[Intel-gfx] [PATCH 0/2] introduce derefenrencing helpers
Daniel Vetter
daniel at ffwll.ch
Wed Mar 17 22:45:55 CET 2010
On Wed, Mar 17, 2010 at 02:15:26PM -0700, Eric Anholt wrote:
> On Mon, 8 Mar 2010 13:35:01 +0100, Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> Informal review on IRC was "wow, that's a lot of churn. Does it do
> something useful?"
>
> Certainly the dev_priv one makes the intel driver look different from
> all other drivers, and I don't think in a good way. If we had a record
> of getting types mixed up, the type safety might make a good argument,
> but when you just copy and paste the line for the lookup every time
> anyway...
Well, type-safety was not the goal of these ... I've intended to make
drm_gem_object subclassable (i.e. obj_priv.base instead of obj_priv->obj).
This should
- save a few bytes and at least one cachelines (the drm bo is no longer somewhere
totally different).
- save the need to carry around two pointers per bo (x86-32 has only 4
truely usable registers ...).
And therefore hopefully make execbuf go faster. The dev_priv was more or
less just a corollary. I've various stages of this lying around but never
really managed to complete & benchmark it. But know that this dreaded i8xx
cache coherency problem is finally making progress (at least here) I plan
to tackle this idea for real.
-Daniel
--
Daniel Vetter
Mail: daniel at ffwll.ch
Mobile: +41 (0)79 365 57 48
More information about the Intel-gfx
mailing list