[rescue] Re: SGI Onyx4 - uses ATI graphics chips (sigh)

[Joshua D. Boyd]

On Fri, Aug 15, 2003 at 08:31:23AM -0400, Sandwich Maker wrote:

> i don't understand why the film industry would care - i'm not sure
> it's possible to distinguish more than 30 bits of color resolution,
> maybe even 24.

First, when they say it is rendering 32bits to the frame buffer, I'm
fairly certain that they mean 8 bits per channel, not two channels of 11
bits and one chanel of 10.  In this case, 24bit and 32bit don't change
how the picture looks, they just change how it is stored and processed.
The extra 8 bits might be an alpha channel, or it might just be padding
for word alignment.  In the case of SGI, I'd think alpha channel.

Now, 30bits virtually always refers to 3 10bit channels, although if
that's what people mean, they usually say 10 bit, not 30 bit graphics. 

48bits typically means 4 12bit channels.  What is the forth channel?
Probably an alpha channel, but why we care about that when it hits the
ramdac, I don't know.  We probably don't, but since the same frame
buffer can be used for both output to the ramdac, and for copying to
some place else for some other reason, we preserve it anyway, I guess. 

Anyway, it isn't that hard to create images where 24bits isn't enough.
If you have an image with enough dynamic range to make 24bits
inadequate, you could just compress it and most people wouldn't notice
the difference.  But, if you clip the image, people can tell that
somethings wrong.  So, 24 or 32bits may be enough to screen a movie for
all but the most sensitive people, but only after the raw data has been
beaten into a shape optimal for presentation using only that amount of
bits. 

Further, when you start compositing images, you need more bits to keep
from having trouble, particularly with clipping.  One can get away with
these bits being used for processing only, not display, but that isn't
exactly optimal.   But most, if not all, of Windows and Linux and MacOS
X compositing programs have to dither down to 8 bit linear graphics from
10 log, 16 linear, 32bit float, or any number of other internal formats.