[Mesa-dev] [PATCH 02/10] i965 gs: Reduce information in key to avoid unnecessary recompiles.

[Eric Anholt]

On Mon, 5 Dec 2011 20:01:52 -0800, Paul Berry <stereotype441 at gmail.com> wrote:
> On 5 December 2011 18:42, Eric Anholt <eric at anholt.net> wrote:
> 
> > On Mon, 5 Dec 2011 15:14:13 -0800, Paul Berry <stereotype441 at gmail.com>
> > wrote:
> > > On 5 December 2011 14:53, Eric Anholt <eric at anholt.net> wrote:
> > >
> > > > On Mon,  5 Dec 2011 09:40:45 -0800, Paul Berry <
> > stereotype441 at gmail.com>
> > > > wrote:
> > > > > Previously, the geometry shader program key was storing all the
> > > > > information necessary to compute the exact structure of the VUE map
> > > > > (attrs and userclip_active).  However, the GS program doesn't depend
> > > > > on the exact structure of the VUE map; only on the size (in 256-bit
> > > > > registers) of the VUE.  So we were recompiling the GS program
> > > > > unnecessarily when the VUE map changed in structure but not in size.
> > > > >
> > > > > This patch changes the key to store just the size of the VUE map, not
> > > > > its structure.
> > > >
> > > > Hmm.  I'm betting that brw_compute_vue_map() is way more expensive than
> > > > the extra couple bytes of key data to compare.
> > > >
> > >
> > > True, but it's way less expensive than unnecessarily compiling (and
> > > uploading to the GPU) a brand new GS program that is identical to the one
> > > that's already there.  That's what this patch is avoiding.
> >
> > That still happens a limited number of times, while key setup is per
> > drawing operation.  And if you've got the same program code, you should
> > even end up associating the same instructions with the identical copies.
> >
> 
> So am I correct in assuming you're ok with this patch, provided that I do a
> follow-on patch that computes the vue map at the top of the pipeline, as
> previously discussed?

Yeah, I just wanted to clarify what is generally true for thinking about
performance in the driver: key setup is called many thousands of times
more frequently than compile.  (If compile happens after the first
frame, something is usually broken).
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 197 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20111206/a27e0e9a/attachment.pgp>