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-inl.h (9222B)

#pragma once

#include <c10/macros/Macros.h>
#include <limits>

namespace c10 {

/// Constructors
inline C10_HOST_DEVICE BFloat16::BFloat16(float value) {
#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000 && \
    defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
  x = __bfloat16_as_ushort(__float2bfloat16(value));
#else
  // RNE by default
  x = detail::round_to_nearest_even(value);
#endif
}

/// Implicit conversions
inline C10_HOST_DEVICE BFloat16::operator float() const {
#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000
  return __bfloat162float(*reinterpret_cast<const __nv_bfloat16*>(&x));
#else
  return detail::f32_from_bits(x);
#endif
}

#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000
inline C10_HOST_DEVICE BFloat16::BFloat16(const __nv_bfloat16& value) {
  x = *reinterpret_cast<const unsigned short*>(&value);
}
inline C10_HOST_DEVICE BFloat16::operator __nv_bfloat16() const {
  return *reinterpret_cast<const __nv_bfloat16*>(&x);
}
#endif

// CUDA intrinsics

#if defined(__CUDACC__) || defined(__HIPCC__)
inline C10_DEVICE BFloat16 __ldg(const BFloat16* ptr) {
#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000 && \
    defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
  return __ldg(reinterpret_cast<const __nv_bfloat16*>(ptr));
#else
  return *ptr;
#endif
}
#endif

/// Arithmetic

inline C10_HOST_DEVICE BFloat16
operator+(const BFloat16& a, const BFloat16& b) {
  return static_cast<float>(a) + static_cast<float>(b);
}

inline C10_HOST_DEVICE BFloat16
operator-(const BFloat16& a, const BFloat16& b) {
  return static_cast<float>(a) - static_cast<float>(b);
}

inline C10_HOST_DEVICE BFloat16
operator*(const BFloat16& a, const BFloat16& b) {
  return static_cast<float>(a) * static_cast<float>(b);
}

inline C10_HOST_DEVICE BFloat16 operator/(const BFloat16& a, const BFloat16& b)
    __ubsan_ignore_float_divide_by_zero__ {
  return static_cast<float>(a) / static_cast<float>(b);
}

inline C10_HOST_DEVICE BFloat16 operator-(const BFloat16& a) {
  return -static_cast<float>(a);
}

inline C10_HOST_DEVICE BFloat16& operator+=(BFloat16& a, const BFloat16& b) {
  a = a + b;
  return a;
}

inline C10_HOST_DEVICE BFloat16& operator-=(BFloat16& a, const BFloat16& b) {
  a = a - b;
  return a;
}

inline C10_HOST_DEVICE BFloat16& operator*=(BFloat16& a, const BFloat16& b) {
  a = a * b;
  return a;
}

inline C10_HOST_DEVICE BFloat16& operator/=(BFloat16& a, const BFloat16& b) {
  a = a / b;
  return a;
}

inline C10_HOST_DEVICE BFloat16& operator|(BFloat16& a, const BFloat16& b) {
  a.x = a.x | b.x;
  return a;
}

inline C10_HOST_DEVICE BFloat16& operator^(BFloat16& a, const BFloat16& b) {
  a.x = a.x ^ b.x;
  return a;
}

inline C10_HOST_DEVICE BFloat16& operator&(BFloat16& a, const BFloat16& b) {
  a.x = a.x & b.x;
  return a;
}

/// Arithmetic with floats

inline C10_HOST_DEVICE float operator+(BFloat16 a, float b) {
  return static_cast<float>(a) + b;
}
inline C10_HOST_DEVICE float operator-(BFloat16 a, float b) {
  return static_cast<float>(a) - b;
}
inline C10_HOST_DEVICE float operator*(BFloat16 a, float b) {
  return static_cast<float>(a) * b;
}
inline C10_HOST_DEVICE float operator/(BFloat16 a, float b) {
  return static_cast<float>(a) / b;
}

inline C10_HOST_DEVICE float operator+(float a, BFloat16 b) {
  return a + static_cast<float>(b);
}
inline C10_HOST_DEVICE float operator-(float a, BFloat16 b) {
  return a - static_cast<float>(b);
}
inline C10_HOST_DEVICE float operator*(float a, BFloat16 b) {
  return a * static_cast<float>(b);
}
inline C10_HOST_DEVICE float operator/(float a, BFloat16 b) {
  return a / static_cast<float>(b);
}

inline C10_HOST_DEVICE float& operator+=(float& a, const BFloat16& b) {
  return a += static_cast<float>(b);
}
inline C10_HOST_DEVICE float& operator-=(float& a, const BFloat16& b) {
  return a -= static_cast<float>(b);
}
inline C10_HOST_DEVICE float& operator*=(float& a, const BFloat16& b) {
  return a *= static_cast<float>(b);
}
inline C10_HOST_DEVICE float& operator/=(float& a, const BFloat16& b) {
  return a /= static_cast<float>(b);
}

/// Arithmetic with doubles

inline C10_HOST_DEVICE double operator+(BFloat16 a, double b) {
  return static_cast<double>(a) + b;
}
inline C10_HOST_DEVICE double operator-(BFloat16 a, double b) {
  return static_cast<double>(a) - b;
}
inline C10_HOST_DEVICE double operator*(BFloat16 a, double b) {
  return static_cast<double>(a) * b;
}
inline C10_HOST_DEVICE double operator/(BFloat16 a, double b) {
  return static_cast<double>(a) / b;
}

inline C10_HOST_DEVICE double operator+(double a, BFloat16 b) {
  return a + static_cast<double>(b);
}
inline C10_HOST_DEVICE double operator-(double a, BFloat16 b) {
  return a - static_cast<double>(b);
}
inline C10_HOST_DEVICE double operator*(double a, BFloat16 b) {
  return a * static_cast<double>(b);
}
inline C10_HOST_DEVICE double operator/(double a, BFloat16 b) {
  return a / static_cast<double>(b);
}

/// Arithmetic with ints

inline C10_HOST_DEVICE BFloat16 operator+(BFloat16 a, int b) {
  return a + static_cast<BFloat16>(b);
}
inline C10_HOST_DEVICE BFloat16 operator-(BFloat16 a, int b) {
  return a - static_cast<BFloat16>(b);
}
inline C10_HOST_DEVICE BFloat16 operator*(BFloat16 a, int b) {
  return a * static_cast<BFloat16>(b);
}
inline C10_HOST_DEVICE BFloat16 operator/(BFloat16 a, int b) {
  return a / static_cast<BFloat16>(b);
}

inline C10_HOST_DEVICE BFloat16 operator+(int a, BFloat16 b) {
  return static_cast<BFloat16>(a) + b;
}
inline C10_HOST_DEVICE BFloat16 operator-(int a, BFloat16 b) {
  return static_cast<BFloat16>(a) - b;
}
inline C10_HOST_DEVICE BFloat16 operator*(int a, BFloat16 b) {
  return static_cast<BFloat16>(a) * b;
}
inline C10_HOST_DEVICE BFloat16 operator/(int a, BFloat16 b) {
  return static_cast<BFloat16>(a) / b;
}

//// Arithmetic with int64_t

inline C10_HOST_DEVICE BFloat16 operator+(BFloat16 a, int64_t b) {
  return a + static_cast<BFloat16>(b);
}
inline C10_HOST_DEVICE BFloat16 operator-(BFloat16 a, int64_t b) {
  return a - static_cast<BFloat16>(b);
}
inline C10_HOST_DEVICE BFloat16 operator*(BFloat16 a, int64_t b) {
  return a * static_cast<BFloat16>(b);
}
inline C10_HOST_DEVICE BFloat16 operator/(BFloat16 a, int64_t b) {
  return a / static_cast<BFloat16>(b);
}

inline C10_HOST_DEVICE BFloat16 operator+(int64_t a, BFloat16 b) {
  return static_cast<BFloat16>(a) + b;
}
inline C10_HOST_DEVICE BFloat16 operator-(int64_t a, BFloat16 b) {
  return static_cast<BFloat16>(a) - b;
}
inline C10_HOST_DEVICE BFloat16 operator*(int64_t a, BFloat16 b) {
  return static_cast<BFloat16>(a) * b;
}
inline C10_HOST_DEVICE BFloat16 operator/(int64_t a, BFloat16 b) {
  return static_cast<BFloat16>(a) / b;
}

// Overloading < and > operators, because std::max and std::min use them.

inline C10_HOST_DEVICE bool operator>(BFloat16& lhs, BFloat16& rhs) {
  return float(lhs) > float(rhs);
}

inline C10_HOST_DEVICE bool operator<(BFloat16& lhs, BFloat16& rhs) {
  return float(lhs) < float(rhs);
}

} // namespace c10

namespace std {

template <>
class numeric_limits<c10::BFloat16> {
 public:
  static constexpr bool is_signed = true;
  static constexpr bool is_specialized = true;
  static constexpr bool is_integer = false;
  static constexpr bool is_exact = false;
  static constexpr bool has_infinity = true;
  static constexpr bool has_quiet_NaN = true;
  static constexpr bool has_signaling_NaN = true;
  static constexpr auto has_denorm = numeric_limits<float>::has_denorm;
  static constexpr auto has_denorm_loss =
      numeric_limits<float>::has_denorm_loss;
  static constexpr auto round_style = numeric_limits<float>::round_style;
  static constexpr bool is_iec559 = false;
  static constexpr bool is_bounded = true;
  static constexpr bool is_modulo = false;
  static constexpr int digits = 8;
  static constexpr int digits10 = 2;
  static constexpr int max_digits10 = 4;
  static constexpr int radix = 2;
  static constexpr int min_exponent = -125;
  static constexpr int min_exponent10 = -37;
  static constexpr int max_exponent = 128;
  static constexpr int max_exponent10 = 38;
  static constexpr auto traps = numeric_limits<float>::traps;
  static constexpr auto tinyness_before =
      numeric_limits<float>::tinyness_before;

static constexpr c10::BFloat16 min() {
    return c10::BFloat16(0x0080, c10::BFloat16::from_bits());
  }
  static constexpr c10::BFloat16 lowest() {
    return c10::BFloat16(0xFF7F, c10::BFloat16::from_bits());
  }
  static constexpr c10::BFloat16 max() {
    return c10::BFloat16(0x7F7F, c10::BFloat16::from_bits());
  }
  static constexpr c10::BFloat16 epsilon() {
    return c10::BFloat16(0x3C00, c10::BFloat16::from_bits());
  }
  static constexpr c10::BFloat16 round_error() {
    return c10::BFloat16(0x3F00, c10::BFloat16::from_bits());
  }
  static constexpr c10::BFloat16 infinity() {
    return c10::BFloat16(0x7F80, c10::BFloat16::from_bits());
  }
  static constexpr c10::BFloat16 quiet_NaN() {
    return c10::BFloat16(0x7FC0, c10::BFloat16::from_bits());
  }
  static constexpr c10::BFloat16 signaling_NaN() {
    return c10::BFloat16(0x7F80, c10::BFloat16::from_bits());
  }
  static constexpr c10::BFloat16 denorm_min() {
    return c10::BFloat16(0x0001, c10::BFloat16::from_bits());
  }
};

} // namespace std

cmd: