xref: /aosp_15_r20/external/XNNPACK/src/f16-dwconv2d-chw/gen/3x3s2p1-minmax-neonfp16arith-4x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-dwconv2d-chw/3x3s2p1-neonfp16arith.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_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_4x4(size_t input_height,size_t input_width,const void * input,const void * weights,const void * zero,void * output,uint32_t padding_top,const union xnn_f16_chw_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_4x4(
19     size_t input_height,
20     size_t input_width,
21     const void* input,
22     const void* weights,
23     const void* zero,
24     void* output,
25     uint32_t padding_top,
26     const union xnn_f16_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(__fp16) == 0);
31   assert(padding_top >= 0);
32   assert(padding_top <= 1);
33 
34   const uint16x4_t vmask_even = vld1_u16(params->neonfp16arith.mask_even);
35   const uint16x4_t vmask_odd  = vld1_u16(params->neonfp16arith.mask_odd);
36   const float16x4_t vmax = vld1_dup_f16(&params->neonfp16arith.max);
37   const float16x4_t vmin = vld1_dup_f16(&params->neonfp16arith.min);
38 
39   const __fp16* w0 = (const __fp16*)weights;
40   const float16x8_t vw01234567 = vld1q_f16(w0);
41   const float16x4_t vw89 = vreinterpret_f16_u32(vld1_lane_u32((const void*)(w0 + 8), vmov_n_u32(0), 0));
42 
43   const size_t input_decrement = round_down_po2(input_width, 4 /* SIMD output width */ * 2 /* subsampling */ * sizeof(__fp16));
44   const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(__fp16)) / 2, sizeof(__fp16));
45 
46   const __fp16* i0 = (const __fp16*) ((uintptr_t) input - ((-padding_top) & input_width));
47   const __fp16* i1 = (const __fp16*) ((uintptr_t) i0 + input_width);
48   if XNN_UNPREDICTABLE(padding_top != 0) {
49     i0 = zero;
50   }
51   const __fp16* i2 = (const __fp16*) ((uintptr_t) i1 + input_width);
52   const __fp16* i3 = (const __fp16*) ((uintptr_t) i2 + input_width);
53   const __fp16* i4 = (const __fp16*) ((uintptr_t) i3 + input_width);
54   const __fp16* i5 = (const __fp16*) ((uintptr_t) i4 + input_width);
55   const __fp16* i6 = (const __fp16*) ((uintptr_t) i5 + input_width);
56   const __fp16* i7 = (const __fp16*) ((uintptr_t) i6 + input_width);
57   const __fp16* i8 = (const __fp16*) ((uintptr_t) i7 + input_width);
58 
59   __fp16* o0 = output;
60   __fp16* o1 = (__fp16*) ((uintptr_t) o0 + output_width);
61   __fp16* o2 = (__fp16*) ((uintptr_t) o1 + output_width);
62   __fp16* o3 = (__fp16*) ((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     float16x4_t vi0x1357 = vmov_n_f16(0);
93     float16x4_t vi1x1357 = vmov_n_f16(0);
94     float16x4_t vi2x1357 = vmov_n_f16(0);
95     float16x4_t vi3x1357 = vmov_n_f16(0);
96     float16x4_t vi4x1357 = vmov_n_f16(0);
97     float16x4_t vi5x1357 = vmov_n_f16(0);
98     float16x4_t vi6x1357 = vmov_n_f16(0);
99     float16x4_t vi7x1357 = vmov_n_f16(0);
100     float16x4_t vi8x1357 = vmov_n_f16(0);
101 
102     size_t w = input_width;
103     for (; w >= 8 * sizeof(__fp16); w -= 8 * sizeof(__fp16)) {
104       float16x4_t vo0p0 = vdup_laneq_f16(vw01234567, 0);
105       float16x4_t vo1p0 = vdup_laneq_f16(vw01234567, 0);
106       float16x4_t vo2p0 = vdup_laneq_f16(vw01234567, 0);
107       float16x4_t vo3p0 = vdup_laneq_f16(vw01234567, 0);
108 
109       const float16x4x2_t vi0x8ACE9BDF = vld2_f16(i0); i0 += 8;
110       const float16x4x2_t vi1x8ACE9BDF = vld2_f16(i1); i1 += 8;
111       const float16x4x2_t vi2x8ACE9BDF = vld2_f16(i2); i2 += 8;
112       const float16x4x2_t vi3x8ACE9BDF = vld2_f16(i3); i3 += 8;
113       const float16x4x2_t vi4x8ACE9BDF = vld2_f16(i4); i4 += 8;
114       const float16x4x2_t vi5x8ACE9BDF = vld2_f16(i5); i5 += 8;
115       const float16x4x2_t vi6x8ACE9BDF = vld2_f16(i6); i6 += 8;
116       const float16x4x2_t vi7x8ACE9BDF = vld2_f16(i7); i7 += 8;
117       const float16x4x2_t vi8x8ACE9BDF = vld2_f16(i8); i8 += 8;
118 
119       vo0p0 = vfma_laneq_f16(vo0p0, vi0x8ACE9BDF.val[0], vw01234567, 2);
120       vo1p0 = vfma_laneq_f16(vo1p0, vi2x8ACE9BDF.val[0], vw01234567, 2);
121       vo2p0 = vfma_laneq_f16(vo2p0, vi4x8ACE9BDF.val[0], vw01234567, 2);
122       vo3p0 = vfma_laneq_f16(vo3p0, vi6x8ACE9BDF.val[0], vw01234567, 2);
123 
124       vo0p0 = vfma_laneq_f16(vo0p0, vi1x8ACE9BDF.val[0], vw01234567, 5);
125       vo1p0 = vfma_laneq_f16(vo1p0, vi3x8ACE9BDF.val[0], vw01234567, 5);
126       vo2p0 = vfma_laneq_f16(vo2p0, vi5x8ACE9BDF.val[0], vw01234567, 5);
127       vo3p0 = vfma_laneq_f16(vo3p0, vi7x8ACE9BDF.val[0], vw01234567, 5);
128 
129       vo0p0 = vfma_lane_f16(vo0p0, vi2x8ACE9BDF.val[0], vw89, 0);
130       vo1p0 = vfma_lane_f16(vo1p0, vi4x8ACE9BDF.val[0], vw89, 0);
131       vo2p0 = vfma_lane_f16(vo2p0, vi6x8ACE9BDF.val[0], vw89, 0);
132       vo3p0 = vfma_lane_f16(vo3p0, vi8x8ACE9BDF.val[0], vw89, 0);
133 
134       const float16x4_t vi0x7BDF = vext_f16(vi0x1357, vi0x8ACE9BDF.val[1], 3);
135       vi0x1357 = vi0x8ACE9BDF.val[1];
136       const float16x4_t vi1x7BDF = vext_f16(vi1x1357, vi1x8ACE9BDF.val[1], 3);
137       vi1x1357 = vi1x8ACE9BDF.val[1];
138       const float16x4_t vi2x7BDF = vext_f16(vi2x1357, vi2x8ACE9BDF.val[1], 3);
139       vi2x1357 = vi2x8ACE9BDF.val[1];
140       const float16x4_t vi3x7BDF = vext_f16(vi3x1357, vi3x8ACE9BDF.val[1], 3);
141       vi3x1357 = vi3x8ACE9BDF.val[1];
142       const float16x4_t vi4x7BDF = vext_f16(vi4x1357, vi4x8ACE9BDF.val[1], 3);
143       vi4x1357 = vi4x8ACE9BDF.val[1];
144       const float16x4_t vi5x7BDF = vext_f16(vi5x1357, vi5x8ACE9BDF.val[1], 3);
145       vi5x1357 = vi5x8ACE9BDF.val[1];
146       const float16x4_t vi6x7BDF = vext_f16(vi6x1357, vi6x8ACE9BDF.val[1], 3);
147       vi6x1357 = vi6x8ACE9BDF.val[1];
148       const float16x4_t vi7x7BDF = vext_f16(vi7x1357, vi7x8ACE9BDF.val[1], 3);
149       vi7x1357 = vi7x8ACE9BDF.val[1];
150       const float16x4_t vi8x7BDF = vext_f16(vi8x1357, vi8x8ACE9BDF.val[1], 3);
151       vi8x1357 = vi8x8ACE9BDF.val[1];
152 
153       vo0p0 = vfma_laneq_f16(vo0p0, vi0x7BDF, vw01234567, 1);
154       vo1p0 = vfma_laneq_f16(vo1p0, vi2x7BDF, vw01234567, 1);
155       vo2p0 = vfma_laneq_f16(vo2p0, vi4x7BDF, vw01234567, 1);
156       vo3p0 = vfma_laneq_f16(vo3p0, vi6x7BDF, vw01234567, 1);
157 
158       vo0p0 = vfma_laneq_f16(vo0p0, vi1x7BDF, vw01234567, 4);
159       vo1p0 = vfma_laneq_f16(vo1p0, vi3x7BDF, vw01234567, 4);
160       vo2p0 = vfma_laneq_f16(vo2p0, vi5x7BDF, vw01234567, 4);
161       vo3p0 = vfma_laneq_f16(vo3p0, vi7x7BDF, vw01234567, 4);
162 
163       vo0p0 = vfma_laneq_f16(vo0p0, vi2x7BDF, vw01234567, 7);
164       vo1p0 = vfma_laneq_f16(vo1p0, vi4x7BDF, vw01234567, 7);
165       vo2p0 = vfma_laneq_f16(vo2p0, vi6x7BDF, vw01234567, 7);
166       vo3p0 = vfma_laneq_f16(vo3p0, vi8x7BDF, vw01234567, 7);
167 
168       vo0p0 = vfma_laneq_f16(vo0p0, vi0x8ACE9BDF.val[1], vw01234567, 3);
169       vo1p0 = vfma_laneq_f16(vo1p0, vi2x8ACE9BDF.val[1], vw01234567, 3);
170       vo2p0 = vfma_laneq_f16(vo2p0, vi4x8ACE9BDF.val[1], vw01234567, 3);
171       vo3p0 = vfma_laneq_f16(vo3p0, vi6x8ACE9BDF.val[1], vw01234567, 3);
172 
173       vo0p0 = vfma_laneq_f16(vo0p0, vi1x8ACE9BDF.val[1], vw01234567, 6);
174       vo1p0 = vfma_laneq_f16(vo1p0, vi3x8ACE9BDF.val[1], vw01234567, 6);
175       vo2p0 = vfma_laneq_f16(vo2p0, vi5x8ACE9BDF.val[1], vw01234567, 6);
176       vo3p0 = vfma_laneq_f16(vo3p0, vi7x8ACE9BDF.val[1], vw01234567, 6);
177 
178       vo0p0 = vfma_lane_f16(vo0p0, vi2x8ACE9BDF.val[1], vw89, 1);
179       vo1p0 = vfma_lane_f16(vo1p0, vi4x8ACE9BDF.val[1], vw89, 1);
180       vo2p0 = vfma_lane_f16(vo2p0, vi6x8ACE9BDF.val[1], vw89, 1);
181       vo3p0 = vfma_lane_f16(vo3p0, vi8x8ACE9BDF.val[1], vw89, 1);
182 
183 
184       float16x4_t vo0 = vmax_f16(vo0p0, vmin);
185       float16x4_t vo1 = vmax_f16(vo1p0, vmin);
186       float16x4_t vo2 = vmax_f16(vo2p0, vmin);
187       float16x4_t vo3 = vmax_f16(vo3p0, vmin);
188 
189       vo0 = vmin_f16(vo0, vmax);
190       vo1 = vmin_f16(vo1, vmax);
191       vo2 = vmin_f16(vo2, vmax);
192       vo3 = vmin_f16(vo3, vmax);
193 
194       vst1_f16(o3, vo3); o3 += 4;
195       vst1_f16(o2, vo2); o2 += 4;
196       vst1_f16(o1, vo1); o1 += 4;
197       vst1_f16(o0, vo0); o0 += 4;
198     }
199     // Last block has 0-7 pixels to process.
200     assert(w < 8 * sizeof(__fp16));
201     if XNN_LIKELY(w != 0) {
202       float16x4_t vo0p0 = vdup_laneq_f16(vw01234567, 0);
203       float16x4_t vo1p0 = vdup_laneq_f16(vw01234567, 0);
204       float16x4_t vo2p0 = vdup_laneq_f16(vw01234567, 0);
205       float16x4_t vo3p0 = vdup_laneq_f16(vw01234567, 0);
206 
207       const float16x4x2_t vi0x8ACE9BDF = vld2_f16(i0);
208       const float16x4x2_t vi1x8ACE9BDF = vld2_f16(i1);
209       const float16x4x2_t vi2x8ACE9BDF = vld2_f16(i2);
210       const float16x4x2_t vi3x8ACE9BDF = vld2_f16(i3);
211       const float16x4x2_t vi4x8ACE9BDF = vld2_f16(i4);
212       const float16x4x2_t vi5x8ACE9BDF = vld2_f16(i5);
213       const float16x4x2_t vi6x8ACE9BDF = vld2_f16(i6);
214       const float16x4x2_t vi7x8ACE9BDF = vld2_f16(i7);
215       const float16x4x2_t vi8x8ACE9BDF = vld2_f16(i8);
216 
217       const float16x4_t vi0x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi0x8ACE9BDF.val[0])));
218       const float16x4_t vi0x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd,  vreinterpret_u16_f16(vi0x8ACE9BDF.val[1])));
219       const float16x4_t vi1x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi1x8ACE9BDF.val[0])));
220       const float16x4_t vi1x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd,  vreinterpret_u16_f16(vi1x8ACE9BDF.val[1])));
221       const float16x4_t vi2x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi2x8ACE9BDF.val[0])));
222       const float16x4_t vi2x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd,  vreinterpret_u16_f16(vi2x8ACE9BDF.val[1])));
223       const float16x4_t vi3x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi3x8ACE9BDF.val[0])));
224       const float16x4_t vi3x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd,  vreinterpret_u16_f16(vi3x8ACE9BDF.val[1])));
225       const float16x4_t vi4x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi4x8ACE9BDF.val[0])));
226       const float16x4_t vi4x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd,  vreinterpret_u16_f16(vi4x8ACE9BDF.val[1])));
227       const float16x4_t vi5x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi5x8ACE9BDF.val[0])));
228       const float16x4_t vi5x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd,  vreinterpret_u16_f16(vi5x8ACE9BDF.val[1])));
229       const float16x4_t vi6x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi6x8ACE9BDF.val[0])));
230       const float16x4_t vi6x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd,  vreinterpret_u16_f16(vi6x8ACE9BDF.val[1])));
231       const float16x4_t vi7x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi7x8ACE9BDF.val[0])));
232       const float16x4_t vi7x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd,  vreinterpret_u16_f16(vi7x8ACE9BDF.val[1])));
233       const float16x4_t vi8x8ACE = vreinterpret_f16_u16(vand_u16(vmask_even, vreinterpret_u16_f16(vi8x8ACE9BDF.val[0])));
234       const float16x4_t vi8x9BDF = vreinterpret_f16_u16(vand_u16(vmask_odd,  vreinterpret_u16_f16(vi8x8ACE9BDF.val[1])));
235 
236       vo0p0 = vfma_laneq_f16(vo0p0, vi0x8ACE, vw01234567, 2);
237       vo1p0 = vfma_laneq_f16(vo1p0, vi2x8ACE, vw01234567, 2);
238       vo2p0 = vfma_laneq_f16(vo2p0, vi4x8ACE, vw01234567, 2);
239       vo3p0 = vfma_laneq_f16(vo3p0, vi6x8ACE, vw01234567, 2);
240 
241       vo0p0 = vfma_laneq_f16(vo0p0, vi1x8ACE, vw01234567, 5);
242       vo1p0 = vfma_laneq_f16(vo1p0, vi3x8ACE, vw01234567, 5);
243       vo2p0 = vfma_laneq_f16(vo2p0, vi5x8ACE, vw01234567, 5);
244       vo3p0 = vfma_laneq_f16(vo3p0, vi7x8ACE, vw01234567, 5);
245 
246       vo0p0 = vfma_lane_f16(vo0p0, vi2x8ACE, vw89, 0);
247       vo1p0 = vfma_lane_f16(vo1p0, vi4x8ACE, vw89, 0);
248       vo2p0 = vfma_lane_f16(vo2p0, vi6x8ACE, vw89, 0);
249       vo3p0 = vfma_lane_f16(vo3p0, vi8x8ACE, vw89, 0);
250 
251       const float16x4_t vi0x7BDF = vext_f16(vi0x1357, vi0x9BDF, 3);
252       const float16x4_t vi1x7BDF = vext_f16(vi1x1357, vi1x9BDF, 3);
253       const float16x4_t vi2x7BDF = vext_f16(vi2x1357, vi2x9BDF, 3);
254       const float16x4_t vi3x7BDF = vext_f16(vi3x1357, vi3x9BDF, 3);
255       const float16x4_t vi4x7BDF = vext_f16(vi4x1357, vi4x9BDF, 3);
256       const float16x4_t vi5x7BDF = vext_f16(vi5x1357, vi5x9BDF, 3);
257       const float16x4_t vi6x7BDF = vext_f16(vi6x1357, vi6x9BDF, 3);
258       const float16x4_t vi7x7BDF = vext_f16(vi7x1357, vi7x9BDF, 3);
259       const float16x4_t vi8x7BDF = vext_f16(vi8x1357, vi8x9BDF, 3);
260 
261       vo0p0 = vfma_laneq_f16(vo0p0, vi0x7BDF, vw01234567, 1);
262       vo1p0 = vfma_laneq_f16(vo1p0, vi2x7BDF, vw01234567, 1);
263       vo2p0 = vfma_laneq_f16(vo2p0, vi4x7BDF, vw01234567, 1);
264       vo3p0 = vfma_laneq_f16(vo3p0, vi6x7BDF, vw01234567, 1);
265 
266       vo0p0 = vfma_laneq_f16(vo0p0, vi1x7BDF, vw01234567, 4);
267       vo1p0 = vfma_laneq_f16(vo1p0, vi3x7BDF, vw01234567, 4);
268       vo2p0 = vfma_laneq_f16(vo2p0, vi5x7BDF, vw01234567, 4);
269       vo3p0 = vfma_laneq_f16(vo3p0, vi7x7BDF, vw01234567, 4);
270 
271       vo0p0 = vfma_laneq_f16(vo0p0, vi2x7BDF, vw01234567, 7);
272       vo1p0 = vfma_laneq_f16(vo1p0, vi4x7BDF, vw01234567, 7);
273       vo2p0 = vfma_laneq_f16(vo2p0, vi6x7BDF, vw01234567, 7);
274       vo3p0 = vfma_laneq_f16(vo3p0, vi8x7BDF, vw01234567, 7);
275 
276       vo0p0 = vfma_laneq_f16(vo0p0, vi0x9BDF, vw01234567, 3);
277       vo1p0 = vfma_laneq_f16(vo1p0, vi2x9BDF, vw01234567, 3);
278       vo2p0 = vfma_laneq_f16(vo2p0, vi4x9BDF, vw01234567, 3);
279       vo3p0 = vfma_laneq_f16(vo3p0, vi6x9BDF, vw01234567, 3);
280 
281       vo0p0 = vfma_laneq_f16(vo0p0, vi1x9BDF, vw01234567, 6);
282       vo1p0 = vfma_laneq_f16(vo1p0, vi3x9BDF, vw01234567, 6);
283       vo2p0 = vfma_laneq_f16(vo2p0, vi5x9BDF, vw01234567, 6);
284       vo3p0 = vfma_laneq_f16(vo3p0, vi7x9BDF, vw01234567, 6);
285 
286       vo0p0 = vfma_lane_f16(vo0p0, vi2x9BDF, vw89, 1);
287       vo1p0 = vfma_lane_f16(vo1p0, vi4x9BDF, vw89, 1);
288       vo2p0 = vfma_lane_f16(vo2p0, vi6x9BDF, vw89, 1);
289       vo3p0 = vfma_lane_f16(vo3p0, vi8x9BDF, vw89, 1);
290 
291 
292       float16x4_t vo0 = vmax_f16(vo0p0, vmin);
293       float16x4_t vo1 = vmax_f16(vo1p0, vmin);
294       float16x4_t vo2 = vmax_f16(vo2p0, vmin);
295       float16x4_t vo3 = vmax_f16(vo3p0, vmin);
296 
297       vo0 = vmin_f16(vo0, vmax);
298       vo1 = vmin_f16(vo1, vmax);
299       vo2 = vmin_f16(vo2, vmax);
300       vo3 = vmin_f16(vo3, vmax);
301 
302       w += 1 * sizeof(__fp16);
303 
304       if XNN_LIKELY(w == 8 * sizeof(__fp16)) {
305         vst1_f16(o3, vo3); o3 += 4;
306         vst1_f16(o2, vo2); o2 += 4;
307         vst1_f16(o1, vo1); o1 += 4;
308         vst1_f16(o0, vo0); o0 += 4;
309       } else {
310         if (w & (4 * sizeof(__fp16))) {
311           vst1_lane_u32((void*) o3, vreinterpret_u32_f16(vo3), 0); o3 += 2;
312           vst1_lane_u32((void*) o2, vreinterpret_u32_f16(vo2), 0); o2 += 2;
313           vst1_lane_u32((void*) o1, vreinterpret_u32_f16(vo1), 0); o1 += 2;
314           vst1_lane_u32((void*) o0, vreinterpret_u32_f16(vo0), 0); o0 += 2;
315 
316           vo0 = vext_f16(vo0, vo0, 2);
317           vo1 = vext_f16(vo1, vo1, 2);
318           vo2 = vext_f16(vo2, vo2, 2);
319           vo3 = vext_f16(vo3, vo3, 2);
320         }
321         if (w & (2 * sizeof(__fp16))) {
322           vst1_lane_f16(o3, vo3, 0); o3 += 1;
323           vst1_lane_f16(o2, vo2, 0); o2 += 1;
324           vst1_lane_f16(o1, vo1, 0); o1 += 1;
325           vst1_lane_f16(o0, vo0, 0); o0 += 1;
326         }
327       }
328 
329     }
330 
331     i0 = (const __fp16*) ((uintptr_t) i8 - input_decrement);
332     i1 = (const __fp16*) ((uintptr_t) i0 + input_width);
333     i2 = (const __fp16*) ((uintptr_t) i1 + input_width);
334     i3 = (const __fp16*) ((uintptr_t) i2 + input_width);
335     i4 = (const __fp16*) ((uintptr_t) i3 + input_width);
336     i5 = (const __fp16*) ((uintptr_t) i4 + input_width);
337     i6 = (const __fp16*) ((uintptr_t) i5 + input_width);
338     i7 = (const __fp16*) ((uintptr_t) i6 + input_width);
339     i8 = (const __fp16*) ((uintptr_t) i7 + input_width);
340 
341     o0 = o3;
342     o1 = (__fp16*) ((uintptr_t) o0 + output_width);
343     o2 = (__fp16*) ((uintptr_t) o1 + output_width);
344     o3 = (__fp16*) ((uintptr_t) o2 + output_width);
345 
346     output_height = doz(output_height, 4);
347     padded_input_height = doz(padded_input_height, 8);
348   } while (output_height != 0);
349 }
350