File size: 7,412 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
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
// Auto-generated file. Do not edit!
//   Template: src/f16-vtanh/avx-expm1minus.c.in
//   Generator: tools/xngen
//
// 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.

#include <assert.h>
#include <stddef.h>
#include <stdint.h>

#include <immintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/microparams.h>
#include <xnnpack/vunary.h>

void xnn_f16_vtanh_ukernel__avx2_expm1minus_rr1_p3h2ts_div_x24(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_tanh_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->avx_expm1minus_rr1_p3h2.sign_mask);
  const __m256 vsat_cutoff = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.sat_cutoff);
  const __m256 vlog2e = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.log2e);
  const __m256 vmagic_bias = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.magic_bias);
  const __m256 vminus_ln2 = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.minus_ln2);
  const __m256 vc3 = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.c3);
  const __m256 vc2 = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.c2);
  const __m256 vtwo = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.two);
  const __m256 vminus_one = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.minus_one);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  for (; batch >= 24 * sizeof(uint16_t); batch -= 24 * sizeof(uint16_t)) {
    const __m128i vx0 = _mm_loadu_si128((const __m128i*) i);
    const __m128i vx1 = _mm_loadu_si128((const __m128i*) (i + 8));
    const __m128i vx2 = _mm_loadu_si128((const __m128i*) (i + 16));
    i += 24;

    const __m128i vabsx0 = _mm_or_si128(vx0, vsign_mask);
    const __m128i vabsx1 = _mm_or_si128(vx1, vsign_mask);
    const __m128i vabsx2 = _mm_or_si128(vx2, vsign_mask);

    __m256 vz0 = _mm256_cvtph_ps(vabsx0);
    const __m128i vinvsignx0 = _mm_xor_si128(vx0, vabsx0);
    __m256 vz1 = _mm256_cvtph_ps(vabsx1);
    const __m128i vinvsignx1 = _mm_xor_si128(vx1, vabsx1);
    __m256 vz2 = _mm256_cvtph_ps(vabsx2);
    const __m128i vinvsignx2 = _mm_xor_si128(vx2, vabsx2);

    vz0 = _mm256_max_ps(vsat_cutoff, vz0);
    __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
    vz1 = _mm256_max_ps(vsat_cutoff, vz1);
    __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
    vz2 = _mm256_max_ps(vsat_cutoff, vz2);
    __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);

    const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
    vn0 = _mm256_sub_ps(vn0, vmagic_bias);
    const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
    vn1 = _mm256_sub_ps(vn1, vmagic_bias);
    const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
    vn2 = _mm256_sub_ps(vn2, vmagic_bias);

    const __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
    const __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
    const __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);

    __m256 vp0 = vc3;
    __m256 vp1 = vc3;
    __m256 vp2 = vc3;
    vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
    vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
    vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
    vp0 = _mm256_fmadd_ps(vp0, vt0, vtwo);
    vp1 = _mm256_fmadd_ps(vp1, vt1, vtwo);
    vp2 = _mm256_fmadd_ps(vp2, vt2, vtwo);

    const __m256 vts0 = _mm256_mul_ps(vt0, vs0);
    const __m256 vsmo0 = _mm256_add_ps(vs0, vminus_one);
    const __m256 vts1 = _mm256_mul_ps(vt1, vs1);
    const __m256 vsmo1 = _mm256_add_ps(vs1, vminus_one);
    const __m256 vts2 = _mm256_mul_ps(vt2, vs2);
    const __m256 vsmo2 = _mm256_add_ps(vs2, vminus_one);
    const __m256 vemo0 = _mm256_fmadd_ps(vp0, vts0, vsmo0);
    const __m256 vemo1 = _mm256_fmadd_ps(vp1, vts1, vsmo1);
    const __m256 vemo2 = _mm256_fmadd_ps(vp2, vts2, vsmo2);

    const __m256 vepo0 = _mm256_add_ps(vemo0, vtwo);
    const __m256 vepo1 = _mm256_add_ps(vemo1, vtwo);
    const __m256 vepo2 = _mm256_add_ps(vemo2, vtwo);

    __m256 vy0 = _mm256_div_ps(vemo0, vepo0);
    __m256 vy1 = _mm256_div_ps(vemo1, vepo1);
    __m256 vy2 = _mm256_div_ps(vemo2, vepo2);


    __m128i vh0 = _mm256_cvtps_ph(vy0, _MM_FROUND_TO_NEAREST_INT);
    __m128i vh1 = _mm256_cvtps_ph(vy1, _MM_FROUND_TO_NEAREST_INT);
    __m128i vh2 = _mm256_cvtps_ph(vy2, _MM_FROUND_TO_NEAREST_INT);
    vh0 = _mm_xor_si128(vh0, vinvsignx0);
    vh1 = _mm_xor_si128(vh1, vinvsignx1);
    vh2 = _mm_xor_si128(vh2, vinvsignx2);

    _mm_storeu_si128((__m128i*) o, vh0);
    _mm_storeu_si128((__m128i*) (o + 8), vh1);
    _mm_storeu_si128((__m128i*) (o + 16), vh2);
    o += 24;
  }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const __m128i vx = _mm_loadu_si128((const __m128i*) i);
    i += 8;

    const __m128i vabsx = _mm_or_si128(vx, vsign_mask);
    __m256 vz = _mm256_cvtph_ps(vabsx);

    const __m128i vinvsignx = _mm_xor_si128(vx, vabsx);
    vz = _mm256_max_ps(vsat_cutoff, vz);

    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);

    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));

    vn = _mm256_sub_ps(vn, vmagic_bias);

    const __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);

    __m256 vp = vc3;
    vp = _mm256_fmadd_ps(vp, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vtwo);

    const __m256 vts = _mm256_mul_ps(vt, vs);
    const __m256 vsmo = _mm256_add_ps(vs, vminus_one);
    const __m256 vemo = _mm256_fmadd_ps(vp, vts, vsmo);

    const __m256 vepo = _mm256_add_ps(vemo, vtwo);

    __m256 vy = _mm256_div_ps(vemo, vepo);


    __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_TO_NEAREST_INT);
    vh = _mm_xor_si128(vh, vinvsignx);

    _mm_storeu_si128((__m128i*) o, vh);
    o += 8;
  }
  if (batch != 0) {
    const __m128i vx = _mm_loadu_si128((const __m128i*) i);

    const __m128i vabsx = _mm_or_si128(vx, vsign_mask);
    __m256 vz = _mm256_cvtph_ps(vabsx);

    const __m128i vinvsignx = _mm_xor_si128(vx, vabsx);
    vz = _mm256_max_ps(vsat_cutoff, vz);

    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);

    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));

    vn = _mm256_sub_ps(vn, vmagic_bias);

    const __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);

    __m256 vp = vc3;
    vp = _mm256_fmadd_ps(vp, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vtwo);

    const __m256 vts = _mm256_mul_ps(vt, vs);
    const __m256 vsmo = _mm256_add_ps(vs, vminus_one);
    const __m256 vemo = _mm256_fmadd_ps(vp, vts, vsmo);

    const __m256 vepo = _mm256_add_ps(vemo, vtwo);

    __m256 vy = _mm256_div_ps(vemo, vepo);


    __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_TO_NEAREST_INT);
    vh = _mm_xor_si128(vh, vinvsignx);

    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storel_epi64((__m128i*) o, vh);
      vh = _mm_unpackhi_epi64(vh, vh);
      o += 4;
    }
    if (batch & (2 * sizeof(uint16_t))) {
      _mm_storeu_si32(o, vh);
      vh = _mm_srli_epi64(vh, 32);
      o += 2;
    }
    if (batch & (1 * sizeof(uint16_t))) {
      *o = (uint16_t) _mm_extract_epi16(vh, 0);
    }
  }
}