Hi,
Richard Spindler schrieb:
2007/12/10, Herman Robak <[EMAIL PROTECTED]>:
YUV, YUVA, RGB, RGBA or floating point presision color format?
My guess is that 90% of the users ought to use YUVA, so let that
be the default.
Is the difference in memory footprint between YUV and YUVA so big
that it warrants a setting? I'm not sure.
Two things: First, I think in most cases, convinience is more
important than trying to save every possible tiny bit of memory.
Second, RGBA is 32bit, RGB is 24bit. Most SSE and Fast vector
operations in CPUs work on 32bit aligned data. My totally naive guess
is that in certain cases, the 32bit alignment of Alpha-Enabled Data
might be faster to process. But I won't claim that until properly
benchmarked. ;-)
Alignment is 128 bit for SSE IIRC.
Theoretically YUVA is 33% larger than YUV, but it can well be, that
YUV is stored in 32 bit words internally (wasting memory but speeding
up processing).
Another quesion are the numerical operations. If you have to process
4 channels instead of 3 things will get slowed down (at least in non-SIMD
routines).
The question of RGB vs YUV depends heavily on what you are doing.
Some effects (especially the ones which change the color)
work in faster in YUV, others in RGB. Then for some operations, you need
an alpha channel, they fail if you don't have one.
IMO the colormodel settings of a project aren't necessary at all. Usually
it's always possible for all elements of a pipeline to negotiate their
colormodels,
if the APIs are designed properly. If one want's to increase precision, one can
insert a "Force colormodel" filter to make subsequent operations work in float
RGBA
instead of 32 bit RGBA.
Furthermore, there is no law, that all video data in all tracks must have the
same colormodel
all the time and thinking so prevents some fundamental optimizations (both
speed and
quality wise): E.g. if YUV 4:2:0 images are read from the source, processed by
some simple
Brightness/Saturation/Contrast filter and written to the rendered file
(again in YUV 4:2:0), they don't have to be upsampled to 4:4:4 at all.
Burkhard
_______________________________________________
Cinelerra mailing list
Cinelerra@skolelinux.no
https://init.linpro.no/mailman/skolelinux.no/listinfo/cinelerra
