From kde-kimageshop  Thu May 17 15:03:46 2007
From: Tom Burdick <tburdi1 () uic ! edu>
Date: Thu, 17 May 2007 15:03:46 +0000
To: kde-kimageshop
Subject: Re: OpenMP and krita
Message-Id: <200705171003.46650.tburdi1 () uic ! edu>
X-MARC-Message: https://marc.info/?l=kde-kimageshop&m=117941422123859

Loading the SIMD registers shouldn't be too costly if the data is already 
pre-aligned with the right instrinsic datatype macros right?
as in movaps instead of movups. Been busy with school and robot project... so 
sorry for the instrusion here.

-Tom


On Thursday 17 May 2007 09:36:41 am Cyrille Berger wrote:
> On Thursday 17 May 2007, Boudewijn Rempt wrote:
> > Yes, that should be cool. Does OpenMP also use MMX and similar where
> > possible?
>
> no, it's not related :)
>
> > Because that's where I see the biggest benefit for compositing.
>
> Well the benefit of SIMD instructions is around 40% and 50% maximum (due to
> the cost of loading in the SIMD registers). While multithreading can reduce
> the CPU cost to a maximum 1/N (with N the number of processor). And some
> composite operation doesn't benefit from SIMD at all, at least that's what
> my experiments show, and it seems to be confirmed by the gimp source code
> ;)
>
> Anyway it needs experimenting, I am still working on SIMD library for Krita
> 2 :)


_______________________________________________
kimageshop mailing list
kimageshop@kde.org
https://mail.kde.org/mailman/listinfo/kimageshop