1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
16
17
xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane_prfm(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane_prfm(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41
42 do {
43 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
44 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
45 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
46 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
47
48 size_t k = kc;
49 while (k >= 8 * sizeof(int8_t)) {
50 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
51 const int16x8_t vxa0 = vmovl_s8(va0);
52
53 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
54 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
55
56 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
57 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
58 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
59 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
60
61 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
62 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
63 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
64 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
65
66 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
67 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
68 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
69 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
70
71 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
72 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
73 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
74 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
75
76 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
77 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
78 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
79 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
80
81 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
82 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
83 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
84 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
85
86 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
87 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
88 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
89 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
90
91 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
92 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
93
94 __builtin_prefetch((const int8_t*) w + 448);
95 __builtin_prefetch((const int8_t*) w + 512);
96
97 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
98 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
99
100 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
101 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
102 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
103 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
104
105 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
106 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
107 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
108 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
109
110 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
111 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
112 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
113 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
114
115 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
116 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
117 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
118 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
119
120 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
121 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
122 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
123 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
124
125 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
126 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
127 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
128 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
129
130 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
131 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
132 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
133 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
134
135 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
136 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
137
138 k -= 8 * sizeof(int8_t);
139 }
140 if XNN_UNLIKELY(k != 0) {
141 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
142 const int16x8_t vxa0 = vmovl_s8(va0);
143
144 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
145 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
146 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
147 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
148
149 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
150 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
151 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
152 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
153
154 if (k >= 2 * sizeof(int8_t)) {
155 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
156 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
157 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
158 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
159
160 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
161 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
162 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
163 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
164
165 if (k > 2 * sizeof(int8_t)) {
166 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
167 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
168 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
169 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
170
171 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
172 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
173 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
174 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
175
176 if (k >= 4 * sizeof(int8_t)) {
177 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
178 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
179 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
180 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
181
182 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
183 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
184 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
185 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
186
187 if (k > 4 * sizeof(int8_t)) {
188 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
189 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
190 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
191 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
192
193 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
194 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
195 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
196 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
197
198 if (k >= 6 * sizeof(int8_t)) {
199 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
200 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
201 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
202 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
203
204 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
205 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
206 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
207 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
208
209 if (k > 6 * sizeof(int8_t)) {
210 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
211 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
212 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
213 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
214
215 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
216 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
217 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
218 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
219 }
220 }
221 }
222 }
223 }
224 }
225 }
226
227 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
228 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
229 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
230 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
231
232 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
233 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
234 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
235 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
236 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
237 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
238 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
239 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
240
241 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
242 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
243 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
244 vacc0x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x89AB, vmagic_bias));
245 vacc0xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc0xCDEF, vmagic_bias));
246
247 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
248 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
249 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
250 vacc0x89AB = vqsubq_s32(vacc0x89AB, vmagic_bias_less_output_zero_point);
251 vacc0xCDEF = vqsubq_s32(vacc0xCDEF, vmagic_bias_less_output_zero_point);
252
253 #if XNN_ARCH_ARM64
254 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
255 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
256
257
258 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
259 #else
260 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
261 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
262
263
264 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
265 #endif
266
267 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
268 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
269
270 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
271 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
272
273 if (nc >= 16) {
274 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
275
276 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
277
278 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
279
280 nc -= 16;
281 } else {
282 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
283 if (nc & 8) {
284 vst1_s8(c0, vout0x01234567); c0 += 8;
285 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
286 }
287 if (nc & 4) {
288 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
289 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
290 }
291 if (nc & 2) {
292 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
293 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
294 }
295 if (nc & 1) {
296 vst1_lane_s8(c0, vout0x01234567, 0);
297 }
298
299 nc = 0;
300 }
301 } while (nc != 0);
302 }
303