FM / util

/usr/local/lib64/python3.6/site-packages/torch/include/c10/util/

/usr/local/lib64/python3.6/site-packages/torch/include/c10/util

Name	Size	Mode	Actions
accumulate.h	4205	0644	edit dl rm
AlignOf.h	4835	0644	edit dl rm
Array.h	11354	0644	edit dl rm
ArrayRef.h	9058	0644	edit dl rm
Backtrace.h	364	0644	edit dl rm
BFloat16-inl.h	9222	0644	edit dl rm
BFloat16-math.h	5217	0644	edit dl rm
BFloat16.h	2372	0644	edit dl rm
Bitset.h	3414	0644	edit dl rm
C++17.h	13329	0644	edit dl rm
complex.h	17542	0644	edit dl rm
complex_math.h	10902	0644	edit dl rm
complex_utils.h	958	0644	edit dl rm
ConstexprCrc.h	6633	0644	edit dl rm
copysign.h	866	0644	edit dl rm
DeadlockDetection.h	1920	0644	edit dl rm
Deprecated.h	3579	0644	edit dl rm
either.h	6423	0644	edit dl rm
env.h	835	0644	edit dl rm
Exception.h	24723	0644	edit dl rm
ExclusivelyOwned.h	4759	0644	edit dl rm
Flags.h	10056	0644	edit dl rm
flat_hash_map.h	61655	0644	edit dl rm
FunctionRef.h	2301	0644	edit dl rm
Half-inl.h	8674	0644	edit dl rm
Half.h	18965	0644	edit dl rm
hash.h	5084	0644	edit dl rm
IdWrapper.h	2348	0644	edit dl rm
intrusive_ptr.h	35563	0644	edit dl rm
in_place.h	350	0644	edit dl rm
irange.h	2679	0644	edit dl rm
LeftRight.h	6016	0644	edit dl rm
llvmMathExtras.h	29168	0644	edit dl rm
Logging.h	11256	0644	edit dl rm
logging_is_google_glog.h	2031	0644	edit dl rm
logging_is_not_google_glog.h	8271	0644	edit dl rm
MathConstants.h	858	0644	edit dl rm
math_compat.h	7296	0644	edit dl rm
MaybeOwned.h	6689	0644	edit dl rm
Metaprogramming.h	15288	0644	edit dl rm
numa.h	696	0644	edit dl rm
Optional.h	35592	0644	edit dl rm
order_preserving_flat_hash_map.h	65482	0644	edit dl rm
overloaded.h	709	0644	edit dl rm
python_stub.h	56	0644	edit dl rm
qint8.h	472	0644	edit dl rm
qint32.h	319	0644	edit dl rm
quint4x2.h	366	0644	edit dl rm
quint8.h	320	0644	edit dl rm
Registry.h	12242	0644	edit dl rm
reverse_iterator.h	8796	0644	edit dl rm
ScopeExit.h	1345	0644	edit dl rm
signal_handler.h	3154	0644	edit dl rm
SmallBuffer.h	1243	0644	edit dl rm
SmallVector.h	34456	0644	edit dl rm
sparse_bitset.h	26511	0644	edit dl rm
StringUtil.h	4538	0644	edit dl rm
string_utils.h	3989	0644	edit dl rm
string_view.h	20189	0644	edit dl rm
tempfile.h	6029	0644	edit dl rm
ThreadLocal.h	3883	0644	edit dl rm
ThreadLocalDebugInfo.h	2603	0644	edit dl rm
thread_name.h	148	0644	edit dl rm
Type.h	607	0644	edit dl rm
TypeCast.h	6972	0644	edit dl rm
typeid.h	18793	0644	edit dl rm
TypeIndex.h	5251	0644	edit dl rm
TypeList.h	16901	0644	edit dl rm
TypeTraits.h	5368	0644	edit dl rm
Unicode.h	295	0644	edit dl rm
UniqueVoidPtr.h	4117	0644	edit dl rm
Unroll.h	667	0644	edit dl rm
variant.h	95182	0644	edit dl rm
win32-headers.h	858	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/c10/util/TypeCast.h (6972B)

#pragma once
#include <c10/core/ScalarType.h>
#include <c10/macros/Macros.h>
#include <c10/util/BFloat16.h>
#include <c10/util/Half.h>

#include <type_traits>

namespace c10 {

template <typename dest_t, typename src_t>
struct needs_real {
  constexpr static bool value =
      (is_complex<src_t>::value && !is_complex<dest_t>::value);
};

template <bool, typename src_t>
struct maybe_real {
  C10_HOST_DEVICE static inline src_t apply(src_t src) {
    return src;
  }
};

template <typename src_t>
struct maybe_real<true, src_t> {
  C10_HOST_DEVICE static inline decltype(auto) apply(src_t src) {
    return src.real();
  }
};

// Note: deliberately ignores undefined behavior, consistent with NumPy.
// PyTorch's type conversions can cause a variety of undefined behavior,
// including float to integral overflow and signed to unsigned integer overflow.
// Some of this undefined behavior is addressed below.
template <typename dest_t, typename src_t>
struct static_cast_with_inter_type {
  C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline dest_t apply(
      src_t src) {
    constexpr bool real = needs_real<dest_t, src_t>::value;
    return static_cast<dest_t>(maybe_real<real, src_t>::apply(src));
  }
};

// Partial template instantiation for casting to uint8.
// Note: Converting from negative float values to unsigned integer types is
// undefined behavior in C++, and current CPU and GPU compilers exhibit
// divergent behavior. Casting from negative float values to signed
// integer types and then to unsigned integer types is not undefined,
// however, so this cast improves the consistency of type conversions
// to uint8 across compilers.
// Further note: Type conversions across compilers still have other undefined
// and divergent behavior.
template <typename src_t>
struct static_cast_with_inter_type<uint8_t, src_t> {
  C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline uint8_t apply(
      src_t src) {
    constexpr bool real = needs_real<uint8_t, src_t>::value;
    return static_cast<uint8_t>(
        static_cast<int64_t>(maybe_real<real, src_t>::apply(src)));
  }
};

// Dynamic type casting utils:
// - fetch_and_cast
// - cast_and_store
//
// fetch_and_cast fetch a value with dynamic type specified by a ScalarType
// from a void pointer and cast it to a static type.
//
// cast_and_store casts a static typed value into dynamic type specified
// by a ScalarType, and store it into a void pointer.
//
// NOTE:
//
// Dynamic casting allows us to support type promotion without blowing up
// the combination space: For example, without dynamic cast, in order to
// implement `add_` with type promotion, we would need something like
//
// AT_DISPATCH_ALL_TYPES(output.dtype(),
//    AT_DISPATCH_ALL_TYPES(input1.dtype(),
//       AT_DISPATCH_ALL_TYPES(input2.dtype(),
//           [](arg0_t a, arg1_t b) -> out_t { return a + b; }
//       )
//    )
// )
//
// If we support N dtypes, the above code would generate the a+b kernel for
// all the N * N * N different supported types, the compilation time and
// binary size would become horrible.
//
// Dynamic casting might sounds like a bad idea in terms of performance.
// Especially if you ever do it in a loop, you are going to do a billion tests.
// But in practice it is not as bad as it might look:
//
// - on CPU, this is a branch that always has the same outcome, therefore
//   hopefully the branch predictor could do the job pretty well
// - on GPU, these branches will not diverge, so we could still have the same
//   warp executing the same line of code
// - Most kernels, like `add`, are bandwidth bound, adding a few clock cycles to
//   check an integer does not hurt the performance much because the ALUs would
//   wait for load instructions anyway.
//
// For the discussion and benchmark, refer to:
// - https://github.com/pytorch/pytorch/pull/28343
// - https://github.com/pytorch/pytorch/pull/28344
// - https://github.com/pytorch/pytorch/pull/28345
//

#ifdef C10_HOST_DEVICE
#define ERROR_UNSUPPORTED_CAST CUDA_KERNEL_ASSERT(false);
#else
#define ERROR_UNSUPPORTED_CAST TORCH_CHECK(false, "Unexpected scalar type");
#endif

// Fetch a value with dynamic type src_type from ptr, and cast it to static type
// dest_t.
#define FETCH_AND_CAST_CASE(type, scalartype) \
  case ScalarType::scalartype:                \
    return static_cast_with_inter_type<dest_t, type>::apply(*(const type*)ptr);
template <typename dest_t>
C10_HOST_DEVICE inline dest_t fetch_and_cast(
    const ScalarType src_type,
    const void* ptr) {
  switch (src_type) {
    AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_EXCEPT_COMPLEX_HALF(FETCH_AND_CAST_CASE)
    default:
      ERROR_UNSUPPORTED_CAST
  }
  return dest_t(0); // just to avoid compiler warning
}

// Cast a value with static type src_t into dynamic dest_type, and store it to
// ptr.
#define CAST_AND_STORE_CASE(type, scalartype)                             \
  case ScalarType::scalartype:                                            \
    *(type*)ptr = static_cast_with_inter_type<type, src_t>::apply(value); \
    return;
template <typename src_t>
C10_HOST_DEVICE inline void cast_and_store(
    const ScalarType dest_type,
    void* ptr,
    src_t value) {
  switch (dest_type) {
    AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_EXCEPT_COMPLEX_HALF(CAST_AND_STORE_CASE)
    default:;
  }
  ERROR_UNSUPPORTED_CAST
}

#define DEFINE_UNCASTABLE(T, scalartype_)                     \
  template <>                                                 \
  C10_HOST_DEVICE inline T fetch_and_cast<T>(                 \
      const ScalarType src_type, const void* ptr) {           \
    CUDA_KERNEL_ASSERT(ScalarType::scalartype_ == src_type);  \
    return *(const T*)ptr;                                    \
  }                                                           \
  template <>                                                 \
  C10_HOST_DEVICE inline void cast_and_store<T>(              \
      const ScalarType dest_type, void* ptr, T value) {       \
    CUDA_KERNEL_ASSERT(ScalarType::scalartype_ == dest_type); \
    *(T*)ptr = value;                                         \
  }

AT_FORALL_QINT_TYPES(DEFINE_UNCASTABLE)

#undef FETCH_AND_CAST_CASE
#undef CAST_AND_STORE_CASE
#undef DEFINE_UNCASTABLE
#undef ERROR_UNSUPPORTED_CAST

template <typename To, typename From>
To convert(From f) {
  return static_cast_with_inter_type<To, From>::apply(f);
}

template <typename To, typename From>
To checked_convert(From f, const char* name) {
  // Converting to bool can't overflow so we exclude this case from checking.
  if (!std::is_same<To, bool>::value && overflows<To, From>(f)) {
    std::ostringstream oss;
    oss << "value cannot be converted to type " << name
        << " without overflow: " << f;
    throw std::runtime_error(
        oss.str()); // rather than domain_error (issue 33562)
  }
  return convert<To, From>(f);
}

} // namespace c10

// Trigger tests for D25440771. TODO: Remove this line any time you want.

cmd: