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
204       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
205       int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
206       int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
207       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
208       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
209       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
210       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
211       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
212       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
213       int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
214       int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
215       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
216       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
217       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
218       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
219       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
220       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
221       int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
222       int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3x0);
223       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
224       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
225       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
226       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
227       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
228       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
229       int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
230       int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3x0);
231       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
232       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
233       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
234       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
235       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
236       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
237       int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
238       int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0);
239       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
240       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
241       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
242       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
243       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
244       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
245       int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
246       int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3x0);
247       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
248       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
249       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
250       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
251 
252       k -= 8 * sizeof(int8_t);
253     } while (k != 0);
254 
255 #if XNN_ARCH_ARM64
256     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
257     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
258     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
259     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
260     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
261     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
262     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
263     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
264     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
265     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
266     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
267     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
268     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
269     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
270     int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB);
271     int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF);
272 #else
273     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
274     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
275     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
276     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
277     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
278     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
279     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
280     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
281     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
282     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
283     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
284     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
285     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
286     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
287     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
288     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
289     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
290     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
291     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
292     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
293     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
294     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
295     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
296     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
297     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
298     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
299     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
300     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
301     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
302     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
303     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
304     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
305     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
306     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
307     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
308     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
309     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
310     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
311     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
312     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
313     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
314     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
315     const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89));
316     const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB));
317     int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB);
318     const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD));
319     const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF));
320     int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF);
321 #endif
322 
323     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
324     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
325     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
326 
327     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
328     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
329     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
330     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
331     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
332     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
333     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
334     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
335     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
336     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
337     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
338     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
339     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
340     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
341     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
342     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
343 
344     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
345     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
346     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
347     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
348     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
349     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
350     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
351     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
352     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
353     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
354     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
355     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
356     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
357     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
358     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
359     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
360 
361     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
362     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
363     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
364     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
365     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
366     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
367     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
368     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
369     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
370     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
371     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
372     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
373     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
374     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
375     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
376     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
377 
378     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
379 #if XNN_ARCH_ARM64
380     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
381     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
382     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
383     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
384     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
385     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
386     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
387     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
388 
389     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
390     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
391     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
392     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
393     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
394     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
395     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
396     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
397 
398     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
399     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
400     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
401     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
402 #else
403     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
404     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
405     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
406     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
407     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
408     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
409     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
410     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
411 
412     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
413     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
414     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
415     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
416     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
417     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
418     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
419     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
420 
421     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
422     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
423     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
424     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
425 #endif
426 
427     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
428     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
429     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
430     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
431     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
432 
433     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
434     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
435     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
436     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
437     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
438 
439     if (nc >= 16) {
440       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
441       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
442       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
443       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
444 
445       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
446       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
447       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
448       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
449 
450       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
451       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
452       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
453       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
454 
455       nc -= 16;
456     } else {
457       // Final case where not all of the 16 columns fit in the destination.
458       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
459       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
460       if (nc & 8) {
461         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
462         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
463         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
464         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
465         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
466         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
467       }
468       if (nc & 4) {
469         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
470         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
471         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
472         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
473         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
474         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
475       }
476       if (nc & 2) {
477         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
478         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
479         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
480         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
481         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
482         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
483       }
484       if (nc & 1) {
485         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
486         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
487         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
488         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
489       }
490 
491       nc = 0;
492     }
493   } while (nc != 0);
494 }
495