Hugging Face's logo

Androidonnxfork
/
test

Model card Files Community
5
test
File size: 8,314 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
 
// Auto-generated file. Do not edit!
//   Template: src/f16-f32-vcvt/wasmsimd-int16.c.in
//   Generator: tools/xngen
//
// Copyright 2021 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 <wasm_simd128.h>

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


void xnn_f16_f32_vcvt_ukernel__wasmrelaxedsimd_int16_x16(
    size_t batch,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_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 v128_t vsign_mask = wasm_v128_load64_splat(params->wasmsimd_int16.sign_mask);
  const v128_t vexp_offset = wasm_v128_load64_splat(params->wasmsimd_int16.exp_offset);
  const v128_t vexp_scale = wasm_v128_load64_splat(params->wasmsimd_int16.exp_scale);
  const v128_t vmagic_mask = wasm_v128_load64_splat(params->wasmsimd_int16.magic_mask);
  const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_int16.magic_bias);
  const v128_t vdenorm_cutoff = wasm_v128_load64_splat(params->wasmsimd_int16.denorm_cutoff);

  const uint16_t* i = (const uint16_t*) input;
  for (; batch >= 16 * sizeof(uint16_t); batch -= 16 * sizeof(uint16_t)) {
    const v128_t vh0 = wasm_v128_load(i);
    const v128_t vh1 = wasm_v128_load(i + 8);
    i += 16;

    const v128_t vsign0 = wasm_v128_and(vh0, vsign_mask);
    const v128_t vsign1 = wasm_v128_and(vh1, vsign_mask);

    const v128_t vnonsign0 = wasm_v128_xor(vh0, vsign0);
    const v128_t vnonsign1 = wasm_v128_xor(vh1, vsign1);

    const v128_t vprenorm0 = wasm_i16x8_shl(vnonsign0, 13);
    const v128_t vprenorm1 = wasm_i16x8_add(wasm_u16x8_shr(vnonsign0, 3), vexp_offset);
    const v128_t vprenorm2 = wasm_i16x8_shl(vnonsign1, 13);
    const v128_t vprenorm3 = wasm_i16x8_add(wasm_u16x8_shr(vnonsign1, 3), vexp_offset);

    const v128_t vnorm0 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm0, vprenorm1, 0,  8, 1,  9, 2, 10, 3, 11), vexp_scale);
    const v128_t vnorm1 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm0, vprenorm1, 4, 12, 5, 13, 6, 14, 7, 15), vexp_scale);
    const v128_t vnorm2 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm2, vprenorm3, 0,  8, 1,  9, 2, 10, 3, 11), vexp_scale);
    const v128_t vnorm3 = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm2, vprenorm3, 4, 12, 5, 13, 6, 14, 7, 15), vexp_scale);

    const v128_t vdenorm0 = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign0, vmagic_mask, 0,  8, 1,  9, 2, 10, 3, 11), vmagic_bias);
    const v128_t vdenorm1 = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign0, vmagic_mask, 4, 12, 5, 13, 6, 14, 7, 15), vmagic_bias);
    const v128_t vdenorm2 = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign1, vmagic_mask, 0,  8, 1,  9, 2, 10, 3, 11), vmagic_bias);
    const v128_t vdenorm3 = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign1, vmagic_mask, 4, 12, 5, 13, 6, 14, 7, 15), vmagic_bias);

    const v128_t vmask0 = wasm_i16x8_gt(vnonsign0, vdenorm_cutoff);
    const v128_t vmask1 = wasm_i16x8_gt(vnonsign1, vdenorm_cutoff);
    const v128_t vzero = wasm_i16x8_const_splat(0);

    const v128_t vxmask0 = wasm_i32x4_extend_low_i16x8(vmask0);
    const v128_t vxmask1 = wasm_i32x4_extend_high_i16x8(vmask0);
    const v128_t vxmask2 = wasm_i32x4_extend_low_i16x8(vmask1);
    const v128_t vxmask3 = wasm_i32x4_extend_high_i16x8(vmask1);

    const v128_t vabsf0 = wasm_i32x4_relaxed_laneselect(vnorm0, vdenorm0, vxmask0);
    const v128_t vsignf0 = wasm_v16x8_shuffle(vzero, vsign0, 0,  8, 1,  9, 2, 10, 3, 11);
    const v128_t vabsf1 = wasm_i32x4_relaxed_laneselect(vnorm1, vdenorm1, vxmask1);
    const v128_t vsignf1 = wasm_v16x8_shuffle(vzero, vsign0, 4, 12, 5, 13, 6, 14, 7, 15);
    const v128_t vabsf2 = wasm_i32x4_relaxed_laneselect(vnorm2, vdenorm2, vxmask2);
    const v128_t vsignf2 = wasm_v16x8_shuffle(vzero, vsign1, 0,  8, 1,  9, 2, 10, 3, 11);
    const v128_t vabsf3 = wasm_i32x4_relaxed_laneselect(vnorm3, vdenorm3, vxmask3);
    const v128_t vsignf3 = wasm_v16x8_shuffle(vzero, vsign1, 4, 12, 5, 13, 6, 14, 7, 15);

    const v128_t vf0 = wasm_v128_or(vsignf0, vabsf0);
    const v128_t vf1 = wasm_v128_or(vsignf1, vabsf1);
    const v128_t vf2 = wasm_v128_or(vsignf2, vabsf2);
    const v128_t vf3 = wasm_v128_or(vsignf3, vabsf3);

    wasm_v128_store(output, vf0);
    wasm_v128_store(output + 4, vf1);
    wasm_v128_store(output + 8, vf2);
    wasm_v128_store(output + 12, vf3);
    output += 16;
  }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const v128_t vh = wasm_v128_load(i);
    i += 8;

    const v128_t vsign = wasm_v128_and(vh, vsign_mask);

    const v128_t vnonsign = wasm_v128_xor(vh, vsign);

    const v128_t vprenorm_lo = wasm_i16x8_shl(vnonsign, 13);
    const v128_t vprenorm_hi = wasm_i16x8_add(wasm_u16x8_shr(vnonsign, 3), vexp_offset);

    const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0,  8, 1,  9, 2, 10, 3, 11), vexp_scale);
    const v128_t vnorm_hi = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 4, 12, 5, 13, 6, 14, 7, 15), vexp_scale);

    const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 0,  8, 1,  9, 2, 10, 3, 11), vmagic_bias);
    const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 4, 12, 5, 13, 6, 14, 7, 15), vmagic_bias);

    const v128_t vmask = wasm_i16x8_gt(vnonsign, vdenorm_cutoff);
    const v128_t vzero = wasm_i16x8_const_splat(0);

    const v128_t vxmask_lo = wasm_i32x4_extend_low_i16x8(vmask);
    const v128_t vxmask_hi = wasm_i32x4_extend_high_i16x8(vmask);

    const v128_t vabsf_lo = wasm_i32x4_relaxed_laneselect(vnorm_lo, vdenorm_lo, vxmask_lo);
    const v128_t vsignf_lo = wasm_v16x8_shuffle(vzero, vsign, 0,  8, 1,  9, 2, 10, 3, 11);
    const v128_t vabsf_hi = wasm_i32x4_relaxed_laneselect(vnorm_hi, vdenorm_hi, vxmask_hi);
    const v128_t vsignf_hi = wasm_v16x8_shuffle(vzero, vsign, 4, 12, 5, 13, 6, 14, 7, 15);

    const v128_t vf_lo = wasm_v128_or(vsignf_lo, vabsf_lo);
    const v128_t vf_hi = wasm_v128_or(vsignf_hi, vabsf_hi);

    wasm_v128_store(output, vf_lo);
    wasm_v128_store(output + 4, vf_hi);
    output += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));
    const v128_t vh = wasm_v128_load(i);

    const v128_t vsign = wasm_v128_and(vh, vsign_mask);

    const v128_t vnonsign = wasm_v128_xor(vh, vsign);

    const v128_t vprenorm_lo = wasm_i16x8_shl(vnonsign, 13);
    const v128_t vprenorm_hi = wasm_i16x8_add(wasm_u16x8_shr(vnonsign, 3), vexp_offset);

    const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0,  8, 1,  9, 2, 10, 3, 11), vexp_scale);
    const v128_t vnorm_hi = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 4, 12, 5, 13, 6, 14, 7, 15), vexp_scale);

    const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 0,  8, 1,  9, 2, 10, 3, 11), vmagic_bias);
    const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 4, 12, 5, 13, 6, 14, 7, 15), vmagic_bias);

    const v128_t vmask = wasm_i16x8_gt(vnonsign, vdenorm_cutoff);
    const v128_t vzero = wasm_i16x8_const_splat(0);

    const v128_t vxmask_lo = wasm_i32x4_extend_low_i16x8(vmask);
    const v128_t vxmask_hi = wasm_i32x4_extend_high_i16x8(vmask);

    const v128_t vabsf_lo = wasm_i32x4_relaxed_laneselect(vnorm_lo, vdenorm_lo, vxmask_lo);
    const v128_t vsignf_lo = wasm_v16x8_shuffle(vzero, vsign, 0,  8, 1,  9, 2, 10, 3, 11);
    const v128_t vabsf_hi = wasm_i32x4_relaxed_laneselect(vnorm_hi, vdenorm_hi, vxmask_hi);
    const v128_t vsignf_hi = wasm_v16x8_shuffle(vzero, vsign, 4, 12, 5, 13, 6, 14, 7, 15);

    v128_t vf = wasm_v128_or(vsignf_lo, vabsf_lo);

    if (batch & (4 * sizeof(uint16_t))) {
      wasm_v128_store(output, vf);
      output += 4;

      vf = wasm_v128_or(vsignf_hi, vabsf_hi);
    }
    if (batch & (2 * sizeof(uint16_t))) {
      wasm_v128_store64_lane(output, vf, 0);
      vf = wasm_v64x2_shuffle(vf, vf, 1, 1);
      output += 2;
    }
    if (batch & (1 * sizeof(uint16_t))) {
      wasm_v128_store32_lane(output, vf, 0);
    }
  }
}