xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c4-neon-mull-dup.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_3x16c4__neon_mull_dup(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_3x16c4__neon_mull_dup(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 3);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 4 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54 
55   do {
56     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
59     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
60     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
61     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
62     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
63     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
64     int32x4_t vacc1x01 = vacc0x01;
65     int32x4_t vacc1x23 = vacc0x23;
66     int32x4_t vacc1x45 = vacc0x45;
67     int32x4_t vacc1x67 = vacc0x67;
68     int32x4_t vacc1x89 = vacc0x89;
69     int32x4_t vacc1xAB = vacc0xAB;
70     int32x4_t vacc1xCD = vacc0xCD;
71     int32x4_t vacc1xEF = vacc0xEF;
72     int32x4_t vacc2x01 = vacc0x01;
73     int32x4_t vacc2x23 = vacc0x23;
74     int32x4_t vacc2x45 = vacc0x45;
75     int32x4_t vacc2x67 = vacc0x67;
76     int32x4_t vacc2x89 = vacc0x89;
77     int32x4_t vacc2xAB = vacc0xAB;
78     int32x4_t vacc2xCD = vacc0xCD;
79     int32x4_t vacc2xEF = vacc0xEF;
80 
81     size_t k = kc;
82 
83 
84     while (k >= 8 * sizeof(int8_t)) {
85       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
86       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
87       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
88 
89       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101       const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102       const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103       const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104       const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 
106       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
107       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
108       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
109 
110       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
111       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
112       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
113       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
114       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
115       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
116       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
117       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
118       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
119       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
120       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
121       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
122       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
123       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
124       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
125       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
126       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
127       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
128       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
129       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
130       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
131       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
132       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
133       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
134       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
135       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
136       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
137       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
138       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
139       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
140       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
141       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
142       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
143       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
144       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
145       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
146       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
147       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
148       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
149       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
150       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
151       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
152       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
153       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
154       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
155       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
156       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
157       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
158       const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
159       const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
160       const int8x8_t va2c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 1));
161 
162       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
163       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
164       const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
165       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
166       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
167       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
168       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
169       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
170       const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
171       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
172       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
173       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
174       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
175       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
176       const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
177       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
178       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
179       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
180       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
181       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
182       const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
183       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
184       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
185       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
186       const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
187       const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
188       const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1);
189       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
190       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
191       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
192       const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
193       const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
194       const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1);
195       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
196       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
197       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
198       const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
199       const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
200       const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1);
201       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
202       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
203       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
204       const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
205       const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
206       const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1);
207       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
208       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
209       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
210 
211       k -= 8 * sizeof(int8_t);
212     }
213 
214     if XNN_UNLIKELY(k != 0) {
215       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
216       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
217       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
218 
219       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
222       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
224       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
226       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227 
228       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
229       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
230       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
231       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
232       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
233       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
234       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
235       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
236       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
237       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
238       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
239       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
240       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
241       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
242       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
243       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
244       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
245       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
246       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
247       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
248       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
249       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
250       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
251       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
252       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
253       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
254       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
255       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
256       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
257       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
258       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
259       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
260       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
261       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
262       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
263       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
264       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
265       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
266       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
267       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
268       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
269       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
270       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
271       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
272       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
273       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
274       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
275       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
276       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
277       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
278       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
279     }
280 
281 #if XNN_ARCH_ARM64
282     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
283     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
284     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
285     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
286     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
287     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
288     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
289     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
290     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
291     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
292     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
293     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
294 #else
295     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
296     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
297     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
298     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
299     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
300     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
301     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
302     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
303     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
304     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
305     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
306     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
307     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
308     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
309     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
310     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
311     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
312     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
313     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
314     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
315     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
316     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
317     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
318     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
319     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
320     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
321     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
322     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
323     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
324     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
325     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
326     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
327     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
328     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
329     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
330     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
331 #endif
332 
333     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
334     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
335     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
336 
337     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
338     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
339     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
340     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
341     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
342     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
343     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
344     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
345     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
346     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
347     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
348     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
349 
350     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
351     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
352     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
353     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
354     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
355     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
356     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
357     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
358     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
359     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
360     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
361     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
362 
363     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
364     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
365     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
366     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
367     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
368     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
369     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
370     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
371     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
372     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
373     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
374     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
375 
376     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
377 #if XNN_ARCH_ARM64
378     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
379     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
380     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
381     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
382     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
383     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
384 
385     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
386     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
387     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
388     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
389     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
390     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
391 
392     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
393     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
394     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
395 #else
396     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
397     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
398     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
399     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
400     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
401     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
402 
403     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
404     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
405     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
406     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
407     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
408     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
409 
410     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
411     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
412     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
413 #endif
414 
415     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
416     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
417     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
418     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
419 
420     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
421     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
422     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
423     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
424 
425     if (nc >= 16) {
426       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
427       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
428       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
429 
430       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
431       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
432       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
433 
434       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
435       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
436       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
437 
438       nc -= 16;
439     } else {
440       // Final case where not all of the 16 columns fit in the destination.
441       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
442       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
443       if (nc & 8) {
444         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
445         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
446         vst1_s8(c2, vout2x01234567); c2 += 8;
447         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
448         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
449       }
450       if (nc & 4) {
451         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
452         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
453         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
454         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
455         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
456       }
457       if (nc & 2) {
458         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
459         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
460         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
461         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
462         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
463       }
464       if (nc & 1) {
465         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
466         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
467         vst1_lane_s8(c2, vout2x01234567, 0);
468       }
469 
470       nc = 0;
471     }
472   } while (nc != 0);
473 }
474