File size: 9,089 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
// Auto-generated file. Do not edit!
//   Template: src/f16-dwconv/unipass-fma3.c.in
//   Generator: tools/xngen
//
// 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 <immintrin.h>

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


void xnn_f16_dwconv_minmax_ukernel_4p16c__fma3(
    size_t channels,
    size_t output_width,
    const void** input,
    const void* weights,
    void* output,
    intptr_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const void* zero,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);

  const __m256 vmax = _mm256_load_ps(params->avx.max);
  const __m256 vmin = _mm256_load_ps(params->avx.min);

  uint16_t* o = (uint16_t*) output;
  do {
    const uint16_t* i0 = input[0];
    assert(i0 != NULL);
    if XNN_UNPREDICTABLE(i0 != zero) {
      i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
    }
    const uint16_t* i1 = input[1];
    assert(i1 != NULL);
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
    }
    const uint16_t* i2 = input[2];
    assert(i2 != NULL);
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
    }
    const uint16_t* i3 = input[3];
    assert(i3 != NULL);
    if XNN_UNPREDICTABLE(i3 != zero) {
      i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
    }
    input = (const void**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const uint16_t* w = weights;
    for (; c >= 16; c -= 16) {
      __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
      __m256 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 8)));


      const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
      const __m256 vi0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 8)));
      i0 += 16;

      const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 16)));
      const __m256 vk0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 24)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));
      vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
      const __m256 vi1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 8)));
      i1 += 16;

      const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 32)));
      const __m256 vk1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 40)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));
      vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi1x89ABCDEF, vk1x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
      const __m256 vi2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 8)));
      i2 += 16;

      const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 48)));
      const __m256 vk2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 56)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));
      vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
      const __m256 vi3x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i3 + 8)));
      i3 += 16;

      const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 64)));
      const __m256 vk3x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 72)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));
      vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x89ABCDEF, vk3x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_TO_NEAREST_INT));

      w += 80;


      __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
      __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin);
      vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
      vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vmax);

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_TO_NEAREST_INT));
      _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vacc89ABCDEF, _MM_FROUND_TO_NEAREST_INT));
      o += 16;
    }
    for (; c >= 8; c -= 8) {
      __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));

      const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
      i0 += 8;

      const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
      i1 += 8;

      const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
      i2 += 8;

      const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 48)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
      i3 += 8;

      const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));

      w += 8;


      __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
      vacc01234567 = _mm256_min_ps(vacc01234567, vmax);

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_TO_NEAREST_INT));
      o += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      assert(c >= 1);
      assert(c <= 7);

      __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));

      const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));

      const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));

      const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));

      const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 48)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));

      const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));

      const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
      vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0), _MM_FROUND_TO_NEAREST_INT));


      __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
      vacc01234567 = _mm256_min_ps(vacc01234567, vmax);

      __m128i vh01234567 = _mm256_cvtps_ph(vacc01234567, _MM_FROUND_TO_NEAREST_INT);
      if (c & 4) {
        _mm_storel_epi64((__m128i*) o, vh01234567);
        vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
        o += 4;
      }
      if (c & 2) {
        _mm_storeu_si32(o, vh01234567);
        vh01234567 = _mm_srli_epi64(vh01234567, 32);
        o += 2;
      }
      if (c & 1) {
        *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
        o += 1;
      }
    }

    o = (uint16_t*) ((uintptr_t) o + output_increment);
  } while (--output_width != 0);
}