1 | #pragma once |
2 | #include <c10/macros/Macros.h> |
3 | #include <c10/util/BFloat16.h> |
4 | #include <c10/util/Half.h> |
5 | |
6 | #include <type_traits> |
7 | |
8 | C10_CLANG_DIAGNOSTIC_PUSH() |
9 | #if C10_CLANG_HAS_WARNING("-Wimplicit-float-conversion") |
10 | C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-float-conversion" ) |
11 | #endif |
12 | #if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion") |
13 | C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion" ) |
14 | #endif |
15 | |
16 | namespace c10 { |
17 | |
18 | template <typename dest_t, typename src_t> |
19 | struct needs_real { |
20 | constexpr static bool value = |
21 | (is_complex<src_t>::value && !is_complex<dest_t>::value); |
22 | }; |
23 | |
24 | template <bool, typename src_t> |
25 | struct maybe_real { |
26 | C10_HOST_DEVICE static inline src_t apply(src_t src) { |
27 | return src; |
28 | } |
29 | }; |
30 | |
31 | template <typename src_t> |
32 | struct maybe_real<true, src_t> { |
33 | C10_HOST_DEVICE static inline decltype(auto) apply(src_t src) { |
34 | return src.real(); |
35 | } |
36 | }; |
37 | |
38 | // Note: deliberately ignores undefined behavior, consistent with NumPy. |
39 | // PyTorch's type conversions can cause a variety of undefined behavior, |
40 | // including float to integral overflow and signed to unsigned integer overflow. |
41 | // Some of this undefined behavior is addressed below. |
42 | template <typename dest_t, typename src_t> |
43 | struct static_cast_with_inter_type { |
44 | C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline dest_t apply( |
45 | src_t src) { |
46 | constexpr bool real = needs_real<dest_t, src_t>::value; |
47 | auto r = maybe_real<real, src_t>::apply(src); |
48 | return static_cast<dest_t>(r); |
49 | } |
50 | }; |
51 | |
52 | // Partial template instantiation for casting to uint8. |
53 | // Note: Converting from negative float values to unsigned integer types is |
54 | // undefined behavior in C++, and current CPU and GPU compilers exhibit |
55 | // divergent behavior. Casting from negative float values to signed |
56 | // integer types and then to unsigned integer types is not undefined, |
57 | // however, so this cast improves the consistency of type conversions |
58 | // to uint8 across compilers. |
59 | // Further note: Type conversions across compilers still have other undefined |
60 | // and divergent behavior. |
61 | template <typename src_t> |
62 | struct static_cast_with_inter_type<uint8_t, src_t> { |
63 | C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline uint8_t apply( |
64 | src_t src) { |
65 | constexpr bool real = needs_real<uint8_t, src_t>::value; |
66 | return static_cast<uint8_t>( |
67 | static_cast<int64_t>(maybe_real<real, src_t>::apply(src))); |
68 | } |
69 | }; |
70 | |
71 | template <> |
72 | struct static_cast_with_inter_type<c10::complex<c10::Half>, c10::BFloat16> { |
73 | C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< |
74 | c10::Half> |
75 | apply(c10::BFloat16 src) { |
76 | return static_cast<c10::complex<c10::Half>>(c10::complex<float>{src}); |
77 | } |
78 | }; |
79 | |
80 | template <> |
81 | struct static_cast_with_inter_type<c10::complex<c10::Half>, c10::Half> { |
82 | C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< |
83 | c10::Half> |
84 | apply(c10::Half src) { |
85 | return static_cast<c10::complex<c10::Half>>(c10::complex<float>{src}); |
86 | } |
87 | }; |
88 | |
89 | template <> |
90 | struct static_cast_with_inter_type< |
91 | c10::complex<c10::Half>, |
92 | c10::complex<double>> { |
93 | C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< |
94 | c10::Half> |
95 | apply(c10::complex<double> src) { |
96 | return static_cast<c10::complex<c10::Half>>( |
97 | static_cast<c10::complex<float>>(src)); |
98 | } |
99 | }; |
100 | |
101 | template <typename To, typename From> |
102 | C10_HOST_DEVICE To convert(From f) { |
103 | return static_cast_with_inter_type<To, From>::apply(f); |
104 | } |
105 | |
106 | // Define separately to avoid being inlined and prevent code-size bloat |
107 | C10_API void report_overflow(const char* name); |
108 | |
109 | template <typename To, typename From> |
110 | To checked_convert(From f, const char* name) { |
111 | // Converting to bool can't overflow so we exclude this case from checking. |
112 | if (!std::is_same<To, bool>::value && overflows<To, From>(f)) { |
113 | report_overflow(name); |
114 | } |
115 | return convert<To, From>(f); |
116 | } |
117 | |
118 | } // namespace c10 |
119 | |
120 | C10_CLANG_DIAGNOSTIC_POP() |
121 | |
122 | // Trigger tests for D25440771. TODO: Remove this line any time you want. |
123 | |