[Intel-gfx] [PATCH 0/3] explicit fencing

[Owain Ainsworth]

On Sun, Apr 11, 2010 at 05:49:58PM +0100, Chris Wilson wrote:
> On Sun, 11 Apr 2010 15:38:21 +0200, Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> > Hi all,
> > 
> > These three patches implement explicit fencing via execbuf2 for
> > xf86-video-intel. Benchmarking results for thes patches on my i855GM are
> > underwhelming: gvim-0 from cairo-perf-traces is 17% faster, everything
> > else +/- a few percentage points the same. So tiled pixmaps still suck for
> > xrender on fence reg constrained hw (up to 2x slowdown on certain
> > cairo-perf traces versus untiled pixmaps).
> > 
> > Tested on my i855GM and i945GM.
> > 
> > If no one raises any issues with this, I'll merge these patches to master
> > in a few days.
> 
> Hmm, odd that this would appear to work without:
> 
> /**
>  * Enable use of fenced reloc type.
>  *
>  * New code should enable this to avoid unnecessary fence register
>  * allocation.  If this option is not enabled, all relocs will have fence
>  * register allocated.
>  */
> void
> drm_intel_bufmgr_gem_enable_fenced_relocs(drm_intel_bufmgr *bufmgr);

looking at libdrm, all the fenced_relocs struct member does is make
drm_intel_gem_bo_emit_reloc not require a fence if the object is fence.
drm_intel_gem_bo_emit_reloc_fence is unconditional. So this probably
does want fixing up.

-0-
-- 
Fine day to throw a party.  Throw him as far as you can.