File size: 8,650 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
// Auto-generated file. Do not edit!
//   Template: src/f16-f32-vcvt/sse-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 <smmintrin.h>

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


void xnn_f16_f32_vcvt_ukernel__avx_int16_x24(
    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 __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse_int16.sign_mask);
  const __m128i vexp_offset = _mm_load_si128((const __m128i*) params->sse_int16.exp_offset);
  const __m128 vexp_scale = _mm_load_ps(params->sse_int16.exp_scale);
  const __m128i vmagic_mask = _mm_load_si128((const __m128i*) params->sse_int16.magic_mask);
  const __m128 vmagic_bias = _mm_load_ps(params->sse_int16.magic_bias);
  const __m128i vdenorm_cutoff = _mm_load_si128((const __m128i*) params->sse_int16.denorm_cutoff);

  const uint16_t* i = (const uint16_t*) input;
  for (; batch >= 24 * sizeof(uint16_t); batch -= 24 * sizeof(uint16_t)) {
    const __m128i vh0 = _mm_loadu_si128((const __m128i*) i);
    const __m128i vh1 = _mm_loadu_si128((const __m128i*) (i + 8));
    const __m128i vh2 = _mm_loadu_si128((const __m128i*) (i + 16));
    i += 24;

    const __m128i vsign0 = _mm_and_si128(vh0, vsign_mask);
    const __m128i vsign1 = _mm_and_si128(vh1, vsign_mask);
    const __m128i vsign2 = _mm_and_si128(vh2, vsign_mask);

    const __m128i vnonsign0 = _mm_xor_si128(vh0, vsign0);
    const __m128i vnonsign1 = _mm_xor_si128(vh1, vsign1);
    const __m128i vnonsign2 = _mm_xor_si128(vh2, vsign2);

    const __m128i vprenorm0 = _mm_slli_epi16(vnonsign0, 13);
    const __m128i vprenorm1 = _mm_add_epi16(_mm_srli_epi16(vnonsign0, 3), vexp_offset);
    const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13);
    const __m128i vprenorm3 = _mm_add_epi16(_mm_srli_epi16(vnonsign1, 3), vexp_offset);
    const __m128i vprenorm4 = _mm_slli_epi16(vnonsign2, 13);
    const __m128i vprenorm5 = _mm_add_epi16(_mm_srli_epi16(vnonsign2, 3), vexp_offset);

    const __m128i vnorm0 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm0, vprenorm1)), vexp_scale));
    const __m128i vnorm1 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm0, vprenorm1)), vexp_scale));
    const __m128i vnorm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), vexp_scale));
    const __m128i vnorm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), vexp_scale));
    const __m128i vnorm4 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm4, vprenorm5)), vexp_scale));
    const __m128i vnorm5 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm4, vprenorm5)), vexp_scale));

    const __m128i vdenorm0 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign0, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm1 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign0, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm3 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign1, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm5 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign2, vmagic_mask)), vmagic_bias));

    const __m128i vmask0 = _mm_cmpgt_epi16(vnonsign0, vdenorm_cutoff);
    const __m128i vmask1 = _mm_cmpgt_epi16(vnonsign1, vdenorm_cutoff);
    const __m128i vmask2 = _mm_cmpgt_epi16(vnonsign2, vdenorm_cutoff);

    const __m128i vf0 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign0),
      _mm_blendv_epi8(vdenorm0, vnorm0, _mm_cvtepi16_epi32(vmask0)));
    const __m128i vf1 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign0),
      _mm_blendv_epi8(vdenorm1, vnorm1, _mm_unpackhi_epi16(vmask0, vmask0)));
    const __m128i vf2 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign1),
      _mm_blendv_epi8(vdenorm2, vnorm2, _mm_cvtepi16_epi32(vmask1)));
    const __m128i vf3 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign1),
      _mm_blendv_epi8(vdenorm3, vnorm3, _mm_unpackhi_epi16(vmask1, vmask1)));
    const __m128i vf4 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign2),
      _mm_blendv_epi8(vdenorm4, vnorm4, _mm_cvtepi16_epi32(vmask2)));
    const __m128i vf5 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign2),
      _mm_blendv_epi8(vdenorm5, vnorm5, _mm_unpackhi_epi16(vmask2, vmask2)));

    _mm_storeu_ps(output, _mm_castsi128_ps(vf0));
    _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf1));
    _mm_storeu_ps(output + 8, _mm_castsi128_ps(vf2));
    _mm_storeu_ps(output + 12, _mm_castsi128_ps(vf3));
    _mm_storeu_ps(output + 16, _mm_castsi128_ps(vf4));
    _mm_storeu_ps(output + 20, _mm_castsi128_ps(vf5));
    output += 24;
  }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const __m128i vh = _mm_loadu_si128((const __m128i*) i);
    i += 8;

    const __m128i vsign = _mm_and_si128(vh, vsign_mask);

    const __m128i vnonsign = _mm_xor_si128(vh, vsign);

    const __m128i vprenorm_lo = _mm_slli_epi16(vnonsign, 13);
    const __m128i vprenorm_hi = _mm_add_epi16(_mm_srli_epi16(vnonsign, 3), vexp_offset);

    const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));
    const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));

    const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign, vmagic_mask)), vmagic_bias));

    const __m128i vmask = _mm_cmpgt_epi16(vnonsign, vdenorm_cutoff);

    const __m128i vf_lo = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign),
      _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask)));

    const __m128i vf_hi = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign),
      _mm_blendv_epi8(vdenorm_hi, vnorm_hi, _mm_unpackhi_epi16(vmask, vmask)));

    _mm_storeu_ps(output, _mm_castsi128_ps(vf_lo));
    _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf_hi));
    output += 8;
  }
  if XNN_UNPREDICTABLE(batch != 0) {
    const __m128i vh = _mm_loadu_si128((const __m128i*) i);

    const __m128i vsign = _mm_and_si128(vh, vsign_mask);

    const __m128i vnonsign = _mm_xor_si128(vh, vsign);

    const __m128i vprenorm_lo = _mm_slli_epi16(vnonsign, 13);
    const __m128i vprenorm_hi = _mm_add_epi16(_mm_srli_epi16(vnonsign, 3), vexp_offset);

    const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));
    const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));

    const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign, vmagic_mask)), vmagic_bias));

    const __m128i vmask = _mm_cmpgt_epi16(vnonsign, vdenorm_cutoff);

    __m128i vf = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign),
      _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask)));

    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storeu_ps(output, _mm_castsi128_ps(vf));
      output += 4;

      vf = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign),
        _mm_blendv_epi8(vdenorm_hi, vnorm_hi, _mm_unpackhi_epi16(vmask, vmask)));
    }
    if (batch & (2 * sizeof(uint16_t))) {
      _mm_storel_pi((__m64*) output, _mm_castsi128_ps(vf));
      output += 2;

      vf = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(vf), _mm_castsi128_ps(vf)));
    }
    if (batch & (1 * sizeof(uint16_t))) {
      _mm_store_ss(output, _mm_castsi128_ps(vf));
    }
  }
}