This branch compares several implementation of the Laplace operator, to understand what is the optimal way to implement time and memory efficient forces in Julia.
The test uses the Gray-Scott model on grid of different sizes to compare the performances of different class of algorithms:
Old_Laplace is our current implementation that involves reshaping
*_c* algorithm that use circshift
stencil uses matrix multiplications for the finite differences
conv uses the convolution with a kernel
*Lux* use Lux framework to perform the convolution as a NN
This is an overview of my results
Overall circhift or Lux seems to perform the best. Before drawing a final conclusion I think it is worth to test the results using GPU
This branch compares several implementation of the Laplace operator, to understand what is the optimal way to implement time and memory efficient forces in Julia. The test uses the Gray-Scott model on grid of different sizes to compare the performances of different class of algorithms:
Old_Laplace
is our current implementation that involves reshaping*_c*
algorithm that usecircshift
stencil
uses matrix multiplications for the finite differencesconv
uses the convolution with a kernel*Lux*
use Lux framework to perform the convolution as a NNThis is an overview of my results![A](https://github.com/DEEPDIP-project/CoupledNODE.jl/assets/58949181/fb8be6da-1621-48ab-adab-0007913c37ee)
Overall
circhift
orLux
seems to perform the best. Before drawing a final conclusion I think it is worth to test the results using GPU