fdo-mirrors/freetype
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/freetype/freetype.git synced 2026-05-07 06:28:02 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into 'master'

Browse source 
Replace some fixed-point maths with floating point maths.  Much faster.

See merge request freetype/freetype!347
This commit is contained in:
Nicholas Chapman 2026-04-12 03:58:11 +00:00
commit 7fffe85dd3
2 changed files with 17 additions and 114 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

95
src/base/ftcalc.c
View file

@ -37,6 +37,7 @@
#include <freetype/internal/ftcalc.h>
#include <freetype/internal/ftdebug.h>
#include <freetype/internal/ftobjs.h>
#include <math.h>
/* cancel inlining macro from internal/ftcalc.h */
#ifdef FT_MulFix
@ -786,91 +787,15 @@
FT_BASE_DEF( FT_UInt32 )
FT_Vector_NormLen( FT_Vector* vector )
{
FT_Int32 x_ = vector->x;
FT_Int32 y_ = vector->y;
FT_Int32 b, z;
FT_UInt32 x, y, u, v, l;
FT_Int sx = 1, sy = 1, shift;
FT_MOVE_SIGN( FT_UInt32, x_, x, sx );
FT_MOVE_SIGN( FT_UInt32, y_, y, sy );
/* trivial cases */
if ( x == 0 )
{
if ( y > 0 )
vector->y = sy * 0x10000;
return y;
}
else if ( y == 0 )
{
if ( x > 0 )
vector->x = sx * 0x10000;
return x;
}
/* Estimate length and prenormalize by shifting so that */
/* the new approximate length is between 2/3 and 4/3. */
/* The magic constant 0xAAAAAAAAUL (2/3 of 2^32) helps */
/* achieve this in 16.16 fixed-point representation. */
l = x > y ? x + ( y >> 1 )
: y + ( x >> 1 );
shift = 31 - FT_MSB( l );
shift -= 15 + ( l >= ( 0xAAAAAAAAUL >> shift ) );
if ( shift > 0 )
{
x <<= shift;
y <<= shift;
/* re-estimate length for tiny vectors */
l = x > y ? x + ( y >> 1 )
: y + ( x >> 1 );
}
else
{
x >>= -shift;
y >>= -shift;
l >>= -shift;
}
/* lower linear approximation for reciprocal length minus one */
b = 0x10000 - (FT_Int32)l;
x_ = (FT_Int32)x;
y_ = (FT_Int32)y;
/* Newton's iterations */
do
{
u = (FT_UInt32)( x_ + ( x_ * b >> 16 ) );
v = (FT_UInt32)( y_ + ( y_ * b >> 16 ) );
/* Normalized squared length in the parentheses approaches 2^32. */
/* On two's complement systems, converting to signed gives the */
/* difference with 2^32 even if the expression wraps around. */
z = -(FT_Int32)( u * u + v * v ) / 0x200;
z = z * ( ( 0x10000 + b ) >> 8 ) / 0x10000;
b += z;
} while ( z > 0 );
vector->x = sx < 0 ? -(FT_Pos)u : (FT_Pos)u;
vector->y = sy < 0 ? -(FT_Pos)v : (FT_Pos)v;
/* Conversion to signed helps to recover from likely wrap around */
/* in calculating the prenormalized length, because it gives the */
/* correct difference with 2^32 on two's complement systems. */
l = (FT_UInt32)( 0x10000 + (FT_Int32)( u * x + v * y ) / 0x10000 );
if ( shift > 0 )
l = ( l + ( 1 << ( shift - 1 ) ) ) >> shift;
else
l <<= -shift;
return l;
// vx' = (x / len) * 65536 = (x * 65536) / len = ((vx / 65536) * 65536) / len = vx / len = vx / (sqrt(vx*vx + vy*vy) / 65536)
// = 65536 * vx / (sqrt(vx*vx + vy*vy)
const float vx = (float)vector->x;
const float vy = (float)vector->y;
const float vlen = sqrtf(vx*vx + vy*vy);
const float scale_factor = 65536.f / vlen;
vector->x = (FT_Pos)(vx * scale_factor);
vector->y = (FT_Pos)(vy * scale_factor);
return (FT_UInt32)vlen;
}

36
src/base/fttrigon.c
View file

@ -32,6 +32,7 @@
#include <freetype/internal/ftobjs.h>
#include <freetype/internal/ftcalc.h>
#include <freetype/fttrigon.h>
#include <math.h>
/* the Cordic shrink factor 0.858785336480436 * 2^32 */
@ -416,35 +417,12 @@
FT_EXPORT_DEF( FT_Fixed )
FT_Vector_Length( FT_Vector* vec )
{
FT_Int shift;
FT_Vector v;
if ( !vec )
return 0;
v = *vec;
/* handle trivial cases */
if ( v.x == 0 )
{
return FT_ABS( v.y );
}
else if ( v.y == 0 )
{
return FT_ABS( v.x );
}
/* general case */
shift = ft_trig_prenorm( &v );
ft_trig_pseudo_polarize( &v );
v.x = ft_trig_downscale( v.x );
if ( shift > 0 )
return ( v.x + ( 1L << ( shift - 1 ) ) ) >> shift;
return (FT_Fixed)( (FT_UInt32)v.x << -shift );
// len * 65536 = sqrtf(x*x + y*y) * 65536 = sqrt((vx / 65536) * (vx / 65536) + (vy / 65536) * (vy / 65536)) * 65536
// = sqrt((vx*vx + vy*vy) / 2^32) * 65536
// = sqrt((vx*vx + vy*vy) / 2^32) * sqrt(2^32)
// = sqrt((vx*vx + vy*vy) / 2^32 * 2^32)
// = sqrt(vx*vx + vy*vy)
return (FT_UInt32)sqrtf((float)vec->x*(float)vec->x + (float)vec->y*(float)vec->y);
}