xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x16c4-minmax-rndnu-neon-mull-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_3x16c4__neon_mull_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_3x16c4__neon_mull_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 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 int32x2_t va00 = vld1_dup_s32((const void*)a0);
86       const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
87       const int32x2_t va10 = vld1_dup_s32((const void*)a1);
88       const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
89       const int32x2_t va20 = vld1_dup_s32((const void*)a2);
90       const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
91 
92       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104       const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105       const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106       const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107       const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108 
109       const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
110       const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
111       const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
112 
113       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
114       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
115       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
116       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
117       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
118       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
119       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
120       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
121       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
122       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
123       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
124       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
125       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
126       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
127       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
128       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
129       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
130       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
131       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
132       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
133       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
134       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
135       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
136       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
137       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
138       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
139       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
140       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
141       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
142       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
143       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
144       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
145       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
146       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
147       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
148       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
149       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
150       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
151       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
152       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
153       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
154       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
155       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
156       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
157       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
158       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
159       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
160       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
161       const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
162       const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
163       const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
164 
165       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
166       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
167       const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
168       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
169       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
170       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
171       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
172       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
173       const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
174       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
175       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
176       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
177       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
178       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
179       const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
180       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
181       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
182       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
183       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
184       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
185       const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
186       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
187       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
188       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
189       const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
190       const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
191       const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1);
192       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
193       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
194       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
195       const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
196       const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
197       const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1);
198       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
199       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
200       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
201       const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
202       const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
203       const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1);
204       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
205       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
206       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
207       const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
208       const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
209       const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1);
210       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
211       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
212       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
213 
214       k -= 8 * sizeof(int8_t);
215     }
216 
217     if XNN_UNLIKELY(k != 0) {
218       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
219       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
220       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
221 
222       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
224       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
226       const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227       const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
228       const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229       const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230 
231       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
232       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
233       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
234       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
235       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
236       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
237       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
238       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
239       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
240       const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
241       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
242       const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
243       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
244       const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
245       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
246       const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
247       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
248       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
249       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
250       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
251       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
252       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
253       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
254       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
255       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
256       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
257       const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
258       vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
259       const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
260       vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
261       const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
262       vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
263       const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
264       vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
265       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
266       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
267       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
268       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
269       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
270       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
271       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
272       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
273       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
274       const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
275       vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
276       const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
277       vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
278       const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
279       vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
280       const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
281       vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
282     }
283 
284 #if XNN_ARCH_ARM64
285     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
286     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
287     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
288     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
289     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
290     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
291     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
292     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
293     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
294     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
295     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
296     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
297 #else
298     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
299     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
300     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
301     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
302     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
303     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
304     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
305     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
306     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
307     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
308     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
309     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
310     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
311     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
312     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
313     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
314     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
315     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
316     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
317     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
318     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
319     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
320     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
321     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
322     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
323     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
324     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
325     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
326     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
327     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
328     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
329     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
330     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
331     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
332     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
333     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
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     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
349     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
350     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
351     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
352 
353     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
354     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
355     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
356     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
357     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
358     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
359     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
360     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
361     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
362     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
363     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
364     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
365 
366     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
367     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
368     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
369     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
370     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
371     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
372     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
373     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
374     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
375     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
376     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
377     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
378 
379     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
380 #if XNN_ARCH_ARM64
381     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
382     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
383     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
384     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
385     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
386     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
387 
388     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
389     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
390     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
391     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
392     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
393     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
394 
395     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
396     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
397     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
398 #else
399     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
400     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
401     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
402     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
403     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
404     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
405 
406     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
407     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
408     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
409     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
410     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
411     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
412 
413     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
414     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
415     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
416 #endif
417 
418     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
419     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
420     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
421     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
422 
423     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
424     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
425     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
426     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
427 
428     if (nc >= 16) {
429       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
430       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
431       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
432 
433       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
434       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
435       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
436 
437       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
438       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
439       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
440 
441       nc -= 16;
442     } else {
443       // Final case where not all of the 16 columns fit in the destination.
444       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
445       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
446       if (nc & 8) {
447         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
448         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
449         vst1_s8(c2, vout2x01234567); c2 += 8;
450         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
451         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
452       }
453       if (nc & 4) {
454         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
455         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
456         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
457         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
458         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
459       }
460       if (nc & 2) {
461         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
462         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
463         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
464         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
465         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
466       }
467       if (nc & 1) {
468         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
469         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
470         vst1_lane_s8(c2, vout2x01234567, 0);
471       }
472 
473       nc = 0;
474     }
475   } while (nc != 0);
476 }
477