Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x16-minmax-neon-mlal-lane.c

raw
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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/igemm.h"
17
18
19 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x16__neon_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 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const int8_t* zero_data,
32 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 1);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (1 * sizeof(void*)) == 0);
41 assert(a_offset % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 float* c0 = c;
47
48 do {
49 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
50 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
53 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
54 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
55 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
56 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
57 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
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 } else {
65 a0 = zero_data;
66 }
67 a += 1;
68
69 size_t k = kc;
70 while (k >= 8 * sizeof(int8_t)) {
71 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
72 const int16x8_t vxa0 = vmovl_s8(va0);
73
74 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
75 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
76
77 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
78 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
79 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
81
82 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
83 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
84 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
85 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
86
87 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
88 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
89 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
90 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
91
92 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
93 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
94 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
95 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
96
97 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
98 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
99 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
100 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
101
102 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
103 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
104 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
105 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
106
107 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
108 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
109 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
110 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
111
112 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
113 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
114
115
116 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
117 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
118
119 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
120 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
121 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
122 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
123
124 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
125 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
126 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
127 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
128
129 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
130 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
131 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
132 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
133
134 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
135 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
136 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
137 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
138
139 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
140 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
141 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
142 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
143
144 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
145 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
146 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
147 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
148
149 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
150 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
151 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const int8_t*) w + 8;
152 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
153
154 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
155 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
156
157 k -= 8 * sizeof(int8_t);
158 }
159 if XNN_UNLIKELY(k != 0) {
160 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
161 const int16x8_t vxa0 = vmovl_s8(va0);
162
163 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
164 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
165 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
166 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
167
168 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
169 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
170 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
171 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
172
173 if (k >= 2 * sizeof(int8_t)) {
174 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
175 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
176 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
177 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
178
179 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
180 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
181 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
182 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
183
184 if (k > 2 * sizeof(int8_t)) {
185 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
186 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
187 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
188 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
189
190 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
191 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
192 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
193 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
194
195 if (k >= 4 * sizeof(int8_t)) {
196 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
197 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
198 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
199 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
200
201 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
202 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
203 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
204 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
205
206 if (k > 4 * sizeof(int8_t)) {
207 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
208 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
209 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
210 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
211
212 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
213 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
214 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
215 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
216
217 if (k >= 6 * sizeof(int8_t)) {
218 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
219 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
220 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
221 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
222
223 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
224 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
225 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
226 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
227
228 if (k > 6 * sizeof(int8_t)) {
229 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
230 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
231 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
232 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
233
234 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
235 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
236 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
237 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
238 }
239 }
240 }
241 }
242 }
243 }
244 }
245 p -= 1 * sizeof(void*);
246 } while (p != 0);
247
248 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
249 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
250 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
251 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
252
253 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
254 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
255 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
256 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
257 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
258
259 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
260 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
261 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
262 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
263
264 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
265 #if XNN_ARCH_ARM64
266 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
267 #else
268 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
269 #endif
270 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
271 #if XNN_ARCH_ARM64
272 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
273 #else
274 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
275 #endif
276 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
277 #if XNN_ARCH_ARM64
278 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
279 #else
280 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
281 #endif
282 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
283 #if XNN_ARCH_ARM64
284 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
285 #else
286 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
287 #endif
288
289 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
290 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
291 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
292 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
293 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
294
295 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
296 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
297 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
298 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
299 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
300
301 if XNN_LIKELY(nc >= 16) {
302 vst1q_f32(&c0[0], vout0x0123);
303 vst1q_f32(&c0[4], vout0x4567);
304 vst1q_f32(&c0[8], vout0x89AB);
305 vst1q_f32(&c0[12], vout0xCDEF);
306
307 a = (const int8_t**restrict) ((uintptr_t) a - ks);
308
309 c0 = (float*) ((uintptr_t) c0 + cn_stride);
310
311 nc -= 16;
312 } else {
313 if (nc & 8) {
314 vst1q_f32(c0, vout0x0123); c0 += 4;
315 vout0x0123 = vout0x89AB;
316 vst1q_f32(c0, vout0x4567); c0 += 4;
317 vout0x4567 = vout0xCDEF;
318 }
319 if (nc & 4) {
320 vst1q_f32(c0, vout0x0123); c0 += 4;
321 vout0x0123 = vout0x4567;
322 }
323 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
324 if (nc & 2) {
325 vst1_f32(c0, vout0x01); c0 += 2;
326 vout0x01 = vget_high_f32(vout0x0123);
327 }
328 if (nc & 1) {
329 vst1_lane_f32(c0, vout0x01, 0);
330 }
331 nc = 0;
332 }
333 } while (nc != 0);
334 }