Bananymous
/
banan-os
1
1
Fork 1
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

BAN: Cleanup math code and add SSE sqrt 

We should prefer SSE instructions when they are easily available. For
other functions x87 is just simpler. It's hard to write faster and close
to as accurate approximations with SSE.

This does not use xmmintrin.h as clangd does not like that file and
starts throwing errors in every file that includes this :)
This commit is contained in:
Bananymous 2026-03-21 03:38:11 +02:00
parent d18a0de879
commit 09292bb87e
1 changed files with 50 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

88
BAN/include/BAN/Math.h
View File

@ -6,6 +6,19 @@
#include <float.h>
// This is ugly but my clangd does not like including
// intrinsic headers at all
#if !defined(__SSE__) || !defined(__SSE2__)
#pragma GCC push_options
#ifndef __SSE__
#pragma GCC target("sse")
#endif
#ifndef __SSE2__
#pragma GCC target("sse2")
#endif
#define BAN_MATH_POP_OPTIONS
#endif
namespace BAN::Math
{
@ -167,7 +180,7 @@ namespace BAN::Math
template<floating_point T>
static T modf(T x, T* iptr)
{
const T frac = BAN::Math::fmod<T>(x, 1);
const T frac = BAN::Math::fmod<T>(x, (T)1.0);
*iptr = x - frac;
return frac;
}
@ -175,15 +188,15 @@ namespace BAN::Math
template<floating_point T>
inline constexpr T frexp(T num, int* exp)
{
if (num == 0.0)
if (num == (T)0.0)
{
*exp = 0;
return 0.0;
return (T)0.0;
}
T _exp;
asm("fxtract" : "+t"(num), "=u"(_exp));
*exp = (int)_exp + 1;
T e;
asm("fxtract" : "+t"(num), "=u"(e));
*exp = (int)e + 1;
return num / (T)2.0;
}
@ -251,6 +264,7 @@ namespace BAN::Math
"fstp %%st(1);"
: "+t"(x)
);
return x;
}
@ -263,18 +277,9 @@ namespace BAN::Math
template<floating_point T>
inline constexpr T pow(T x, T y)
{
asm(
"fyl2x;"
"fld1;"
"fld %%st(1);"
"fprem;"
"f2xm1;"
"faddp;"
"fscale;"
: "+t"(x), "+u"(y)
);
return x;
if (x == (T)0.0)
return (T)0.0;
return exp2<T>(y * log2<T>(x));
}
template<floating_point T>
@ -310,16 +315,27 @@ namespace BAN::Math
template<floating_point T>
inline constexpr T sqrt(T x)
{
asm("fsqrt" : "+t"(x));
return x;
if constexpr(BAN::is_same_v<T, float>)
{
using v4sf = float __attribute__((vector_size(16)));
return __builtin_ia32_sqrtss((v4sf) { x, 0.0f, 0.0f, 0.0f })[0];
}
else if constexpr(BAN::is_same_v<T, double>)
{
using v2df = double __attribute__((vector_size(16)));
return __builtin_ia32_sqrtsd((v2df) { x, 0.0 })[0];
}
else if constexpr(BAN::is_same_v<T, long double>)
{
asm("fsqrt" : "+t"(x));
return x;
}
}
template<floating_point T>
inline constexpr T cbrt(T value)
{
if (value == 0.0)
return 0.0;
return pow<T>(value, 1.0 / 3.0);
return pow<T>(value, (T)1.0 / (T)3.0);
}
template<floating_point T>
@ -346,30 +362,21 @@ namespace BAN::Math
inline constexpr T tan(T x)
{
T one, ret;
asm(
"fptan"
: "=t"(one), "=u"(ret)
: "0"(x)
);
asm("fptan" : "=t"(one), "=u"(ret) : "0"(x));
return ret;
}
template<floating_point T>
inline constexpr T atan2(T y, T x)
{
asm(
"fpatan"
: "+t"(x)
: "u"(y)
: "st(1)"
);
asm("fpatan" : "+t"(x) : "u"(y) : "st(1)");
return x;
}
template<floating_point T>
inline constexpr T atan(T x)
{
return atan2<T>(x, 1.0);
return atan2<T>(x, (T)1.0);
}
template<floating_point T>
@ -378,10 +385,10 @@ namespace BAN::Math
if (x == (T)0.0)
return (T)0.0;
if (x == (T)1.0)
return numbers::pi_v<T> / (T)2.0;
return +numbers::pi_v<T> / (T)2.0;
if (x == (T)-1.0)
return -numbers::pi_v<T> / (T)2.0;
return (T)2.0 * atan<T>(x / (T(1.0) + sqrt<T>((T)1.0 - x * x)));
return (T)2.0 * atan<T>(x / ((T)1.0 + sqrt<T>((T)1.0 - x * x)));
}
template<floating_point T>
@ -411,7 +418,7 @@ namespace BAN::Math
template<floating_point T>
inline constexpr T tanh(T x)
{
const T exp_px = exp<T>(x);
const T exp_px = exp<T>(+x);
const T exp_nx = exp<T>(-x);
return (exp_px - exp_nx) / (exp_px + exp_nx);
}
@ -441,3 +448,8 @@ namespace BAN::Math
}
}
#ifdef BAN_MATH_POP_OPTIONS
#undef BAN_MATH_POP_OPTIONS
#pragma GCC pop_options
#endif