AK: Move rounding function to the top of AK/Math.h

These are useful in other algorithms, so lets move them up
This commit is contained in:
Hediadyoin1 2023-05-11 11:23:17 +02:00 committed by Andreas Kling
parent 47c074ab11
commit 6573ace8f1
Notes: sideshowbarker 2024-07-17 05:58:46 +09:00

408
AK/Math.h
View file

@ -81,6 +81,214 @@ constexpr size_t product_odd() { return value * product_odd<value - 2>(); }
return res; \ return res; \
} }
namespace Rounding {
template<FloatingPoint T>
constexpr T ceil(T num)
{
if (is_constant_evaluated()) {
if (num < NumericLimits<i64>::min() || num > NumericLimits<i64>::max())
return num;
return (static_cast<T>(static_cast<i64>(num)) == num)
? static_cast<i64>(num)
: static_cast<i64>(num) + ((num > 0) ? 1 : 0);
}
#if ARCH(AARCH64)
AARCH64_INSTRUCTION(frintp, num);
#else
return __builtin_ceil(num);
#endif
}
template<FloatingPoint T>
constexpr T floor(T num)
{
if (is_constant_evaluated()) {
if (num < NumericLimits<i64>::min() || num > NumericLimits<i64>::max())
return num;
return (static_cast<T>(static_cast<i64>(num)) == num)
? static_cast<i64>(num)
: static_cast<i64>(num) - ((num > 0) ? 0 : 1);
}
#if ARCH(AARCH64)
AARCH64_INSTRUCTION(frintm, num);
#else
return __builtin_floor(num);
#endif
}
template<FloatingPoint T>
constexpr T round(T x)
{
CONSTEXPR_STATE(round, x);
// Note: This is break-tie-away-from-zero, so not the hw's understanding of
// "nearest", which would be towards even.
if (x == 0.)
return x;
if (x > 0.)
return floor(x + .5);
return ceil(x - .5);
}
template<Integral I, FloatingPoint P>
ALWAYS_INLINE I round_to(P value);
#if ARCH(X86_64)
template<Integral I>
ALWAYS_INLINE I round_to(long double value)
{
// Note: fistps outputs into a signed integer location (i16, i32, i64),
// so lets be nice and tell the compiler that.
Conditional<sizeof(I) >= sizeof(i16), MakeSigned<I>, i16> ret;
if constexpr (sizeof(I) == sizeof(i64)) {
asm("fistpll %0"
: "=m"(ret)
: "t"(value)
: "st");
} else if constexpr (sizeof(I) == sizeof(i32)) {
asm("fistpl %0"
: "=m"(ret)
: "t"(value)
: "st");
} else {
asm("fistps %0"
: "=m"(ret)
: "t"(value)
: "st");
}
return static_cast<I>(ret);
}
template<Integral I>
ALWAYS_INLINE I round_to(float value)
{
// FIXME: round_to<u64> might will cause issues, aka the indefinite value being set,
// if the value surpasses the i64 limit, even if the result could fit into an u64
// To solve this we would either need to detect that value or do a range check and
// then do a more specialized conversion, which might include a division (which is expensive)
if constexpr (sizeof(I) == sizeof(i64) || IsSame<I, u32>) {
i64 ret;
asm("cvtss2si %1, %0"
: "=r"(ret)
: "xm"(value));
return static_cast<I>(ret);
}
i32 ret;
asm("cvtss2si %1, %0"
: "=r"(ret)
: "xm"(value));
return static_cast<I>(ret);
}
template<Integral I>
ALWAYS_INLINE I round_to(double value)
{
// FIXME: round_to<u64> might will cause issues, aka the indefinite value being set,
// if the value surpasses the i64 limit, even if the result could fit into an u64
// To solve this we would either need to detect that value or do a range check and
// then do a more specialized conversion, which might include a division (which is expensive)
if constexpr (sizeof(I) == sizeof(i64) || IsSame<I, u32>) {
i64 ret;
asm("cvtsd2si %1, %0"
: "=r"(ret)
: "xm"(value));
return static_cast<I>(ret);
}
i32 ret;
asm("cvtsd2si %1, %0"
: "=r"(ret)
: "xm"(value));
return static_cast<I>(ret);
}
#elif ARCH(AARCH64)
template<Signed I>
ALWAYS_INLINE I round_to(float value)
{
if constexpr (sizeof(I) <= sizeof(u32)) {
i32 res;
asm("fcvtns %w0, %s1"
: "=r"(res)
: "w"(value));
return static_cast<I>(res);
}
i64 res;
asm("fcvtns %0, %s1"
: "=r"(res)
: "w"(value));
return static_cast<I>(res);
}
template<Signed I>
ALWAYS_INLINE I round_to(double value)
{
if constexpr (sizeof(I) <= sizeof(u32)) {
i32 res;
asm("fcvtns %w0, %d1"
: "=r"(res)
: "w"(value));
return static_cast<I>(res);
}
i64 res;
asm("fcvtns %0, %d1"
: "=r"(res)
: "w"(value));
return static_cast<I>(res);
}
template<Unsigned U>
ALWAYS_INLINE U round_to(float value)
{
if constexpr (sizeof(U) <= sizeof(u32)) {
u32 res;
asm("fcvtnu %w0, %s1"
: "=r"(res)
: "w"(value));
return static_cast<U>(res);
}
i64 res;
asm("fcvtnu %0, %s1"
: "=r"(res)
: "w"(value));
return static_cast<U>(res);
}
template<Unsigned U>
ALWAYS_INLINE U round_to(double value)
{
if constexpr (sizeof(U) <= sizeof(u32)) {
u32 res;
asm("fcvtns %w0, %d1"
: "=r"(res)
: "w"(value));
return static_cast<U>(res);
}
i64 res;
asm("fcvtns %0, %d1"
: "=r"(res)
: "w"(value));
return static_cast<U>(res);
}
#else
template<Integral I, FloatingPoint P>
ALWAYS_INLINE I round_to(P value)
{
if constexpr (IsSame<P, long double>)
return static_cast<I>(__builtin_llrintl(value));
if constexpr (IsSame<P, double>)
return static_cast<I>(__builtin_llrint(value));
if constexpr (IsSame<P, float>)
return static_cast<I>(__builtin_llrintf(value));
}
#endif
}
using Rounding::ceil;
using Rounding::floor;
using Rounding::round;
using Rounding::round_to;
namespace Division { namespace Division {
template<FloatingPoint T> template<FloatingPoint T>
constexpr T fmod(T x, T y) constexpr T fmod(T x, T y)
@ -680,159 +888,6 @@ using Hyperbolic::cosh;
using Hyperbolic::sinh; using Hyperbolic::sinh;
using Hyperbolic::tanh; using Hyperbolic::tanh;
template<Integral I, FloatingPoint P>
ALWAYS_INLINE I round_to(P value);
#if ARCH(X86_64)
template<Integral I>
ALWAYS_INLINE I round_to(long double value)
{
// Note: fistps outputs into a signed integer location (i16, i32, i64),
// so lets be nice and tell the compiler that.
Conditional<sizeof(I) >= sizeof(i16), MakeSigned<I>, i16> ret;
if constexpr (sizeof(I) == sizeof(i64)) {
asm("fistpll %0"
: "=m"(ret)
: "t"(value)
: "st");
} else if constexpr (sizeof(I) == sizeof(i32)) {
asm("fistpl %0"
: "=m"(ret)
: "t"(value)
: "st");
} else {
asm("fistps %0"
: "=m"(ret)
: "t"(value)
: "st");
}
return static_cast<I>(ret);
}
template<Integral I>
ALWAYS_INLINE I round_to(float value)
{
// FIXME: round_to<u64> might will cause issues, aka the indefinite value being set,
// if the value surpasses the i64 limit, even if the result could fit into an u64
// To solve this we would either need to detect that value or do a range check and
// then do a more specialized conversion, which might include a division (which is expensive)
if constexpr (sizeof(I) == sizeof(i64) || IsSame<I, u32>) {
i64 ret;
asm("cvtss2si %1, %0"
: "=r"(ret)
: "xm"(value));
return static_cast<I>(ret);
}
i32 ret;
asm("cvtss2si %1, %0"
: "=r"(ret)
: "xm"(value));
return static_cast<I>(ret);
}
template<Integral I>
ALWAYS_INLINE I round_to(double value)
{
// FIXME: round_to<u64> might will cause issues, aka the indefinite value being set,
// if the value surpasses the i64 limit, even if the result could fit into an u64
// To solve this we would either need to detect that value or do a range check and
// then do a more specialized conversion, which might include a division (which is expensive)
if constexpr (sizeof(I) == sizeof(i64) || IsSame<I, u32>) {
i64 ret;
asm("cvtsd2si %1, %0"
: "=r"(ret)
: "xm"(value));
return static_cast<I>(ret);
}
i32 ret;
asm("cvtsd2si %1, %0"
: "=r"(ret)
: "xm"(value));
return static_cast<I>(ret);
}
#elif ARCH(AARCH64)
template<Signed I>
ALWAYS_INLINE I round_to(float value)
{
if constexpr (sizeof(I) <= sizeof(u32)) {
i32 res;
asm("fcvtns %w0, %s1"
: "=r"(res)
: "w"(value));
return static_cast<I>(res);
}
i64 res;
asm("fcvtns %0, %s1"
: "=r"(res)
: "w"(value));
return static_cast<I>(res);
}
template<Signed I>
ALWAYS_INLINE I round_to(double value)
{
if constexpr (sizeof(I) <= sizeof(u32)) {
i32 res;
asm("fcvtns %w0, %d1"
: "=r"(res)
: "w"(value));
return static_cast<I>(res);
}
i64 res;
asm("fcvtns %0, %d1"
: "=r"(res)
: "w"(value));
return static_cast<I>(res);
}
template<Unsigned U>
ALWAYS_INLINE U round_to(float value)
{
if constexpr (sizeof(U) <= sizeof(u32)) {
u32 res;
asm("fcvtnu %w0, %s1"
: "=r"(res)
: "w"(value));
return static_cast<U>(res);
}
i64 res;
asm("fcvtnu %0, %s1"
: "=r"(res)
: "w"(value));
return static_cast<U>(res);
}
template<Unsigned U>
ALWAYS_INLINE U round_to(double value)
{
if constexpr (sizeof(U) <= sizeof(u32)) {
u32 res;
asm("fcvtns %w0, %d1"
: "=r"(res)
: "w"(value));
return static_cast<U>(res);
}
i64 res;
asm("fcvtns %0, %d1"
: "=r"(res)
: "w"(value));
return static_cast<U>(res);
}
#else
template<Integral I, FloatingPoint P>
ALWAYS_INLINE I round_to(P value)
{
if constexpr (IsSame<P, long double>)
return static_cast<I>(__builtin_llrintl(value));
if constexpr (IsSame<P, double>)
return static_cast<I>(__builtin_llrint(value));
if constexpr (IsSame<P, float>)
return static_cast<I>(__builtin_llrintf(value));
}
#endif
template<FloatingPoint T> template<FloatingPoint T>
constexpr T pow(T x, T y) constexpr T pow(T x, T y)
{ {
@ -859,53 +914,6 @@ constexpr T pow(T x, T y)
return exp2<T>(y * log2<T>(x)); return exp2<T>(y * log2<T>(x));
} }
template<FloatingPoint T>
constexpr T ceil(T num)
{
if (is_constant_evaluated()) {
if (num < NumericLimits<i64>::min() || num > NumericLimits<i64>::max())
return num;
return (static_cast<T>(static_cast<i64>(num)) == num)
? static_cast<i64>(num)
: static_cast<i64>(num) + ((num > 0) ? 1 : 0);
}
#if ARCH(AARCH64)
AARCH64_INSTRUCTION(frintp, num);
#else
return __builtin_ceil(num);
#endif
}
template<FloatingPoint T>
constexpr T floor(T num)
{
if (is_constant_evaluated()) {
if (num < NumericLimits<i64>::min() || num > NumericLimits<i64>::max())
return num;
return (static_cast<T>(static_cast<i64>(num)) == num)
? static_cast<i64>(num)
: static_cast<i64>(num) - ((num > 0) ? 0 : 1);
}
#if ARCH(AARCH64)
AARCH64_INSTRUCTION(frintm, num);
#else
return __builtin_floor(num);
#endif
}
template<FloatingPoint T>
constexpr T round(T x)
{
CONSTEXPR_STATE(round, x);
// Note: This is break-tie-away-from-zero, so not the hw's understanding of
// "nearest", which would be towards even.
if (x == 0.)
return x;
if (x > 0.)
return floor(x + .5);
return ceil(x - .5);
}
template<typename T> template<typename T>
constexpr int clamp_to_int(T value) constexpr int clamp_to_int(T value)
{ {