xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x16c2s4-minmax-rndnu-neon-mlal.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_3x16c2s4__neon_mlal(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_3x16c2s4__neon_mlal(
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 vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
57     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
58     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
59     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
60     int32x4_t vacc1x0123 = vacc0x0123;
61     int32x4_t vacc1x4567 = vacc0x4567;
62     int32x4_t vacc1x89AB = vacc0x89AB;
63     int32x4_t vacc1xCDEF = vacc0xCDEF;
64     int32x4_t vacc2x0123 = vacc0x0123;
65     int32x4_t vacc2x4567 = vacc0x4567;
66     int32x4_t vacc2x89AB = vacc0x89AB;
67     int32x4_t vacc2xCDEF = vacc0xCDEF;
68 
69     size_t k = kc;
70     while (k >= 16 * sizeof(int8_t)) {
71       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
72       int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
73       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
74       int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
75       int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
76       int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
77 
78       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94 
95       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
96       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
97       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
98       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
99       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
100       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
101       vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2x1);
102       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
103       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
104       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
105       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
106       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
107       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
108       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
109       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
110       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
111       vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2x1);
112       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
113       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
114       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
115       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
116       int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
117       int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2x0);
118       const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
119       vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0x1);
120       vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1x1);
121       vprod2x89ABc0 = vmlal_s8(vprod2x89ABc0, vb89ABc0x1, va2x1);
122       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
123       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
124       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
125       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
126       int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
127       int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2x0);
128       const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
129       vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0x1);
130       vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1x1);
131       vprod2xCDEFc0 = vmlal_s8(vprod2xCDEFc0, vbCDEFc0x1, va2x1);
132       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
133       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
134       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
135       va0x0 = vext_s8(va0x0, va0x0, 2);
136       va0x1 = vext_s8(va0x1, va0x1, 2);
137       va1x0 = vext_s8(va1x0, va1x0, 2);
138       va1x1 = vext_s8(va1x1, va1x1, 2);
139       va2x0 = vext_s8(va2x0, va2x0, 2);
140       va2x1 = vext_s8(va2x1, va2x1, 2);
141       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
142       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
143       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
144       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
145       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
146       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
147       vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2x1);
148       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
149       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
150       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
151       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
152       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
153       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
154       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
155       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
156       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
157       vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2x1);
158       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
159       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
160       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
161       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
162       int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
163       int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2x0);
164       const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
165       vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0x1);
166       vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1x1);
167       vprod2x89ABc1 = vmlal_s8(vprod2x89ABc1, vb89ABc1x1, va2x1);
168       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
169       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
170       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
171       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
172       int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
173       int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2x0);
174       const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
175       vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0x1);
176       vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1x1);
177       vprod2xCDEFc1 = vmlal_s8(vprod2xCDEFc1, vbCDEFc1x1, va2x1);
178       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
179       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
180       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
181       va0x0 = vext_s8(va0x0, va0x0, 2);
182       va0x1 = vext_s8(va0x1, va0x1, 2);
183       va1x0 = vext_s8(va1x0, va1x0, 2);
184       va1x1 = vext_s8(va1x1, va1x1, 2);
185       va2x0 = vext_s8(va2x0, va2x0, 2);
186       va2x1 = vext_s8(va2x1, va2x1, 2);
187       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
188       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
189       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
190       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
191       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
192       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
193       vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2x1);
194       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
195       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
196       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
197       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
198       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
199       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
200       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
201       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
202       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
203       vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2x1);
204       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
205       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
206       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
207       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
208       int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
209       int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2x0);
210       const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
211       vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0x1);
212       vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1x1);
213       vprod2x89ABc2 = vmlal_s8(vprod2x89ABc2, vb89ABc2x1, va2x1);
214       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
215       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
216       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
217       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
218       int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
219       int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2x0);
220       const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
221       vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0x1);
222       vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1x1);
223       vprod2xCDEFc2 = vmlal_s8(vprod2xCDEFc2, vbCDEFc2x1, va2x1);
224       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
225       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
226       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
227       va0x0 = vext_s8(va0x0, va0x0, 2);
228       va0x1 = vext_s8(va0x1, va0x1, 2);
229       va1x0 = vext_s8(va1x0, va1x0, 2);
230       va1x1 = vext_s8(va1x1, va1x1, 2);
231       va2x0 = vext_s8(va2x0, va2x0, 2);
232       va2x1 = vext_s8(va2x1, va2x1, 2);
233       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
234       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
235       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
236       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
237       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
238       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
239       vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2x1);
240       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
241       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
242       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
243       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
244       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
245       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
246       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
247       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
248       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
249       vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2x1);
250       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
251       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
252       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
253       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
254       int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
255       int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0);
256       const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
257       vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0x1);
258       vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1x1);
259       vprod2x89ABc3 = vmlal_s8(vprod2x89ABc3, vb89ABc3x1, va2x1);
260       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
261       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
262       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
263       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
264       int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
265       int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2x0);
266       const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
267       vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0x1);
268       vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1x1);
269       vprod2xCDEFc3 = vmlal_s8(vprod2xCDEFc3, vbCDEFc3x1, va2x1);
270       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
271       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
272       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
273 
274       k -= 16 * sizeof(int8_t);
275     }
276     if (k != 0) {
277       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
278       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
279       int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
280 
281       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
282       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
283       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
284       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
285       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
286       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
287       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
288       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
289       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
290       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
291       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
292       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
293       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
294       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
295       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
296       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
297 
298       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
299       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
300       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
301       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
302       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
303       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
304       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
305       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
306       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
307       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
308       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
309       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
310       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
311       int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
312       int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2x0);
313       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
314       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
315       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
316       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
317       int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
318       int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2x0);
319       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
320       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
321       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
322       va0x0 = vext_s8(va0x0, va0x0, 2);
323       va1x0 = vext_s8(va1x0, va1x0, 2);
324       va2x0 = vext_s8(va2x0, va2x0, 2);
325       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
326       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
327       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
328       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
329       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
330       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
331       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
332       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
333       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
334       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
335       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
336       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
337       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
338       int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
339       int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2x0);
340       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
341       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
342       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
343       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
344       int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
345       int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2x0);
346       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
347       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
348       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
349       va0x0 = vext_s8(va0x0, va0x0, 2);
350       va1x0 = vext_s8(va1x0, va1x0, 2);
351       va2x0 = vext_s8(va2x0, va2x0, 2);
352       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
353       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
354       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
355       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
356       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
357       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
358       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
359       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
360       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
361       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
362       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
363       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
364       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
365       int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
366       int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2x0);
367       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
368       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
369       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
370       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
371       int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
372       int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2x0);
373       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
374       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
375       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
376       va0x0 = vext_s8(va0x0, va0x0, 2);
377       va1x0 = vext_s8(va1x0, va1x0, 2);
378       va2x0 = vext_s8(va2x0, va2x0, 2);
379       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
380       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
381       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
382       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
383       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
384       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
385       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
386       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
387       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
388       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
389       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
390       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
391       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
392       int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
393       int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0);
394       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
395       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
396       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
397       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
398       int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
399       int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2x0);
400       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
401       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
402       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
403 
404     }
405 
406     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
407     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
408     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
409 
410     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
411     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
412     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
413     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
414     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
415     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
416     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
417     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
418     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
419     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
420     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
421     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
422 
423     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
424     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
425     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
426     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
427     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
428     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
429     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
430     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
431     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
432     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
433     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
434     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
435 
436     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
437     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
438     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
439     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
440     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
441     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
442     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
443     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
444     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
445     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
446     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
447     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
448 
449     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
450 #if XNN_ARCH_ARM64
451     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
452     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
453     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
454     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
455     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
456     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
457 
458     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
459     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
460     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
461     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
462     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
463     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
464 
465     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
466     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
467     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
468 #else
469     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
470     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
471     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
472     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
473     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
474     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
475 
476     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
477     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
478     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
479     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
480     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
481     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
482 
483     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
484     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
485     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
486 #endif
487 
488     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
489     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
490     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
491     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
492 
493     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
494     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
495     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
496     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
497 
498     if (nc >= 16) {
499       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
500       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
501       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
502 
503       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
504       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
505       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
506 
507       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
508       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
509       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
510 
511       nc -= 16;
512     } else {
513       // Final case where not all of the 16 columns fit in the destination.
514       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
515       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
516       if (nc & 8) {
517         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
518         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
519         vst1_s8(c2, vout2x01234567); c2 += 8;
520         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
521         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
522       }
523       if (nc & 4) {
524         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
525         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
526         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
527         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
528         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
529       }
530       if (nc & 2) {
531         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
532         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
533         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
534         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
535         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
536       }
537       if (nc & 1) {
538         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
539         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
540         vst1_lane_s8(c2, vout2x01234567, 0);
541       }
542 
543       nc = 0;
544     }
545   } while (nc != 0);
546 }
547