Mailinglist Archive: radeonhd (93 mails)

[Matthias Hopf <mhopf@xxxxxxx>]

On Feb 02, 10 20:37:22 +1300, Michael Cree wrote:
> On 26/01/10 11:50, Christian König wrote:
> > IDCT and MC should be pretty easy, IDCT are just a bunch of cos and mul
> > operations, and MC scripted blitter operations.
> Yeah, on the surface it's straightforward, but often CPU implementations of 
> the DCT for JPEG/MPEG trade off precision for efficiency by using fixed 
> point or even integer arithmetic.  With hardware implementation it might be 
> nice to aim for precision to maintain best quality assuming sufficient 
> speedup?  I'm not familiar with shaders; do they provide floating point?

Actually, integer arithmetics has only been added *after* floating point
was already defined and working :-)

> > The real tricky part is
> > VLD, i have no clue how this works.
> VLD?  What's that?  I'm guessing - variable length decoding?

Guess so. Anything that requires an inherent serial workflow (Huffman
etc.) is truly bad to parallelize, especially in a SIMD array like GPUs.

Matthias

-- 
Matthias Hopf <mhopf@xxxxxxx>      __        __   __
Maxfeldstr. 5 / 90409 Nuernberg   (_   | |  (_   |__          mat@xxxxxxxxx
Phone +49-911-74053-715           __)  |_|  __)  |__  R & D   www.mshopf.de
-- 
To unsubscribe, e-mail: radeonhd+unsubscribe@xxxxxxxxxxxx
For additional commands, e-mail: radeonhd+help@xxxxxxxxxxxx