xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/4x8c4-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_4x8c4__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_4x8c4__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 <= 4);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (4 * 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   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54   if XNN_UNPREDICTABLE(mr != 4) {
55     c3 = c2;
56   }
57 
58   do {
59     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
60     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
61     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
62     int32x4_t vacc0x67 = 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 vacc2x01 = vacc0x01;
68     int32x4_t vacc2x23 = vacc0x23;
69     int32x4_t vacc2x45 = vacc0x45;
70     int32x4_t vacc2x67 = vacc0x67;
71     int32x4_t vacc3x01 = vacc0x01;
72     int32x4_t vacc3x23 = vacc0x23;
73     int32x4_t vacc3x45 = vacc0x45;
74     int32x4_t vacc3x67 = vacc0x67;
75 
76     size_t p = ks;
77     do {
78       const int8_t* restrict a0 = a[0];
79       if XNN_UNPREDICTABLE(a0 != zero) {
80         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
81       }
82       const int8_t* restrict a1 = a[1];
83       if XNN_UNPREDICTABLE(a1 != zero) {
84         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
85       }
86       const int8_t* restrict a2 = a[2];
87       if XNN_UNPREDICTABLE(a2 != zero) {
88         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
89       }
90       const int8_t* restrict a3 = a[3];
91       if XNN_UNPREDICTABLE(a3 != zero) {
92         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
93       }
94       a += 4;
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         const int32x2_t va30 = vld1_dup_s32((const void*)a3);
107         const int32x2_t va31 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8;
108 
109         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 
118         const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
119         const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
120         const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
121         const int8x8_t va3c0 = vreinterpret_s8_s32(va30);
122 
123         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
124         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
125         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
126         const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
127         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
128         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
129         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
130         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
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         const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
135         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
136         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
137         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
138         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
139         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
140         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
141         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
142         const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
143         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
144         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
145         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
146         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
147         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
148         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
149         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
150         const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
151         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
152         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
153         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
154         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
155         const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
156         const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
157         const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
158         const int8x8_t va3c1 = vreinterpret_s8_s32(va31);
159 
160         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
161         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
162         const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
163         const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
164         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
165         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
166         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
167         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
168         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
169         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
170         const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
171         const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
172         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
173         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
174         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
175         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
176         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
177         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
178         const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
179         const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
180         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
181         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
182         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
183         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
184         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
185         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
186         const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
187         const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
188         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
189         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
190         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
191         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
192 
193         k -= 8 * sizeof(int8_t);
194       }
195 
196       if XNN_UNLIKELY(k != 0) {
197         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
198         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
199         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
200         const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
201 
202         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
203         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
204         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206 
207         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
208         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
209         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
210         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
211         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
212         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
213         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
214         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
215         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
216         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
217         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
218         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
219         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
220         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
221         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
222         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
223         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
224         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
225         const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
226         const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
227         vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
228         const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
229         vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
230         const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
231         vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
232         const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
233         vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
234         const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
235         const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
236         vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
237         const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
238         vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
239         const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
240         vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
241         const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
242         vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
243       }
244       p -= 4 * sizeof(void*);
245     } while (p != 0);
246 
247 #if XNN_ARCH_ARM64
248     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
249     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
250     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
251     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
252     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
253     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
254     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
255     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
256 #else
257     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
258     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
259     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
260     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
261     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
262     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
263     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
264     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
265     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
266     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
267     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
268     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
269     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
270     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
271     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
272     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
273     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
274     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
275     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
276     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
277     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
278     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
279     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
280     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
281 #endif
282 
283     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
284     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
285     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
286 
287     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
288     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
289     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
290     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
291     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
292     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
293     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
294     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
295 
296     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
297     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
298     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
299     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
300     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
301     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
302     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
303     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
304 
305     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
306     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
307     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
308     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
309     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
310     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
311     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
312     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
313 
314     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
315 #if XNN_ARCH_ARM64
316     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
317     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
318     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
319     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
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     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
325 
326     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
327     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
328 #else
329     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
330     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
331     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
332     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
333 
334     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
335     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
336     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
337     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
338 
339     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
340     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
341 #endif
342 
343     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
344     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
345     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
346 
347     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
348     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
349     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
350 
351     if (nc >= 8) {
352       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
353       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
354       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
355       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
356 
357       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
358       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
359       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
360       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
361 
362       a = (const int8_t**restrict) ((uintptr_t) a - ks);
363 
364       nc -= 8;
365     } else {
366       if (nc & 4) {
367         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
368         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
369         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
370         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
371         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
372         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
373       }
374       if (nc & 2) {
375         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
376         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
377         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
378         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
379         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
380         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
381       }
382       if (nc & 1) {
383         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
384         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
385         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
386         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
387       }
388 
389       nc = 0;
390     }
391   } while (nc != 0);
392 }
393