Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x16-minmax-neon-mlal-lane-prfm.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 #include "xnnpack/prefetch.h"
18
19
20 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x16__neon_mlal_lane_prfm(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 1);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (1 * sizeof(void*)) == 0);
42 assert(a_offset % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 float* c0 = c;
48
49 do {
50 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
54 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
55 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
56 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
57 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
58 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
59
60 size_t p = ks;
61 do {
62 const int8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 } else {
66 a0 = zero_data;
67 }
68 a += 1;
69
70 size_t k = kc;
71 while (k >= 8 * sizeof(int8_t)) {
72 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
73 const int16x8_t vxa0 = vmovl_s8(va0);
74
75 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
76 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
77
78 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
79 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
80 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
82
83 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
84 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
85 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
87
88 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
89 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
90 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
92
93 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
94 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
95 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
96 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
97
98 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
99 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
100 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
101 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
102
103 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
104 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
105 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
106 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
107
108 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
109 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
110 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
111 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
112
113 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
114 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
115
116 xnn_prefetch_to_l1((const int8_t*) w + 448);
117 xnn_prefetch_to_l1((const int8_t*) w + 512);
118
119 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
120 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
121
122 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
123 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
124 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
125 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
126
127 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
128 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
129 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
130 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
131
132 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
133 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
134 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
135 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
136
137 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
138 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
139 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
140 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
141
142 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
143 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
144 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
145 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
146
147 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
148 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
149 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
150 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
151
152 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
153 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
154 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const int8_t*) w + 8;
155 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
156
157 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
158 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
159
160 k -= 8 * sizeof(int8_t);
161 }
162 if XNN_UNLIKELY(k != 0) {
163 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
164 const int16x8_t vxa0 = vmovl_s8(va0);
165
166 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
167 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
168 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
169 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
170
171 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
172 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
173 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
174 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
175
176 if (k >= 2 * sizeof(int8_t)) {
177 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
178 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
179 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
180 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
181
182 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
183 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
184 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
185 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
186
187 if (k > 2 * sizeof(int8_t)) {
188 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
189 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
190 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
191 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
192
193 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
194 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
195 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
196 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
197
198 if (k >= 4 * sizeof(int8_t)) {
199 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
200 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
201 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
202 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
203
204 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
205 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
206 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
207 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
208
209 if (k > 4 * sizeof(int8_t)) {
210 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
211 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
212 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
213 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
214
215 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
216 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
217 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
218 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
219
220 if (k >= 6 * sizeof(int8_t)) {
221 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
222 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
223 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
224 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
225
226 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
227 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
228 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
229 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
230
231 if (k > 6 * sizeof(int8_t)) {
232 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
233 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
234 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
235 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
236
237 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
238 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
239 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
240 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
241 }
242 }
243 }
244 }
245 }
246 }
247 }
248 p -= 1 * sizeof(void*);
249 } while (p != 0);
250
251 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
252 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
253 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
254 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
255
256 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
257 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
258 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
259 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
260 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
261
262 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
263 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
264 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
265 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
266
267 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
268 #if XNN_ARCH_ARM64
269 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
270 #else
271 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
272 #endif
273 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
274 #if XNN_ARCH_ARM64
275 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
276 #else
277 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
278 #endif
279 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
280 #if XNN_ARCH_ARM64
281 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
282 #else
283 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
284 #endif
285 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
286 #if XNN_ARCH_ARM64
287 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
288 #else
289 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
290 #endif
291
292 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
293 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
294 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
295 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
296 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
297
298 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
299 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
300 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
301 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
302 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
303
304 if XNN_LIKELY(nc >= 16) {
305 vst1q_f32(&c0[0], vout0x0123);
306 vst1q_f32(&c0[4], vout0x4567);
307 vst1q_f32(&c0[8], vout0x89AB);
308 vst1q_f32(&c0[12], vout0xCDEF);
309
310 a = (const int8_t**restrict) ((uintptr_t) a - ks);
311
312 c0 = (float*) ((uintptr_t) c0 + cn_stride);
313
314 nc -= 16;
315 } else {
316 if (nc & 8) {
317 vst1q_f32(c0, vout0x0123); c0 += 4;
318 vout0x0123 = vout0x89AB;
319 vst1q_f32(c0, vout0x4567); c0 += 4;
320 vout0x4567 = vout0xCDEF;
321 }
322 if (nc & 4) {
323 vst1q_f32(c0, vout0x0123); c0 += 4;
324 vout0x0123 = vout0x4567;
325 }
326 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
327 if (nc & 2) {
328 vst1_f32(c0, vout0x01); c0 += 2;
329 vout0x01 = vget_high_f32(vout0x0123);
330 }
331 if (nc & 1) {
332 vst1_lane_f32(c0, vout0x01, 0);
333 }
334 nc = 0;
335 }
336 } while (nc != 0);
337 }