File size: 5,855 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
// Auto-generated file. Do not edit!
//   Template: src/f16-gavgpool/unipass-f16c.c.in
//   Generator: tools/xngen
//
// 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 <immintrin.h>

#include <xnnpack/gavgpool.h>
#include <xnnpack/intrinsics-polyfill.h>


void xnn_f16_gavgpool_minmax_ukernel_7x__f16c_c8(
    size_t rows,
    size_t channels,
    const void* input,
    size_t input_stride,
    const void* zero,
    void* output,
    const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(rows != 0);
  assert(rows <= 7);
  assert(channels != 0);

  const uint16_t* i0 = input;
  const uint16_t* i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
  if XNN_UNPREDICTABLE(rows < 2) {
    i1 = (const uint16_t*) zero;
  }
  const uint16_t* i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
  if XNN_UNPREDICTABLE(rows <= 2) {
    i2 = (const uint16_t*) zero;
  }
  const uint16_t* i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
  if XNN_UNPREDICTABLE(rows < 4) {
    i3 = (const uint16_t*) zero;
  }
  const uint16_t* i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
  if XNN_UNPREDICTABLE(rows <= 4) {
    i4 = (const uint16_t*) zero;
  }
  const uint16_t* i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
  if XNN_UNPREDICTABLE(rows < 6) {
    i5 = (const uint16_t*) zero;
  }
  const uint16_t* i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
  if XNN_UNPREDICTABLE(rows <= 6) {
    i6 = (const uint16_t*) zero;
  }
  uint16_t* o = (uint16_t*) output;

  const __m256 vscale = _mm256_load_ps(params->avx.scale);
  const __m256 vmin = _mm256_load_ps(params->avx.min);
  const __m256 vmax = _mm256_load_ps(params->avx.max);
  for (; channels >= 8; channels -= 8) {
    const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
    i0 += 8;
    const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
    i1 += 8;

    const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
    __m128i vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(vi0x01234567, vi1x01234567), _MM_FROUND_TO_NEAREST_INT);
    i2 += 8;

    const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
    i3 += 8;
    vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi2x01234567), _MM_FROUND_TO_NEAREST_INT);
    const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
    i4 += 8;
    vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi3x01234567), _MM_FROUND_TO_NEAREST_INT);
    const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
    i5 += 8;
    vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi4x01234567), _MM_FROUND_TO_NEAREST_INT);
    const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
    i6 += 8;
    vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi5x01234567), _MM_FROUND_TO_NEAREST_INT);
    vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi6x01234567), _MM_FROUND_TO_NEAREST_INT);

    vacc01234567 = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc01234567), vscale), _MM_FROUND_TO_NEAREST_INT);

    __m256 vout01234567 = _mm256_max_ps(_mm256_cvtph_ps(vacc01234567), vmin);

    vout01234567 = _mm256_min_ps(vout01234567, vmax);

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vout01234567, _MM_FROUND_TO_NEAREST_INT));
    o += 8;
  }
  if XNN_UNLIKELY(channels != 0) {
    {
      const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
      const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));

      const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
      __m128i vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(vi0x01234567, vi1x01234567), _MM_FROUND_TO_NEAREST_INT);

      const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
      vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi2x01234567), _MM_FROUND_TO_NEAREST_INT);
      const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
      vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi3x01234567), _MM_FROUND_TO_NEAREST_INT);
      const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
      vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi4x01234567), _MM_FROUND_TO_NEAREST_INT);
      const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
      vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi5x01234567), _MM_FROUND_TO_NEAREST_INT);
      vacc01234567 = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc01234567), vi6x01234567), _MM_FROUND_TO_NEAREST_INT);

      vacc01234567 = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc01234567), vscale), _MM_FROUND_TO_NEAREST_INT);
      __m256 vout01234567 = _mm256_max_ps(_mm256_cvtph_ps(vacc01234567), vmin);
      vout01234567 = _mm256_min_ps(vout01234567, vmax);

      __m128i vh01234567 = _mm256_cvtps_ph(vout01234567, _MM_FROUND_TO_NEAREST_INT);
      if (channels & 4) {
        _mm_storel_epi64((__m128i*) o, vh01234567);
        o += 4;
        vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
      }
      if (channels & 2) {
        _mm_storeu_si32(o, vh01234567);
        o += 2;
        vh01234567 = _mm_srli_epi64(vh01234567, 32);
      }
      if (channels & 1) {
        *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
      }
    }
  }
}