xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/3x16c4-minmax-rndnu-neon-mull-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t** restrict a,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 3);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (3 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(int8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   kc = round_up_po2(kc, 4 * sizeof(int8_t));
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr < 2) {
47     c1 = c0;
48   }
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     c2 = c1;
52   }
53 
54   do {
55     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
56     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
59     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
60     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
61     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
62     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
63     int32x4_t vacc1x01 = vacc0x01;
64     int32x4_t vacc1x23 = vacc0x23;
65     int32x4_t vacc1x45 = vacc0x45;
66     int32x4_t vacc1x67 = vacc0x67;
67     int32x4_t vacc1x89 = vacc0x89;
68     int32x4_t vacc1xAB = vacc0xAB;
69     int32x4_t vacc1xCD = vacc0xCD;
70     int32x4_t vacc1xEF = vacc0xEF;
71     int32x4_t vacc2x01 = vacc0x01;
72     int32x4_t vacc2x23 = vacc0x23;
73     int32x4_t vacc2x45 = vacc0x45;
74     int32x4_t vacc2x67 = vacc0x67;
75     int32x4_t vacc2x89 = vacc0x89;
76     int32x4_t vacc2xAB = vacc0xAB;
77     int32x4_t vacc2xCD = vacc0xCD;
78     int32x4_t vacc2xEF = vacc0xEF;
79 
80     size_t p = ks;
81     do {
82       const int8_t* restrict a0 = a[0];
83       if XNN_UNPREDICTABLE(a0 != zero) {
84         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
85       }
86       const int8_t* restrict a1 = a[1];
87       if XNN_UNPREDICTABLE(a1 != zero) {
88         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
89       }
90       const int8_t* restrict a2 = a[2];
91       if XNN_UNPREDICTABLE(a2 != zero) {
92         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93       }
94       a += 3;
95 
96       size_t k = kc;
97 
98 
99       while (k >= 8 * sizeof(int8_t)) {
100         const int32x2_t va00 = vld1_dup_s32((const void*)a0);
101         const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
102         const int32x2_t va10 = vld1_dup_s32((const void*)a1);
103         const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
104         const int32x2_t va20 = vld1_dup_s32((const void*)a2);
105         const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
106 
107         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119         const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120         const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121         const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122         const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 
124         const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
125         const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
126         const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
127 
128         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
129         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
130         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
131         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
132         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
133         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
134         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
135         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
136         const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
137         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
138         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
139         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
140         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
141         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
142         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
143         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
144         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
145         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
146         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
147         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
148         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
149         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
150         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
151         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
152         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
153         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
154         const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
155         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
156         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
157         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
158         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
159         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
160         const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
161         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
162         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
163         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
164         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
165         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
166         const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
167         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
168         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
169         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
170         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
171         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
172         const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
173         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
174         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
175         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
176         const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
177         const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
178         const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
179 
180         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
181         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
182         const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
183         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
184         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
185         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
186         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
187         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
188         const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
189         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
190         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
191         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
192         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
193         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
194         const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
195         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
196         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
197         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
198         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
199         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
200         const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
201         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
202         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
203         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
204         const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
205         const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
206         const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1);
207         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
208         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
209         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
210         const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
211         const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
212         const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1);
213         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
214         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
215         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
216         const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
217         const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
218         const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1);
219         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
220         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
221         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
222         const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
223         const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
224         const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1);
225         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
226         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
227         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
228 
229         k -= 8 * sizeof(int8_t);
230       }
231 
232       if XNN_UNLIKELY(k != 0) {
233         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
234         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
235         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
236 
237         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
239         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
240         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245 
246         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
247         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
248         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
249         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
250         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
251         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
252         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
253         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
254         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
255         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
256         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
257         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
258         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
259         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
260         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
261         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
262         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
263         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
264         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
265         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
266         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
267         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
268         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
269         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
270         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
271         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
272         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
273         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
274         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
275         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
276         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
277         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
278         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
279         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
280         const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
281         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
282         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
283         const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
284         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
285         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
286         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
287         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
288         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
289         const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
290         vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
291         const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
292         vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
293         const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
294         vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
295         const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
296         vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
297       }
298       p -= 3 * sizeof(void*);
299     } while (p != 0);
300 
301 #if XNN_ARCH_ARM64
302     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
303     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
304     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
305     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
306     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
307     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
308     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
309     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
310     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
311     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
312     int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
313     int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
314 #else
315     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
316     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
317     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
318     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
319     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
320     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
321     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
322     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
323     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
324     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
325     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
326     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
327     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
328     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
329     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
330     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
331     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
332     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
333     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
334     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
335     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
336     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
337     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
338     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
339     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
340     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
341     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
342     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
343     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
344     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
345     const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
346     const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
347     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
348     const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
349     const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
350     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
351 #endif
352 
353     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
354     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
355     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
356 
357     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
358     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
359     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
360     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
361     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
362     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
363     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
364     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
365     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
366     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
367     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
368     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
369 
370     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
371     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
372     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
373     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
374     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
375     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
376     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
377     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
378     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
379     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
380     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
381     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
382 
383     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
384     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
385     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
386     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
387     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
388     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
389     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
390     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
391     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
392     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
393     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
394     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
395 
396     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
397 #if XNN_ARCH_ARM64
398     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
399     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
400     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
401     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
402     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
403     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
404 
405     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
406     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
407     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
408     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
409     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
410     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
411 
412     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
413     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
414     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
415 #else
416     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
417     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
418     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
419     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
420     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
421     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
422 
423     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
424     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
425     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
426     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
427     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
428     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
429 
430     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
431     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
432     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
433 #endif
434 
435     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
436     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
437     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
438     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
439 
440     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
441     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
442     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
443     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
444 
445     if (nc >= 16) {
446       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
447       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
448       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
449 
450       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
451       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
452       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
453 
454       a = (const int8_t**restrict) ((uintptr_t) a - ks);
455 
456       nc -= 16;
457     } else {
458       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
459       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
460       if (nc & 8) {
461         vst1_s8(c2, vout2x01234567); c2 += 8;
462         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
463         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
464         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
465         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
466       }
467       if (nc & 4) {
468         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
469         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
470         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
471         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
472         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
473       }
474       if (nc & 2) {
475         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
476         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
477         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
478         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
479         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
480       }
481       if (nc & 1) {
482         vst1_lane_s8(c2, vout2x01234567, 0);
483         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
484         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
485       }
486 
487       nc = 0;
488     }
489   } while (nc != 0);
490 }
491