Hugging Face's logo

Androidonnxfork
/
test

Model card Files Community
6
test
File size: 14,876 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
 
// Auto-generated file. Do not edit!
//   Template: src/f32-f16-vcvt/sse.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 <emmintrin.h>

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


void xnn_f32_f16_vcvt_ukernel__sse2_x32(
    size_t batch,
    const float* input,
    void* output,
    const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const __m128 vnonsign_mask = _mm_load_ps((const float*) params->sse2.nonsign_mask);
  const __m128i vexp_bias = _mm_load_si128((const __m128i*) params->sse2.exp_bias);
  const __m128 vscale_to_inf = _mm_load_ps(params->sse2.scale_to_inf);
  const __m128i vexpw_max = _mm_load_si128((const __m128i*) params->sse2.expw_max);
  const __m128 vscale_to_zero = _mm_load_ps(params->sse2.scale_to_zero);
  const __m128i vbias_min = _mm_load_si128((const __m128i*) params->sse2.bias_min);
  const __m128i vmanth_mask = _mm_load_si128((const __m128i*) params->sse2.manth_mask);
  const __m128i vexph_mask = _mm_load_si128((const __m128i*) params->sse2.exph_mask);
  const __m128i vnanh = _mm_load_si128((const __m128i*) params->sse2.nanh);

  uint16_t* o = (uint16_t*) output;
  for (; batch >= 32 * sizeof(float); batch -= 32 * sizeof(float)) {
    const __m128 vx0 = _mm_loadu_ps(input);
    const __m128 vx1 = _mm_loadu_ps(input + 4);
    const __m128 vx2 = _mm_loadu_ps(input + 8);
    const __m128 vx3 = _mm_loadu_ps(input + 12);
    const __m128 vx4 = _mm_loadu_ps(input + 16);
    const __m128 vx5 = _mm_loadu_ps(input + 20);
    const __m128 vx6 = _mm_loadu_ps(input + 24);
    const __m128 vx7 = _mm_loadu_ps(input + 28);
    input += 32;

    const __m128 vabsx0 = _mm_and_ps(vx0, vnonsign_mask);
    const __m128 vabsx1 = _mm_and_ps(vx1, vnonsign_mask);
    const __m128 vabsx2 = _mm_and_ps(vx2, vnonsign_mask);
    const __m128 vabsx3 = _mm_and_ps(vx3, vnonsign_mask);
    const __m128 vabsx4 = _mm_and_ps(vx4, vnonsign_mask);
    const __m128 vabsx5 = _mm_and_ps(vx5, vnonsign_mask);
    const __m128 vabsx6 = _mm_and_ps(vx6, vnonsign_mask);
    const __m128 vabsx7 = _mm_and_ps(vx7, vnonsign_mask);

    const __m128 vsignx0 = _mm_xor_ps(vx0, vabsx0);
    const __m128 vsignx1 = _mm_xor_ps(vx1, vabsx1);
    const __m128 vsignx2 = _mm_xor_ps(vx2, vabsx2);
    const __m128 vsignx3 = _mm_xor_ps(vx3, vabsx3);
    const __m128 vsignx4 = _mm_xor_ps(vx4, vabsx4);
    const __m128 vsignx5 = _mm_xor_ps(vx5, vabsx5);
    const __m128 vsignx6 = _mm_xor_ps(vx6, vabsx6);
    const __m128 vsignx7 = _mm_xor_ps(vx7, vabsx7);

    __m128i vbias0 = _mm_add_epi32(_mm_castps_si128(vabsx0), vexp_bias);
    __m128i vbias1 = _mm_add_epi32(_mm_castps_si128(vabsx1), vexp_bias);
    __m128i vbias2 = _mm_add_epi32(_mm_castps_si128(vabsx2), vexp_bias);
    __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias);
    __m128i vbias4 = _mm_add_epi32(_mm_castps_si128(vabsx4), vexp_bias);
    __m128i vbias5 = _mm_add_epi32(_mm_castps_si128(vabsx5), vexp_bias);
    __m128i vbias6 = _mm_add_epi32(_mm_castps_si128(vabsx6), vexp_bias);
    __m128i vbias7 = _mm_add_epi32(_mm_castps_si128(vabsx7), vexp_bias);

    __m128 vf0 = _mm_mul_ps(vabsx0, vscale_to_inf);
    __m128 vf1 = _mm_mul_ps(vabsx1, vscale_to_inf);
    __m128 vf2 = _mm_mul_ps(vabsx2, vscale_to_inf);
    __m128 vf3 = _mm_mul_ps(vabsx3, vscale_to_inf);
    __m128 vf4 = _mm_mul_ps(vabsx4, vscale_to_inf);
    __m128 vf5 = _mm_mul_ps(vabsx5, vscale_to_inf);
    __m128 vf6 = _mm_mul_ps(vabsx6, vscale_to_inf);
    __m128 vf7 = _mm_mul_ps(vabsx7, vscale_to_inf);

    const __m128i vnanmaskw0 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx0), vexpw_max);
    const __m128i vnanmaskw1 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx1), vexpw_max);
    const __m128i vnanmaskw2 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx2), vexpw_max);
    const __m128i vnanmaskw3 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx3), vexpw_max);
    const __m128i vnanmaskw4 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx4), vexpw_max);
    const __m128i vnanmaskw5 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx5), vexpw_max);
    const __m128i vnanmaskw6 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx6), vexpw_max);
    const __m128i vnanmaskw7 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx7), vexpw_max);

    vbias0 = _mm_and_si128(vbias0, vexpw_max);
    vbias1 = _mm_and_si128(vbias1, vexpw_max);
    vbias2 = _mm_and_si128(vbias2, vexpw_max);
    vbias3 = _mm_and_si128(vbias3, vexpw_max);
    vbias4 = _mm_and_si128(vbias4, vexpw_max);
    vbias5 = _mm_and_si128(vbias5, vexpw_max);
    vbias6 = _mm_and_si128(vbias6, vexpw_max);
    vbias7 = _mm_and_si128(vbias7, vexpw_max);

    vf0 = _mm_mul_ps(vf0, vscale_to_zero);
    vf1 = _mm_mul_ps(vf1, vscale_to_zero);
    vf2 = _mm_mul_ps(vf2, vscale_to_zero);
    vf3 = _mm_mul_ps(vf3, vscale_to_zero);
    vf4 = _mm_mul_ps(vf4, vscale_to_zero);
    vf5 = _mm_mul_ps(vf5, vscale_to_zero);
    vf6 = _mm_mul_ps(vf6, vscale_to_zero);
    vf7 = _mm_mul_ps(vf7, vscale_to_zero);

    const __m128i vnanmaskh0 = _mm_packs_epi32(vnanmaskw0, vnanmaskw1);
    const __m128i vnanmaskh1 = _mm_packs_epi32(vnanmaskw2, vnanmaskw3);
    const __m128i vnanmaskh2 = _mm_packs_epi32(vnanmaskw4, vnanmaskw5);
    const __m128i vnanmaskh3 = _mm_packs_epi32(vnanmaskw6, vnanmaskw7);

    const __m128i vsignh0 = _mm_packs_epi32(_mm_castps_si128(vsignx0), _mm_castps_si128(vsignx1));
    const __m128i vsignh1 = _mm_packs_epi32(_mm_castps_si128(vsignx2), _mm_castps_si128(vsignx3));
    const __m128i vsignh2 = _mm_packs_epi32(_mm_castps_si128(vsignx4), _mm_castps_si128(vsignx5));
    const __m128i vsignh3 = _mm_packs_epi32(_mm_castps_si128(vsignx6), _mm_castps_si128(vsignx7));

    vbias0 = _mm_max_epi16(vbias0, vbias_min);
    vbias1 = _mm_max_epi16(vbias1, vbias_min);
    vbias2 = _mm_max_epi16(vbias2, vbias_min);
    vbias3 = _mm_max_epi16(vbias3, vbias_min);
    vbias4 = _mm_max_epi16(vbias4, vbias_min);
    vbias5 = _mm_max_epi16(vbias5, vbias_min);
    vbias6 = _mm_max_epi16(vbias6, vbias_min);
    vbias7 = _mm_max_epi16(vbias7, vbias_min);

    __m128i vh0 = _mm_and_si128(vnanh, vnanmaskh0);
    __m128i vh1 = _mm_and_si128(vnanh, vnanmaskh1);
    __m128i vh2 = _mm_and_si128(vnanh, vnanmaskh2);
    __m128i vh3 = _mm_and_si128(vnanh, vnanmaskh3);

    vf0 = _mm_add_ps(vf0, _mm_castsi128_ps(vbias0));
    vf1 = _mm_add_ps(vf1, _mm_castsi128_ps(vbias1));
    vf2 = _mm_add_ps(vf2, _mm_castsi128_ps(vbias2));
    vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3));
    vf4 = _mm_add_ps(vf4, _mm_castsi128_ps(vbias4));
    vf5 = _mm_add_ps(vf5, _mm_castsi128_ps(vbias5));
    vf6 = _mm_add_ps(vf6, _mm_castsi128_ps(vbias6));
    vf7 = _mm_add_ps(vf7, _mm_castsi128_ps(vbias7));

    vh0 = _mm_or_si128(vh0, vsignh0);
    vh1 = _mm_or_si128(vh1, vsignh1);
    vh2 = _mm_or_si128(vh2, vsignh2);
    vh3 = _mm_or_si128(vh3, vsignh3);

    __m128i vexpw0 = _mm_srli_epi32(_mm_castps_si128(vf0), 13);
    __m128i vexpw1 = _mm_srli_epi32(_mm_castps_si128(vf1), 13);
    __m128i vexpw2 = _mm_srli_epi32(_mm_castps_si128(vf2), 13);
    __m128i vexpw3 = _mm_srli_epi32(_mm_castps_si128(vf3), 13);
    __m128i vexpw4 = _mm_srli_epi32(_mm_castps_si128(vf4), 13);
    __m128i vexpw5 = _mm_srli_epi32(_mm_castps_si128(vf5), 13);
    __m128i vexpw6 = _mm_srli_epi32(_mm_castps_si128(vf6), 13);
    __m128i vexpw7 = _mm_srli_epi32(_mm_castps_si128(vf7), 13);

    const __m128i vmantw0 = _mm_and_si128(_mm_castps_si128(vf0), vmanth_mask);
    const __m128i vmantw1 = _mm_and_si128(_mm_castps_si128(vf1), vmanth_mask);
    const __m128i vmantw2 = _mm_and_si128(_mm_castps_si128(vf2), vmanth_mask);
    const __m128i vmantw3 = _mm_and_si128(_mm_castps_si128(vf3), vmanth_mask);
    const __m128i vmantw4 = _mm_and_si128(_mm_castps_si128(vf4), vmanth_mask);
    const __m128i vmantw5 = _mm_and_si128(_mm_castps_si128(vf5), vmanth_mask);
    const __m128i vmantw6 = _mm_and_si128(_mm_castps_si128(vf6), vmanth_mask);
    const __m128i vmantw7 = _mm_and_si128(_mm_castps_si128(vf7), vmanth_mask);

    vexpw0 = _mm_and_si128(vexpw0, vexph_mask);
    vexpw1 = _mm_and_si128(vexpw1, vexph_mask);
    vexpw2 = _mm_and_si128(vexpw2, vexph_mask);
    vexpw3 = _mm_and_si128(vexpw3, vexph_mask);
    vexpw4 = _mm_and_si128(vexpw4, vexph_mask);
    vexpw5 = _mm_and_si128(vexpw5, vexph_mask);
    vexpw6 = _mm_and_si128(vexpw6, vexph_mask);
    vexpw7 = _mm_and_si128(vexpw7, vexph_mask);

    const __m128i vnonsignw0 = _mm_add_epi32(vmantw0, vexpw0);
    const __m128i vnonsignw1 = _mm_add_epi32(vmantw1, vexpw1);
    const __m128i vnonsignw2 = _mm_add_epi32(vmantw2, vexpw2);
    const __m128i vnonsignw3 = _mm_add_epi32(vmantw3, vexpw3);
    const __m128i vnonsignw4 = _mm_add_epi32(vmantw4, vexpw4);
    const __m128i vnonsignw5 = _mm_add_epi32(vmantw5, vexpw5);
    const __m128i vnonsignw6 = _mm_add_epi32(vmantw6, vexpw6);
    const __m128i vnonsignw7 = _mm_add_epi32(vmantw7, vexpw7);

    const __m128i vnonsignh0 = _mm_packs_epi32(vnonsignw0, vnonsignw1);
    const __m128i vnonsignh1 = _mm_packs_epi32(vnonsignw2, vnonsignw3);
    const __m128i vnonsignh2 = _mm_packs_epi32(vnonsignw4, vnonsignw5);
    const __m128i vnonsignh3 = _mm_packs_epi32(vnonsignw6, vnonsignw7);

    vh0 = _mm_or_si128(vh0, _mm_andnot_si128(vnanmaskh0, vnonsignh0));
    vh1 = _mm_or_si128(vh1, _mm_andnot_si128(vnanmaskh1, vnonsignh1));
    vh2 = _mm_or_si128(vh2, _mm_andnot_si128(vnanmaskh2, vnonsignh2));
    vh3 = _mm_or_si128(vh3, _mm_andnot_si128(vnanmaskh3, vnonsignh3));

    _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);
    o += 32;
  }
  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    const __m128 vx_lo = _mm_loadu_ps(input);
    const __m128 vx_hi = _mm_loadu_ps(input + 4);
    input += 8;

    const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask);
    const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask);

    const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo);
    const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi);
    __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias);
    __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias);
    __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf);
    __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf);
    const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max);
    const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max);

    vbias_lo = _mm_and_si128(vbias_lo, vexpw_max);
    vbias_hi = _mm_and_si128(vbias_hi, vexpw_max);
    vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero);
    vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero);
    const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi);
    const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi));

    vbias_lo = _mm_max_epi16(vbias_lo, vbias_min);
    vbias_hi = _mm_max_epi16(vbias_hi, vbias_min);
    __m128i vh = _mm_and_si128(vnanh, vnanmaskh);

    vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo));
    vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi));
    vh = _mm_or_si128(vh, vsignh);

    __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13);
    __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13);
    const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask);
    const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask);

    vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask);
    vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask);

    const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo);
    const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi);

    const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi);

    vh = _mm_or_si128(vh, _mm_andnot_si128(vnanmaskh, vnonsignh));

    _mm_storeu_si128((__m128i*) o, vh);
    o += 8;
  }
  if XNN_UNPREDICTABLE(batch != 0) {
    const __m128 vx_lo = _mm_loadu_ps(input);
    const float* input_hi = (const float*) ((uintptr_t) input + (batch & (4 * sizeof(float))));
    const __m128 vx_hi = _mm_loadu_ps(input_hi);

    const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask);
    const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask);

    const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo);
    const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi);
    __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias);
    __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias);
    __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf);
    __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf);
    const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max);
    const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max);

    vbias_lo = _mm_and_si128(vbias_lo, vexpw_max);
    vbias_hi = _mm_and_si128(vbias_hi, vexpw_max);
    vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero);
    vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero);
    const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi);
    const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi));

    vbias_lo = _mm_max_epi16(vbias_lo, vbias_min);
    vbias_hi = _mm_max_epi16(vbias_hi, vbias_min);
    __m128i vh = _mm_and_si128(vnanh, vnanmaskh);

    vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo));
    vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi));
    vh = _mm_or_si128(vh, vsignh);

    __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13);
    __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13);
    const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask);
    const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask);

    vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask);
    vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask);

    const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo);
    const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi);

    const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi);

    vh = _mm_or_si128(vh, _mm_andnot_si128(vnanmaskh, vnonsignh));

    if (batch & (4 * sizeof(float))) {
      _mm_storel_epi64((__m128i*) o, vh);
      vh = _mm_unpackhi_epi64(vh, vh);
      o += 4;
    }
    if (batch & (2 * sizeof(float))) {
      unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vh));
      vh = _mm_srli_epi64(vh, 32);
      o += 2;
    }
    if (batch & (1 * sizeof(float))) {
      *o = (uint16_t) _mm_cvtsi128_si32(vh);
    }
  }
}