Properly optimize Julia code

ODINN-SciML / ODINN.jl

Global glacier model using Universal Differential Equations for climate-glacier interactions

MIT License

71 stars 11 forks source link

Properly optimize Julia code #68

Open JordiBolibar opened 1 year ago

JordiBolibar commented 1 year ago

After some discussions in the Julia Discourse regarding Float64 vs Float32, some adjustments should be made to the code in order to avoid some performance pitfalls. Those are:

Move all code back to Float64. Apparently, unless one is using GPUs, there's no benefit on that for CPUs.
Use Int in exponents (e.g. in the flow law). See explanations in Discourse.
Avoid memory allocation in util functions following example in iceflow_sandbox model.

JordiBolibar commented 1 year ago

Partially implemented in SIA!. There are still some operations remaining for which the pre-initialized size of the matrices needs to be adjusted. This is not urgent for the scientific applications, so to be updated later on.

This could be a good task for Lucille in her internship to optimize the SIA in a forward model.

JordiBolibar commented 1 year ago

More optimizations have been implemented. There are still quite a lot of memory allocations, though. Removing almost all of them would results in a relatively unreadable code, and it would imply passing even more initialized context matrices to SIA!().

To be discussed if it's worth it.

JordiBolibar commented 1 year ago

This still requires a bit more work. Once we have some time, we should perform a proper code optimization following this: https://viralinstruction.com/posts/optimise/.

Another really important optimization will come from making all parameters local, following #36