onnxruntime

1
#include "lib/Api.Image/pch.h"
2
#include "inc/NominalRangeConverter.h"
3

4
namespace _winml {
5
NominalRangeConverter::NominalRangeConverter(winml::LearningModelPixelRange pixelRange) {
6
  // For Normalization: the formula is input_range[min, max] / scale - shift
7
  // For DeNormalization: the formula is (input_range[min, max] + shift) * scale
8
  if (pixelRange == winml::LearningModelPixelRange::ZeroTo255) {
9
    scale = 1.f;
10
    shift = 0;
11
  } else if (pixelRange == winml::LearningModelPixelRange::ZeroToOne) {
12
    scale = 255.f;
13
    shift = 0;
14
  } else if (pixelRange == winml::LearningModelPixelRange::MinusOneToOne) {
15
    scale = (255.f / 2.f);
16
    shift = 1;
17
  }
18
};
19

20
// [0, 255] --> [0, 255]
21
// [0, 255] / 255 --> [0, 1]
22
// [0, 255] * 2 / 255 - 1 --> [-1, 1]
23
float NominalRangeConverter::Normalize(float val) const {
24
  return val / scale - shift;
25
}
26

27
DirectX::PackedVector::HALF NominalRangeConverter::Normalize(DirectX::PackedVector::HALF val) const {
28
  return static_cast<DirectX::PackedVector::HALF>(val / scale - shift);
29
}
30

31
#if defined(_M_AMD64) || defined(_M_IX86)
32
__m128 NominalRangeConverter::Normalize(__m128 sse_data) const {
33
  __m128 sse_shift = _mm_set1_ps(static_cast<float>(shift));
34
  __m128 sse_scale = _mm_set1_ps(scale);
35

36
  auto sse_dived = _mm_div_ps(sse_data, sse_scale);
37
  return _mm_sub_ps(sse_dived, sse_shift);
38
}
39
#endif
40

41
// [0, 255] --> [0, 255]
42
// ([0, 1] + 0 ) * 255 -> [0, 1]
43
// ([-1, 1] + 1) * 255 / 2 --> [-1, 1]
44
float NominalRangeConverter::Denormalize(float val) const {
45
  return scale * (val + shift);
46
}
47

48
DirectX::PackedVector::HALF NominalRangeConverter::Denormalize(DirectX::PackedVector::HALF val) const {
49
  return static_cast<DirectX::PackedVector::HALF>(scale * (val + shift));
50
}
51

52
#if defined(_M_AMD64) || defined(_M_IX86)
53
__m128 NominalRangeConverter::Denormalize(__m128 sse_data) const {
54
  __m128 sse_shift = _mm_set1_ps(static_cast<float>(shift));
55
  __m128 sse_scale = _mm_set1_ps(scale);
56

57
  auto sse_added = _mm_add_ps(sse_data, sse_shift);
58
  return _mm_mul_ps(sse_added, sse_scale);
59
}
60
#endif
61
}  // namespace _winml
62