On 11/24/05, Benjamin Herrenschmidt <[EMAIL PROTECTED]> wrote:

> If your cannot do efficient DMAs to/from system memory, you'll have very
> poor performances with 3D texture uploads/downloads and with modern
> graphics APIs like EXA. I think the ability for the blitter to blit
> to/from system memory is critical.
>
> I think the "security" issue can be addressed differently.
>

We'll have DMA, for sure.  I think it's vital.  Since our GPU won't be
that fast, we have to make up for it by (a) keeping it fed all the
time, and (b) minimizing the host CPU work.  Of course, all of our
vertex shading is done by the host, but the fact that commands are
moved by DMA will help considerably.

As for images, there are special image upload/download commands. 
They're just privileged.  And the GPU can't directly access host
memory.  Instead, you have to upload an image (via DMA) before you can
use it.  The advantage (besides not allowing the GPU to look at kernel
memory) is that uploading an image is more efficient than accessing it
nonlinearly like you would get with a texture lookup.  The
disadvantage is that texture lookups may not look at all pixels in the
texture.  You win some, you lose some.  But our major goal is to make
the design small, resulting in some tradeoffs.

_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)

Reply via email to