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