Re: [fluid-dev] parallel loading of soundfont samples

[Hamish Moffatt]

On 30/10/20 5:42 am, Marcus Weseloh wrote:
> Hi,
>
> for anybody who wants to try the OpenMP parallel loading, I've just
> pushed the fix for the vorbis load to use separate file pointers to
> Toms branch here:
> https://github.com/FluidSynth/fluidsynth/tree/parallelize-sf3-loading
>
> Hamish, would be great to see how that performs on your 6 core
> machine. Like I said, I get a little over twice the speed on four
> cores.
> And the loaded font does actually play a test midi file... so at least
> it's not completely broken :-)

Beautiful! Over 6 times faster. Here's master:

./fluidsynth -i -a file *.sf3  9.13s user 0.17s system 99% cpu 9.298 total

And here's the parallel version:

./fluidsynth -i -a file *.sf3  15.22s user 0.38s system 1105% cpu 1.411 
total

Sounds good too.


To answer your question, is it enough? Yes, but now that I am loading in 
the background I will keep doing that too. Here's what I have done over 
time.

1. Block while load soundfonts. I used to use 1xSF2, 1xSF3 so is was ~5 
seconds.

2. Start loading soundfonts (2xSF3) in a background thread (~10 
seconds), do other initialisations, wait for soundfonts to finish, 
unblock application.

3. Start loading soundfonts, do other initialisation, unblock 
application, and only block again if the user tries to use the synth 
before loading is finished (not common).


So now I have #3 it's not really important for me to speed up the 
loading any further, but I posted this just for interest's sake and I'm 
glad it's generated some discussion, and I think other applications can 
benefit too. In order to get the best out of modern CPUs we have to use 
more threads as the cores aren't getting much faster.


cheers

Hamish