Rust version of SOLWEIG

[songololo]

@biglimp This repo sub-classes the solweig_runner classes to add Rust version of SOLWEIG.

The only change is that this version swaps-in `solweig.py. It is similar to existing solweig calc scripts but swaps-in Rust functions:

• Shadows
• GVF
• Anisotropic
• L veg
• K veg

These functions were identified through performance profiling and appear to work as intended per test_rustalgos.py

Demo here: Athens-demo.py

Individually these functions run between 3 - 20x faster with overall speedup of around 5x for the Athens demo. Larger rasters or more cores probably more.

[biglimp]

Nice! We are also working on implementation of GPU-calculation for the shadow casting algorithms. I wonder if a combination of rust and GPU is possible and if so, what the performance would be like...

[paolodep36]

what about this? https://github.com/xiaojianggis/gpusvf?tab=readme-ov-file. I am also wondering if numba could be another option. what you think?

[biglimp]

what about this? https://github.com/xiaojianggis/gpusvf?tab=readme-ov-file. I am also wondering if numba could be another option. what you think?

We are working on a gpu-version solweig. Should be available this autumn. We have tried numba implementation. The decrease in computation time was not great.

[songololo]

Nice! We are also working on implementation of GPU-calculation for the shadow casting algorithms. I wonder if a combination of rust and GPU is possible and if so, what the performance would be like...

@biglimp as an experiment I added a version of GPU shadowing to the rust branch. For the test dataset it is about 30x faster for shadowing calls but the speedup will vary widely.

PS - full disclosure the GPU version was mostly "vibe coded" from the rust version but it seems to be passing the tests I was using for the rust code.

[songololo]

Another update that larger files sometimes run out of RAM or GPU buffers so a new tiling workflow has been added as a branch in umep-core and is now the version that umepr (rust) is based on

[biglimp]

Just remember when you tile your data that you need an overlap between tiles. See e.g. this publication: Lindberg et al. 2025