xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/1x16-minmax-rndnu-neon-mlal-lane.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/neon-mlal-lane.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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/common.h>
15 #include <xnnpack/igemm.h>
16 
17 
xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane(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_1x16__neon_mlal_lane(
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 <= 1);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (1 * 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 
45   do {
46     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
47     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
48     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
49     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
50 
51     size_t p = ks;
52     do {
53       const int8_t* restrict a0 = a[0];
54       if XNN_UNPREDICTABLE(a0 != zero) {
55         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
56       }
57       a += 1;
58 
59       size_t k = kc;
60       while (k >= 8 * sizeof(int8_t)) {
61         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
62         const int16x8_t vxa0 = vmovl_s8(va0);
63 
64         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
65         const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
66 
67         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
68         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
69         const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
70         const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
71 
72         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
73         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
74         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
75         const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
76 
77         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
78         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
79         const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
80         const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
81 
82         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
83         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
84         const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
85         const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
86 
87         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
88         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
89         const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
90         const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
91 
92         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
93         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
94         const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
95         const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
96 
97         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
98         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
99         const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
100         const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
101 
102         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
103         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
104 
105 
106         const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
107         const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
108 
109         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
110         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
111         const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
112         const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
113 
114         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
115         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
116         const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
117         const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
118 
119         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
120         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
121         const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
122         const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
123 
124         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
125         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
126         const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
127         const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
128 
129         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
130         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
131         const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
132         const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
133 
134         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
135         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
136         const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
137         const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
138 
139         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
140         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
141         const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
142         const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
143 
144         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
145         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
146 
147         k -= 8 * sizeof(int8_t);
148       }
149       if XNN_UNLIKELY(k != 0) {
150         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
151         const int16x8_t vxa0 = vmovl_s8(va0);
152 
153         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
154         const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
155         const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
156         const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
157 
158         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
159         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
160         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
161         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
162 
163         if (k >= 2 * sizeof(int8_t)) {
164           const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
165           const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
166           const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
167           const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
168 
169           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
170           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
171           vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
172           vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
173 
174           if (k > 2 * sizeof(int8_t)) {
175             const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
176             const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
177             const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
178             const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
179 
180             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
181             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
182             vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
183             vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
184 
185             if (k >= 4 * sizeof(int8_t)) {
186               const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
187               const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
188               const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
189               const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
190 
191               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
192               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
193               vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
194               vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
195 
196               if (k > 4 * sizeof(int8_t)) {
197                 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
198                 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
199                 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
200                 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
201 
202                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
203                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
204                 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
205                 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
206 
207                 if (k >= 6 * sizeof(int8_t)) {
208                   const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
209                   const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
210                   const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
211                   const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
212 
213                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
214                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
215                   vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
216                   vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
217 
218                   if (k > 6 * sizeof(int8_t)) {
219                     const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
220                     const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
221                     const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
222                     const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
223 
224                     vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
225                     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
226                     vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
227                     vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
228                   }
229                 }
230               }
231             }
232           }
233         }
234       }
235       p -= 1 * sizeof(void*);
236     } while (p != 0);
237 
238     // Post-accumulation work
239     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
240     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
241     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
242 
243     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
244     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
245     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
246     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
247 
248     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
249     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
250     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
251     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
252 
253     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
254     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
255     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
256     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
257 
258     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
259 #if XNN_ARCH_ARM64
260     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
261     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
262 
263     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
264     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
265 
266     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
267 #else
268     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
269     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
270 
271     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
272     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
273 
274     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
275 #endif
276 
277     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
278     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
279 
280     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
281     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
282 
283     if (nc >= 16) {
284       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
285 
286       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
287 
288       a = (const int8_t**restrict) ((uintptr_t) a - ks);
289 
290       nc -= 16;
291     } else {
292       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
293       if (nc & 8) {
294         vst1_s8(c0, vout0x01234567); c0 += 8;
295         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
296       }
297       if (nc & 4) {
298         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
299         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
300       }
301       if (nc & 2) {
302         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
303         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
304       }
305       if (nc & 1) {
306         vst1_lane_s8(c0, vout0x01234567, 0);
307       }
308 
309       nc = 0;
310     }
311   } while (nc != 0);
312 }
313