On Monday 21 January 2008 23:55:24 Timothy Normand Miller wrote:
> On 1/21/08, Lourens Veen <[EMAIL PROTECTED]> wrote:
> > On Monday 21 January 2008 20:51:02 Kenneth Ostby wrote:
> > > Lourens Veen:
> > > >The S3 part, I think I have everything in the XP10 covered.
> > > >
> > > >Looking at it again, I guess I mainly don't understand the
> > > > connection between the bridge and the arbiter. There is only
> > > > one command queue, but the read end seems to consist of four
> > > > separate parallel FIFOs, and I don't understand why.
> > >
> > > Is it the the four different memory banks you wonder about?
> > > "arb2br_rdata_outA" and so forth?
> >
> > Yes. I understand that there are our memory banks, so it makes
> > sense to have four FIFOs with return data, one for each of them.
> > But then how come we only have one command queue? I can see sending
> > the same command to each bank, but you'd have to send different
> > data upon writing...
>
> There's one set of reads to request, but they get divvied up between
> each memory controller.  Not much point in more than one queue,
> because we can't enqueue them fast enough to make any meaningful
> difference.  We use the lower two address bits in the command to
> select the memory controller, when we pop a command out.

Right, that makes sense.

> For the return, each controller is running independently, so they
> can't put their return data into the same queue.  When returning the
> data over the bridge, we just pop words from the queues in the right
> order (when they arrive).

What about write commands though? Or don't we write to memory through 
the bridge? We'll have to upload textures somehow I guess.

Lourens

Attachment: pgpZdzdInVqnY.pgp
Description: PGP signature

_______________________________________________
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