Feectools with cupy - #74
Conversation
# Conflicts: # feectools/core/bsplines.py # feectools/ddm/blocking_data_exchanger.py # feectools/ddm/tests/test_cart_1d.py # feectools/ddm/tests/test_cart_2d.py # feectools/ddm/tests/test_cart_3d.py # feectools/feec/global_geometric_projectors.py # feectools/fem/tensor.py # feectools/linalg/basic.py Co-authored-by: max-models <181350288+max-models@users.noreply.github.com>
|
@max-models is this ready? |
No, there are still tests in struphy which are failing. So I need to do a few more iterations before merging this. You can merge your branches first. |
I can wait. The error in the test seems to be related to the move of |
Yes, but also the tests don't pass locally when running with the cupy backend yet... Working on it |
|
@spossann struphy tests are finished now. |
Redo of #25 See corresponding PR in feectools here: struphy-hub/feectools#74 Notes: This MR is still pointing towards `318`, but it will be merged into devel once finished (which will be after 318 has already been merged). To run on my PC, I still need to do: ``` export ARRAY_BACKEND=cupy export CUDA_PATH=/usr/local/cuda export PATH=$CUDA_PATH/bin:$PATH ``` I am then running the Vlasov standard parameters ``` struphy params Vlasov -y python params_Vlasov.py ``` To do - [ ] Fix the need for setting CUDA path - [ ] We have many cases of `from numpy import xyz`, this makes sense in pyccelized kernels since we will not be allowed to use cupy there. But in my current testing, but otherwise not. - [ ] When calling a pyccelized kernel while using cupy, we can convert the arrays to numpy arrays, they should also be converted back to cupy arrays afterwards. --------- Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: max-models <181350288+max-models@users.noreply.github.com>
See corresponding PR in struphy here: struphy-hub/struphy#93
Wait for #44 to merge first