• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_2x16c4s2__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_2x16c4s2__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 <= 2);
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 
48   kc = round_up_po2(kc, 8 * sizeof(int8_t));
49   do {
50     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
51     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
52     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
53     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
54     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
55     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
56     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58     int32x4_t vacc1x01 = vacc0x01;
59     int32x4_t vacc1x23 = vacc0x23;
60     int32x4_t vacc1x45 = vacc0x45;
61     int32x4_t vacc1x67 = vacc0x67;
62     int32x4_t vacc1x89 = vacc0x89;
63     int32x4_t vacc1xAB = vacc0xAB;
64     int32x4_t vacc1xCD = vacc0xCD;
65     int32x4_t vacc1xEF = vacc0xEF;
66 
67     size_t k = kc;
68     while (k >= 16 * sizeof(int8_t)) {
69       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
70       int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
71       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
72       int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
73 
74       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
75       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90 
91       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
92       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
93       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
94       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
95       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
96       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
97       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
98       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
99       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
100       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
101       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
102       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
103       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
104       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
105       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
106       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
107       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
108       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
109       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
110       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
111       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
112       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
113       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
114       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
115       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
116       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
117       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
118       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
119       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
120       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
121       const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
122       vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1);
123       vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1);
124       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
125       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
126       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
127       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
128       const int8x8_t vbABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
129       vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0x1);
130       vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1x1);
131       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
132       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
133       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
134       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
135       const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
136       vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0x1);
137       vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1x1);
138       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
139       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
140       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
141       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
142       const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
143       vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0x1);
144       vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1x1);
145       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
146       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
147       va0x0 = vext_s8(va0x0, va0x0, 4);
148       va0x1 = vext_s8(va0x1, va0x1, 4);
149       va1x0 = vext_s8(va1x0, va1x0, 4);
150       va1x1 = vext_s8(va1x1, va1x1, 4);
151       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
152       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
153       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
154       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
155       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
156       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
157       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
158       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
159       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
160       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
161       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
162       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
163       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
164       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
165       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
166       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
167       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
168       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
169       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
170       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
171       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
172       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
173       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
174       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
175       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
176       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
177       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
178       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
179       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
180       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
181       const int8x8_t vb89c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
182       vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0x1);
183       vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1);
184       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
185       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
186       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
187       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
188       const int8x8_t vbABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
189       vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0x1);
190       vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1);
191       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
192       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
193       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
194       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
195       const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
196       vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0x1);
197       vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1x1);
198       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
199       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
200       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
201       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
202       const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
203       vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0x1);
204       vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1x1);
205       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
206       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
207 
208       k -= 16 * sizeof(int8_t);
209     }
210     if (k != 0) {
211       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
212       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
213 
214       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
215       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
216       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
217       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
218       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
219       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
220       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
221       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
222       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
223       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
224       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
225       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
226       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
227       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
228       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
229       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
230 
231       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
232       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
233       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
234       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
235       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
236       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
237       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
238       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
239       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
240       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
241       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
242       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
243       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
244       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
245       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
246       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
247       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
248       int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
249       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
250       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
251       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
252       int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
253       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
254       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
255       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
256       int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
257       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
258       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
259       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
260       int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
261       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
262       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
263       va0x0 = vext_s8(va0x0, va0x0, 4);
264       va1x0 = vext_s8(va1x0, va1x0, 4);
265       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
266       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
267       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
268       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
269       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
270       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
271       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
272       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
273       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
274       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
275       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
276       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
277       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
278       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
279       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
280       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
281       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
282       int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
283       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
284       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
285       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
286       int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
287       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
288       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
289       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
290       int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
291       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
292       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
293       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
294       int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
295       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
296       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
297 
298     }
299 
300 #if XNN_ARCH_ARM64
301     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
302     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
303     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
304     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
305     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
306     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
307     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
308     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
309 #else
310     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
311     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
312     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
313     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
314     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
315     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
316     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
317     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
318     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
319     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
320     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
321     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
322     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
323     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
324     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
325     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
326     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
327     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
328     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
329     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
330     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
331     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
332     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
333     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
334 #endif
335 
336     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
337     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
338     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
339 
340     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
341     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
342     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
343     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
344     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
345     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
346     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
347     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
348 
349     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
350     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
351     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
352     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
353     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
354     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
355     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
356     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
357 
358     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
359     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
360     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
361     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
362     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
363     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
364     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
365     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
366 
367     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
368 #if XNN_ARCH_ARM64
369     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
370     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
371     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
372     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
373 
374     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
375     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
376     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
377     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
378 
379     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
380     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
381 #else
382     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
383     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
384     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
385     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
386 
387     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
388     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
389     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
390     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
391 
392     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
393     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
394 #endif
395 
396     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
397     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
398     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
399 
400     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
401     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
402     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
403 
404     if (nc >= 16) {
405       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
406       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
407 
408       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
409       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
410 
411       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
412       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
413 
414       nc -= 16;
415     } else {
416       // Final case where not all of the 16 columns fit in the destination.
417       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
418       if (nc & 8) {
419         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
420         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
421         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
422       }
423       if (nc & 4) {
424         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
425         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
426         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
427       }
428       if (nc & 2) {
429         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
430         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
431         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
432       }
433       if (nc & 1) {
434         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
435         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
436       }
437 
438       nc = 0;
439     }
440   } while (nc != 0);
441 }
442