[truetype] Improve accuracy of normalization of short vectors.

Unit vector components are stored as 2.14 fixed-point numbers. In
order to calculate all 14 bits accurately, a short vector to be
normalized has to be upscaled to at least 14 bits before its length
is calculated. This has been safe since accurate CORDIC algorithms
were adopted.

* src/truetype/ttinterp.c (Normalize): Scale short vectors by 0x4000.
This commit is contained in:
Alexei Podtelezhnikov 2013-01-12 23:05:55 -05:00
parent 081aba390a
commit 0e0fdc5dc8
2 changed files with 16 additions and 6 deletions

View File

@ -1,3 +1,15 @@
2013-01-12 Alexei Podtelezhnikov <apodtele@gmail.com>
[truetype] Improve accuracy of normalization of short vectors.
Unit vector components are stored as 2.14 fixed-point numbers. In
order to calculate all 14 bits accurately, a short vector to be
normalized has to be upscaled to at least 14 bits before its length
is calculated. This has been safe since accurate CORDIC algorithms
were adopted.
* src/truetype/ttinterp.c (Normalize): Scale short vectors by 0x4000.
2013-01-12 Alexei Podtelezhnikov <apodtele@gmail.com>
[truetype] Kill very old vector normalization hacks.

View File

@ -2635,8 +2635,6 @@
/* In case Vx and Vy are both zero, Normalize() returns SUCCESS, and */
/* R is undefined. */
/* */
static FT_Bool
Normalize( EXEC_OP_ FT_F26Dot6 Vx,
FT_F26Dot6 Vy,
@ -2647,17 +2645,17 @@
FT_UNUSED_EXEC;
if ( FT_ABS( Vx ) < 0x10000L && FT_ABS( Vy ) < 0x10000L )
if ( FT_ABS( Vx ) < 0x4000L && FT_ABS( Vy ) < 0x4000L )
{
Vx *= 0x100;
Vy *= 0x100;
if ( Vx == 0 && Vy == 0 )
{
/* XXX: UNDOCUMENTED! It seems that it is possible to try */
/* to normalize the vector (0,0). Return immediately. */
return SUCCESS;
}
Vx *= 0x4000;
Vy *= 0x4000;
}
W = TT_VecLen( Vx, Vy );