[PROJ] Vector/SIMD acceleration

[Even Rouault]

On vendredi 17 avril 2020 15:13:12 CEST Thomas Knudsen wrote:
> yes, but in my case, it calls 4 times the 4-double, the last time
> recomputing the
> 5th and 6th-last coordinates once more (since I assume this is faster than
> calling
> the serial version twice), but ignores the repeated computations, using
> only the
> two first elements of the paralleldata structure in the last trip through
> the loop...

Hum, sounds a bit complicated. If you have a 4-double version, maybe your strategy can be 
worth for (N % 4) == 2 or 3, but for (N % 4) == 1, the serial version would be faster. From my 
quick benchmarking, the implementation of the (serial) trancendent functions in Linux libm is 
more efficient than the one of sleef. But that's expected since sleef must use branch-less 
implementations whereas libm can use all sort of tricks more efficient for a serial version.
Anyway, that's really a tiny implementation detail.

-- 
Spatialys - Geospatial professional services
http://www.spatialys.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/proj/attachments/20200417/2ea21c9f/attachment.html>