xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x16c2-minmax-rndnu-neon-mull-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c2-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 
xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
29     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 2);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (2 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
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 
49   do {
50     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
53     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
54     int32x4_t vacc1x0123 = vacc0x0123;
55     int32x4_t vacc1x4567 = vacc0x4567;
56     int32x4_t vacc1x89AB = vacc0x89AB;
57     int32x4_t vacc1xCDEF = vacc0xCDEF;
58 
59     size_t p = ks;
60     do {
61       const int8_t* restrict a0 = a[0];
62       if XNN_UNPREDICTABLE(a0 != zero) {
63         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
64       }
65       const int8_t* restrict a1 = a[1];
66       if XNN_UNPREDICTABLE(a1 != zero) {
67         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
68       }
69       a += 2;
70 
71       size_t k = kc;
72 
73 
74       while (k >= 8 * sizeof(int8_t)) {
75         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
76         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
77 
78         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 
95         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
96         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
97 
98         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
99         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
100         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
101         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
102         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
103         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
104         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
105         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
106         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
107         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
108         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
109         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
110         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
111         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
112         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
113         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
114         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
115         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
116 
117         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
118         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
119         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
120         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
121         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
122         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
123         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
124         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
125         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
126         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
127         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
128         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
129         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
130         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
131         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
132         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
133         const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
134         const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
135 
136         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
137         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
138         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
139         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
140         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
141         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
142         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
143         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
144         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
145         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
146         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
147         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
148         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
149         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
150         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
151         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
152         const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
153         const int8x8_t va1c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3));
154 
155         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
156         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
157         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
158         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
159         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
160         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
161         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
162         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
163         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
164         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
165         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
166         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
167         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
168         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
169         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
170         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
171 
172         k -= 8 * sizeof(int8_t);
173       }
174 
175       if XNN_UNLIKELY(k != 0) {
176         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
177         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
178 
179         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183 
184         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
185         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
186         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
187         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
188         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
189         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
190         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
191         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
192         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
193         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
194         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
195         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
196         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
197         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
198         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
199         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
200         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
201         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
202 
203         if (k > 2 * sizeof(int8_t)) {
204           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
208 
209           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
210           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
211           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
212           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
213           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
214           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
215           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
216           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
217           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
218           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
219           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
220           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
221           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
222           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
223           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
224           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
225           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
226           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
227 
228           if (k > 4 * sizeof(int8_t)) {
229             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
233 
234             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
235             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
236             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
237             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
238             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
239             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
240             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
241             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
242             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
243             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
244             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
245             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
246             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
247             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
248             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
249             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
250             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
251             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
252           }
253         }
254       }
255       p -= 2 * sizeof(void*);
256     } while (p != 0);
257 
258     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
259     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
260     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
261 
262     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
263     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
264     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
265     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
266     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
267     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
268     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
269     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
270 
271     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
272     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
273     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
274     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
275     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
276     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
277     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
278     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
279 
280     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
281     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
282     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
283     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
284     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
285     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
286     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
287     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
288 
289     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
290 #if XNN_ARCH_ARM64
291     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
292     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
293     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
294     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
295 
296     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
297     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
298     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
299     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
300 
301     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
302     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
303 #else
304     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
305     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
306     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
307     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
308 
309     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
310     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
311     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
312     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
313 
314     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
315     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
316 #endif
317 
318     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
319     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
320     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
321 
322     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
323     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
324     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
325 
326     if (nc >= 16) {
327       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
328       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
329 
330       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
331       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
332 
333       a = (const int8_t**restrict) ((uintptr_t) a - ks);
334 
335       nc -= 16;
336     } else {
337       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
338       if (nc & 8) {
339         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
340         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
341         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
342       }
343       if (nc & 4) {
344         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
345         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
346         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
347       }
348       if (nc & 2) {
349         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
350         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
351         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
352       }
353       if (nc & 1) {
354         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
355         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
356       }
357 
358       nc = 0;
359     }
360   } while (nc != 0);
361 }
362