On 30/03/2016 18:02, Pavel Demin wrote: > Thanks for your suggestion to switch to the single precision FFTW3. > > I've replaced 'fftw_' with 'fftwf_' in wsprd_exp.c and tested > 'wsprd_exp -J -w' with 410 .wav files. This modification does not > affect the number of decoded messages and improves the decoding speed > by a couple of percents. > > Then I've replaced almost all the remaining doubles with floats in > wsprd_exp.c. This modification improves the decoding speed by another > couple of percents. > > Attached is a patch file with all the modifications. This patch also > fixes a small bug in the writec2file function that I described in my > previous e-mail. > > After applying this patch, I observe the following improvements in the > decoding speed: > ~10% on Intel Pentium > ~20% on ARM Cortex-A9 Hi Pavel,
those gains seem significant. Steve K9AN will probably review them. I wonder if there are any critical areas of calculation where the required FP precision might still warrant a double precision variable. 73 Bill G4WJS. ------------------------------------------------------------------------------ Transform Data into Opportunity. Accelerate data analysis in your applications with Intel Data Analytics Acceleration Library. Click to learn more. http://pubads.g.doubleclick.net/gampad/clk?id=278785471&iu=/4140 _______________________________________________ wsjt-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/wsjt-devel
