[Bf-cycles] regarding OSL texture perforamnce

[Martijn Berger]

Hi Brecht,

regarding further performance for OSL I was looking at making the
structures align more closely to get rid of some of the cost associated
with the shaderdata_to_shaderglobals() calls.

But I would need to factor out the  differential3 and differential to be
more just float3 and float like.
Also I am not sure about the alignment of Imath::Vec3<Float> vs float3 in
all cases / platforms.

An other thing I am already trying to implement is a shim so the svm
texture-backend gets the OIIO interface that should yield the biggest
performance gain it seems from my simple testing. ( testing Barcelona with
missing textures on svm vs osl )

Not sure what your thought's on this subject are but would love to hear
them.

Martijn


On Mon, Aug 5, 2013 at 2:50 PM, Brecht Van Lommel <
brechtvanlommel at pandora.be> wrote:

> Good find, I wasn't aware there was any per thread info we could
> cache. I've committed this patch now with some tweaks, found about a
> 10% performance improvement on images.blend here.
>
> Thanks,
> Brecht.
>
> On Sun, Aug 4, 2013 at 5:17 PM, Martijn Berger <martijn.berger at gmail.com>
> wrote:
> > Hi fellow cyclists,
> >
> > I did a little experiment to try and improve the performance of
> cycles/OSL
> > in scenes that use image textures. I tested the Barcelona scene from
> ba.org
> > and tried a bit of an ugly hack to cache the OIIO perthread info.
> >
> > I only have an ancient macbook pro here with me on holiday and would
> need to
> > investigate further but initial results are a ~ 15% improvement on
> overall
> > render time on this scene.
> >
> > I expect the improvement to be bigger for machines with more cores but I
> > cannot test that from here.
> >
> > I attached the patch
> >
> > Martijn
> >
> > _______________________________________________
> > Bf-cycles mailing list
> > Bf-cycles at blender.org
> > http://lists.blender.org/mailman/listinfo/bf-cycles
> >
> _______________________________________________
> Bf-cycles mailing list
> Bf-cycles at blender.org
> http://lists.blender.org/mailman/listinfo/bf-cycles
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.blender.org/pipermail/bf-cycles/attachments/20130805/d33e1961/attachment.htm