Hugging Face's logo

Androidonnxfork
/
test

Model card Files Community
5
test
File size: 7,072 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
 
// Copyright 2023 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 (P, H) in [(3, 2)]
$assert not PS or (P, H) == (4, 3)
$assert DIV in ["DIV", "RECPEADJ", "NR1RECPS", "NR1FMA"]
$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

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

$POLY_SUFFIX = "p%dh%d%s" % (P, H, "ps" if PS else "ts")
$DIV_SUFFIX = DIV.lower()
$ISA = "aarch64_neonfp16arith" if DIV == "DIV" else "neonfp16arith"
void xnn_f16_vtanh_ukernel__${ISA}_expm1minus_rr1_${POLY_SUFFIX}_${DIV_SUFFIX}_x${BATCH_TILE}(
    size_t n,
    const void* input,
    void* output,
    const union xnn_f16_tanh_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const float16x8_t vsat_cutoff = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x4482)));
  const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x620F)));
  const float16x8_t vminus_log2e = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0xBDC5)));
  const float16x8_t vln2 = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x398C)));
  const float16x8_t vc3 = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0xBD5B)));
  const float16x8_t vc2 = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x4008)));
  const float16x8_t vtwo = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x4000)));
  const float16x8_t vminus_one = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0xBC00)));
  const uint16x8_t vsign_mask = vmovq_n_u16(UINT16_C(0x8000));

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  $for SIMD_TILE_LEVEL in (SIMD_TILE, 1) if SIMD_TILE > 1 else (1,):
    $if SIMD_TILE_LEVEL == 1:
      $ABC =  ["" for a in ABC]
    for (; n >= ${SIMD_TILE_LEVEL} * sizeof(float16x8_t); n -= ${SIMD_TILE_LEVEL} * sizeof(float16x8_t)) {
      $for N in range(SIMD_TILE_LEVEL):
        const float16x8_t vx${ABC[N]} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

      $for N in range(SIMD_TILE_LEVEL):
        float16x8_t vz${ABC[N]} = vabsq_f16(vx${ABC[N]});

      $for N in range(SIMD_TILE_LEVEL):
        vz${ABC[N]} = vminq_f16(vz${ABC[N]}, vsat_cutoff);

      $for N in range(SIMD_TILE_LEVEL):
        float16x8_t vn${ABC[N]} = vfmaq_f16(vmagic_bias, vz${ABC[N]}, vminus_log2e);

      $for N in range(SIMD_TILE_LEVEL):
        const float16x8_t vs${ABC[N]} = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn${ABC[N]}), 10));
        vn${ABC[N]} = vsubq_f16(vn${ABC[N]}, vmagic_bias);
      $for N in range(SIMD_TILE_LEVEL):
        const float16x8_t vt${ABC[N]} = vfmaq_f16(vz${ABC[N]}, vn${ABC[N]}, vln2);
      $for N in range(SIMD_TILE_LEVEL):
        float16x8_t vp${ABC[N]} = vfmaq_f16(vc${P-1}, vc${P}, vt${ABC[N]});
      $for i in reversed(range(2, P - 1)):
        $for N in range(SIMD_TILE_LEVEL):
          vp${ABC[N]} = vfmaq_f16(vc${i}, vp${ABC[N]}, vt${ABC[N]});
      $for N in range(SIMD_TILE_LEVEL):
        vp${ABC[N]} = vfmsq_f16(vtwo, vp${ABC[N]}, vt${ABC[N]});

      $for N in range(SIMD_TILE_LEVEL):
        const float16x8_t vts${ABC[N]} = vmulq_f16(vt${ABC[N]}, vs${ABC[N]});
        const float16x8_t vsmo${ABC[N]} = vaddq_f16(vs${ABC[N]}, vminus_one);
      $for N in range(SIMD_TILE_LEVEL):
        const float16x8_t vemo${ABC[N]} = vfmsq_f16(vsmo${ABC[N]}, vp${ABC[N]}, vts${ABC[N]});
      $for N in range(SIMD_TILE_LEVEL):
        const float16x8_t vepo${ABC[N]} = vaddq_f16(vemo${ABC[N]}, vtwo);

      $if DIV == "DIV":
        $for N in range(SIMD_TILE_LEVEL):
          float16x8_t vy${ABC[N]} = vdivq_f16(vemo${ABC[N]}, vepo${ABC[N]});
      $else:
        $for N in range(SIMD_TILE_LEVEL):
          float16x8_t vrepo${ABC[N]} = vrecpeq_f16(vepo${ABC[N]});
        $if DIV.startswith("NR1RECPS"):
          $for N in range(SIMD_TILE_LEVEL):
            const float16x8_t verepo${ABC[N]} = vrecpsq_f16(vrepo${ABC[N]}, vepo${ABC[N]});
          $for N in range(SIMD_TILE_LEVEL):
            vrepo${ABC[N]} = vmulq_f16(vrepo${ABC[N]}, verepo${ABC[N]});
        $elif DIV.startswith("NR1FMA"):
          $for N in range(SIMD_TILE_LEVEL):
            const float16x8_t verepo${ABC[N]} = vfmaq_f16(vminus_one, vrepo${ABC[N]}, vepo${ABC[N]});
          $for N in range(SIMD_TILE_LEVEL):
            vrepo${ABC[N]} = vfmsq_f16(vrepo${ABC[N]}, vrepo${ABC[N]}, verepo${ABC[N]});

        $for N in range(SIMD_TILE_LEVEL):
          float16x8_t vy${ABC[N]} = vmulq_f16(vemo${ABC[N]}, vrepo${ABC[N]});

        $if DIV.endswith("ADJ"):
          $for N in range(SIMD_TILE_LEVEL):
            const float16x8_t vey${ABC[N]} = vfmsq_f16(vemo${ABC[N]}, vy${ABC[N]}, vepo${ABC[N]});
          $for N in range(SIMD_TILE_LEVEL):
            vy${ABC[N]} = vfmaq_f16(vy${ABC[N]}, vey${ABC[N]}, vrepo${ABC[N]});

      $for N in range(SIMD_TILE_LEVEL):
        vy${ABC[N]} = vbslq_f16(vsign_mask, vx${ABC[N]}, vy${ABC[N]});

      $for N in range(SIMD_TILE_LEVEL):
        vst1q_u16(o, vreinterpretq_u16_f16(vy${ABC[N]})); o += 8;
    }

  if (n != 0) {
    const float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    float16x8_t vz = vabsq_f16(vx);
    vz = vminq_f16(vz, vsat_cutoff);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
    const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
    vn = vsubq_f16(vn, vmagic_bias);
    const float16x8_t vt = vfmaq_f16(vz, vn, vln2);
    float16x8_t vp = vfmaq_f16(vc${P-1}, vc${P}, vt);
    $for i in reversed(range(2, P - 1)):
      vp = vfmaq_f16(vc${i}, vp, vt);
    vp = vfmsq_f16(vtwo, vp, vt);

    const float16x8_t vts = vmulq_f16(vt, vs);
    const float16x8_t vsmo = vaddq_f16(vs, vminus_one);
    const float16x8_t vemo = vfmsq_f16(vsmo, vp, vts);
    const float16x8_t vepo = vaddq_f16(vemo, vtwo);

    $if DIV == "DIV":
      float16x8_t vy = vdivq_f16(vemo, vepo);
    $else:
      float16x8_t vrepo = vrecpeq_f16(vepo);
      $if DIV.startswith("NR1RECPS"):
        const float16x8_t verepo = vrecpsq_f16(vrepo, vepo);
        vrepo = vmulq_f16(vrepo, verepo);
      $elif DIV.startswith("NR1FMA"):
        const float16x8_t verepo = vfmaq_f16(vminus_one, vrepo, vepo);
        vrepo = vfmsq_f16(vrepo, vrepo, verepo);

      float16x8_t vy = vmulq_f16(vemo, vrepo);

      $if DIV.endswith("ADJ"):
        const float16x8_t vey = vfmsq_f16(vemo, vy, vepo);
        vy = vfmaq_f16(vy, vey, vrepo);

    vy = vbslq_f16(vsign_mask, vx, vy);

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