File size: 6,410 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
// Auto-generated file. Do not edit!
//   Template: src/f16-dwconv/unipass-neonfp16arith.c.in
//   Generator: tools/xngen
//
// Copyright 2020 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/dwconv.h>


void xnn_f16_dwconv_minmax_ukernel_3p16c__neonfp16arith(
    size_t channels,
    size_t output_width,
    const void** input,
    const void* weights,
    void* output_ptr,
    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);

  uint16_t* output = (uint16_t*) output_ptr;
  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);
    if XNN_UNPREDICTABLE(i0 != (const uint16_t*) zero) {
      i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
    }
    const uint16_t* i1 = (const uint16_t*) input[1];
    assert(i1 != NULL);
    if XNN_UNPREDICTABLE(i1 != (const uint16_t*) zero) {
      i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
    }
    const uint16_t* i2 = (const uint16_t*) input[2];
    assert(i2 != NULL);
    if XNN_UNPREDICTABLE(i2 != (const uint16_t*) zero) {
      i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
    }

    input = (const void**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const uint16_t* w = (const uint16_t*) weights;
    for (; c >= 16; c -= 16) {
      float16x8_t vacc01234567p0 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
      float16x8_t vacc89ABCDEFp0 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;


      const float16x8_t vi0x01234567 = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
      const float16x8_t vi0x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
      const float16x8_t vk0x01234567 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
      const float16x8_t vk0x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
      vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
      vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF);

      const float16x8_t vi1x01234567 = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
      const float16x8_t vi1x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
      const float16x8_t vk1x01234567 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
      const float16x8_t vk1x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
      vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);
      vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi1x89ABCDEF, vk1x89ABCDEF);

      const float16x8_t vi2x01234567 = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
      const float16x8_t vi2x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
      const float16x8_t vk2x01234567 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
      const float16x8_t vk2x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
      vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
      vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF);


      float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
      float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin);
      vacc01234567 = vminq_f16(vacc01234567, vmax);
      vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);

      vst1q_u16(output, vreinterpretq_u16_f16(vacc01234567)); output += 8;
      vst1q_u16(output, vreinterpretq_u16_f16(vacc89ABCDEF)); output += 8;
    }
    for (; c >= 8; c -= 8) {
      float16x8_t vacc01234567p0 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;


      const float16x8_t vi0x01234567 = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
      const float16x8_t vk0x01234567 = vreinterpretq_f16_u16(vld1q_u16(w + 8));
      vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);

      const float16x8_t vi1x01234567 = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
      const float16x8_t vk1x01234567 = vreinterpretq_f16_u16(vld1q_u16(w + 24));
      vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);

      const float16x8_t vi2x01234567 = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
      const float16x8_t vk2x01234567 = vreinterpretq_f16_u16(vld1q_u16(w + 40));
      vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);


      float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
      vacc01234567 = vminq_f16(vacc01234567, vmax);

      vst1q_u16(output, vreinterpretq_u16_f16(vacc01234567)); output += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      float16x8_t vacc01234567p0 = vreinterpretq_f16_u16(vld1q_u16(w));


      const float16x8_t vi0x01234567 = vreinterpretq_f16_u16(vld1q_u16(i0));
      const float16x8_t vk0x01234567 = vreinterpretq_f16_u16(vld1q_u16(w + 16));
      vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);

      const float16x8_t vi1x01234567 = vreinterpretq_f16_u16(vld1q_u16(i1));
      const float16x8_t vk1x01234567 = vreinterpretq_f16_u16(vld1q_u16(w + 32));
      vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);

      const float16x8_t vi2x01234567 = vreinterpretq_f16_u16(vld1q_u16(i2));
      const float16x8_t vk2x01234567 = vreinterpretq_f16_u16(vld1q_u16(w + 48));
      vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);


      float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
      vacc01234567 = vminq_f16(vacc01234567, vmax);

      float16x4_t vacc0123 = vget_low_f16(vacc01234567);
      if (c & 4) {
        vst1_u16(output, vreinterpret_u16_f16(vacc0123)); output += 4;
        vacc0123 = vget_high_f16(vacc01234567);
      }
      if (c & 2) {
        vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
        vacc0123 = vext_f16(vacc0123, vacc0123, 2);
      }
      if (c & 1) {
        vst1_lane_u16(output, vreinterpret_u16_f16(vacc0123), 0); output += 1;
      }
    }

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