Re: [AD] clipping line algorithm

[Sven Sandberg <svens@xxxxxxxxxx>]

Eric wrote:
> > So, basically, if this new flag is not set, the library would be 
> > allowed to bypass entirely the vtable of the bitmap? I'm still a bit 
> > uncomfortable with this idea.

Bob replied:
> I don't like the idea of bypassing the vtable. If we can, we should have 
> a putpixel and _putpixel vtable entry.

I understand it's uglier to not use the vtable, but a bit of profiling 
indicates that it really speeds things up. The following are the cases I 
tested.

 (1) rotate_sprite() uses bmp->vtable->putpixel and bmp->vtable-getpixel.
 (2) rotate_sprite() uses _putpixel and _getpixel, but as function 
pointers (should be equivalent to having a vtable entry for _putpixel, 
in terms of speed).
 (3) rotate_sprite() as it is now, using bmp_write*() for writing and 
BITMAP->line[y][x] for reading.

Method (1) was surprisingly a bit faster than method (2); however, 
method (3) was (in color depths 8, 15, 16, 24, and 32, respectively) 
about 8, 7, 7, 3.4, and 3.8 times faster than any of the other ones. (I 
profiled 10000 calls to rotate_sprite() from a 320x200 memory sprite to 
the centre of a 640x480 memory bitmap, with angles evenly distributed 
around the circle, and trying all color depths. See 
http://peg.it.uu.se/~sven/rotate-profile.zip )

Although bypassing the vtable is a little bit ugly, is there any 
particular situation where it is really significantly bad? As long as we 
have a version that uses the vtable, and BITMAP->id is able to indicate 
when the optimized version can be used instead, I think we should be fine?

--
Sven