File size: 4,861 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
// 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.

$assert CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert PIXEL_TILE == 1
$ABC = "456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/ibilinear.h>


void xnn_f16_ibilinear_ukernel__neonfp16arith_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}(
    size_t output_pixels,
    size_t channels,
    const void** restrict input,
    size_t input_offset,
    const void* restrict weights,
    void* restrict output,
    size_t output_increment) XNN_OOB_READS
{
  assert(output_pixels != 0);
  assert(channels != 0);
  assert(channels % sizeof(uint16_t) == 0);

  uint16_t* o = (uint16_t*) output;
  do {
    const uint16_t* i0 = (const uint16_t*) ((uintptr_t) input[0] + input_offset);
    const uint16_t* i1 = (const uint16_t*) ((uintptr_t) input[1] + input_offset);
    const uint16_t* i2 = (const uint16_t*) ((uintptr_t) input[2] + input_offset);
    const uint16_t* i3 = (const uint16_t*) ((uintptr_t) input[3] + input_offset);
    input += 4;

    const float16x8_t valphah = vreinterpretq_f16_u16(vld1q_dup_u16(weights)); weights = (const uint16_t*) weights + 1;
    const float16x8_t valphav = vreinterpretq_f16_u16(vld1q_dup_u16(weights)); weights = (const uint16_t*) weights + 1;

    size_t c = channels;
    $if CHANNEL_TILE > 8:
      for (; c >= ${CHANNEL_TILE} * sizeof(uint16_t); c -= ${CHANNEL_TILE} * sizeof(uint16_t)) {
        $for C in range(0, CHANNEL_TILE, 8):
          const float16x8_t vtl${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
          const float16x8_t vtr${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
          const float16x8_t vbl${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
          const float16x8_t vbr${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;

        $for C in range(0, CHANNEL_TILE, 8):
          const float16x8_t vtd${ABC[C:C+8]} = vsubq_f16(vtr${ABC[C:C+8]}, vtl${ABC[C:C+8]});
          const float16x8_t vbd${ABC[C:C+8]} = vsubq_f16(vbr${ABC[C:C+8]}, vbl${ABC[C:C+8]});

        $for C in range(0, CHANNEL_TILE, 8):
          const float16x8_t vt${ABC[C:C+8]} = vfmaq_f16(vtl${ABC[C:C+8]}, vtd${ABC[C:C+8]}, valphah);
          const float16x8_t vb${ABC[C:C+8]} = vfmaq_f16(vbl${ABC[C:C+8]}, vbd${ABC[C:C+8]}, valphah);

        $for C in range(0, CHANNEL_TILE, 8):
          const float16x8_t vd${ABC[C:C+8]} = vsubq_f16(vb${ABC[C:C+8]}, vt${ABC[C:C+8]});

        $for C in range(0, CHANNEL_TILE, 8):
          const float16x8_t vo${ABC[C:C+8]} = vfmaq_f16(vt${ABC[C:C+8]}, vd${ABC[C:C+8]}, valphav);

        $for C in range(0, CHANNEL_TILE, 8):
          vst1q_u16(o, vreinterpretq_u16_f16(vo${ABC[C:C+8]})); o += 8;
      }
    for (; c >= 8 * sizeof(uint16_t); c -= 8 * sizeof(uint16_t)) {
      const float16x8_t vtl = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
      const float16x8_t vtr = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
      const float16x8_t vbl = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
      const float16x8_t vbr = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;

      const float16x8_t vtd = vsubq_f16(vtr, vtl);
      const float16x8_t vbd = vsubq_f16(vbr, vbl);

      const float16x8_t vt = vfmaq_f16(vtl, vtd, valphah);
      const float16x8_t vb = vfmaq_f16(vbl, vbd, valphah);

      const float16x8_t vd = vsubq_f16(vb, vt);

      const float16x8_t vo = vfmaq_f16(vt, vd, valphav);

      vst1q_u16(o, vreinterpretq_u16_f16(vo)); o += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      const float16x8_t vtl = vreinterpretq_f16_u16(vld1q_u16(i0));
      const float16x8_t vtr = vreinterpretq_f16_u16(vld1q_u16(i1));
      const float16x8_t vbl = vreinterpretq_f16_u16(vld1q_u16(i2));
      const float16x8_t vbr = vreinterpretq_f16_u16(vld1q_u16(i3));

      const float16x8_t vtd = vsubq_f16(vtr, vtl);
      const float16x8_t vbd = vsubq_f16(vbr, vbl);

      const float16x8_t vt = vfmaq_f16(vtl, vtd, valphah);
      const float16x8_t vb = vfmaq_f16(vbl, vbd, valphah);

      const float16x8_t vd = vsubq_f16(vb, vt);

      float16x8_t vo = vfmaq_f16(vt, vd, valphav);

      float16x4_t vo_lo = vget_low_f16(vo);
      if (c & (4 * sizeof(uint16_t))) {
        vst1_u16(o, vreinterpret_u16_f16(vo_lo)); o += 4;
        vo_lo = vget_high_f16(vo);
      }
      if (c & (2 * sizeof(uint16_t))) {
        vst1_lane_u32((void*) o, vreinterpret_u32_f16(vo_lo), 0); o += 2;
        vo_lo = vext_f16(vo_lo, vo_lo, 2);
      }
      if (c & (1 * sizeof(uint16_t))) {
        vst1_lane_u16(o, vreinterpret_u16_f16(vo_lo), 0); o += 1;
      }
    }

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