Optimize number of threads per block for each kernel about fbpic HOT 3 CLOSED

I will have a look at it. In my fork I have a branch /soerenjalas/fbpic/tree/test_tpb where the 1D and 2D threads per block can be set via environment variables (if you want I can do a WIP pull request). I will do some benchmarks similar to what @RemiLehe did with the threading stuff.

from fbpic.

That sounds great!! Thanks a lot for taking care of this ; I think we can really some non-negligible speed-up from this!

from fbpic.

The test of the 1D threads per block on several GPU types show that the deposition an gathering methods could benefit from an adjustment of the tpb settings. A reasonable value for all GPUs would be 64.
Note that even though it's usually recommended to use a multiple of 32 tpb, the deposition kernels perform best with 48 tpb on the K80 and P100. I don't know why.

The GPU used below was actually a K80 on Jureca

from fbpic.