xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x16-minmax-fp32-neonv8-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 #include <xnnpack/intrinsics-polyfill.h>
17 
18 
xnn_qc8_igemm_minmax_fp32_ukernel_1x16__neonv8_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const int8_t** restrict a,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const int8_t* zero,
31     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 1);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (1 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(int8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   int8_t* c0 = c;
45 
46   do {
47     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
48     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
49     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
50     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
51 
52     size_t p = ks;
53     do {
54       const int8_t* restrict a0 = a[0];
55       if XNN_UNPREDICTABLE(a0 != zero) {
56         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
57       }
58       a += 1;
59 
60       size_t k = kc;
61       while (k >= 8 * sizeof(int8_t)) {
62         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
63         const int16x8_t vxa0 = vmovl_s8(va0);
64 
65         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
66         const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
67 
68         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
69         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
70         const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
71         const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
72 
73         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
74         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
75         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
76         const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
77 
78         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
79         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
80         const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
81         const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
82 
83         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
84         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
85         const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
86         const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
87 
88         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
89         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
90         const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
91         const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
92 
93         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
94         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
95         const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
96         const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
97 
98         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
99         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
100         const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
101         const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
102 
103         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
104         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
105 
106 
107         const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
108         const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
109 
110         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
111         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
112         const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
113         const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
114 
115         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
116         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
117         const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
118         const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
119 
120         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
121         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
122         const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
123         const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
124 
125         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
126         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
127         const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
128         const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
129 
130         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
131         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
132         const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
133         const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
134 
135         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
136         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
137         const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
138         const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
139 
140         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
141         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
142         const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
143         const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
144 
145         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
146         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
147 
148         k -= 8 * sizeof(int8_t);
149       }
150       if XNN_UNLIKELY(k != 0) {
151         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
152         const int16x8_t vxa0 = vmovl_s8(va0);
153 
154         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
155         const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
156         const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
157         const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
158 
159         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
160         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
161         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
162         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
163 
164         if (k >= 2 * sizeof(int8_t)) {
165           const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
166           const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
167           const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
168           const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
169 
170           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
171           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
172           vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
173           vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
174 
175           if (k > 2 * sizeof(int8_t)) {
176             const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
177             const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
178             const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
179             const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
180 
181             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
182             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
183             vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
184             vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
185 
186             if (k >= 4 * sizeof(int8_t)) {
187               const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
188               const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
189               const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
190               const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
191 
192               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
193               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
194               vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
195               vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
196 
197               if (k > 4 * sizeof(int8_t)) {
198                 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
199                 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
200                 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
201                 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
202 
203                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
204                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
205                 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
206                 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
207 
208                 if (k >= 6 * sizeof(int8_t)) {
209                   const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
210                   const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
211                   const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
212                   const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
213 
214                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
215                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
216                   vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
217                   vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
218 
219                   if (k > 6 * sizeof(int8_t)) {
220                     const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
221                     const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
222                     const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
223                     const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
224 
225                     vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
226                     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
227                     vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
228                     vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
229                   }
230                 }
231               }
232             }
233           }
234         }
235       }
236       p -= 1 * sizeof(void*);
237     } while (p != 0);
238 
239     // Post-accumulation work
240     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
241     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
242     float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
243     float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
244 
245     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
246     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
247     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
248     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
249     const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
250     vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
251     const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
252     vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
253 
254     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
255     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
256     vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
257     vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
258 
259     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
260 #if XNN_ARCH_ARM64
261     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
262     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
263 
264     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
265     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
266 
267     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
268 #else
269     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
270     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
271 
272     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
273     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
274 
275     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
276 #endif
277 
278     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
279     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
280 
281     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
282     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
283 
284     if (nc >= 16) {
285       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
286 
287       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
288 
289       a = (const int8_t**restrict) ((uintptr_t) a - ks);
290 
291       nc -= 16;
292     } else {
293       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
294       if (nc & 8) {
295         vst1_s8(c0, vout0x01234567); c0 += 8;
296         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
297       }
298       if (nc & 4) {
299         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
300         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
301       }
302       if (nc & 2) {
303         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
304         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
305       }
306       if (nc & 1) {
307         vst1_lane_s8(c0, vout0x01234567, 0);
308       }
309 
310       nc = 0;
311     }
312   } while (nc != 0);
313 }
314