Hugging Face's logo

Androidonnxfork
/
test

Model card Files Community
6
test
File size: 3,359 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
 
// Copyright 2019 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/gavgpool.h>


void xnn_f32_gavgpool_minmax_ukernel_7x__neon_c4(
    size_t rows,
    size_t channels,
    const float* input,
    size_t input_stride,
    const float* zero,
    float* output,
    const union xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(rows != 0);
  assert(rows <= 7);
  assert(channels != 0);

  const float* i0 = input;
  const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
  if (rows < 2) {
    i1 = zero;
  }
  const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
  if (rows <= 2) {
    i2 = zero;
  }
  const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
  if (rows < 4) {
    i3 = zero;
  }
  const float* i4 = (const float*) ((uintptr_t) i3 + input_stride);
  if (rows <= 4) {
    i4 = zero;
  }
  const float* i5 = (const float*) ((uintptr_t) i4 + input_stride);
  if (rows < 6) {
    i5 = zero;
  }
  const float* i6 = (const float*) ((uintptr_t) i5 + input_stride);
  if (rows <= 6) {
    i6 = zero;
  }
  const float32x4_t vscale = vld1q_dup_f32(&params->scalar.scale);
  const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
  const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);

  while (channels >= 4) {
    const float32x4_t vi0 = vld1q_f32(i0); i0 += 4;
    const float32x4_t vi1 = vld1q_f32(i1); i1 += 4;
    const float32x4_t vi2 = vld1q_f32(i2); i2 += 4;
    const float32x4_t vi3 = vld1q_f32(i3); i3 += 4;
    const float32x4_t vi4 = vld1q_f32(i4); i4 += 4;
    const float32x4_t vi5 = vld1q_f32(i5); i5 += 4;
    const float32x4_t vi6 = vld1q_f32(i6); i6 += 4;

    const float32x4_t vsum01 = vaddq_f32(vi0, vi1);
    const float32x4_t vsum23 = vaddq_f32(vi2, vi3);
    const float32x4_t vsum45 = vaddq_f32(vi4, vi5);

    const float32x4_t vsum016 = vaddq_f32(vsum01, vi6);
    const float32x4_t vsum2345 = vaddq_f32(vsum23, vsum45);

    const float32x4_t vsum = vaddq_f32(vsum016, vsum2345);

    float32x4_t vout = vmulq_f32(vsum, vscale);
    vout = vmaxq_f32(vout, vmin);
    vout = vminq_f32(vout, vmax);

    vst1q_f32(output, vout); output += 4;

    channels -= 4;
  }
  if (channels != 0) {
    const float32x4_t vi0 = vld1q_f32(i0);
    const float32x4_t vi1 = vld1q_f32(i1);
    const float32x4_t vi2 = vld1q_f32(i2);
    const float32x4_t vi3 = vld1q_f32(i3);
    const float32x4_t vi4 = vld1q_f32(i4);
    const float32x4_t vi5 = vld1q_f32(i5);
    const float32x4_t vi6 = vld1q_f32(i6);

    const float32x4_t vsum01 = vaddq_f32(vi0, vi1);
    const float32x4_t vsum23 = vaddq_f32(vi2, vi3);
    const float32x4_t vsum45 = vaddq_f32(vi4, vi5);

    const float32x4_t vsum016 = vaddq_f32(vsum01, vi6);
    const float32x4_t vsum2345 = vaddq_f32(vsum23, vsum45);

    const float32x4_t vsum = vaddq_f32(vsum016, vsum2345);

    float32x4_t vout = vmulq_f32(vsum, vscale);
    vout = vmaxq_f32(vout, vmin);
    vout = vminq_f32(vout, vmax);

    float32x2_t vout_lo = vget_low_f32(vout);
    if (channels & 2) {
      vst1_f32(output, vout_lo); output += 2;
      vout_lo = vget_high_f32(vout);
    }
    if (channels & 1) {
      vst1_lane_f32(output, vout_lo, 0);
    }
  }
}