xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mlal-ld1r.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_3x8c4__neon_mlal_ld1r(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_3x8c4__neon_mlal_ld1r(
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 vacc1x01 = vacc0x01;
61     int32x4_t vacc1x23 = vacc0x23;
62     int32x4_t vacc1x45 = vacc0x45;
63     int32x4_t vacc1x67 = vacc0x67;
64     int32x4_t vacc2x01 = vacc0x01;
65     int32x4_t vacc2x23 = vacc0x23;
66     int32x4_t vacc2x45 = vacc0x45;
67     int32x4_t vacc2x67 = vacc0x67;
68 
69     size_t k = kc;
70 
71     while (k >= 16 * sizeof(int8_t)) {
72       const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
73       const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
74       const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
75       const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
76       const int32x2_t va10x0 = vld1_dup_s32((const void*)a1);
77       const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
78       const int32x2_t va10x1 = vld1_dup_s32((const void*)a1);
79       const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
80       const int32x2_t va20x0 = vld1_dup_s32((const void*)a2);
81       const int32x2_t va21x0 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
82       const int32x2_t va20x1 = vld1_dup_s32((const void*)a2);
83       const int32x2_t va21x1 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
84 
85       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 
94       const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
95       const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
96       const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0);
97       const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1);
98       const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0);
99       const int8x8_t va2c0x1 = vreinterpret_s8_s32(va20x1);
100 
101       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
102       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
103       int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
104       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
106       vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
107       vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
108       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
109       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
110       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
111       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
112       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
113       int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
114       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
116       vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
117       vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
118       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
119       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
120       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
121       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
122       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
123       int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
124       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
126       vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
127       vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
128       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
129       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
130       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
131       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
132       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
133       int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
134       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
136       vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
137       vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
138       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
139       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
140       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
141       const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
142       const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
143       const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0);
144       const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1);
145       const int8x8_t va2c1x0 = vreinterpret_s8_s32(va21x0);
146       const int8x8_t va2c1x1 = vreinterpret_s8_s32(va21x1);
147 
148       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
149       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
150       int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
151       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
153       vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
154       vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
155       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
156       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
157       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
158       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
159       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
160       int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
161       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
163       vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
164       vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
165       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
166       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
167       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
168       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
169       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
170       int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
171       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
173       vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
174       vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
175       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
176       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
177       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
178       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
179       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
180       int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
181       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
183       vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
184       vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
185       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
186       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
187       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
188 
189       k -= 16 * sizeof(int8_t);
190     }
191 
192     if (k >= 8 * sizeof(int8_t)) {
193       const int32x2_t va00 = vld1_dup_s32((const void*)a0);
194       const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
195       const int32x2_t va10 = vld1_dup_s32((const void*)a1);
196       const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
197       const int32x2_t va20 = vld1_dup_s32((const void*)a2);
198       const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
199 
200       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
201       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
202       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
203       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
204       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
208 
209       const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
210       const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
211       const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
212 
213       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
214       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
215       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
216       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
217       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
218       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
219       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
220       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
221       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
222       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
223       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
224       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
225       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
226       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
227       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
228       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
229       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
230       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
231       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
232       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
233       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
234       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
235       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
236       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
237       const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
238       const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
239       const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
240 
241       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
242       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
243       const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
244       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
245       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
246       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
247       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
248       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
249       const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
250       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
251       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
252       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
253       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
254       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
255       const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
256       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
257       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
258       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
259       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
260       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
261       const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
262       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
263       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
264       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
265 
266       k -= 8 * sizeof(int8_t);
267     }
268 
269     if XNN_UNLIKELY(k != 0) {
270       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
271       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
272       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
273 
274       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
275       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
276       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
277       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
278 
279       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
280       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
281       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
282       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
283       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
284       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
285       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
286       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
287       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
288       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
289       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
290       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
291       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
292       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
293       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
294       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
295       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
296       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
297       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
298       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
299       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
300       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
301       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
302       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
303       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
304       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
305       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
306     }
307 
308 #if XNN_ARCH_ARM64
309     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
310     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
311     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
312     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
313     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
314     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
315 #else
316     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
317     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
318     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
319     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
320     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
321     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
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 vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
329     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
330     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
331     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
332     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
333     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
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     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
343     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
344     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
345     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
346 
347     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
348     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
349     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
350     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
351     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
352     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
353 
354     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
355     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
356     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
357     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
358     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
359     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
360 
361     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
362 #if XNN_ARCH_ARM64
363     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
364     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
365     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
366 
367     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
368     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
369     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
370 
371     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
372     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
373 #else
374     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
375     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
376     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
377 
378     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
379     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
380     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
381 
382     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
383     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
384 #endif
385 
386     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
387     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
388     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
389 
390     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
391     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
392     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
393 
394     if (nc >= 8) {
395       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
396       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
397       vst1_s8(c2 + 0, vout2x01234567);
398 
399       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
400       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
401       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
402 
403       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
404       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
405       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
406 
407       nc -= 8;
408     } else {
409       // Final case where not all of the 8 columns fit in the destination.
410       if (nc & 4) {
411         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
412         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
413         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
414         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
415         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
416       }
417       if (nc & 2) {
418         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
419         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
420         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
421         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
422         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
423       }
424       if (nc & 1) {
425         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
426         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
427         vst1_lane_s8(c2, vout2x01234567, 0);
428       }
429 
430       nc = 0;
431     }
432   } while (nc != 0);
433 }
434