forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
NumericUtils.h
102 lines (84 loc) · 2.49 KB
/
NumericUtils.h
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
#pragma once
#ifdef __HIPCC__
#include <hip/hip_runtime.h>
#endif
#include <cmath>
#include <complex>
#include <type_traits>
#include <c10/util/BFloat16.h>
#include <c10/util/Half.h>
#include <c10/macros/Macros.h>
namespace at {
// std::isnan isn't performant to use on integral types; it will
// (uselessly) convert to floating point and then do the test.
// This function is.
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
inline C10_HOST_DEVICE bool _isnan(T val) {
return false;
}
template <typename T,
typename std::enable_if<std::is_floating_point<T>::value, int>::type = 0>
inline C10_HOST_DEVICE bool _isnan(T val) {
#if defined(__CUDACC__) || defined(__HIPCC__)
return ::isnan(val);
#else
return std::isnan(val);
#endif
}
template <typename T,
typename std::enable_if<c10::is_complex_t<T>::value, int>::type = 0>
inline bool _isnan(T val) {
return std::isnan(val.real()) || std::isnan(val.imag());
}
template <typename T,
typename std::enable_if<std::is_same<T, at::Half>::value, int>::type = 0>
inline C10_HOST_DEVICE bool _isnan(T val) {
return at::_isnan(float(val));
}
inline C10_HOST_DEVICE bool _isnan(at::BFloat16 val) {
return at::_isnan(float(val));
}
template <typename T>
C10_HOST_DEVICE inline T exp(T x) {
static_assert(!std::is_same<T, double>::value, "this template must be used with float or less precise type");
#if defined(__CUDA_ARCH__) || defined(__HIP_ARCH__)
// use __expf fast approximation for peak bandwidth
return __expf(x);
#else
return ::exp(x);
#endif
}
template <>
C10_HOST_DEVICE inline double exp<double>(double x) {
return ::exp(x);
}
template <typename T>
C10_HOST_DEVICE inline T log(T x) {
static_assert(!std::is_same<T, double>::value, "this template must be used with float or less precise type");
#if defined(__CUDA_ARCH__) || defined(__HIP_ARCH__)
// use __logf fast approximation for peak bandwidth
return __logf(x);
#else
return ::log(x);
#endif
}
template <>
C10_HOST_DEVICE inline double log<double>(double x) {
return ::log(x);
}
template <typename T>
C10_HOST_DEVICE inline T tan(T x) {
static_assert(!std::is_same<T, double>::value, "this template must be used with float or less precise type");
#if defined(__CUDA_ARCH__) || defined(__HIP_ARCH__)
// use __tanf fast approximation for peak bandwidth
return __tanf(x);
#else
return ::tan(x);
#endif
}
template <>
C10_HOST_DEVICE inline double tan<double>(double x) {
return ::tan(x);
}
} // namespace at