1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/5x5p2-neon.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/dwconv.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2(
19     size_t input_height,
20     size_t input_width,
21     const float* input,
22     const float* weights,
23     const float* zero,
24     float* output,
25     uint32_t padding_top,
26     const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28   assert(input_height != 0);
29   assert(input_width != 0);
30   assert(input_width % sizeof(float) == 0);
31   assert(padding_top == 2);
32 
33   const uint32x4_t vmask = vld1q_u32(params->neon.mask);
34   const float32x4_t vmax = vld1q_dup_f32(&params->neon.max);
35   const float32x4_t vmin = vld1q_dup_f32(&params->neon.min);
36 
37   const float32x4_t vw0123 = vld1q_f32(weights);
38   const float32x4_t vw4567 = vld1q_f32(weights + 4);
39   const float32x4_t vw89AB = vld1q_f32(weights + 8);
40   const float32x4_t vwCDEF = vld1q_f32(weights + 12);
41   const float32x4_t vwGHIJ = vld1q_f32(weights + 16);
42   const float32x4_t vwKLMN = vld1q_f32(weights + 20);
43   const float32x2_t vwOP = vld1_f32(weights + 24);
44 
45   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
46 
47   const float* i0 = zero;
48   const float* i1 = zero;
49   const float* i2 = input;
50   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
51   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
52 
53   float* o0 = output;
54 
55   size_t output_height = input_height;
56   do {
57     if XNN_UNPREDICTABLE(output_height < 2) {
58       i3 = zero;
59     }
60     if XNN_UNPREDICTABLE(output_height < 3) {
61       i4 = zero;
62     }
63 
64     float32x4_t vi0x0123 = vmovq_n_f32(0.0f);
65     float32x4_t vi1x0123 = vmovq_n_f32(0.0f);
66     float32x4_t vi2x0123 = vmovq_n_f32(0.0f);
67     float32x4_t vi3x0123 = vmovq_n_f32(0.0f);
68     float32x4_t vi4x0123 = vmovq_n_f32(0.0f);
69 
70     float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
71     float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
72     float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
73     float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
74     float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;
75 
76     size_t w = input_width;
77     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
78       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
79 
80       const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
81       const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
82       const float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
83       const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
84       const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4;
85 
86       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);
87 
88       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0);
89 
90       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1);
91 
92       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0);
93 
94       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);
95 
96       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
97       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
98       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
99       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
100       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
101 
102       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0);
103 
104       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1);
105 
106       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x3456, vget_low_f32(vwCDEF), 0);
107 
108       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
109 
110       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);
111 
112       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
113       vi0x0123 = vi0x4567;
114       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
115       vi1x0123 = vi1x4567;
116       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
117       vi2x0123 = vi2x4567;
118       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
119       vi3x0123 = vi3x4567;
120       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
121       vi4x0123 = vi4x4567;
122 
123       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1);
124 
125       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x2345, vget_high_f32(vw4567), 0);
126 
127       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1);
128 
129       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x2345, vget_low_f32(vwGHIJ), 0);
130 
131       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1);
132 
133       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1);
134       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1);
135       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1);
136       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);
137       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);
138 
139       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x5678, vget_low_f32(vw4567), 0);
140 
141       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1);
142 
143       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x5678, vget_high_f32(vwCDEF), 0);
144 
145       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
146 
147       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);
148 
149       const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2);
150       vi0x4567 = vi0x89AB;
151       const float32x4_t vi1x6789 = vextq_f32(vi1x4567, vi1x89AB, 2);
152       vi1x4567 = vi1x89AB;
153       const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2);
154       vi2x4567 = vi2x89AB;
155       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);
156       vi3x4567 = vi3x89AB;
157       const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2);
158       vi4x4567 = vi4x89AB;
159 
160       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1);
161 
162       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x6789, vget_high_f32(vw89AB), 0);
163 
164       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1);
165 
166       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x6789, vget_low_f32(vwKLMN), 0);
167 
168       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1);
169 
170       vo0p0 = vaddq_f32(vo0p0, vo0p1);
171 
172       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
173 
174       vo0 = vminq_f32(vo0, vmax);
175 
176       vst1q_f32(o0, vo0); o0 += 4;
177     }
178     // Always process the last block of 5..8 pixels.
179     if XNN_LIKELY(w > 4 * sizeof(float)) {
180       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
181 
182       float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
183       float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
184       float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
185       float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
186       float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4;
187 
188       vi0x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x89AB)));
189       vi1x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x89AB)));
190       vi2x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x89AB)));
191       vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB)));
192       vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB)));
193 
194       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);
195 
196       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0);
197 
198       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1);
199 
200       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0);
201 
202       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);
203 
204       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
205       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
206       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
207       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
208       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
209 
210       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0);
211 
212       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1);
213 
214       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x3456, vget_low_f32(vwCDEF), 0);
215 
216       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
217 
218       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);
219 
220       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
221       vi0x0123 = vi0x4567;
222       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
223       vi1x0123 = vi1x4567;
224       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
225       vi2x0123 = vi2x4567;
226       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
227       vi3x0123 = vi3x4567;
228       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
229       vi4x0123 = vi4x4567;
230 
231       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1);
232 
233       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x2345, vget_high_f32(vw4567), 0);
234 
235       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1);
236 
237       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x2345, vget_low_f32(vwGHIJ), 0);
238 
239       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1);
240 
241       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1);
242       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1);
243       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1);
244       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);
245       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);
246 
247       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x5678, vget_low_f32(vw4567), 0);
248 
249       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1);
250 
251       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x5678, vget_high_f32(vwCDEF), 0);
252 
253       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
254 
255       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);
256 
257       const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2);
258       vi0x4567 = vi0x89AB;
259       const float32x4_t vi1x6789 = vextq_f32(vi1x4567, vi1x89AB, 2);
260       vi1x4567 = vi1x89AB;
261       const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2);
262       vi2x4567 = vi2x89AB;
263       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);
264       vi3x4567 = vi3x89AB;
265       const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2);
266       vi4x4567 = vi4x89AB;
267 
268       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1);
269 
270       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x6789, vget_high_f32(vw89AB), 0);
271 
272       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1);
273 
274       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x6789, vget_low_f32(vwKLMN), 0);
275 
276       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1);
277 
278       vo0p0 = vaddq_f32(vo0p0, vo0p1);
279 
280       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
281 
282       vo0 = vminq_f32(vo0, vmax);
283 
284       vst1q_f32(o0, vo0); o0 += 4;
285 
286       w -= 4 * sizeof(float);
287     }
288     assert(w >= 1 * sizeof(float));
289     assert(w <= 4 * sizeof(float));
290     {
291       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
292 
293       vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567)));
294       vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567)));
295       vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567)));
296       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));
297       vi4x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x4567)));
298 
299       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);
300 
301       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0);
302 
303       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1);
304 
305       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0);
306 
307       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);
308 
309       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
310       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
311       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
312       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
313       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
314 
315       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0);
316 
317       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1);
318 
319       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x3456, vget_low_f32(vwCDEF), 0);
320 
321       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
322 
323       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);
324 
325       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
326       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
327       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
328       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
329       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
330 
331       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1);
332 
333       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x2345, vget_high_f32(vw4567), 0);
334 
335       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1);
336 
337       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x2345, vget_low_f32(vwGHIJ), 0);
338 
339       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1);
340 
341       const float32x4_t vzero = vmovq_n_f32(0.0f);
342       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1);
343       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vzero, 1);
344       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vzero, 1);
345       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1);
346       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1);
347 
348       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x5678, vget_low_f32(vw4567), 0);
349 
350       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1);
351 
352       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x5678, vget_high_f32(vwCDEF), 0);
353 
354       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
355 
356       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);
357 
358       const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1);
359       const float32x4_t vi1x6789 = vextq_f32(vi1x5678, vzero, 1);
360       const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1);
361       const float32x4_t vi3x6789 = vextq_f32(vi3x5678, vzero, 1);
362       const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1);
363 
364       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1);
365 
366       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x6789, vget_high_f32(vw89AB), 0);
367 
368       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1);
369 
370       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x6789, vget_low_f32(vwKLMN), 0);
371 
372       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1);
373 
374       vo0p0 = vaddq_f32(vo0p0, vo0p1);
375 
376       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
377 
378       vo0 = vminq_f32(vo0, vmax);
379 
380       if XNN_LIKELY(w & (4 * sizeof(float))) {
381         vst1q_f32(o0, vo0); o0 += 4;
382       } else {
383         float32x2_t vo0_lo = vget_low_f32(vo0);
384         if (w & (2 * sizeof(float))) {
385           vst1_f32(o0, vo0_lo); o0 += 2;
386 
387           vo0_lo = vget_high_f32(vo0);
388         }
389         if (w & (1 * sizeof(float))) {
390           vst1_lane_f32(o0, vo0_lo, 0); o0 += 1;
391         }
392       }
393     }
394 
395     i0 = (const float*) ((uintptr_t) i1 - input_decrement);
396     i1 = (const float*) ((uintptr_t) i2 - input_decrement);
397     i2 = (const float*) ((uintptr_t) i1 + input_width);
398     i3 = (const float*) ((uintptr_t) i2 + input_width);
399     i4 = (const float*) ((uintptr_t) i3 + input_width);
400 
401 
402   } while (--output_height != 0);
403 }
404