Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Another set of floating point optimisation attemps #828

Open
wants to merge 3 commits into
base: master
Choose a base branch
from
Open

Another set of floating point optimisation attemps #828

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c8ec8e4
Introduce bit_cast_constexpr
axic May 28, 2022
71d4b4a
Use bit_cast_constexpr
axic May 28, 2022
a5edb7c
Introduce isnan helper
axic May 28, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions lib/fizzy/cxx20/bit.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,8 @@ namespace fizzy
using std::bit_cast;
} // namespace fizzy

#define bit_cast_constexpr constexpr

#else

namespace fizzy
Expand All @@ -41,6 +43,8 @@ template <class To, class From>
}
} // namespace fizzy

#define bit_cast_constexpr

#endif /* __cpp_lib_bit_cast */

#ifdef __cpp_lib_bitops
Expand Down
51 changes: 35 additions & 16 deletions lib/fizzy/execute.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,10 @@ constexpr uint32_t F32AbsMask = 0x7fffffff;
constexpr uint32_t F32SignMask = ~F32AbsMask;
constexpr uint64_t F64AbsMask = 0x7fffffffffffffff;
constexpr uint64_t F64SignMask = ~F64AbsMask;
constexpr uint32_t F32ExpMask = 0x7f800000;
constexpr uint32_t F32FracMask = 0x007fffff;
constexpr uint64_t F64ExpMask = 0x7ff0000000000000;
constexpr uint64_t F64FracMask = 0x000fffffffffffff;

template <typename T>
inline T read(const uint8_t*& input) noexcept
Expand Down Expand Up @@ -336,27 +340,42 @@ inline constexpr T popcnt(T value) noexcept
template <typename T>
T signbit(T value) noexcept = delete;

inline bool signbit(float value) noexcept
inline bit_cast_constexpr bool signbit(float value) noexcept
{
return (bit_cast<uint32_t>(value) & F32SignMask) != 0;
}

inline bool signbit(double value) noexcept
inline bit_cast_constexpr bool signbit(double value) noexcept
{
return (bit_cast<uint64_t>(value) & F64SignMask) != 0;
}

template <typename T>
T isnan(T value) noexcept = delete;

inline bit_cast_constexpr bool isnan(float value) noexcept
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The creation of this was motivated by making it constexpr. This could be the default in C++23: https://open-std.org/JTC1/SC22/WG21/docs/papers/2019/p0533r5.pdf

{
const auto v = bit_cast<uint32_t>(value);
return ((v & F32ExpMask) == F32ExpMask) && ((v & F32FracMask) != 0);
}

inline bit_cast_constexpr bool isnan(double value) noexcept
{
const auto v = bit_cast<uint64_t>(value);
return ((v & F64ExpMask) == F64ExpMask) && ((v & F64FracMask) != 0);
}

template <typename T>
T fabs(T value) noexcept = delete;

template <>
inline float fabs(float value) noexcept
inline bit_cast_constexpr float fabs(float value) noexcept
{
return bit_cast<float>(bit_cast<uint32_t>(value) & F32AbsMask);
}

template <>
inline double fabs(double value) noexcept
inline bit_cast_constexpr double fabs(double value) noexcept
{
return bit_cast<double>(bit_cast<uint64_t>(value) & F64AbsMask);
}
Expand All @@ -365,13 +384,13 @@ template <typename T>
T fneg(T value) noexcept = delete;

template <>
inline float fneg(float value) noexcept
inline bit_cast_constexpr float fneg(float value) noexcept
{
return bit_cast<float>(bit_cast<uint32_t>(value) ^ F32SignMask);
}

template <>
inline double fneg(double value) noexcept
inline bit_cast_constexpr double fneg(double value) noexcept
{
return bit_cast<double>(bit_cast<uint64_t>(value) ^ F64SignMask);
}
Expand All @@ -380,15 +399,15 @@ template <typename T>
T copysign(T a, T b) noexcept = delete;

template <>
inline float copysign(float a, float b) noexcept
inline bit_cast_constexpr float copysign(float a, float b) noexcept
{
const auto a_u = bit_cast<uint32_t>(a);
const auto b_u = bit_cast<uint32_t>(b);
return bit_cast<float>((a_u & F32AbsMask) | (b_u & F32SignMask));
}

template <>
inline double copysign(double a, double b) noexcept
inline bit_cast_constexpr double copysign(double a, double b) noexcept
{
const auto a_u = bit_cast<uint64_t>(a);
const auto b_u = bit_cast<uint64_t>(b);
Expand All @@ -399,7 +418,7 @@ template <typename T>
inline T fceil(T value) noexcept
{
static_assert(std::is_floating_point_v<T>);
if (std::isnan(value))
if (isnan(value))
return std::numeric_limits<T>::quiet_NaN(); // Positive canonical NaN.

// The FE_INEXACT error is ignored (whenever the implementation reports it at all).
Expand All @@ -410,7 +429,7 @@ template <typename T>
inline T ffloor(T value) noexcept
{
static_assert(std::is_floating_point_v<T>);
if (std::isnan(value))
if (isnan(value))
return std::numeric_limits<T>::quiet_NaN(); // Positive canonical NaN.

// The FE_INEXACT error is ignored (whenever the implementation reports it at all).
Expand All @@ -429,7 +448,7 @@ template <typename T>
inline T ftrunc(T value) noexcept
{
static_assert(std::is_floating_point_v<T>);
if (std::isnan(value))
if (isnan(value))
return std::numeric_limits<T>::quiet_NaN(); // Positive canonical NaN.

// The FE_INEXACT error is ignored (whenever the implementation reports it at all).
Expand All @@ -441,7 +460,7 @@ T fnearest(T value) noexcept
{
static_assert(std::is_floating_point_v<T>);

if (std::isnan(value))
if (isnan(value))
return std::numeric_limits<T>::quiet_NaN(); // Positive canonical NaN.

// Check if the input integer (as floating-point type) is even.
Expand All @@ -467,11 +486,11 @@ __attribute__((no_sanitize("float-divide-by-zero"))) inline constexpr T fdiv(T a
}

template <typename T>
inline T fmin(T a, T b) noexcept
inline bit_cast_constexpr T fmin(T a, T b) noexcept
{
static_assert(std::is_floating_point_v<T>);

if (std::isnan(a) || std::isnan(b))
if (isnan(a) || isnan(b))
return std::numeric_limits<T>::quiet_NaN(); // Positive canonical NaN.

if (a == 0 && b == 0 && (signbit(a) || signbit(b)))
Expand All @@ -481,11 +500,11 @@ inline T fmin(T a, T b) noexcept
}

template <typename T>
inline T fmax(T a, T b) noexcept
inline bit_cast_constexpr T fmax(T a, T b) noexcept
{
static_assert(std::is_floating_point_v<T>);

if (std::isnan(a) || std::isnan(b))
if (isnan(a) || isnan(b))
return std::numeric_limits<T>::quiet_NaN(); // Positive canonical NaN.

if (a == 0 && b == 0 && (!signbit(a) || !signbit(b)))
Expand Down