OpenMP parallelization improvements

[tjof2]

Parallelization over frames / frame blocks would be better than the current inner-loop implementation.

[tjof2]

Also figuring out the best way to run on distributed computing architectures. MPI or an alternative?

[hentr]

yeah, I just tested with an 4core/8thread intel i7-6700 @3.4GHz, 512x512 sized frames.

2 threads: 36 sec/fr 4 threads: 30 sec/fr 8 threads: 27 sec/fr

So there's a good room for improvement, is this something that is not too much work? I guess this would be in the C++ part of the code, not python

[tjof2]

I think its a case of commenting out

• https://github.com/tjof2/pgure-svt/blob/master/src/svt.hpp#L129
• https://github.com/tjof2/pgure-svt/blob/master/src/svt.hpp#L165

And adding in the OpenMP #pragma here:

• https://github.com/tjof2/pgure-svt/blob/master/src/lib-PGURE-SVT.cpp#L161

But this approach would need to take care over:

• https://github.com/tjof2/pgure-svt/blob/master/src/lib-PGURE-SVT.cpp#L232

To avoid competing threads from trying to overwrite the same frame.