[Sidefx-houdini-list] Film at 16 bit
malexander at sidefx.com
Wed Mar 12 13:46:09 EDT 2008
Well, I can't speak for other compositing applications and what they do
internally. However, because there is no native 16b FP support on any
current CPU, it has to be handled specially. Some operations that don't
care about the color of the pixel (like cropping) would conceivably be
just as quick as they can treat the pixel as a 8B block. The easiest way
to handle the other cases is to simply convert 16b FP to 32b FP (which
is native to all CPUs), and I assume this is what most applications do.
You can do some very simple math with 16b FP, but not a whole lot before
it becomes faster just to convert, do the op in 32b FP, and convert back.
GPUs, on the other hand, support 16b FP natively, and have just fairly
recently added full 32b FP support. So, if there's any sort of
GPU-assisted processing going on in an application, presumably 16b FP
could be handled natively by the GPU. However, I don't see 16b FP ever
being added to the already insanely complex x86 instruction set, at
least until CPUs & GPUs merge, and even that's a bit of a longshot. I
can see the CPU designers wanting to optimize 32b FP rather than support
a somewhat limited format like 16b FP.
More information about the Sidefx-houdini-list