#1 IDCT selection simplification

[Anonymous]

Although it is not serious, I think that this would be
a good change:
Instead of a number of check boxes for all of the
different IDCTs, there was a pull-down menu that
had them.
perhaps something like:
reference IDCT 100%
standard IDCT 99.99999%
Altivec IDCT 99.998%

or something like that. Perhaps also put up a
warning if the person selects reference IDCT..
saying that the difference in quality is negligible,
but the difference in speed will be significant.

The main issue with this at the moment is that
there are two check boxes that affect the inverse
discrete cosine transform. It seems more
complicated than it needs to be, and also
misleading (that you can use an altivec reference
IDCT)

[Anonymous]

Logged In: YES
user_id=350001

Hi

I don't know who has tested the quality of the different idct routines, but in my tests the fast-idct (non-altivec) is 100% ieee1180 conform. So maybe we can remove the ref-idct preference checkbox and use it for something else.

Denis

[Colin Mckellar]

Logged In: YES
user_id=369030

Denis, I remember a post ages ago (probably by vlich) that mentioned the quality.. it was about what I mentioned in the first post. (I wasn't a member then)

I think altivec was a bit better.

the post was about the number of pixels in 1000 that was off by one, in one of the YUV scales... for the standard IDCT, it was about 15, Altivec had about 70...

but if you think about it, each pixel has 3 bytes (or so) assigned to it... this is seriously simplistic (and probably fundamentaly wrong), but if you multiply 1000 by 24 bits, you get 24000bits... divide 15 by that, and you get 0.0000625... take away one, and take the abosolute value, you get 99.9365%

do the same thing for 70 out of 24000, and you get 99.708%

of course this is a bullshit scale, as for one, it doesn't get any bit incorrect, it gets the least significant one incorrect...

anywho, to stop playing with numbers, and actually comment on your suggestion: yes, you can probably get rid of the reference IDCT altogether... there is no reason to use it, unless you are insane... the act of scaling (let alone encoding it into something else) would do more damage than using even the altivec IDCT would do.
Also, would it be possible to get rid of the altivec option, and just have it automatically use altivec, if its available?

Blibbler

[Stuart Espey]

Logged In: YES
user_id=206022

yes, if the fast one is 100% conformant then there is not reason to keep the reference one.