File size: 6,582 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
// 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/avgpool.h>


void xnn_f16_avgpool_minmax_ukernel_9x__neonfp16arith_c8(
    size_t output_pixels,
    size_t kernel_elements,
    size_t channels,
    const void** input,
    size_t input_offset,
    const void* zero,
    void* output,
    size_t input_increment,
    size_t output_increment,
    const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(output_pixels != 0);
  assert(kernel_elements != 0);
  assert(kernel_elements <= 9);
  assert(channels != 0);

  const float16x8_t vscale = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.scale));
  const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.min));
  const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.max));

  do {
    const uint16_t* i0 = (const uint16_t*) input[0];
    assert(i0 != NULL);
    const uint16_t* i1 = (const uint16_t*) input[1];
    const uint16_t* i2 = (const uint16_t*) input[2];
    const uint16_t* i3 = (const uint16_t*) input[3];
    const uint16_t* i4 = (const uint16_t*) input[4];
    const uint16_t* i5 = (const uint16_t*) input[5];
    const uint16_t* i6 = (const uint16_t*) input[6];
    const uint16_t* i7 = (const uint16_t*) input[7];
    const uint16_t* i8 = (const uint16_t*) input[8];
    input = (const void**) ((uintptr_t) input + input_increment);
    if (kernel_elements < 2) {
      i1 = (const uint16_t*) zero;
    }
    assert(i1 != NULL);
    if (kernel_elements <= 2) {
      i2 = (const uint16_t*) zero;
    }
    assert(i2 != NULL);
    if (kernel_elements < 4) {
      i3 = (const uint16_t*) zero;
    }
    assert(i3 != NULL);
    if (kernel_elements <= 4) {
      i4 = (const uint16_t*) zero;
    }
    assert(i4 != NULL);
    if (kernel_elements < 6) {
      i5 = (const uint16_t*) zero;
    }
    assert(i5 != NULL);
    if (kernel_elements <= 6) {
      i6 = (const uint16_t*) zero;
    }
    assert(i6 != NULL);
    if (kernel_elements < 8) {
      i7 = (const uint16_t*) zero;
    }
    assert(i7 != NULL);
    if (kernel_elements <= 8) {
      i8 = (const uint16_t*) zero;
    }
    assert(i8 != NULL);
    if XNN_UNPREDICTABLE(i0 != zero) {
      i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
    }
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
    }
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
    }
    if XNN_UNPREDICTABLE(i3 != zero) {
      i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
    }
    if XNN_UNPREDICTABLE(i4 != zero) {
      i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
    }
    if XNN_UNPREDICTABLE(i5 != zero) {
      i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
    }
    if XNN_UNPREDICTABLE(i6 != zero) {
      i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
    }
    if XNN_UNPREDICTABLE(i7 != zero) {
      i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
    }
    if XNN_UNPREDICTABLE(i8 != zero) {
      i8 = (const uint16_t*) ((uintptr_t) i8 + input_offset);
    }

    size_t c = channels;
    while (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 vsum01 = vaddq_f16(vi0, vi1);
      const float16x8_t vsum23 = vaddq_f16(vi2, vi3);
      const float16x8_t vsum45 = vaddq_f16(vi4, vi5);
      const float16x8_t vsum67 = vaddq_f16(vi6, vi7);
      const float16x8_t vsum018 = vaddq_f16(vsum01, vi8);
      const float16x8_t vsum2345 = vaddq_f16(vsum23, vsum45);
      const float16x8_t vsum01678 = vaddq_f16(vsum018, vsum67);
      const float16x8_t vsum = vaddq_f16(vsum2345, vsum01678);

      float16x8_t vout = vmulq_f16(vsum, vscale);
      vout = vmaxq_f16(vout, vmin);
      vout = vminq_f16(vout, vmax);

      vst1q_u16(output, vreinterpretq_u16_f16(vout)); output = (uint16_t*) output + 8;

      c -= 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 vi8 = vreinterpretq_f16_u16(vld1q_u16(i8));

      const float16x8_t vsum01 = vaddq_f16(vi0, vi1);
      const float16x8_t vsum23 = vaddq_f16(vi2, vi3);
      const float16x8_t vsum45 = vaddq_f16(vi4, vi5);
      const float16x8_t vsum67 = vaddq_f16(vi6, vi7);
      const float16x8_t vsum018 = vaddq_f16(vsum01, vi8);
      const float16x8_t vsum2345 = vaddq_f16(vsum23, vsum45);
      const float16x8_t vsum01678 = vaddq_f16(vsum018, vsum67);
      const float16x8_t vsum = vaddq_f16(vsum2345, vsum01678);

      float16x8_t vout = vmulq_f16(vsum, vscale);
      vout = vmaxq_f16(vout, vmin);
      vout = vminq_f16(vout, vmax);

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