Re: Floating point in the kernel

[Paul M. <le...@li...>]

On Fri, Oct 05, 2007 at 11:51:09PM +0100, Adrian McMenamin wrote:
> I am still toying with the idea of writing some accelerate 2D stuff for
> the Dreamcast PVR2 (essentially using the 3D capabilities to accelerate
> 2D).
> 
> But the 3D hardware appears to natively rely on FP maths.
> 
> Is there anyway round this? (As I have always understood SH4 kernel code
> does not use FP - though this won't touch the SH4 as such).
> 
> Or is this a classic X11 v kernel issue?
> 
Well, using any FP from the driver itself is certainly not doable. The
basic framebuffer optimizations should extend to fast writes (via DMA and
so) and not much else. There are fillrect/copyarea/imageblit primitives
that you could presumably interface the TA for, but if you can't find a
way to do it without FP, then you'll have to stick with the generic
implementations.

At this point you're stuck with doing all of your extensive optimizations
in something like X or DirectFB. It's still worth doing deferred I/O for
bulk page write-out via the DMAC for the framebuffer, which should at
least give you some more bandwidth for things like X and mplayer which
mmap() the device.