File size: 1,961 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
// Auto-generated file. Do not edit!
//   Template: src/f16-f32-vcvt/f16c.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 <immintrin.h>

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


void xnn_f16_f32_vcvt_ukernel__f16c_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 uint16_t* i = (const uint16_t*) input;
  for (; batch >= 16 * sizeof(uint16_t); batch -= 16 * sizeof(uint16_t)) {
    const __m256 vacc0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    const __m256 vacc1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 8)));
    i += 16;

    _mm256_storeu_ps(output, vacc0);
    _mm256_storeu_ps(output + 8, vacc1);
    output += 16;
  }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    i += 8;

    _mm256_storeu_ps(output, vacc);
    output += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));
    const __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));

    __m128 vacc_lo = _mm256_castps256_ps128(vacc);
    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storeu_ps(output, vacc_lo);
      vacc_lo = _mm256_extractf128_ps(vacc, 1);
      output += 4;
    }
    if (batch & (2 * sizeof(uint16_t))) {
      _mm_storel_pi((__m64*) output, vacc_lo);
      vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
      output += 2;
    }
    if (batch & (1 * sizeof(uint16_t))) {
      _mm_store_ss(output, vacc_lo);
    }
  }
}