Initial attempts at i965 text batching
Keith Whitwell
keith at tungstengraphics.com
Wed Dec 19 10:47:41 PST 2007
Carl Worth wrote:
> On Wed, 19 Dec 2007 11:41:04 -0500, "=?UTF-8?Q?Kristian_H=C3=B8gsberg?=" wrote:
>> Which reminds me, I did a couple of DRM patches to use MI_SET_CONTEXT
>> that I promised to send out. I've not benchmarked them, and there is
>> still work to do to finish them (they leak memory) but if there's a
>> lot of overhead in emitting the invariant state, these patches might
>> help.
>
> Thanks for the starting point.
>
> And thanks, Michel, for pointing out how I can fix opreport.
>
> Even without opreport though, it's obvious that there's still tons of
> stuff being emitted in every prepare_composite uselessly. One piece is
> the "invariant" state Kristian mentions. This is followed by a "long
> sequence" of commands most of which are the same for all composite
> operations.
>
> So it should be a fairly simple matter to simply emit those at the
> beginning of each batch rather than with every composite operation
> within the batch. (In fact, Keith and I have even done that work once
> or twice before on earlier branches.)
>
> That will still leave some per-batch overhead of course, but we should
> be able to amortize most of that with increasing the number of
> operations in each batch.
Indeed -- even if you only have two operations per batch, you get 50% of
the theoretical best-possible optimization of this state emit by just
removing one of them.
Keith
More information about the xorg
mailing list