Hugging Face's logo

Androidonnxfork
/
test

Model card Files Community
5
test
File size: 9,669 Bytes 
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
 
// Copyright 2022 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/maxpool.h>


void xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8(
    size_t output_pixels,
    size_t kernel_elements,
    size_t channels,
    const void** input,
    size_t input_offset,
    void* output,
    size_t input_increment,
    size_t output_increment,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(output_pixels != 0);
  assert(kernel_elements != 0);
  assert(channels != 0);

  const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.min));
  const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.max));
  do {
    uint16_t* o = output;
    {
      const uint16_t* i0 = *input++;
      const uint16_t* i1 = *input++;
      const uint16_t* i2 = *input++;
      const uint16_t* i3 = *input++;
      const uint16_t* i4 = *input++;
      const uint16_t* i5 = *input++;
      const uint16_t* i6 = *input++;
      const uint16_t* i7 = *input++;
      const uint16_t* i8 = *input++;
      i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
      i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
      i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
      i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
      i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
      i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
      i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
      i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
      i8 = (const uint16_t*) ((uintptr_t) i8 + input_offset);
      if (kernel_elements < 2) {
        i1 = i0;
      }
      if (kernel_elements <= 2) {
        i2 = i0;
      }
      if (kernel_elements < 4) {
        i3 = i0;
      }
      if (kernel_elements <= 4) {
        i4 = i0;
      }
      if (kernel_elements < 6) {
        i5 = i0;
      }
      if (kernel_elements <= 6) {
        i6 = i0;
      }
      if (kernel_elements < 8) {
        i7 = i0;
      }
      if (kernel_elements <= 8) {
        i8 = i0;
      }

      size_t c = channels;
      for (; c >= 8; c -= 8) {
        const float16x8_t vi0 = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
        const float16x8_t vi1 = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
        const float16x8_t vi2 = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
        const float16x8_t vi3 = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;
        const float16x8_t vi4 = vreinterpretq_f16_u16(vld1q_u16(i4)); i4 += 8;
        const float16x8_t vi5 = vreinterpretq_f16_u16(vld1q_u16(i5)); i5 += 8;
        const float16x8_t vi6 = vreinterpretq_f16_u16(vld1q_u16(i6)); i6 += 8;
        const float16x8_t vi7 = vreinterpretq_f16_u16(vld1q_u16(i7)); i7 += 8;
        const float16x8_t vi8 = vreinterpretq_f16_u16(vld1q_u16(i8)); i8 += 8;

        const float16x8_t vmax018 = vmaxq_f16(vmaxq_f16(vi0, vi1), vi8);
        const float16x8_t vmax23 = vmaxq_f16(vi2, vi3);
        const float16x8_t vmax45 = vmaxq_f16(vi4, vi5);
        const float16x8_t vmax67 = vmaxq_f16(vi6, vi7);

        const float16x8_t vmax2345 = vmaxq_f16(vmax23, vmax45);
        const float16x8_t vmax01678 = vmaxq_f16(vmax018, vmax67);
        const float16x8_t vmax = vmaxq_f16(vmax2345, vmax01678);
        const float16x8_t vout = vmaxq_f16(vminq_f16(vmax, voutput_max), voutput_min);

        vst1q_u16(o, vreinterpretq_u16_f16(vout)); o += 8;
      }
      if (c != 0) {
        const float16x8_t vi0 = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
        const float16x8_t vi1 = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
        const float16x8_t vi2 = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
        const float16x8_t vi3 = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;
        const float16x8_t vi4 = vreinterpretq_f16_u16(vld1q_u16(i4)); i4 += 8;
        const float16x8_t vi5 = vreinterpretq_f16_u16(vld1q_u16(i5)); i5 += 8;
        const float16x8_t vi6 = vreinterpretq_f16_u16(vld1q_u16(i6)); i6 += 8;
        const float16x8_t vi7 = vreinterpretq_f16_u16(vld1q_u16(i7)); i7 += 8;
        const float16x8_t vi8 = vreinterpretq_f16_u16(vld1q_u16(i8)); i8 += 8;

        const float16x8_t vmax018 = vmaxq_f16(vmaxq_f16(vi0, vi1), vi8);
        const float16x8_t vmax23 = vmaxq_f16(vi2, vi3);
        const float16x8_t vmax45 = vmaxq_f16(vi4, vi5);
        const float16x8_t vmax67 = vmaxq_f16(vi6, vi7);

        const float16x8_t vmax2345 = vmaxq_f16(vmax23, vmax45);
        const float16x8_t vmax01678 = vmaxq_f16(vmax018, vmax67);
        const float16x8_t vmax = vmaxq_f16(vmax2345, vmax01678);
        float16x8_t vout = vmaxq_f16(vminq_f16(vmax, voutput_max), voutput_min);

        float16x4_t vout_lo = vget_low_f16(vout);
        if (c & 4) {
          vst1_u16(o, vreinterpret_u16_f16(vout_lo)); o += 4;
          vout_lo = vget_high_f16(vout);
        }
        if (c & 2) {
          vst1_lane_u32((void*) o, vreinterpret_u32_f16(vout_lo), 0); o += 2;
          vout_lo = vext_f16(vout_lo, vout_lo, 2);
        }
        if (c & 1) {
          vst1_lane_u16(o, vreinterpret_u16_f16(vout_lo), 0); o += 1;
        }
      }
    }

    for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) {
      const uint16_t* i0 = *input++;
      const uint16_t* i1 = *input++;
      const uint16_t* i2 = *input++;
      const uint16_t* i3 = *input++;
      const uint16_t* i4 = *input++;
      const uint16_t* i5 = *input++;
      const uint16_t* i6 = *input++;
      const uint16_t* i7 = *input++;
      i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
      i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
      i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
      i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
      i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
      i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
      i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
      i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
      if (k < 2) {
        i1 = i0;
      }
      if (k <= 2) {
        i2 = i0;
      }
      if (k < 4) {
        i3 = i0;
      }
      if (k <= 4) {
        i4 = i0;
      }
      if (k < 6) {
        i5 = i0;
      }
      if (k <= 6) {
        i6 = i0;
      }
      if (k < 8) {
        i7 = i0;
      }

      o = output;
      size_t c = channels;
      for (; c >= 8; c -= 8) {
        const float16x8_t vi0 = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
        const float16x8_t vi1 = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
        const float16x8_t vi2 = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
        const float16x8_t vi3 = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;
        const float16x8_t vi4 = vreinterpretq_f16_u16(vld1q_u16(i4)); i4 += 8;
        const float16x8_t vi5 = vreinterpretq_f16_u16(vld1q_u16(i5)); i5 += 8;
        const float16x8_t vi6 = vreinterpretq_f16_u16(vld1q_u16(i6)); i6 += 8;
        const float16x8_t vi7 = vreinterpretq_f16_u16(vld1q_u16(i7)); i7 += 8;
        const float16x8_t vo = vreinterpretq_f16_u16(vld1q_u16(o));

        const float16x8_t vmax01 = vmaxq_f16(vmaxq_f16(vi0, vi1), vo);
        const float16x8_t vmax23 = vmaxq_f16(vi2, vi3);
        const float16x8_t vmax45 = vmaxq_f16(vi4, vi5);
        const float16x8_t vmax67 = vmaxq_f16(vi6, vi7);

        const float16x8_t vmax2345 = vmaxq_f16(vmax23, vmax45);
        const float16x8_t vmax0167 = vmaxq_f16(vmax01, vmax67);
        const float16x8_t vmax = vmaxq_f16(vmax2345, vmax0167);
        const float16x8_t vout = vmaxq_f16(vminq_f16(vmax, voutput_max), voutput_min);

        vst1q_u16(o, vreinterpretq_u16_f16(vout)); o += 8;
      }
      if (c != 0) {
        const float16x8_t vi0 = vreinterpretq_f16_u16(vld1q_u16(i0));
        const float16x8_t vi1 = vreinterpretq_f16_u16(vld1q_u16(i1));
        const float16x8_t vi2 = vreinterpretq_f16_u16(vld1q_u16(i2));
        const float16x8_t vi3 = vreinterpretq_f16_u16(vld1q_u16(i3));
        const float16x8_t vi4 = vreinterpretq_f16_u16(vld1q_u16(i4));
        const float16x8_t vi5 = vreinterpretq_f16_u16(vld1q_u16(i5));
        const float16x8_t vi6 = vreinterpretq_f16_u16(vld1q_u16(i6));
        const float16x8_t vi7 = vreinterpretq_f16_u16(vld1q_u16(i7));
        const float16x8_t vo = vreinterpretq_f16_u16(vld1q_u16(o));

        const float16x8_t vmax01 = vmaxq_f16(vmaxq_f16(vi0, vi1), vo);
        const float16x8_t vmax23 = vmaxq_f16(vi2, vi3);
        const float16x8_t vmax45 = vmaxq_f16(vi4, vi5);
        const float16x8_t vmax67 = vmaxq_f16(vi6, vi7);

        const float16x8_t vmax2345 = vmaxq_f16(vmax23, vmax45);
        const float16x8_t vmax0167 = vmaxq_f16(vmax01, vmax67);
        const float16x8_t vmax = vmaxq_f16(vmax2345, vmax0167);
        float16x8_t vout = vmaxq_f16(vminq_f16(vmax, voutput_max), voutput_min);

        float16x4_t vout_lo = vget_low_f16(vout);
        if (c & 4) {
          vst1_u16(o, vreinterpret_u16_f16(vout_lo)); o += 4;
          vout_lo = vget_high_f16(vout);
        }
        if (c & 2) {
          vst1_lane_u32((void*) o, vreinterpret_u32_f16(vout_lo), 0); o += 2;
          vout_lo = vext_f16(vout_lo, vout_lo, 2);
        }
        if (c & 1) {
          vst1_lane_u16(o, vreinterpret_u16_f16(vout_lo), 0); o += 1;
        }
      }
    }
    input = (const void**) ((uintptr_t) input + input_increment);
    output = (uint16_t*) ((uintptr_t) o + output_increment);
  } while (--output_pixels != 0);
}