Create out-of-place workflow

[JordiBolibar]

Right now everything is implemented in-place in order to be memory efficient for the PDE solving. This is great for the forward PDE runs, but we need an out-of-place version to be run with AD, particularly for ReverseDiff.jl.

This implies to basically reproduce the same workflow as in place, changing all functions like SIA2D! -> SIA2D, to indicate that it's out-of-place. All the code is available in the current ODINN.jl repo.

[facusapienza21]

I implemented this is my Huginn.jl fork using the out-of-place implementation of the PDE (no UDE) forward model. This consisted in creating the same functions for the in-place case but without with out-of-place evaluation. Now two set of separate test are being run

@testset "Halfar Solution (in-place)" halfar_test(; rtol=0.02, atol=1.0, inplace=true)

@testset "Halfar Solution (out-of-place)" halfar_test(; rtol=0.02, atol=1.0, inplace=false)

We can see the difference in performance between the two methods using BenchmarkTools.jl. This first example is using the in-place option:

And the same ODE problem but with out-of-place allocation memory:

We can see the difference in memory storage. However, what I find odd is that the in-place method is just slightly faster than the out-of-place... I was expecting the difference between the two to be more notorious.

Between the functions I implemented for the out-of-place we have

• run\_0(). Ideally this should be called just run`, but that is an existing function in Julia base. Any suggestion for a better name?
• simulate_iceflow_PDE()
• initialize_iceflow_model()
• SIA()
• avg_surface_velocity()
• surface_V()

[JordiBolibar]

@facusapienza21 We should post this on Discourse in order to better understand why the in-place workflow doesn't provide a significant speed up wrt the out-of-place. There are clear differences in memory usage, though.

[JordiBolibar]

@facusapienza21 What's the status on this? Have you created the Discourse post?