xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x16c2s4-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_4x16c2s4__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_4x16c2s4__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 <= 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     while (k >= 16 * sizeof(int8_t)) {
81       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
82       int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
83       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
84       int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
85       int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
86       int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
87       int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
88       int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
89 
90       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
96       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
97       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
98       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
99       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
100       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
101       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
102       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
103       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
104       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
105       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
106 
107       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
108       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
109       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
110       int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);
111       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
112       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
113       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
114       vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2x1);
115       vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, va3x1);
116       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
117       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
118       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
119       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
120       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
121       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
122       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
123       int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);
124       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
125       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
126       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
127       vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2x1);
128       vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, va3x1);
129       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
130       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
131       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
132       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
133       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
134       int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
135       int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2x0);
136       int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0x0, va3x0);
137       const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
138       vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0x1);
139       vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1x1);
140       vprod2x89ABc0 = vmlal_s8(vprod2x89ABc0, vb89ABc0x1, va2x1);
141       vprod3x89ABc0 = vmlal_s8(vprod3x89ABc0, vb89ABc0x1, va3x1);
142       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
143       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
144       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
145       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
146       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
147       int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
148       int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2x0);
149       int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0x0, va3x0);
150       const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
151       vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0x1);
152       vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1x1);
153       vprod2xCDEFc0 = vmlal_s8(vprod2xCDEFc0, vbCDEFc0x1, va2x1);
154       vprod3xCDEFc0 = vmlal_s8(vprod3xCDEFc0, vbCDEFc0x1, va3x1);
155       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
156       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
157       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
158       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
159       va0x0 = vext_s8(va0x0, va0x0, 2);
160       va0x1 = vext_s8(va0x1, va0x1, 2);
161       va1x0 = vext_s8(va1x0, va1x0, 2);
162       va1x1 = vext_s8(va1x1, va1x1, 2);
163       va2x0 = vext_s8(va2x0, va2x0, 2);
164       va2x1 = vext_s8(va2x1, va2x1, 2);
165       va3x0 = vext_s8(va3x0, va3x0, 2);
166       va3x1 = vext_s8(va3x1, va3x1, 2);
167       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
168       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
169       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
170       int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);
171       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
172       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
173       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
174       vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2x1);
175       vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3x1);
176       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
177       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
178       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
179       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
180       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
181       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
182       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
183       int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);
184       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
185       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
186       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
187       vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2x1);
188       vprod3x4567c1 = vmlal_s8(vprod3x4567c1, vb4567c1x1, va3x1);
189       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
190       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
191       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
192       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
193       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
194       int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
195       int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2x0);
196       int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3x0);
197       const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
198       vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0x1);
199       vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1x1);
200       vprod2x89ABc1 = vmlal_s8(vprod2x89ABc1, vb89ABc1x1, va2x1);
201       vprod3x89ABc1 = vmlal_s8(vprod3x89ABc1, vb89ABc1x1, va3x1);
202       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
203       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
204       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
205       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
206       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
207       int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
208       int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2x0);
209       int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0);
210       const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
211       vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0x1);
212       vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1x1);
213       vprod2xCDEFc1 = vmlal_s8(vprod2xCDEFc1, vbCDEFc1x1, va2x1);
214       vprod3xCDEFc1 = vmlal_s8(vprod3xCDEFc1, vbCDEFc1x1, va3x1);
215       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
216       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
217       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
218       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
219       va0x0 = vext_s8(va0x0, va0x0, 2);
220       va0x1 = vext_s8(va0x1, va0x1, 2);
221       va1x0 = vext_s8(va1x0, va1x0, 2);
222       va1x1 = vext_s8(va1x1, va1x1, 2);
223       va2x0 = vext_s8(va2x0, va2x0, 2);
224       va2x1 = vext_s8(va2x1, va2x1, 2);
225       va3x0 = vext_s8(va3x0, va3x0, 2);
226       va3x1 = vext_s8(va3x1, va3x1, 2);
227       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
228       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
229       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
230       int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);
231       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
232       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
233       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
234       vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2x1);
235       vprod3x0123c2 = vmlal_s8(vprod3x0123c2, vb0123c2x1, va3x1);
236       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
237       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
238       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
239       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
240       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
241       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
242       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
243       int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);
244       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
245       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
246       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
247       vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2x1);
248       vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3x1);
249       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
250       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
251       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
252       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
253       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
254       int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
255       int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2x0);
256       int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2x0, va3x0);
257       const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
258       vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0x1);
259       vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1x1);
260       vprod2x89ABc2 = vmlal_s8(vprod2x89ABc2, vb89ABc2x1, va2x1);
261       vprod3x89ABc2 = vmlal_s8(vprod3x89ABc2, vb89ABc2x1, va3x1);
262       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
263       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
264       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
265       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
266       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
267       int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
268       int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2x0);
269       int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2x0, va3x0);
270       const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
271       vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0x1);
272       vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1x1);
273       vprod2xCDEFc2 = vmlal_s8(vprod2xCDEFc2, vbCDEFc2x1, va2x1);
274       vprod3xCDEFc2 = vmlal_s8(vprod3xCDEFc2, vbCDEFc2x1, va3x1);
275       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
276       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
277       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
278       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
279       va0x0 = vext_s8(va0x0, va0x0, 2);
280       va0x1 = vext_s8(va0x1, va0x1, 2);
281       va1x0 = vext_s8(va1x0, va1x0, 2);
282       va1x1 = vext_s8(va1x1, va1x1, 2);
283       va2x0 = vext_s8(va2x0, va2x0, 2);
284       va2x1 = vext_s8(va2x1, va2x1, 2);
285       va3x0 = vext_s8(va3x0, va3x0, 2);
286       va3x1 = vext_s8(va3x1, va3x1, 2);
287       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
288       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
289       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
290       int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);
291       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
292       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
293       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
294       vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2x1);
295       vprod3x0123c3 = vmlal_s8(vprod3x0123c3, vb0123c3x1, va3x1);
296       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
297       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
298       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
299       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
300       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
301       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
302       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
303       int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);
304       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
305       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
306       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
307       vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2x1);
308       vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3x1);
309       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
310       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
311       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
312       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
313       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
314       int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
315       int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0);
316       int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3x0, va3x0);
317       const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
318       vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0x1);
319       vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1x1);
320       vprod2x89ABc3 = vmlal_s8(vprod2x89ABc3, vb89ABc3x1, va2x1);
321       vprod3x89ABc3 = vmlal_s8(vprod3x89ABc3, vb89ABc3x1, va3x1);
322       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
323       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
324       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
325       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
326       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
327       int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
328       int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2x0);
329       int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3x0, va3x0);
330       const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
331       vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0x1);
332       vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1x1);
333       vprod2xCDEFc3 = vmlal_s8(vprod2xCDEFc3, vbCDEFc3x1, va2x1);
334       vprod3xCDEFc3 = vmlal_s8(vprod3xCDEFc3, vbCDEFc3x1, va3x1);
335       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
336       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
337       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
338       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
339 
340       k -= 16 * sizeof(int8_t);
341     }
342     if (k != 0) {
343       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
344       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
345       int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
346       int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
347 
348       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
349       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
350       const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
351       const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
352       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
353       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
354       const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
355       const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
356       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
357       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
358       const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
359       const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
360       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
361       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
362       const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
363       const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
364 
365       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
366       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
367       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
368       int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);
369       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
370       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
371       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
372       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
373       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
374       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
375       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
376       int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);
377       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
378       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
379       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
380       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
381       int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
382       int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
383       int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2x0);
384       int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0x0, va3x0);
385       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
386       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
387       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
388       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
389       int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
390       int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
391       int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2x0);
392       int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0x0, va3x0);
393       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
394       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
395       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
396       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
397       va0x0 = vext_s8(va0x0, va0x0, 2);
398       va1x0 = vext_s8(va1x0, va1x0, 2);
399       va2x0 = vext_s8(va2x0, va2x0, 2);
400       va3x0 = vext_s8(va3x0, va3x0, 2);
401       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
402       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
403       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
404       int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);
405       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
406       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
407       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
408       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
409       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
410       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
411       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
412       int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);
413       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
414       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
415       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
416       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
417       int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
418       int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
419       int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2x0);
420       int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3x0);
421       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
422       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
423       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
424       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
425       int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
426       int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
427       int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2x0);
428       int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0);
429       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
430       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
431       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
432       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
433       va0x0 = vext_s8(va0x0, va0x0, 2);
434       va1x0 = vext_s8(va1x0, va1x0, 2);
435       va2x0 = vext_s8(va2x0, va2x0, 2);
436       va3x0 = vext_s8(va3x0, va3x0, 2);
437       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
438       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
439       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
440       int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);
441       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
442       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
443       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
444       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
445       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
446       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
447       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
448       int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);
449       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
450       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
451       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
452       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
453       int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
454       int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
455       int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2x0);
456       int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2x0, va3x0);
457       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
458       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
459       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
460       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
461       int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
462       int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
463       int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2x0);
464       int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2x0, va3x0);
465       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
466       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
467       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
468       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
469       va0x0 = vext_s8(va0x0, va0x0, 2);
470       va1x0 = vext_s8(va1x0, va1x0, 2);
471       va2x0 = vext_s8(va2x0, va2x0, 2);
472       va3x0 = vext_s8(va3x0, va3x0, 2);
473       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
474       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
475       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
476       int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);
477       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
478       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
479       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
480       vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
481       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
482       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
483       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
484       int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);
485       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
486       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
487       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
488       vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
489       int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
490       int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
491       int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0);
492       int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3x0, va3x0);
493       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
494       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
495       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
496       vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
497       int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
498       int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
499       int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2x0);
500       int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3x0, va3x0);
501       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
502       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
503       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
504       vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
505 
506     }
507 
508     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
509     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
510     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
511 
512     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
513     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
514     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
515     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
516     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
517     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
518     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
519     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
520     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
521     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
522     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
523     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
524     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
525     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
526     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
527     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
528 
529     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
530     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
531     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
532     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
533     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
534     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
535     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
536     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
537     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
538     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
539     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
540     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
541     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
542     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
543     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
544     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
545 
546     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
547     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
548     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
549     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
550     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
551     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
552     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
553     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
554     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
555     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
556     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
557     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
558     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
559     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
560     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
561     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
562 
563     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
564 #if XNN_ARCH_ARM64
565     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
566     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
567     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
568     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
569     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
570     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
571     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
572     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
573 
574     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
575     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
576     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
577     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
578     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
579     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
580     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
581     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
582 
583     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
584     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
585     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
586     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
587 #else
588     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
589     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
590     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
591     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
592     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
593     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
594     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
595     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
596 
597     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
598     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
599     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
600     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
601     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
602     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
603     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
604     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
605 
606     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
607     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
608     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
609     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
610 #endif
611 
612     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
613     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
614     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
615     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
616     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
617 
618     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
619     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
620     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
621     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
622     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
623 
624     if (nc >= 16) {
625       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
626       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
627       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
628       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
629 
630       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
631       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
632       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
633       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
634 
635       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
636       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
637       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
638       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
639 
640       nc -= 16;
641     } else {
642       // Final case where not all of the 16 columns fit in the destination.
643       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
644       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
645       if (nc & 8) {
646         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
647         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
648         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
649         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
650         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
651         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
652       }
653       if (nc & 4) {
654         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
655         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
656         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
657         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
658         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
659         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
660       }
661       if (nc & 2) {
662         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
663         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
664         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
665         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
666         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
667         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
668       }
669       if (nc & 1) {
670         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
671         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
672         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
673         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
674       }
675 
676       nc = 0;
677     }
678   } while (nc != 0);
679 }
680