1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/neon-shuffle.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/igemm.h>
15
16
xnn_f32_igemm_minmax_ukernel_4x8s4__neon(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_4x8s4__neon(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 do {
58 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
59 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
60 float32x4_t vacc1x0123 = vacc0x0123;
61 float32x4_t vacc1x4567 = vacc0x4567;
62 float32x4_t vacc2x0123 = vacc0x0123;
63 float32x4_t vacc2x4567 = vacc0x4567;
64 float32x4_t vacc3x0123 = vacc0x0123;
65 float32x4_t vacc3x4567 = vacc0x4567;
66
67 size_t p = ks;
68 do {
69 const float* restrict a0 = a[0];
70 assert(a0 != NULL);
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const float*) ((uintptr_t) a0 + a_offset);
73 }
74 const float* restrict a1 = a[1];
75 assert(a1 != NULL);
76 if XNN_UNPREDICTABLE(a1 != zero) {
77 a1 = (const float*) ((uintptr_t) a1 + a_offset);
78 }
79 const float* restrict a2 = a[2];
80 assert(a2 != NULL);
81 if XNN_UNPREDICTABLE(a2 != zero) {
82 a2 = (const float*) ((uintptr_t) a2 + a_offset);
83 }
84 const float* restrict a3 = a[3];
85 assert(a3 != NULL);
86 if XNN_UNPREDICTABLE(a3 != zero) {
87 a3 = (const float*) ((uintptr_t) a3 + a_offset);
88 }
89 a += 4;
90
91 size_t k = kc;
92 while (k >= 4 * sizeof(float)) {
93 float32x4_t va0 = vld1q_f32(a0); a0 += 4;
94 float32x4_t va1 = vld1q_f32(a1); a1 += 4;
95 float32x4_t va2 = vld1q_f32(a2); a2 += 4;
96 float32x4_t va3 = vld1q_f32(a3); a3 += 4;
97
98
99 const float32x4_t vb0123c0 = vld1q_f32(w + 0);
100 const float32x4_t vb4567c0 = vld1q_f32(w + 4);
101
102 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c0);
103 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c0);
104 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c0);
105 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c0);
106 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c0);
107 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c0);
108 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c0);
109 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c0);
110
111 va0 = vextq_f32(va0, va0, 1);
112 va1 = vextq_f32(va1, va1, 1);
113 va2 = vextq_f32(va2, va2, 1);
114 va3 = vextq_f32(va3, va3, 1);
115
116 const float32x4_t vb0123c1 = vld1q_f32(w + 8);
117 const float32x4_t vb4567c1 = vld1q_f32(w + 12);
118
119 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c1);
120 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c1);
121 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c1);
122 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c1);
123 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c1);
124 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c1);
125 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c1);
126 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c1);
127
128 va0 = vextq_f32(va0, va0, 1);
129 va1 = vextq_f32(va1, va1, 1);
130 va2 = vextq_f32(va2, va2, 1);
131 va3 = vextq_f32(va3, va3, 1);
132
133 const float32x4_t vb0123c2 = vld1q_f32(w + 16);
134 const float32x4_t vb4567c2 = vld1q_f32(w + 20);
135
136 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c2);
137 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c2);
138 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c2);
139 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c2);
140 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c2);
141 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c2);
142 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c2);
143 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c2);
144
145 va0 = vextq_f32(va0, va0, 1);
146 va1 = vextq_f32(va1, va1, 1);
147 va2 = vextq_f32(va2, va2, 1);
148 va3 = vextq_f32(va3, va3, 1);
149
150 const float32x4_t vb0123c3 = vld1q_f32(w + 24);
151 const float32x4_t vb4567c3 = vld1q_f32(w + 28);
152
153 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c3);
154 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c3);
155 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c3);
156 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c3);
157 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c3);
158 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c3);
159 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c3);
160 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c3);
161
162
163 w += 32;
164 k -= 4 * sizeof(float);
165 }
166 if XNN_UNLIKELY(k != 0) {
167 float32x4_t va0 = vld1q_f32(a0); a0 = (const float*) ((uintptr_t) a0 + k);
168 float32x4_t va1 = vld1q_f32(a1); a1 = (const float*) ((uintptr_t) a1 + k);
169 float32x4_t va2 = vld1q_f32(a2); a2 = (const float*) ((uintptr_t) a2 + k);
170 float32x4_t va3 = vld1q_f32(a3); a3 = (const float*) ((uintptr_t) a3 + k);
171
172
173 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
174 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
175
176 const float32x4_t vmska0x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
177 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c0, vb0123c0);
178 const float32x4_t vmska1x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
179 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c0, vb0123c0);
180 const float32x4_t vmska2x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
181 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c0, vb0123c0);
182 const float32x4_t vmska3x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
183 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c0, vb0123c0);
184 const float32x4_t vmska0x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
185 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c0, vb4567c0);
186 const float32x4_t vmska1x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
187 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c0, vb4567c0);
188 const float32x4_t vmska2x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
189 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c0, vb4567c0);
190 const float32x4_t vmska3x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
191 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c0, vb4567c0);
192
193 va0 = vextq_f32(va0, va0, 1);
194 va1 = vextq_f32(va1, va1, 1);
195 va2 = vextq_f32(va2, va2, 1);
196 va3 = vextq_f32(va3, va3, 1);
197
198 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
199 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
200
201 const float32x4_t vmska0x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
202 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c1, vb0123c1);
203 const float32x4_t vmska1x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
204 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c1, vb0123c1);
205 const float32x4_t vmska2x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
206 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c1, vb0123c1);
207 const float32x4_t vmska3x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
208 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c1, vb0123c1);
209 const float32x4_t vmska0x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
210 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c1, vb4567c1);
211 const float32x4_t vmska1x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
212 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c1, vb4567c1);
213 const float32x4_t vmska2x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
214 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c1, vb4567c1);
215 const float32x4_t vmska3x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
216 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c1, vb4567c1);
217
218 va0 = vextq_f32(va0, va0, 1);
219 va1 = vextq_f32(va1, va1, 1);
220 va2 = vextq_f32(va2, va2, 1);
221 va3 = vextq_f32(va3, va3, 1);
222
223 const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
224 const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
225
226 const float32x4_t vmska0x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
227 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c2, vb0123c2);
228 const float32x4_t vmska1x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
229 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c2, vb0123c2);
230 const float32x4_t vmska2x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
231 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c2, vb0123c2);
232 const float32x4_t vmska3x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
233 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c2, vb0123c2);
234 const float32x4_t vmska0x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
235 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c2, vb4567c2);
236 const float32x4_t vmska1x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
237 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c2, vb4567c2);
238 const float32x4_t vmska2x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
239 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c2, vb4567c2);
240 const float32x4_t vmska3x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
241 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c2, vb4567c2);
242
243 va0 = vextq_f32(va0, va0, 1);
244 va1 = vextq_f32(va1, va1, 1);
245 va2 = vextq_f32(va2, va2, 1);
246 va3 = vextq_f32(va3, va3, 1);
247
248 const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
249 const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
250
251 const float32x4_t vmska0x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
252 vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c3, vb0123c3);
253 const float32x4_t vmska1x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
254 vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c3, vb0123c3);
255 const float32x4_t vmska2x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
256 vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c3, vb0123c3);
257 const float32x4_t vmska3x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
258 vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c3, vb0123c3);
259 const float32x4_t vmska0x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
260 vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c3, vb4567c3);
261 const float32x4_t vmska1x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
262 vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c3, vb4567c3);
263 const float32x4_t vmska2x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
264 vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c3, vb4567c3);
265 const float32x4_t vmska3x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
266 vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c3, vb4567c3);
267
268 }
269
270 p -= 4 * sizeof(void*);
271 } while (p != 0);
272
273 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
274 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
275 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
276 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
277 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
278 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
279 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
280 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
281 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
282
283 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
284 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
285 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
286 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
287 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
288 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
289 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
290 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
291 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
292
293 if XNN_LIKELY(nc >= 8) {
294 vst1q_f32(c3, vacc3x0123);
295 vst1q_f32(c3 + 4, vacc3x4567);
296 c3 = (float*) ((uintptr_t) c3 + cn_stride);
297 vst1q_f32(c2, vacc2x0123);
298 vst1q_f32(c2 + 4, vacc2x4567);
299 c2 = (float*) ((uintptr_t) c2 + cn_stride);
300 vst1q_f32(c1, vacc1x0123);
301 vst1q_f32(c1 + 4, vacc1x4567);
302 c1 = (float*) ((uintptr_t) c1 + cn_stride);
303 vst1q_f32(c0, vacc0x0123);
304 vst1q_f32(c0 + 4, vacc0x4567);
305 c0 = (float*) ((uintptr_t) c0 + cn_stride);
306
307 a = (const float**restrict) ((uintptr_t) a - ks);
308 nc -= 8;
309 } else {
310 if (nc & 4) {
311 vst1q_f32(c3, vacc3x0123); c3 += 4;
312 vst1q_f32(c2, vacc2x0123); c2 += 4;
313 vst1q_f32(c1, vacc1x0123); c1 += 4;
314 vst1q_f32(c0, vacc0x0123); c0 += 4;
315
316 vacc3x0123 = vacc3x4567;
317 vacc2x0123 = vacc2x4567;
318 vacc1x0123 = vacc1x4567;
319 vacc0x0123 = vacc0x4567;
320 }
321 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
322 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
323 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
324 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
325 if (nc & 2) {
326 vst1_f32(c3, vacc3x01); c3 += 2;
327 vst1_f32(c2, vacc2x01); c2 += 2;
328 vst1_f32(c1, vacc1x01); c1 += 2;
329 vst1_f32(c0, vacc0x01); c0 += 2;
330
331 vacc3x01 = vget_high_f32(vacc3x0123);
332 vacc2x01 = vget_high_f32(vacc2x0123);
333 vacc1x01 = vget_high_f32(vacc1x0123);
334 vacc0x01 = vget_high_f32(vacc0x0123);
335 }
336 if (nc & 1) {
337 vst1_lane_f32(c3, vacc3x01, 0);
338 vst1_lane_f32(c2, vacc2x01, 0);
339 vst1_lane_f32(c1, vacc1x01, 0);
340 vst1_lane_f32(c0, vacc0x01, 0);
341 }
342
343 nc = 0;
344 }
345 } while (nc != 0);
346 }
347