• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c4-neon-mull-shuffle.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_3x8c4s2__neon_mlal(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_3x8c4s2__neon_mlal(
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   int8_t* c0 = c;
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52 
53   kc = round_up_po2(kc, 8 * sizeof(int8_t));
54   do {
55     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
56     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
59     int32x4_t vacc1x01 = vacc0x01;
60     int32x4_t vacc1x23 = vacc0x23;
61     int32x4_t vacc1x45 = vacc0x45;
62     int32x4_t vacc1x67 = vacc0x67;
63     int32x4_t vacc2x01 = vacc0x01;
64     int32x4_t vacc2x23 = vacc0x23;
65     int32x4_t vacc2x45 = vacc0x45;
66     int32x4_t vacc2x67 = vacc0x67;
67 
68     size_t p = ks;
69     do {
70       const int8_t* restrict a0 = a[0];
71       if XNN_UNPREDICTABLE(a0 != zero) {
72         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73       }
74       const int8_t* restrict a1 = a[1];
75       if XNN_UNPREDICTABLE(a1 != zero) {
76         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77       }
78       const int8_t* restrict a2 = a[2];
79       if XNN_UNPREDICTABLE(a2 != zero) {
80         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
81       }
82       a += 3;
83 
84       size_t k = kc;
85       while (k >= 16 * sizeof(int8_t)) {
86         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
87         int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
88         int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
89         int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
90         int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
91         int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
92 
93         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
96         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
97         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
98         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
99         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
100         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
101 
102         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
103         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
104         int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
105         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
106         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
107         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
108         vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2x1);
109         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
110         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
111         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
112         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
113         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
114         int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
115         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
116         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
117         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
118         vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1);
119         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
120         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
121         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
122         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
123         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
124         int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
125         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
126         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
127         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
128         vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2x1);
129         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
130         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
131         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
132         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
133         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
134         int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
135         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
136         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
137         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
138         vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1);
139         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
140         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
141         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
142         va0x0 = vext_s8(va0x0, va0x0, 4);
143         va0x1 = vext_s8(va0x1, va0x1, 4);
144         va1x0 = vext_s8(va1x0, va1x0, 4);
145         va1x1 = vext_s8(va1x1, va1x1, 4);
146         va2x0 = vext_s8(va2x0, va2x0, 4);
147         va2x1 = vext_s8(va2x1, va2x1, 4);
148         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
149         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
150         int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
151         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
152         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
153         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
154         vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2x1);
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, va0x0);
159         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
160         int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
161         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
162         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
163         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
164         vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2x1);
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, va0x0);
169         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
170         int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
171         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
172         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
173         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
174         vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2x1);
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, va0x0);
179         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
180         int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
181         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
182         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
183         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
184         vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2x1);
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       if (k != 0) {
192         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
193         int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
194         int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
195 
196         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
197         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
198         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
199         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
200         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
201         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
202         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
203         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
204 
205         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
206         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
207         int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
208         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
209         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
210         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
211         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
212         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
213         int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
214         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
215         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
216         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
217         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
218         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
219         int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
220         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
221         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
222         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
223         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
224         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
225         int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
226         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
227         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
228         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
229         va0x0 = vext_s8(va0x0, va0x0, 4);
230         va1x0 = vext_s8(va1x0, va1x0, 4);
231         va2x0 = vext_s8(va2x0, va2x0, 4);
232         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
233         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
234         int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
235         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
236         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
237         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
238         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
239         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
240         int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
241         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
242         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
243         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
244         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
245         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
246         int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
247         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
248         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
249         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
250         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
251         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
252         int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
253         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
254         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
255         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
256 
257       }
258 
259       p -= 3 * sizeof(void*);
260     } while (p != 0);
261 
262 #if XNN_ARCH_ARM64
263     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
264     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
265     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
266     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
267     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
268     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
269 #else
270     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
271     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
272     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
273     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
274     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
275     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
276     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
277     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
278     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
279     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
280     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
281     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
282     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
283     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
284     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
285     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
286     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
287     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
288 #endif
289 
290     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
291     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
292     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
293 
294     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
295     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
296     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
297     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
298     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
299     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
300 
301     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
302     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
303     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
304     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
305     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
306     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
307 
308     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
309     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
310     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
311     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
312     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
313     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
314 
315     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
316 #if XNN_ARCH_ARM64
317     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
318     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
319     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
320 
321     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
322     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
323     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
324 
325     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
326     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
327 #else
328     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
329     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
330     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
331 
332     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
333     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
334     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
335 
336     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
337     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
338 #endif
339 
340     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
341     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
342     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
343 
344     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
345     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
346     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
347 
348     if (nc >= 8) {
349       vst1_s8(c2 + 0, vout2x01234567);
350       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
351       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
352 
353       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
354       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
355       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
356 
357       a = (const int8_t**restrict) ((uintptr_t) a - ks);
358 
359       nc -= 8;
360     } else {
361       if (nc & 4) {
362         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
363         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
364         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
365         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
366         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
367       }
368       if (nc & 2) {
369         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
370         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
371         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
372         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
373         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
374       }
375       if (nc & 1) {
376         vst1_lane_s8(c2, vout2x01234567, 0);
377         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
378         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
379       }
380 
381       nc = 0;
382     }
383   } while (nc != 0);
384 }
385