Re: [Dri-devel] Matrox G400/450 PCI support

[F. <j_r...@ya...>]

On 2002.06.10 18:32 Ville Syrjälä wrote:
> Here's my understanding of the situation. Someone more enlightened should
> correct me if I'm wrong.
> 
> On Mon, Jun 10, 2002 at 04:22:11PM +0000, Gareth Knight wrote:
> > I am interested in adding DRI support for an Algorithmics P6064 mips
> > development board.  I already have the standard X server working OK
> >
> > ...
> 
> This might be nasty as the card doesn't do scatter-gather DMA. You'd need
> a physically contiguous chunk of memory which the Linux kernel can't
> guarantee. I'm not sure how much memory the DMA buffers require. Putting
> textures to PCI memory might be useless speedwise, but maybe the driver
> would be cleaner as the card would fetch the textures by itself so the
> driver wouldn't have to swap them constatly. utah-glx required the user
> to permanently steal memory with the kernel mem= option. This approach
> would also mess up mtrr setup. I can think of one nice advantage: It
> would
> work on every card from G200 to G550. People with AGP cards could also
> test it so you'd have "lots" of testers.

I'm not sure about the Matrox design, but in the absence of AGP and PCI 
scatter-gather DMA, generally a card will only allow to have textures on 
the onboard memory. This it's what happens on the 3DFX, Gamma, and Mach64 
drivers when handling the PCI versions. This is also why you don't find 
PCI cards with so little memory as the AGP counterparts - they can't rely 
on the system memory to fill in the rest..

The only physical memory you need is for vertex buffers. For that you use 
many medium sized buffers which don't really need to be contiguous between 
themselves, just only when considered individually.

> ...

José Fonseca