Open mohd-afeef-badri opened 2 days ago
Matrix format | Dimension | Processor | PR |
---|---|---|---|
Legacy | 2D / 3D | CPU | |
CSR | 2D / 3D | CPU | |
#164 | |||
COO | 2D | CPU | #165 |
COO Sort | 2D | CPU | #165 |
COO GPU | 2D | CPU/GPU | |
NodeWise CSR | 2D | CPU | |
BuildLess CSR | 2D / 3D | CPU/GPU | |
CSR GPU | 2D | CPU/GPU (atomic ope) |
@mohd-afeef-badri I will open a PR (with a new branch associated to it in my forked repo of arcaneFEM) for these modifications, is it OK ?
Also, concerning the CuSPARSE format, you were saying that we could remove it, should I do it within this PR ?
Finally, do you think it's a good idea to do some refactoring in Poisson module to make the code closer of what is done in Laplace module and so make it clearer (also within this same PR) ?
@toutane thanks for the table.
yes you can open a branch for these devs. Cusparse needs to be purged, so start with that as a separate PR.
I guess refactoring here is not needed right away, as this module is the testing ground for now. eventually this would be a separate PR.
in a nutshell. you can open multiple PRs for each of the tasks here.
I opened a PR (#164) where I removed all code related to CuSPARSE, all CI tests pass.
@toutane can you please add table with 3D/2D compatibility resume here. Then next step could be to target one by one the matrix formats that are not 3D compatible.