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/BFloat16.h (2372B)

#pragma once

// Defines the bloat16 type (brain floating-point). This representation uses
// 1 bit for the sign, 8 bits for the exponent and 7 bits for the mantissa.

#include <c10/macros/Macros.h>
#include <cmath>
#include <cstring>

#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000
#include <cuda_bf16.h>
#endif

namespace c10 {

namespace detail {
inline C10_HOST_DEVICE float f32_from_bits(uint16_t src) {
  float res = 0;
  uint32_t tmp = src;
  tmp <<= 16;

#ifdef __HIP_PLATFORM_HCC__
  float* tempRes;

// We should be using memcpy in order to respect the strict aliasing rule
  // but it fails in the HIP environment.
  tempRes = reinterpret_cast<float*>(&tmp);
  res = *tempRes;
#else
  std::memcpy(&res, &tmp, sizeof(tmp));
#endif

return res;
}

inline C10_HOST_DEVICE uint16_t bits_from_f32(float src) {
  uint32_t res = 0;

#ifdef __HIP_PLATFORM_HCC__
  // We should be using memcpy in order to respect the strict aliasing rule
  // but it fails in the HIP environment.
  uint32_t* tempRes = reinterpret_cast<uint32_t*>(&src);
  res = *tempRes;
#else
  std::memcpy(&res, &src, sizeof(res));
#endif

return res >> 16;
}

inline C10_HOST_DEVICE uint16_t round_to_nearest_even(float src) {
#if defined(__HIP_PLATFORM_HCC__)
  if (src != src) {
#elif defined(_MSC_VER)
  if (isnan(src)) {
#else
  if (std::isnan(src)) {
#endif
    return UINT16_C(0x7FC0);
  } else {
    union {
      uint32_t U32;
      float F32;
    };

F32 = src;
    uint32_t rounding_bias = ((U32 >> 16) & 1) + UINT32_C(0x7FFF);
    return static_cast<uint16_t>((U32 + rounding_bias) >> 16);
  }
}
} // namespace detail

struct alignas(2) BFloat16 {
  uint16_t x;

// HIP wants __host__ __device__ tag, CUDA does not
#ifdef __HIP_PLATFORM_HCC__
  C10_HOST_DEVICE BFloat16() = default;
#else
  BFloat16() = default;
#endif

struct from_bits_t {};
  static constexpr C10_HOST_DEVICE from_bits_t from_bits() {
    return from_bits_t();
  }

constexpr C10_HOST_DEVICE BFloat16(unsigned short bits, from_bits_t)
      : x(bits){};
  inline C10_HOST_DEVICE BFloat16(float value);
  inline C10_HOST_DEVICE operator float() const;

#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000
  inline C10_HOST_DEVICE BFloat16(const __nv_bfloat16& value);
  explicit inline C10_HOST_DEVICE operator __nv_bfloat16() const;
#endif
};

} // namespace c10

#include <c10/util/BFloat16-inl.h>

cmd: