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__neonfma_1x4_acc5(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__neonfma_1x4_acc5(
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       float32x4_t vo0p2 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw89AB), 0);
89 
90       float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1);
91 
92       float32x4_t vo0p4 = vmulq_lane_f32(vi3x4567, vget_high_f32(vwGHIJ), 0);
93 
94       vo0p1 = vfmaq_lane_f32(vo0p1, 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       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x3456, vget_high_f32(vw0123), 0);
103 
104       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x3456, vget_high_f32(vw4567), 1);
105 
106       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x3456, vget_low_f32(vwCDEF), 0);
107 
108       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
109 
110       vo0p1 = vfmaq_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       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x2345, vget_low_f32(vw0123), 1);
124 
125       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x2345, vget_high_f32(vw4567), 0);
126 
127       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x2345, vget_high_f32(vw89AB), 1);
128 
129       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0);
130 
131       vo0p1 = vfmaq_lane_f32(vo0p1, 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       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x5678, vget_low_f32(vw4567), 0);
140 
141       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x5678, vget_low_f32(vw89AB), 1);
142 
143       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x5678, vget_high_f32(vwCDEF), 0);
144 
145       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
146 
147       vo0p1 = vfmaq_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       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x6789, vget_low_f32(vw4567), 1);
161 
162       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x6789, vget_high_f32(vw89AB), 0);
163 
164       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x6789, vget_high_f32(vwCDEF), 1);
165 
166       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);
167 
168       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x6789, vwOP, 1);
169 
170       vo0p0 = vaddq_f32(vo0p0, vo0p1);
171       vo0p2 = vaddq_f32(vo0p2, vo0p3);
172       vo0p0 = vaddq_f32(vo0p0, vo0p2);
173       vo0p0 = vaddq_f32(vo0p0, vo0p4);
174 
175       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
176 
177       vo0 = vminq_f32(vo0, vmax);
178 
179       vst1q_f32(o0, vo0); o0 += 4;
180     }
181     // Always process the last block of 5..8 pixels.
182     if XNN_LIKELY(w > 4 * sizeof(float)) {
183       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
184 
185       float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
186       float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
187       float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
188       float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
189       float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4;
190 
191       vi0x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x89AB)));
192       vi1x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x89AB)));
193       vi2x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x89AB)));
194       vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB)));
195       vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB)));
196 
197       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);
198 
199       float32x4_t vo0p2 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw89AB), 0);
200 
201       float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1);
202 
203       float32x4_t vo0p4 = vmulq_lane_f32(vi3x4567, vget_high_f32(vwGHIJ), 0);
204 
205       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x4567, vget_high_f32(vwKLMN), 1);
206 
207       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
208       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
209       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
210       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
211       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
212 
213       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x3456, vget_high_f32(vw0123), 0);
214 
215       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x3456, vget_high_f32(vw4567), 1);
216 
217       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x3456, vget_low_f32(vwCDEF), 0);
218 
219       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
220 
221       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);
222 
223       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
224       vi0x0123 = vi0x4567;
225       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
226       vi1x0123 = vi1x4567;
227       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
228       vi2x0123 = vi2x4567;
229       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
230       vi3x0123 = vi3x4567;
231       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
232       vi4x0123 = vi4x4567;
233 
234       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x2345, vget_low_f32(vw0123), 1);
235 
236       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x2345, vget_high_f32(vw4567), 0);
237 
238       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x2345, vget_high_f32(vw89AB), 1);
239 
240       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0);
241 
242       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x2345, vget_low_f32(vwKLMN), 1);
243 
244       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1);
245       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1);
246       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1);
247       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);
248       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);
249 
250       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x5678, vget_low_f32(vw4567), 0);
251 
252       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x5678, vget_low_f32(vw89AB), 1);
253 
254       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x5678, vget_high_f32(vwCDEF), 0);
255 
256       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
257 
258       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);
259 
260       const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2);
261       vi0x4567 = vi0x89AB;
262       const float32x4_t vi1x6789 = vextq_f32(vi1x4567, vi1x89AB, 2);
263       vi1x4567 = vi1x89AB;
264       const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2);
265       vi2x4567 = vi2x89AB;
266       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);
267       vi3x4567 = vi3x89AB;
268       const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2);
269       vi4x4567 = vi4x89AB;
270 
271       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x6789, vget_low_f32(vw4567), 1);
272 
273       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x6789, vget_high_f32(vw89AB), 0);
274 
275       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x6789, vget_high_f32(vwCDEF), 1);
276 
277       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);
278 
279       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x6789, vwOP, 1);
280 
281       vo0p0 = vaddq_f32(vo0p0, vo0p1);
282       vo0p2 = vaddq_f32(vo0p2, vo0p3);
283       vo0p0 = vaddq_f32(vo0p0, vo0p2);
284       vo0p0 = vaddq_f32(vo0p0, vo0p4);
285 
286       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
287 
288       vo0 = vminq_f32(vo0, vmax);
289 
290       vst1q_f32(o0, vo0); o0 += 4;
291 
292       w -= 4 * sizeof(float);
293     }
294     assert(w >= 1 * sizeof(float));
295     assert(w <= 4 * sizeof(float));
296     {
297       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
298 
299       vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567)));
300       vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567)));
301       vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567)));
302       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));
303       vi4x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x4567)));
304 
305       float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1);
306 
307       float32x4_t vo0p2 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw89AB), 0);
308 
309       float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1);
310 
311       float32x4_t vo0p4 = vmulq_lane_f32(vi3x4567, vget_high_f32(vwGHIJ), 0);
312 
313       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x4567, vget_high_f32(vwKLMN), 1);
314 
315       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
316       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
317       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
318       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
319       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
320 
321       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x3456, vget_high_f32(vw0123), 0);
322 
323       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x3456, vget_high_f32(vw4567), 1);
324 
325       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x3456, vget_low_f32(vwCDEF), 0);
326 
327       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
328 
329       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x3456, vget_high_f32(vwKLMN), 0);
330 
331       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
332       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
333       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
334       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
335       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
336 
337       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x2345, vget_low_f32(vw0123), 1);
338 
339       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x2345, vget_high_f32(vw4567), 0);
340 
341       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x2345, vget_high_f32(vw89AB), 1);
342 
343       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0);
344 
345       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x2345, vget_low_f32(vwKLMN), 1);
346 
347       const float32x4_t vzero = vmovq_n_f32(0.0f);
348       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1);
349       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vzero, 1);
350       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vzero, 1);
351       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1);
352       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1);
353 
354       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x5678, vget_low_f32(vw4567), 0);
355 
356       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x5678, vget_low_f32(vw89AB), 1);
357 
358       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x5678, vget_high_f32(vwCDEF), 0);
359 
360       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
361 
362       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x5678, vwOP, 0);
363 
364       const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1);
365       const float32x4_t vi1x6789 = vextq_f32(vi1x5678, vzero, 1);
366       const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1);
367       const float32x4_t vi3x6789 = vextq_f32(vi3x5678, vzero, 1);
368       const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1);
369 
370       vo0p2 = vfmaq_lane_f32(vo0p2, vi0x6789, vget_low_f32(vw4567), 1);
371 
372       vo0p3 = vfmaq_lane_f32(vo0p3, vi1x6789, vget_high_f32(vw89AB), 0);
373 
374       vo0p4 = vfmaq_lane_f32(vo0p4, vi2x6789, vget_high_f32(vwCDEF), 1);
375 
376       vo0p0 = vfmaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);
377 
378       vo0p1 = vfmaq_lane_f32(vo0p1, vi4x6789, vwOP, 1);
379 
380       vo0p0 = vaddq_f32(vo0p0, vo0p1);
381       vo0p2 = vaddq_f32(vo0p2, vo0p3);
382       vo0p0 = vaddq_f32(vo0p0, vo0p2);
383       vo0p0 = vaddq_f32(vo0p0, vo0p4);
384 
385       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
386 
387       vo0 = vminq_f32(vo0, vmax);
388 
389       if XNN_LIKELY(w & (4 * sizeof(float))) {
390         vst1q_f32(o0, vo0); o0 += 4;
391       } else {
392         float32x2_t vo0_lo = vget_low_f32(vo0);
393         if (w & (2 * sizeof(float))) {
394           vst1_f32(o0, vo0_lo); o0 += 2;
395 
396           vo0_lo = vget_high_f32(vo0);
397         }
398         if (w & (1 * sizeof(float))) {
399           vst1_lane_f32(o0, vo0_lo, 0); o0 += 1;
400         }
401       }
402     }
403 
404     i0 = (const float*) ((uintptr_t) i1 - input_decrement);
405     i1 = (const float*) ((uintptr_t) i2 - input_decrement);
406     i2 = (const float*) ((uintptr_t) i1 + input_width);
407     i3 = (const float*) ((uintptr_t) i2 + input_width);
408     i4 = (const float*) ((uintptr_t) i3 + input_width);
409 
410 
411   } while (--output_height != 0);
412 }
413