Re: [Devel] Getting accurate advances

[David Bevan]

Werner LEMBERG wrote:
> > So I tried linearHoriAdvance which was better but, despite what the
> > tutorial says, still not accurate enough - because it is calculated
> > using x_ppem (ftobjs.c line 554) which has been rounded to the
> > nearest pixel (x_scale is similarly rounded). The cumulative error
> > can be several mm with a text string of 15cm or so.
>
>
> This is definitely a bug.
>
>
> > Perhaps the tutorial could be changed so that it doesn't claim that
> > linearHoriAdvance is accurate to 1/65536th of a pixel - or, better,
> > linearHoriAdvance could be calculated accurately (I can't see how to
> > do this without storing an additional unrounded value in a new
> > field).
>
>
> The necessary data is already available in metrics->x_scale, with
> sufficient accuracy I think.  We have to say
>
>   slot->linearHoriAdvance =
>     FT_MulDiv( slot->linearHoriAdvance, metrics->x_scale, 64 );
>
> instead of
>
>   slot->linearHoriAdvance =
>     FT_MulDiv( slot->linearHoriAdvance,
>                (FT_Long)metrics->x_ppem << 16, EM );
>
> in function FT_Load_Glyph.
>
> Please test.  If this works as expected I'll apply the change to the
> CVS.

As I mentioned in my first paragraph, metrics->x_scale is also rounded - 
in fact it's simply 2048 * metrics->x_ppem with font I'm using, so the 
rsults are the same as using metrics->x_ppem.

The relevant code is from line 2019 in ftobjs.c:

  dim_x = ( ( char_width * horz_resolution + (36+32*72) ) / 72 ) & -64;

  metrics->x_ppem  = (FT_UShort)( dim_x >> 6 );

  metrics->x_scale = FT_DivFix( dim_x, face->units_per_EM );

Note that the rounding occurs in the calculation for dim_x. Changing 
this to:

  dim_x = ( char_width * horz_resolution + 36 ) / 72;

  metrics->x_ppem  = (FT_UShort)( ( dim_x + 32 ) >> 6 );

  metrics->x_scale = FT_DivFix( dim_x, face->units_per_EM );

produces the correct advances when used with your suggested change - 
*but* this changes the x_scale value used throughout all resizing, 
rendering, hinting, etc. Is this change correct or dangerous? Has the 
x_scale value always been wrong or does it need to be rounded for some 
reason? Since I don't know the answer to these questions, I am very 
reluctant to recommend the above change. (A parallel change would also 
be necessary for y_scale and linearVertAdvance).

Thanks.

David %^>

--
 ________________________________________________________
 David Bevan, Emtex Ltd.
 Emtex House, Station Road, Kings Langley, Herts. WD4 8LH England
 Tel:    +44 (0)1923 270882
 E-mail: address@hidden
 Web:    www.emtex.com