Speed up CPU applications?

[edwardlavender]

I wonder if we can speed up applications on the CPU e.g., via multithreading? I was trying to implement the code on the CPU (to test whether some convergence issues could be related to floating point accuracy on the GPU) but for our bathymetry grid the CPU code is slow (~2 hours for a filter run for one month with 2 min time resolution). But it may be tricky to parallelise operations such as the convolution without breaking the GPU code...

[scheidan]

I think we first need to identify the slow parts.

For example for the filter run what is time i) the convolution, ii) the observations, and iii) memory handling and garbage collection take?

For i) we may be able to use https://github.com/JuliaImages/ImageFiltering.jl but I think I remember that this was rather slow.
ii) should be rather easy to multi-thread.
iii) is tricky, because the GPU does not like inplace operation so much