Pipe vs. plane fixes (hopefully third time's the charm)
jbarnes at virtuousgeek.org
Mon Sep 17 11:36:20 PDT 2007
On Thursday, September 13, 2007, Roland Dreier wrote:
> > We already have to read a few registers on this path though, so any
> > added cost should be hard to measure...
> I would assume that register reads are not posted, so every one ends
> up getting serialized and every additional read would hurt a lot,
> unless you're on a path where performance doesn't matter anyway.
Right, but my point was that this path already has a fairly high fixed
overhead--adding a single additional register read shouldn't double the
cost or anything. Moreover, it's not a performance critical path for
the most part: vblank swaps just have to get done before the next
frame is displayed, which is a fairly long time in terms of CPU
instructions. But again, if it turns out to be a problem we can always
use another scheme.
More information about the xorg