File size: 16,642 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
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
// 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__f16c_expm1minus_rr1_p3h2ts_rcp_x64(
    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 >= 64 * sizeof(uint16_t); batch -= 64 * 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));
    const __m128i vx3 = _mm_loadu_si128((const __m128i*) (i + 24));
    const __m128i vx4 = _mm_loadu_si128((const __m128i*) (i + 32));
    const __m128i vx5 = _mm_loadu_si128((const __m128i*) (i + 40));
    const __m128i vx6 = _mm_loadu_si128((const __m128i*) (i + 48));
    const __m128i vx7 = _mm_loadu_si128((const __m128i*) (i + 56));
    i += 64;

    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);
    const __m128i vabsx3 = _mm_or_si128(vx3, vsign_mask);
    const __m128i vabsx4 = _mm_or_si128(vx4, vsign_mask);
    const __m128i vabsx5 = _mm_or_si128(vx5, vsign_mask);
    const __m128i vabsx6 = _mm_or_si128(vx6, vsign_mask);
    const __m128i vabsx7 = _mm_or_si128(vx7, 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);
    __m256 vz3 = _mm256_cvtph_ps(vabsx3);
    const __m128i vinvsignx3 = _mm_xor_si128(vx3, vabsx3);
    __m256 vz4 = _mm256_cvtph_ps(vabsx4);
    const __m128i vinvsignx4 = _mm_xor_si128(vx4, vabsx4);
    __m256 vz5 = _mm256_cvtph_ps(vabsx5);
    const __m128i vinvsignx5 = _mm_xor_si128(vx5, vabsx5);
    __m256 vz6 = _mm256_cvtph_ps(vabsx6);
    const __m128i vinvsignx6 = _mm_xor_si128(vx6, vabsx6);
    __m256 vz7 = _mm256_cvtph_ps(vabsx7);
    const __m128i vinvsignx7 = _mm_xor_si128(vx7, vabsx7);

    const __m256 vm0 = _mm256_cmp_ps(vz0, vsat_cutoff, _CMP_LE_OS);
    __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
    const __m256 vm1 = _mm256_cmp_ps(vz1, vsat_cutoff, _CMP_LE_OS);
    __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
    const __m256 vm2 = _mm256_cmp_ps(vz2, vsat_cutoff, _CMP_LE_OS);
    __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
    const __m256 vm3 = _mm256_cmp_ps(vz3, vsat_cutoff, _CMP_LE_OS);
    __m256 vn3 = _mm256_add_ps(_mm256_mul_ps(vz3, vlog2e), vmagic_bias);
    const __m256 vm4 = _mm256_cmp_ps(vz4, vsat_cutoff, _CMP_LE_OS);
    __m256 vn4 = _mm256_add_ps(_mm256_mul_ps(vz4, vlog2e), vmagic_bias);
    const __m256 vm5 = _mm256_cmp_ps(vz5, vsat_cutoff, _CMP_LE_OS);
    __m256 vn5 = _mm256_add_ps(_mm256_mul_ps(vz5, vlog2e), vmagic_bias);
    const __m256 vm6 = _mm256_cmp_ps(vz6, vsat_cutoff, _CMP_LE_OS);
    __m256 vn6 = _mm256_add_ps(_mm256_mul_ps(vz6, vlog2e), vmagic_bias);
    const __m256 vm7 = _mm256_cmp_ps(vz7, vsat_cutoff, _CMP_LE_OS);
    __m256 vn7 = _mm256_add_ps(_mm256_mul_ps(vz7, vlog2e), vmagic_bias);

    const __m128 vn0_hi = _mm256_extractf128_ps(vn0, 1);
    __m256 vs0 = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23)));
    vn0 = _mm256_sub_ps(vn0, vmagic_bias);
    const __m128 vn1_hi = _mm256_extractf128_ps(vn1, 1);
    __m256 vs1 = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23)));
    vn1 = _mm256_sub_ps(vn1, vmagic_bias);
    const __m128 vn2_hi = _mm256_extractf128_ps(vn2, 1);
    __m256 vs2 = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)), 23)));
    vn2 = _mm256_sub_ps(vn2, vmagic_bias);
    const __m128 vn3_hi = _mm256_extractf128_ps(vn3, 1);
    __m256 vs3 = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn3)), 23)));
    vn3 = _mm256_sub_ps(vn3, vmagic_bias);
    const __m128 vn4_hi = _mm256_extractf128_ps(vn4, 1);
    __m256 vs4 = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn4)), 23)));
    vn4 = _mm256_sub_ps(vn4, vmagic_bias);
    const __m128 vn5_hi = _mm256_extractf128_ps(vn5, 1);
    __m256 vs5 = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn5)), 23)));
    vn5 = _mm256_sub_ps(vn5, vmagic_bias);
    const __m128 vn6_hi = _mm256_extractf128_ps(vn6, 1);
    __m256 vs6 = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn6)), 23)));
    vn6 = _mm256_sub_ps(vn6, vmagic_bias);
    const __m128 vn7_hi = _mm256_extractf128_ps(vn7, 1);
    __m256 vs7 = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn7)), 23)));
    vn7 = _mm256_sub_ps(vn7, vmagic_bias);

    const __m128 vs0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0_hi), 23));
    const __m128 vs1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn1_hi), 23));
    const __m128 vs2_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn2_hi), 23));
    const __m128 vs3_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn3_hi), 23));
    const __m128 vs4_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4_hi), 23));
    const __m128 vs5_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn5_hi), 23));
    const __m128 vs6_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn6_hi), 23));
    const __m128 vs7_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn7_hi), 23));

    vs0 = _mm256_insertf128_ps(vs0, vs0_hi, 1);
    vs1 = _mm256_insertf128_ps(vs1, vs1_hi, 1);
    vs2 = _mm256_insertf128_ps(vs2, vs2_hi, 1);
    vs3 = _mm256_insertf128_ps(vs3, vs3_hi, 1);
    vs4 = _mm256_insertf128_ps(vs4, vs4_hi, 1);
    vs5 = _mm256_insertf128_ps(vs5, vs5_hi, 1);
    vs6 = _mm256_insertf128_ps(vs6, vs6_hi, 1);
    vs7 = _mm256_insertf128_ps(vs7, vs7_hi, 1);

    const __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2), vz0);
    const __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2), vz1);
    const __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2), vz2);
    const __m256 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2), vz3);
    const __m256 vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2), vz4);
    const __m256 vt5 = _mm256_add_ps(_mm256_mul_ps(vn5, vminus_ln2), vz5);
    const __m256 vt6 = _mm256_add_ps(_mm256_mul_ps(vn6, vminus_ln2), vz6);
    const __m256 vt7 = _mm256_add_ps(_mm256_mul_ps(vn7, vminus_ln2), vz7);

    __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc3, vt0), vc2);
    __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc3, vt1), vc2);
    __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc3, vt2), vc2);
    __m256 vp3 = _mm256_add_ps(_mm256_mul_ps(vc3, vt3), vc2);
    __m256 vp4 = _mm256_add_ps(_mm256_mul_ps(vc3, vt4), vc2);
    __m256 vp5 = _mm256_add_ps(_mm256_mul_ps(vc3, vt5), vc2);
    __m256 vp6 = _mm256_add_ps(_mm256_mul_ps(vc3, vt6), vc2);
    __m256 vp7 = _mm256_add_ps(_mm256_mul_ps(vc3, vt7), vc2);
    vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vtwo);
    vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vtwo);
    vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vtwo);
    vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vtwo);
    vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vtwo);
    vp5 = _mm256_add_ps(_mm256_mul_ps(vp5, vt5), vtwo);
    vp6 = _mm256_add_ps(_mm256_mul_ps(vp6, vt6), vtwo);
    vp7 = _mm256_add_ps(_mm256_mul_ps(vp7, vt7), 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 vts3 = _mm256_mul_ps(vt3, vs3);
    const __m256 vsmo3 = _mm256_add_ps(vs3, vminus_one);
    const __m256 vts4 = _mm256_mul_ps(vt4, vs4);
    const __m256 vsmo4 = _mm256_add_ps(vs4, vminus_one);
    const __m256 vts5 = _mm256_mul_ps(vt5, vs5);
    const __m256 vsmo5 = _mm256_add_ps(vs5, vminus_one);
    const __m256 vts6 = _mm256_mul_ps(vt6, vs6);
    const __m256 vsmo6 = _mm256_add_ps(vs6, vminus_one);
    const __m256 vts7 = _mm256_mul_ps(vt7, vs7);
    const __m256 vsmo7 = _mm256_add_ps(vs7, vminus_one);
    const __m256 vemo0 = _mm256_add_ps(_mm256_mul_ps(vp0, vts0), vsmo0);
    const __m256 vemo1 = _mm256_add_ps(_mm256_mul_ps(vp1, vts1), vsmo1);
    const __m256 vemo2 = _mm256_add_ps(_mm256_mul_ps(vp2, vts2), vsmo2);
    const __m256 vemo3 = _mm256_add_ps(_mm256_mul_ps(vp3, vts3), vsmo3);
    const __m256 vemo4 = _mm256_add_ps(_mm256_mul_ps(vp4, vts4), vsmo4);
    const __m256 vemo5 = _mm256_add_ps(_mm256_mul_ps(vp5, vts5), vsmo5);
    const __m256 vemo6 = _mm256_add_ps(_mm256_mul_ps(vp6, vts6), vsmo6);
    const __m256 vemo7 = _mm256_add_ps(_mm256_mul_ps(vp7, vts7), vsmo7);

    const __m256 vepo0 = _mm256_add_ps(vemo0, vtwo);
    const __m256 vepo1 = _mm256_add_ps(vemo1, vtwo);
    const __m256 vepo2 = _mm256_add_ps(vemo2, vtwo);
    const __m256 vepo3 = _mm256_add_ps(vemo3, vtwo);
    const __m256 vepo4 = _mm256_add_ps(vemo4, vtwo);
    const __m256 vepo5 = _mm256_add_ps(vemo5, vtwo);
    const __m256 vepo6 = _mm256_add_ps(vemo6, vtwo);
    const __m256 vepo7 = _mm256_add_ps(vemo7, vtwo);

    __m256 vrepo0 = _mm256_rcp_ps(vepo0);
    __m256 vrepo1 = _mm256_rcp_ps(vepo1);
    __m256 vrepo2 = _mm256_rcp_ps(vepo2);
    __m256 vrepo3 = _mm256_rcp_ps(vepo3);
    __m256 vrepo4 = _mm256_rcp_ps(vepo4);
    __m256 vrepo5 = _mm256_rcp_ps(vepo5);
    __m256 vrepo6 = _mm256_rcp_ps(vepo6);
    __m256 vrepo7 = _mm256_rcp_ps(vepo7);

    __m256 vy0 = _mm256_mul_ps(vemo0, vrepo0);
    __m256 vy1 = _mm256_mul_ps(vemo1, vrepo1);
    __m256 vy2 = _mm256_mul_ps(vemo2, vrepo2);
    __m256 vy3 = _mm256_mul_ps(vemo3, vrepo3);
    __m256 vy4 = _mm256_mul_ps(vemo4, vrepo4);
    __m256 vy5 = _mm256_mul_ps(vemo5, vrepo5);
    __m256 vy6 = _mm256_mul_ps(vemo6, vrepo6);
    __m256 vy7 = _mm256_mul_ps(vemo7, vrepo7);

    vy0 = _mm256_blendv_ps(vy0, vminus_one, vm0);
    vy1 = _mm256_blendv_ps(vy1, vminus_one, vm1);
    vy2 = _mm256_blendv_ps(vy2, vminus_one, vm2);
    vy3 = _mm256_blendv_ps(vy3, vminus_one, vm3);
    vy4 = _mm256_blendv_ps(vy4, vminus_one, vm4);
    vy5 = _mm256_blendv_ps(vy5, vminus_one, vm5);
    vy6 = _mm256_blendv_ps(vy6, vminus_one, vm6);
    vy7 = _mm256_blendv_ps(vy7, vminus_one, vm7);

    __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);
    __m128i vh3 = _mm256_cvtps_ph(vy3, _MM_FROUND_TO_NEAREST_INT);
    __m128i vh4 = _mm256_cvtps_ph(vy4, _MM_FROUND_TO_NEAREST_INT);
    __m128i vh5 = _mm256_cvtps_ph(vy5, _MM_FROUND_TO_NEAREST_INT);
    __m128i vh6 = _mm256_cvtps_ph(vy6, _MM_FROUND_TO_NEAREST_INT);
    __m128i vh7 = _mm256_cvtps_ph(vy7, _MM_FROUND_TO_NEAREST_INT);
    vh0 = _mm_xor_si128(vh0, vinvsignx0);
    vh1 = _mm_xor_si128(vh1, vinvsignx1);
    vh2 = _mm_xor_si128(vh2, vinvsignx2);
    vh3 = _mm_xor_si128(vh3, vinvsignx3);
    vh4 = _mm_xor_si128(vh4, vinvsignx4);
    vh5 = _mm_xor_si128(vh5, vinvsignx5);
    vh6 = _mm_xor_si128(vh6, vinvsignx6);
    vh7 = _mm_xor_si128(vh7, vinvsignx7);

    _mm_storeu_si128((__m128i*) o, vh0);
    _mm_storeu_si128((__m128i*) (o + 8), vh1);
    _mm_storeu_si128((__m128i*) (o + 16), vh2);
    _mm_storeu_si128((__m128i*) (o + 24), vh3);
    _mm_storeu_si128((__m128i*) (o + 32), vh4);
    _mm_storeu_si128((__m128i*) (o + 40), vh5);
    _mm_storeu_si128((__m128i*) (o + 48), vh6);
    _mm_storeu_si128((__m128i*) (o + 56), vh7);
    o += 64;
  }
  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);
    const __m256 vm = _mm256_cmp_ps(vz, vsat_cutoff, _CMP_LE_OS);

    __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);

    const __m128 vn_hi = _mm256_extractf128_ps(vn, 1);
    __m256 vs = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23)));
    const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn_hi), 23));
    vs = _mm256_insertf128_ps(vs, vs_hi, 1);

    vn = _mm256_sub_ps(vn, vmagic_bias);

    const __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2), vz);

    __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc3, vt), vc2);
    vp = _mm256_add_ps(_mm256_mul_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_add_ps(_mm256_mul_ps(vp, vts), vsmo);

    const __m256 vepo = _mm256_add_ps(vemo, vtwo);

    __m256 vrepo = _mm256_rcp_ps(vepo);

    __m256 vy = _mm256_mul_ps(vemo, vrepo);

    vy = _mm256_blendv_ps(vy, vminus_one, vm);

    __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);
    const __m256 vm = _mm256_cmp_ps(vz, vsat_cutoff, _CMP_LE_OS);

    __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);

    const __m128 vn_hi = _mm256_extractf128_ps(vn, 1);
    __m256 vs = _mm256_castps128_ps256(_mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23)));
    const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn_hi), 23));
    vs = _mm256_insertf128_ps(vs, vs_hi, 1);

    vn = _mm256_sub_ps(vn, vmagic_bias);

    const __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2), vz);

    __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc3, vt), vc2);
    vp = _mm256_add_ps(_mm256_mul_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_add_ps(_mm256_mul_ps(vp, vts), vsmo);

    const __m256 vepo = _mm256_add_ps(vemo, vtwo);

    __m256 vrepo = _mm256_rcp_ps(vepo);

    __m256 vy = _mm256_mul_ps(vemo, vrepo);

    vy = _mm256_blendv_ps(vy, vminus_one, vm);

    __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);
    }
  }
}