FloatingPointHelpers Namespace Reference

Namespaces
namespace	BitCastHelper
namespace	LeadingZerosPortability
namespace	OperatorsHelper

Classes
struct	ConversionHelper
struct	ConversionHelper< Format, Format >
struct	ConversionHelper< StandardDouble, StandardFloat >
class	IEEE754_like
	Specifies a floating point format like those described in IEEE-754, with an adjustable number of bits in the exponent and mantissa. More...

Typedefs
using	StandardFloat = IEEE754_like<23, 8>
using	StandardDouble = IEEE754_like<52, 11>
using	CUDAHalfFloat = IEEE754_like<10, 5>
using	CUDABFloat16 = IEEE754_like<7, 8>

Enumerations
enum class	RoundingModes {   ToPlusInfinity , ToMinusInfinity , ToZero , ToNearestEven ,   ToNearestAwayFromZero , Default = ToNearestEven }
	Specifies the rounding mode to use for the operations. More...

Functions
template<class T, class FDest, class FSource>
static constexpr T	down_convert (const T pattern, const RoundingModes rt=RoundingModes::Default)
	Converts pattern from the larger floating point format FSource to FDest.
template<class T, class FDest, class FSource>
static constexpr T	up_convert (const T pattern, const RoundingModes rt=RoundingModes::Default)
	Converts pattern from the smaller floating point format FSource to FDest.

Typedef Documentation

CUDABFloat16

using FloatingPointHelpers::CUDABFloat16 = IEEE754_like<7, 8>

Definition at line 916 of file FPHelpers.h.

CUDAHalfFloat

using FloatingPointHelpers::CUDAHalfFloat = IEEE754_like<10, 5>

Definition at line 915 of file FPHelpers.h.

StandardDouble

using FloatingPointHelpers::StandardDouble = IEEE754_like<52, 11>

Definition at line 834 of file FPHelpers.h.

StandardFloat

using FloatingPointHelpers::StandardFloat = IEEE754_like<23, 8>

Definition at line 832 of file FPHelpers.h.

Enumeration Type Documentation

RoundingModes

enum class FloatingPointHelpers::RoundingModes

strong

Specifies the rounding mode to use for the operations.

Enumerator
ToPlusInfinity
ToMinusInfinity
ToZero
ToNearestEven
ToNearestAwayFromZero
Default

Definition at line 57 of file FPHelpers.h.

  {
    ToPlusInfinity, ToMinusInfinity, ToZero, ToNearestEven, ToNearestAwayFromZero, Default = ToNearestEven
  };

Function Documentation

down_convert()

template<class T, class FDest, class FSource>

constexpr T FloatingPointHelpers::down_convert ( const T pattern, const RoundingModes rt = RoundingModes::Default )

inlinestaticconstexpr

Converts pattern from the larger floating point format FSource to FDest.

Note: A few tests show that, under some circumstances, compilers may make this branchless. Clang seemed to do it, but not GCC. NVCC (for the GPU) surely did. There's probably still room for improvement, but...

Definition at line 807 of file FPHelpers.h.

  {
    static_assert(FDest::mantissa_size_bits() <= FSource::mantissa_size_bits() &&
                  FDest::exponent_size_bits() <= FSource::exponent_size_bits(),
                  "The destination type must not be a larger floating point type than the source one.");
 
    return ConversionHelper<FSource, FDest>::template down_convert<T>(pattern, rt);
  }

up_convert()

template<class T, class FDest, class FSource>

constexpr T FloatingPointHelpers::up_convert ( const T pattern, const RoundingModes rt = RoundingModes::Default )

inlinestaticconstexpr

Converts pattern from the smaller floating point format FSource to FDest.

Warning

: This could be re-written in a more branchless way, so performance might be sub-optimal...

Definition at line 823 of file FPHelpers.h.

  {
    static_assert(FDest::mantissa_size_bits() >= FSource::mantissa_size_bits() &&
                  FDest::exponent_size_bits() >= FSource::exponent_size_bits(),
                  "The source type must not be a larger floating point type than the destination one.");
    return ConversionHelper<FDest, FSource>::template up_convert<T>(pattern, rt);
  }