File size: 3,495 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
// Auto-generated file. Do not edit!
//   Template: src/f16-vhswish/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/common.h>
#include <xnnpack/vunary.h>


void xnn_f16_vhswish_ukernel__neonfp16arith_x16(
    size_t batch,
    const void* restrict input,
    void* restrict output,
    const union xnn_f16_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;

  const float16x8_t vsixth = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.sixth));
  const float16x8_t vthree = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.three));
  const int16x8_t vsix = vreinterpretq_s16_u16(vld1q_dup_u16(&params->fp16arith.six));
  const int16x8_t vzero = vdupq_n_s16(0);

  for (; batch >= 16 * sizeof(uint16_t); batch -= 16 * sizeof(uint16_t)) {
    float16x8_t vx01234567 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    float16x8_t vx89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

    float16x8_t vacc01234567 = vaddq_f16(vx01234567, vthree);
    vx01234567 = vmulq_f16(vx01234567, vsixth);
    float16x8_t vacc89ABCDEF = vaddq_f16(vx89ABCDEF, vthree);
    vx89ABCDEF = vmulq_f16(vx89ABCDEF, vsixth);

    vacc01234567 = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc01234567), vzero));
    vacc89ABCDEF = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc89ABCDEF), vzero));

    vacc01234567 = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc01234567), vsix));
    vacc89ABCDEF = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc89ABCDEF), vsix));

    vacc01234567 = vmulq_f16(vacc01234567, vx01234567);
    vacc89ABCDEF = vmulq_f16(vacc89ABCDEF, vx89ABCDEF);

    vst1q_u16(o, vreinterpretq_u16_f16(vacc01234567)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vacc89ABCDEF)); o += 8;
  }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    float16x8_t vacc = vaddq_f16(vx, vthree);
    vx = vmulq_f16(vx, vsixth);
    vacc = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc), vzero));
    vacc = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc), vsix));
    vacc = vmulq_f16(vacc, vx);
    vst1q_u16(o, vreinterpretq_u16_f16(vacc)); o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i));
    float16x8_t vacc = vaddq_f16(vx, vthree);
    vx = vmulq_f16(vx, vsixth);
    vacc = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc), vzero));
    vacc = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc), vsix));
    vacc = vmulq_f16(vacc, vx);

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