Ben tried to write an SSE version of the convolution routine, but did not get a working one. As I recall the problem is that SSE lacks a critical operation to make it efficient (or perhaps even possible) to run as vector code. You can always just set the compiler flags to the extreme levels and see if that helps.
I wrote Altivec code for partconv~ a few years ago, and built a PPC version with an optimized FFTW, but I don't know if that is what is included with pd-extended or not. I guess if you can load it on Intel machines it is running a more generic version. FFTW needs to have at least an architecture, or ideally a CPU, specific build to run well.
hello,
I would appreciate knowing if the build of partconv~ v0.2 included in
0.40.3-extended-macosx104 is currently taking advantage of sse
optimization?
thanks
Gregorio
--
Gregorio García Karman
Runzstraße 71.
79102 Freiburg - Germany
home: 0049 (0)761 2852 808
office: 0049 (0)761 3808 276
handy: 0049 (0)1515 3292 098
_______________________________________________
Pd-list@iem.at mailing list
UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list