ANUGA GPU support about anuga_core HOT 4 OPEN

@samcom12 good idea. I have been thinking about this as well. The first step I think is to convert the inner timestepping procedures to cython C and then invoking openmp in the various expensive procedures. Hopefully we could then move over to offloading to the GPU.

from anuga_core.

Thanks @stoiver .
I just tried profiling for offloading benefits with sample rectangular testcase. The results are promising showing almost 4X speedup which can further increase with real datasets we are using.

You can refer the attached report.
offload_model.tar.gz

from anuga_core.

@samcom12, the results look promising. It is good that the offload advisor just picks up 2 procedures, _compute_fluxes_central and _extrapolate_second_order_edge_sw to look at. I think it also suggests that we should look at just using standard openmp on those two procedures. Does intel have an advisor for studying likely improvement using openmp?

from anuga_core.

Hello @stoiver

There is a tool which converts serial C,C++, and Fortran code to OpenMP based code.

http://rosecompiler.org/

And, with Intel VTune we can do threading analysis to see how effectively threads are being used in the code.

https://www.intel.com/content/www/us/en/develop/documentation/vtune-help/top/analyze-performance/parallelism-analysis-group/threading-analysis.html

from anuga_core.