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_2x4(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_2x4(
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   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
53 
54   float* o0 = output;
55   float* o1 = (float*) ((uintptr_t) o0 + input_width);
56 
57   size_t output_height = input_height;
58   do {
59     if XNN_UNPREDICTABLE(output_height < 2) {
60       i3 = zero;
61       o1 = o0;
62     }
63     if XNN_UNPREDICTABLE(output_height < 3) {
64       i4 = zero;
65     }
66     if XNN_UNPREDICTABLE(output_height < 4) {
67       i5 = zero;
68     }
69 
70     float32x4_t vi0x0123 = vmovq_n_f32(0.0f);
71     float32x4_t vi1x0123 = vmovq_n_f32(0.0f);
72     float32x4_t vi2x0123 = vmovq_n_f32(0.0f);
73     float32x4_t vi3x0123 = vmovq_n_f32(0.0f);
74     float32x4_t vi4x0123 = vmovq_n_f32(0.0f);
75     float32x4_t vi5x0123 = vmovq_n_f32(0.0f);
76 
77     float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
78     float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
79     float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
80     float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
81     float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;
82     float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4;
83 
84     size_t w = input_width;
85     for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
86       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
87       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
88 
89       const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
90       const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
91       const float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
92       const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
93       const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4;
94       const float32x4_t vi5x89AB = vld1q_f32(i5); i5 += 4;
95 
96       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1);
97       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1);
98 
99       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0);
100       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0);
101 
102       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1);
103       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1);
104 
105       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0);
106       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0);
107 
108       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);
109       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1);
110 
111       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
112       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
113       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
114       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
115       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
116       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);
117 
118       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0);
119       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x3456, vget_high_f32(vw0123), 0);
120 
121       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1);
122       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1);
123 
124       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0);
125       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0);
126 
127       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
128       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1);
129 
130       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0);
131       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0);
132 
133       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
134       vi0x0123 = vi0x4567;
135       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
136       vi1x0123 = vi1x4567;
137       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
138       vi2x0123 = vi2x4567;
139       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
140       vi3x0123 = vi3x4567;
141       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
142       vi4x0123 = vi4x4567;
143       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);
144       vi5x0123 = vi5x4567;
145 
146       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1);
147       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1);
148 
149       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0);
150       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0);
151 
152       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1);
153       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x2345, vget_high_f32(vw89AB), 1);
154 
155       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0);
156       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0);
157 
158       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1);
159       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1);
160 
161       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1);
162       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1);
163       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1);
164       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);
165       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);
166       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);
167 
168       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0);
169       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0);
170 
171       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1);
172       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1);
173 
174       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0);
175       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vget_high_f32(vwCDEF), 0);
176 
177       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
178       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x5678, vget_high_f32(vwGHIJ), 1);
179 
180       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0);
181       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0);
182 
183       const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2);
184       vi0x4567 = vi0x89AB;
185       const float32x4_t vi1x6789 = vextq_f32(vi1x4567, vi1x89AB, 2);
186       vi1x4567 = vi1x89AB;
187       const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2);
188       vi2x4567 = vi2x89AB;
189       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);
190       vi3x4567 = vi3x89AB;
191       const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2);
192       vi4x4567 = vi4x89AB;
193       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);
194       vi5x4567 = vi5x89AB;
195 
196       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1);
197       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1);
198 
199       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0);
200       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0);
201 
202       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1);
203       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x6789, vget_high_f32(vwCDEF), 1);
204 
205       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);
206       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0);
207 
208       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1);
209       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1);
210 
211 
212       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
213       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
214 
215       vo0 = vminq_f32(vo0, vmax);
216       vo1 = vminq_f32(vo1, vmax);
217 
218       vst1q_f32(o1, vo1); o1 += 4;
219       vst1q_f32(o0, vo0); o0 += 4;
220     }
221     // Always process the last block of 5..8 pixels.
222     if XNN_LIKELY(w > 4 * sizeof(float)) {
223       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
224       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
225 
226       float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
227       float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
228       float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
229       float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
230       float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4;
231       float32x4_t vi5x89AB = vld1q_f32(i5); i5 += 4;
232 
233       vi0x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x89AB)));
234       vi1x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x89AB)));
235       vi2x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x89AB)));
236       vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB)));
237       vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB)));
238       vi5x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x89AB)));
239 
240       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1);
241       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1);
242 
243       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0);
244       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0);
245 
246       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1);
247       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1);
248 
249       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0);
250       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0);
251 
252       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);
253       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1);
254 
255       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
256       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
257       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
258       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
259       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
260       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);
261 
262       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0);
263       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x3456, vget_high_f32(vw0123), 0);
264 
265       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1);
266       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1);
267 
268       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0);
269       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0);
270 
271       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
272       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1);
273 
274       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0);
275       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0);
276 
277       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
278       vi0x0123 = vi0x4567;
279       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
280       vi1x0123 = vi1x4567;
281       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
282       vi2x0123 = vi2x4567;
283       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
284       vi3x0123 = vi3x4567;
285       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
286       vi4x0123 = vi4x4567;
287       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);
288       vi5x0123 = vi5x4567;
289 
290       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1);
291       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1);
292 
293       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0);
294       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0);
295 
296       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1);
297       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x2345, vget_high_f32(vw89AB), 1);
298 
299       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0);
300       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0);
301 
302       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1);
303       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1);
304 
305       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1);
306       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1);
307       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1);
308       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);
309       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);
310       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);
311 
312       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0);
313       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0);
314 
315       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1);
316       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1);
317 
318       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0);
319       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vget_high_f32(vwCDEF), 0);
320 
321       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
322       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x5678, vget_high_f32(vwGHIJ), 1);
323 
324       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0);
325       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0);
326 
327       const float32x4_t vi0x6789 = vextq_f32(vi0x4567, vi0x89AB, 2);
328       vi0x4567 = vi0x89AB;
329       const float32x4_t vi1x6789 = vextq_f32(vi1x4567, vi1x89AB, 2);
330       vi1x4567 = vi1x89AB;
331       const float32x4_t vi2x6789 = vextq_f32(vi2x4567, vi2x89AB, 2);
332       vi2x4567 = vi2x89AB;
333       const float32x4_t vi3x6789 = vextq_f32(vi3x4567, vi3x89AB, 2);
334       vi3x4567 = vi3x89AB;
335       const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2);
336       vi4x4567 = vi4x89AB;
337       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);
338       vi5x4567 = vi5x89AB;
339 
340       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1);
341       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1);
342 
343       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0);
344       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0);
345 
346       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1);
347       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x6789, vget_high_f32(vwCDEF), 1);
348 
349       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);
350       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0);
351 
352       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1);
353       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1);
354 
355 
356       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
357       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
358 
359       vo0 = vminq_f32(vo0, vmax);
360       vo1 = vminq_f32(vo1, vmax);
361 
362       vst1q_f32(o1, vo1); o1 += 4;
363       vst1q_f32(o0, vo0); o0 += 4;
364 
365       w -= 4 * sizeof(float);
366     }
367     assert(w >= 1 * sizeof(float));
368     assert(w <= 4 * sizeof(float));
369     {
370       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
371       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
372 
373       vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567)));
374       vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567)));
375       vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567)));
376       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));
377       vi4x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x4567)));
378       vi5x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x4567)));
379 
380       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1);
381       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1);
382 
383       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0);
384       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw89AB), 0);
385 
386       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1);
387       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x4567, vget_low_f32(vwCDEF), 1);
388 
389       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0);
390       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0);
391 
392       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1);
393       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1);
394 
395       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
396       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
397       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
398       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
399       const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
400       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);
401 
402       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0);
403       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x3456, vget_high_f32(vw0123), 0);
404 
405       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1);
406       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1);
407 
408       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0);
409       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x3456, vget_low_f32(vwCDEF), 0);
410 
411       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1);
412       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x3456, vget_low_f32(vwGHIJ), 1);
413 
414       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0);
415       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0);
416 
417       const float32x4_t vi0x2345 = vextq_f32(vi0x0123, vi0x4567, 2);
418       const float32x4_t vi1x2345 = vextq_f32(vi1x0123, vi1x4567, 2);
419       const float32x4_t vi2x2345 = vextq_f32(vi2x0123, vi2x4567, 2);
420       const float32x4_t vi3x2345 = vextq_f32(vi3x0123, vi3x4567, 2);
421       const float32x4_t vi4x2345 = vextq_f32(vi4x0123, vi4x4567, 2);
422       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);
423 
424       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1);
425       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x2345, vget_low_f32(vw0123), 1);
426 
427       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0);
428       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x2345, vget_high_f32(vw4567), 0);
429 
430       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1);
431       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x2345, vget_high_f32(vw89AB), 1);
432 
433       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0);
434       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x2345, vget_low_f32(vwGHIJ), 0);
435 
436       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1);
437       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x2345, vget_low_f32(vwKLMN), 1);
438 
439       const float32x4_t vzero = vmovq_n_f32(0.0f);
440       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1);
441       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vzero, 1);
442       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vzero, 1);
443       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1);
444       const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1);
445       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1);
446 
447       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0);
448       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x5678, vget_low_f32(vw4567), 0);
449 
450       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1);
451       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x5678, vget_low_f32(vw89AB), 1);
452 
453       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0);
454       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x5678, vget_high_f32(vwCDEF), 0);
455 
456       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1);
457       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x5678, vget_high_f32(vwGHIJ), 1);
458 
459       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0);
460       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0);
461 
462       const float32x4_t vi0x6789 = vextq_f32(vi0x5678, vzero, 1);
463       const float32x4_t vi1x6789 = vextq_f32(vi1x5678, vzero, 1);
464       const float32x4_t vi2x6789 = vextq_f32(vi2x5678, vzero, 1);
465       const float32x4_t vi3x6789 = vextq_f32(vi3x5678, vzero, 1);
466       const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1);
467       const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1);
468 
469       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1);
470       vo1p0 = vmlaq_lane_f32(vo1p0, vi1x6789, vget_low_f32(vw4567), 1);
471 
472       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0);
473       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x6789, vget_high_f32(vw89AB), 0);
474 
475       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1);
476       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x6789, vget_high_f32(vwCDEF), 1);
477 
478       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0);
479       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0);
480 
481       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1);
482       vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1);
483 
484 
485       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
486       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
487 
488       vo0 = vminq_f32(vo0, vmax);
489       vo1 = vminq_f32(vo1, vmax);
490 
491       if XNN_LIKELY(w & (4 * sizeof(float))) {
492         vst1q_f32(o1, vo1); o1 += 4;
493         vst1q_f32(o0, vo0); o0 += 4;
494       } else {
495         float32x2_t vo0_lo = vget_low_f32(vo0);
496         float32x2_t vo1_lo = vget_low_f32(vo1);
497         if (w & (2 * sizeof(float))) {
498           vst1_f32(o1, vo1_lo); o1 += 2;
499           vst1_f32(o0, vo0_lo); o0 += 2;
500 
501           vo0_lo = vget_high_f32(vo0);
502           vo1_lo = vget_high_f32(vo1);
503         }
504         if (w & (1 * sizeof(float))) {
505           vst1_lane_f32(o1, vo1_lo, 0); o1 += 1;
506           vst1_lane_f32(o0, vo0_lo, 0); o0 += 1;
507         }
508       }
509     }
510 
511     i0 = (const float*) ((uintptr_t) i2 - input_decrement);
512     i1 = (const float*) ((uintptr_t) i3 - input_decrement);
513     i2 = (const float*) ((uintptr_t) i1 + input_width);
514     i3 = (const float*) ((uintptr_t) i2 + input_width);
515     i4 = (const float*) ((uintptr_t) i3 + input_width);
516     i5 = (const float*) ((uintptr_t) i4 + input_width);
517 
518     o0 = o1;
519     o1 = (float*) ((uintptr_t) o0 + input_width);
520 
521     output_height = doz(output_height, 2);
522   } while (output_height != 0);
523 }
524