xref: /aosp_15_r20/external/XNNPACK/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/3x3s2p1-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_3x3s2p1__neon_4x4(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_3x3s2p1__neon_4x4(
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)]) XNN_OOB_READS
27 {
28   assert(input_height != 0);
29   assert(input_width != 0);
30   assert(input_width % sizeof(float) == 0);
31   assert(padding_top >= 0);
32   assert(padding_top <= 1);
33 
34   const uint32x4_t vmask_even = vld1q_u32(params->neon.mask_even);
35   const uint32x4_t vmask_odd  = vld1q_u32(params->neon.mask_odd);
36   const float32x4_t vmax = vld1q_dup_f32(&params->neon.max);
37   const float32x4_t vmin = vld1q_dup_f32(&params->neon.min);
38 
39   const float32x4_t vw0123 = vld1q_f32(weights);
40   const float32x4_t vw4567 = vld1q_f32(weights + 4);
41   const float32x2_t vw89 = vld1_f32(weights + 8);
42 
43   const size_t input_decrement = round_down_po2(input_width, 4 /* SIMD output width */ * 2 /* subsampling */ * sizeof(float));
44   const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
45 
46   const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
47   const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
48   if XNN_UNPREDICTABLE(padding_top != 0) {
49     i0 = zero;
50   }
51   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
52   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
53   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
54   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
55   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
56   const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
57   const float* i8 = (const float*) ((uintptr_t) i7 + input_width);
58 
59   float* o0 = output;
60   float* o1 = (float*) ((uintptr_t) o0 + output_width);
61   float* o2 = (float*) ((uintptr_t) o1 + output_width);
62   float* o3 = (float*) ((uintptr_t) o2 + output_width);
63 
64   size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
65   size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
66   do {
67     if XNN_UNPREDICTABLE(padded_input_height < 4) {
68       i2 = zero;
69     }
70     if XNN_UNPREDICTABLE(padded_input_height < 5) {
71       i3 = zero;
72       o1 = o0;
73     }
74     if XNN_UNPREDICTABLE(padded_input_height < 6) {
75       i4 = zero;
76     }
77     if XNN_UNPREDICTABLE(padded_input_height < 7) {
78       i5 = zero;
79       o2 = o1;
80     }
81     if XNN_UNPREDICTABLE(padded_input_height < 8) {
82       i6 = zero;
83     }
84     if XNN_UNPREDICTABLE(padded_input_height < 9) {
85       i7 = zero;
86       o3 = o2;
87     }
88     if XNN_UNPREDICTABLE(padded_input_height < 10) {
89       i8 = zero;
90     }
91 
92     float32x4_t vi0x1357 = vmovq_n_f32(0.0f);
93     float32x4_t vi1x1357 = vmovq_n_f32(0.0f);
94     float32x4_t vi2x1357 = vmovq_n_f32(0.0f);
95     float32x4_t vi3x1357 = vmovq_n_f32(0.0f);
96     float32x4_t vi4x1357 = vmovq_n_f32(0.0f);
97     float32x4_t vi5x1357 = vmovq_n_f32(0.0f);
98     float32x4_t vi6x1357 = vmovq_n_f32(0.0f);
99     float32x4_t vi7x1357 = vmovq_n_f32(0.0f);
100     float32x4_t vi8x1357 = vmovq_n_f32(0.0f);
101 
102     size_t w = input_width;
103     for (; w >= 8 * sizeof(float); w -= 8 * sizeof(float)) {
104       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
105       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
106       float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
107       float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
108 
109       const float32x4x2_t vi0x8ACE9BDF = vld2q_f32(i0); i0 += 8;
110       const float32x4x2_t vi1x8ACE9BDF = vld2q_f32(i1); i1 += 8;
111       const float32x4x2_t vi2x8ACE9BDF = vld2q_f32(i2); i2 += 8;
112       const float32x4x2_t vi3x8ACE9BDF = vld2q_f32(i3); i3 += 8;
113       const float32x4x2_t vi4x8ACE9BDF = vld2q_f32(i4); i4 += 8;
114       const float32x4x2_t vi5x8ACE9BDF = vld2q_f32(i5); i5 += 8;
115       const float32x4x2_t vi6x8ACE9BDF = vld2q_f32(i6); i6 += 8;
116       const float32x4x2_t vi7x8ACE9BDF = vld2q_f32(i7); i7 += 8;
117       const float32x4x2_t vi8x8ACE9BDF = vld2q_f32(i8); i8 += 8;
118 
119       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x8ACE9BDF.val[0], vget_high_f32(vw0123), 0);
120       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x8ACE9BDF.val[0], vget_high_f32(vw0123), 0);
121       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x8ACE9BDF.val[0], vget_high_f32(vw0123), 0);
122       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x8ACE9BDF.val[0], vget_high_f32(vw0123), 0);
123 
124       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x8ACE9BDF.val[0], vget_low_f32(vw4567), 1);
125       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x8ACE9BDF.val[0], vget_low_f32(vw4567), 1);
126       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x8ACE9BDF.val[0], vget_low_f32(vw4567), 1);
127       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x8ACE9BDF.val[0], vget_low_f32(vw4567), 1);
128 
129       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x8ACE9BDF.val[0], vw89, 0);
130       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x8ACE9BDF.val[0], vw89, 0);
131       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x8ACE9BDF.val[0], vw89, 0);
132       vo3p0 = vmlaq_lane_f32(vo3p0, vi8x8ACE9BDF.val[0], vw89, 0);
133 
134       const float32x4_t vi0x7BDF = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3);
135       vi0x1357 = vi0x8ACE9BDF.val[1];
136       const float32x4_t vi1x7BDF = vextq_f32(vi1x1357, vi1x8ACE9BDF.val[1], 3);
137       vi1x1357 = vi1x8ACE9BDF.val[1];
138       const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3);
139       vi2x1357 = vi2x8ACE9BDF.val[1];
140       const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3);
141       vi3x1357 = vi3x8ACE9BDF.val[1];
142       const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3);
143       vi4x1357 = vi4x8ACE9BDF.val[1];
144       const float32x4_t vi5x7BDF = vextq_f32(vi5x1357, vi5x8ACE9BDF.val[1], 3);
145       vi5x1357 = vi5x8ACE9BDF.val[1];
146       const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3);
147       vi6x1357 = vi6x8ACE9BDF.val[1];
148       const float32x4_t vi7x7BDF = vextq_f32(vi7x1357, vi7x8ACE9BDF.val[1], 3);
149       vi7x1357 = vi7x8ACE9BDF.val[1];
150       const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x8ACE9BDF.val[1], 3);
151       vi8x1357 = vi8x8ACE9BDF.val[1];
152 
153       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x7BDF, vget_low_f32(vw0123), 1);
154       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x7BDF, vget_low_f32(vw0123), 1);
155       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1);
156       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x7BDF, vget_low_f32(vw0123), 1);
157 
158       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x7BDF, vget_low_f32(vw4567), 0);
159       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0);
160       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x7BDF, vget_low_f32(vw4567), 0);
161       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x7BDF, vget_low_f32(vw4567), 0);
162 
163       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1);
164       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1);
165       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x7BDF, vget_high_f32(vw4567), 1);
166       vo3p0 = vmlaq_lane_f32(vo3p0, vi8x7BDF, vget_high_f32(vw4567), 1);
167 
168       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x8ACE9BDF.val[1], vget_high_f32(vw0123), 1);
169       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x8ACE9BDF.val[1], vget_high_f32(vw0123), 1);
170       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x8ACE9BDF.val[1], vget_high_f32(vw0123), 1);
171       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x8ACE9BDF.val[1], vget_high_f32(vw0123), 1);
172 
173       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x8ACE9BDF.val[1], vget_high_f32(vw4567), 0);
174       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x8ACE9BDF.val[1], vget_high_f32(vw4567), 0);
175       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x8ACE9BDF.val[1], vget_high_f32(vw4567), 0);
176       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x8ACE9BDF.val[1], vget_high_f32(vw4567), 0);
177 
178       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x8ACE9BDF.val[1], vw89, 1);
179       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x8ACE9BDF.val[1], vw89, 1);
180       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x8ACE9BDF.val[1], vw89, 1);
181       vo3p0 = vmlaq_lane_f32(vo3p0, vi8x8ACE9BDF.val[1], vw89, 1);
182 
183 
184       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
185       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
186       float32x4_t vo2 = vmaxq_f32(vo2p0, vmin);
187       float32x4_t vo3 = vmaxq_f32(vo3p0, vmin);
188 
189       vo0 = vminq_f32(vo0, vmax);
190       vo1 = vminq_f32(vo1, vmax);
191       vo2 = vminq_f32(vo2, vmax);
192       vo3 = vminq_f32(vo3, vmax);
193 
194       vst1q_f32(o3, vo3); o3 += 4;
195       vst1q_f32(o2, vo2); o2 += 4;
196       vst1q_f32(o1, vo1); o1 += 4;
197       vst1q_f32(o0, vo0); o0 += 4;
198     }
199     // Last block has 0-7 pixels to process.
200     assert(w < 8 * sizeof(float));
201     if XNN_LIKELY(w != 0) {
202       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
203       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
204       float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
205       float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
206 
207       const float32x4x2_t vi0x8ACE9BDF = vld2q_f32(i0);
208       const float32x4x2_t vi1x8ACE9BDF = vld2q_f32(i1);
209       const float32x4x2_t vi2x8ACE9BDF = vld2q_f32(i2);
210       const float32x4x2_t vi3x8ACE9BDF = vld2q_f32(i3);
211       const float32x4x2_t vi4x8ACE9BDF = vld2q_f32(i4);
212       const float32x4x2_t vi5x8ACE9BDF = vld2q_f32(i5);
213       const float32x4x2_t vi6x8ACE9BDF = vld2q_f32(i6);
214       const float32x4x2_t vi7x8ACE9BDF = vld2q_f32(i7);
215       const float32x4x2_t vi8x8ACE9BDF = vld2q_f32(i8);
216 
217       const float32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9BDF.val[0])));
218       const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi0x8ACE9BDF.val[1])));
219       const float32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9BDF.val[0])));
220       const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi1x8ACE9BDF.val[1])));
221       const float32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9BDF.val[0])));
222       const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi2x8ACE9BDF.val[1])));
223       const float32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9BDF.val[0])));
224       const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi3x8ACE9BDF.val[1])));
225       const float32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9BDF.val[0])));
226       const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi4x8ACE9BDF.val[1])));
227       const float32x4_t vi5x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi5x8ACE9BDF.val[0])));
228       const float32x4_t vi5x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi5x8ACE9BDF.val[1])));
229       const float32x4_t vi6x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi6x8ACE9BDF.val[0])));
230       const float32x4_t vi6x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi6x8ACE9BDF.val[1])));
231       const float32x4_t vi7x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi7x8ACE9BDF.val[0])));
232       const float32x4_t vi7x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi7x8ACE9BDF.val[1])));
233       const float32x4_t vi8x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi8x8ACE9BDF.val[0])));
234       const float32x4_t vi8x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi8x8ACE9BDF.val[1])));
235 
236       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x8ACE, vget_high_f32(vw0123), 0);
237       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x8ACE, vget_high_f32(vw0123), 0);
238       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x8ACE, vget_high_f32(vw0123), 0);
239       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x8ACE, vget_high_f32(vw0123), 0);
240 
241       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x8ACE, vget_low_f32(vw4567), 1);
242       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x8ACE, vget_low_f32(vw4567), 1);
243       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x8ACE, vget_low_f32(vw4567), 1);
244       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x8ACE, vget_low_f32(vw4567), 1);
245 
246       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x8ACE, vw89, 0);
247       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x8ACE, vw89, 0);
248       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x8ACE, vw89, 0);
249       vo3p0 = vmlaq_lane_f32(vo3p0, vi8x8ACE, vw89, 0);
250 
251       const float32x4_t vi0x7BDF = vextq_f32(vi0x1357, vi0x9BDF, 3);
252       const float32x4_t vi1x7BDF = vextq_f32(vi1x1357, vi1x9BDF, 3);
253       const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3);
254       const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3);
255       const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3);
256       const float32x4_t vi5x7BDF = vextq_f32(vi5x1357, vi5x9BDF, 3);
257       const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3);
258       const float32x4_t vi7x7BDF = vextq_f32(vi7x1357, vi7x9BDF, 3);
259       const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x9BDF, 3);
260 
261       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x7BDF, vget_low_f32(vw0123), 1);
262       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x7BDF, vget_low_f32(vw0123), 1);
263       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1);
264       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x7BDF, vget_low_f32(vw0123), 1);
265 
266       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x7BDF, vget_low_f32(vw4567), 0);
267       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0);
268       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x7BDF, vget_low_f32(vw4567), 0);
269       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x7BDF, vget_low_f32(vw4567), 0);
270 
271       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1);
272       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1);
273       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x7BDF, vget_high_f32(vw4567), 1);
274       vo3p0 = vmlaq_lane_f32(vo3p0, vi8x7BDF, vget_high_f32(vw4567), 1);
275 
276       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x9BDF, vget_high_f32(vw0123), 1);
277       vo1p0 = vmlaq_lane_f32(vo1p0, vi2x9BDF, vget_high_f32(vw0123), 1);
278       vo2p0 = vmlaq_lane_f32(vo2p0, vi4x9BDF, vget_high_f32(vw0123), 1);
279       vo3p0 = vmlaq_lane_f32(vo3p0, vi6x9BDF, vget_high_f32(vw0123), 1);
280 
281       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x9BDF, vget_high_f32(vw4567), 0);
282       vo1p0 = vmlaq_lane_f32(vo1p0, vi3x9BDF, vget_high_f32(vw4567), 0);
283       vo2p0 = vmlaq_lane_f32(vo2p0, vi5x9BDF, vget_high_f32(vw4567), 0);
284       vo3p0 = vmlaq_lane_f32(vo3p0, vi7x9BDF, vget_high_f32(vw4567), 0);
285 
286       vo0p0 = vmlaq_lane_f32(vo0p0, vi2x9BDF, vw89, 1);
287       vo1p0 = vmlaq_lane_f32(vo1p0, vi4x9BDF, vw89, 1);
288       vo2p0 = vmlaq_lane_f32(vo2p0, vi6x9BDF, vw89, 1);
289       vo3p0 = vmlaq_lane_f32(vo3p0, vi8x9BDF, vw89, 1);
290 
291 
292       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
293       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
294       float32x4_t vo2 = vmaxq_f32(vo2p0, vmin);
295       float32x4_t vo3 = vmaxq_f32(vo3p0, vmin);
296 
297       vo0 = vminq_f32(vo0, vmax);
298       vo1 = vminq_f32(vo1, vmax);
299       vo2 = vminq_f32(vo2, vmax);
300       vo3 = vminq_f32(vo3, vmax);
301 
302       w += 1 * sizeof(float);
303       if (w & (8 * sizeof(float))) {
304         vst1q_f32(o3, vo3); o3 += 4;
305         vst1q_f32(o2, vo2); o2 += 4;
306         vst1q_f32(o1, vo1); o1 += 4;
307         vst1q_f32(o0, vo0); o0 += 4;
308       } else {
309         float32x2_t vo0_lo = vget_low_f32(vo0);
310         float32x2_t vo1_lo = vget_low_f32(vo1);
311         float32x2_t vo2_lo = vget_low_f32(vo2);
312         float32x2_t vo3_lo = vget_low_f32(vo3);
313         if (w & (4 * sizeof(float))) {
314           vst1_f32(o3, vo3_lo); o3 += 2;
315           vst1_f32(o2, vo2_lo); o2 += 2;
316           vst1_f32(o1, vo1_lo); o1 += 2;
317           vst1_f32(o0, vo0_lo); o0 += 2;
318 
319           vo0_lo = vget_high_f32(vo0);
320           vo1_lo = vget_high_f32(vo1);
321           vo2_lo = vget_high_f32(vo2);
322           vo3_lo = vget_high_f32(vo3);
323         }
324         if (w & (2 * sizeof(float))) {
325           vst1_lane_f32(o3, vo3_lo, 0); o3 += 1;
326           vst1_lane_f32(o2, vo2_lo, 0); o2 += 1;
327           vst1_lane_f32(o1, vo1_lo, 0); o1 += 1;
328           vst1_lane_f32(o0, vo0_lo, 0); o0 += 1;
329         }
330       }
331     }
332 
333     i0 = (const float*) ((uintptr_t) i8 - input_decrement);
334     i1 = (const float*) ((uintptr_t) i0 + input_width);
335     i2 = (const float*) ((uintptr_t) i1 + input_width);
336     i3 = (const float*) ((uintptr_t) i2 + input_width);
337     i4 = (const float*) ((uintptr_t) i3 + input_width);
338     i5 = (const float*) ((uintptr_t) i4 + input_width);
339     i6 = (const float*) ((uintptr_t) i5 + input_width);
340     i7 = (const float*) ((uintptr_t) i6 + input_width);
341     i8 = (const float*) ((uintptr_t) i7 + input_width);
342 
343     o0 = o3;
344     o1 = (float*) ((uintptr_t) o0 + output_width);
345     o2 = (float*) ((uintptr_t) o1 + output_width);
346     o3 = (float*) ((uintptr_t) o2 + output_width);
347 
348     output_height = doz(output_height, 4);
349     padded_input_height = doz(padded_input_height, 8);
350   } while (output_height != 0);
351 }
352