xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x16c4s2-minmax-rndnu-neon-mlal.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_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_4x16c4s2__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 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     while (k >= 16 * sizeof(int8_t)) {
97       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
98       int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
99       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
100       int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
101       int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
102       int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
103       int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
104       int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
105 
106       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
107       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
112       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
113       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
114       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
115       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
116       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
117       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
118       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
119       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
120       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
121       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
122 
123       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
124       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
125       int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
126       int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
127       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
128       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
129       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
130       vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2x1);
131       vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3x1);
132       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
133       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
134       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
135       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
136       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
137       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
138       int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
139       int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
140       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
141       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
142       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
143       vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1);
144       vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3x1);
145       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
146       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
147       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
148       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
149       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
150       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
151       int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
152       int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
153       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
154       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
155       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
156       vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2x1);
157       vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3x1);
158       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
159       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
160       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
161       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
162       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
163       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
164       int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
165       int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
166       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
167       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
168       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
169       vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1);
170       vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3x1);
171       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
172       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
173       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
174       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
175       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
176       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
177       int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
178       int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3x0);
179       const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
180       vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1);
181       vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1);
182       vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2x1);
183       vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3x1);
184       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
185       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
186       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
187       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
188       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
189       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
190       int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
191       int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3x0);
192       const int8x8_t vbABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
193       vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0x1);
194       vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1x1);
195       vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2x1);
196       vprod3xABc0 = vmlal_s8(vprod3xABc0, vbABc0x1, va3x1);
197       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
198       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
199       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
200       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
201       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
202       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
203       int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
204       int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3x0);
205       const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
206       vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0x1);
207       vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1x1);
208       vprod2xCDc0 = vmlal_s8(vprod2xCDc0, vbCDc0x1, va2x1);
209       vprod3xCDc0 = vmlal_s8(vprod3xCDc0, vbCDc0x1, va3x1);
210       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
211       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
212       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
213       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
214       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
215       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
216       int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
217       int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0);
218       const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
219       vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0x1);
220       vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1x1);
221       vprod2xEFc0 = vmlal_s8(vprod2xEFc0, vbEFc0x1, va2x1);
222       vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3x1);
223       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
224       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
225       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
226       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
227       va0x0 = vext_s8(va0x0, va0x0, 4);
228       va0x1 = vext_s8(va0x1, va0x1, 4);
229       va1x0 = vext_s8(va1x0, va1x0, 4);
230       va1x1 = vext_s8(va1x1, va1x1, 4);
231       va2x0 = vext_s8(va2x0, va2x0, 4);
232       va2x1 = vext_s8(va2x1, va2x1, 4);
233       va3x0 = vext_s8(va3x0, va3x0, 4);
234       va3x1 = vext_s8(va3x1, va3x1, 4);
235       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
236       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
237       int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
238       int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
239       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
240       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
241       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
242       vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2x1);
243       vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3x1);
244       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
245       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
246       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
247       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
248       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
249       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
250       int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
251       int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
252       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
253       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
254       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
255       vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2x1);
256       vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3x1);
257       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
258       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
259       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
260       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
261       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
262       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
263       int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
264       int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
265       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
266       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
267       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
268       vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2x1);
269       vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3x1);
270       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
271       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
272       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
273       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
274       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
275       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
276       int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
277       int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
278       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
279       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
280       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
281       vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2x1);
282       vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3x1);
283       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
284       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
285       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
286       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
287       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
288       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
289       int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
290       int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3x0);
291       const int8x8_t vb89c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
292       vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0x1);
293       vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1);
294       vprod2x89c1 = vmlal_s8(vprod2x89c1, vb89c1x1, va2x1);
295       vprod3x89c1 = vmlal_s8(vprod3x89c1, vb89c1x1, va3x1);
296       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
297       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
298       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
299       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
300       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
301       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
302       int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
303       int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3x0);
304       const int8x8_t vbABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
305       vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0x1);
306       vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1);
307       vprod2xABc1 = vmlal_s8(vprod2xABc1, vbABc1x1, va2x1);
308       vprod3xABc1 = vmlal_s8(vprod3xABc1, vbABc1x1, va3x1);
309       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
310       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
311       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
312       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
313       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
314       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
315       int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
316       int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0);
317       const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
318       vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0x1);
319       vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1x1);
320       vprod2xCDc1 = vmlal_s8(vprod2xCDc1, vbCDc1x1, va2x1);
321       vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3x1);
322       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
323       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
324       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
325       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
326       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
327       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
328       int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
329       int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3x0);
330       const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
331       vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0x1);
332       vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1x1);
333       vprod2xEFc1 = vmlal_s8(vprod2xEFc1, vbEFc1x1, va2x1);
334       vprod3xEFc1 = vmlal_s8(vprod3xEFc1, vbEFc1x1, va3x1);
335       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
336       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
337       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
338       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
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 vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
349       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
350       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
351       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
352       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
353       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
354       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
355       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
356       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
357       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
358       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
359       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
360       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
361       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
362       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
363       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
364 
365       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
366       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
367       int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
368       int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
369       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
370       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
371       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
372       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
373       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
374       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
375       int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
376       int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
377       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
378       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
379       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
380       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
381       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
382       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
383       int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
384       int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
385       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
386       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
387       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
388       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
389       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
390       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
391       int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
392       int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
393       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
394       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
395       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
396       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
397       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
398       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
399       int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
400       int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3x0);
401       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
402       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
403       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
404       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
405       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
406       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
407       int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
408       int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3x0);
409       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
410       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
411       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
412       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
413       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
414       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
415       int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
416       int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3x0);
417       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
418       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
419       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
420       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
421       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
422       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
423       int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
424       int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0);
425       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
426       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
427       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
428       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
429       va0x0 = vext_s8(va0x0, va0x0, 4);
430       va1x0 = vext_s8(va1x0, va1x0, 4);
431       va2x0 = vext_s8(va2x0, va2x0, 4);
432       va3x0 = vext_s8(va3x0, va3x0, 4);
433       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
434       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
435       int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
436       int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
437       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
438       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
439       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
440       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
441       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
442       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
443       int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
444       int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
445       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
446       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
447       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
448       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
449       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
450       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
451       int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
452       int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
453       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
454       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
455       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
456       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
457       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
458       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
459       int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
460       int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
461       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
462       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
463       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
464       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
465       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
466       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
467       int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
468       int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3x0);
469       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
470       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
471       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
472       vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
473       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
474       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
475       int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
476       int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3x0);
477       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
478       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
479       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
480       vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
481       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
482       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
483       int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
484       int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0);
485       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
486       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
487       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
488       vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
489       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
490       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
491       int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
492       int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3x0);
493       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
494       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
495       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
496       vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
497 
498     }
499 
500 #if XNN_ARCH_ARM64
501     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
502     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
503     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
504     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
505     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
506     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
507     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
508     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
509     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
510     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
511     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
512     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
513     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
514     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
515     int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB);
516     int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF);
517 #else
518     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
519     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
520     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
521     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
522     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
523     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
524     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
525     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
526     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
527     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
528     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
529     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
530     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
531     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
532     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
533     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
534     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
535     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
536     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
537     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
538     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
539     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
540     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
541     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
542     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
543     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
544     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
545     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
546     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
547     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
548     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
549     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
550     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
551     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
552     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
553     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
554     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
555     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
556     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
557     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
558     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
559     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
560     const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89));
561     const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB));
562     int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB);
563     const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD));
564     const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF));
565     int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF);
566 #endif
567 
568     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
569     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
570     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
571 
572     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
573     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
574     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
575     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
576     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
577     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
578     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
579     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
580     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
581     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
582     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
583     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
584     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
585     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
586     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
587     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
588 
589     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
590     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
591     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
592     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
593     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
594     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
595     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
596     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
597     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
598     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
599     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
600     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
601     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
602     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
603     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
604     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
605 
606     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
607     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
608     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
609     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
610     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
611     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
612     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
613     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
614     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
615     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
616     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
617     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
618     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
619     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
620     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
621     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
622 
623     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
624 #if XNN_ARCH_ARM64
625     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
626     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
627     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
628     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
629     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
630     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
631     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
632     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
633 
634     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
635     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
636     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
637     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
638     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
639     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
640     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
641     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
642 
643     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
644     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
645     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
646     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
647 #else
648     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
649     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
650     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
651     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
652     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
653     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
654     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
655     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
656 
657     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
658     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
659     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
660     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
661     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
662     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
663     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
664     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
665 
666     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
667     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
668     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
669     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
670 #endif
671 
672     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
673     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
674     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
675     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
676     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
677 
678     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
679     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
680     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
681     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
682     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
683 
684     if (nc >= 16) {
685       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
686       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
687       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
688       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
689 
690       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
691       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
692       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
693       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
694 
695       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
696       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
697       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
698       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
699 
700       nc -= 16;
701     } else {
702       // Final case where not all of the 16 columns fit in the destination.
703       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
704       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
705       if (nc & 8) {
706         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
707         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
708         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
709         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
710         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
711         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
712       }
713       if (nc & 4) {
714         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
715         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
716         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
717         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
718         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
719         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
720       }
721       if (nc & 2) {
722         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
723         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
724         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
725         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
726         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
727         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
728       }
729       if (nc & 1) {
730         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
731         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
732         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
733         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
734       }
735 
736       nc = 0;
737     }
738   } while (nc != 0);
739 }
740