[Lang] Load cuSparse and cuSolver library for preparing sparse matrix on GPU #4904
Conversation
✅ Deploy Preview for docsite-preview ready!
To edit notification comments on pull requests, go to your Netlify site settings. |
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
|
Sorry for the delayed update. Now, the BTW, should we add flags to indicate whether we should load those libraries, like |
|
Hi @Hanke98, thanks very much.
Maybe we could use another approach. When using the sparse matrix with How do you guys think @k-ye, @strongoier ? |
| } | ||
| } | ||
|
|
||
| CUSPARSEDriver::CUSPARSEDriver() { |
There was a problem hiding this comment.
nit: I'd suggest not to do this in the constructor, becuase if the lib fails to load, we get a broken/zombie object. A safer approach could be to load DLL in a function. If that fails, returns the exception/error to Python. Otherwise move that loaded lib into CuSolverDriver.
Also, we need to make sure not accidentally invoke CUSPARSEDRiver::get_instance() when we don't need the sparse matrix, otherwise Taichi could be broken if cuSparse is not installed.
I know this is not a newly introduced problem, so this is just something to think about :-)
SG! TBH I was initially thinking something like torch's tags, e.g. |
Related issue = #2906
Start to support sparse matrix on CUDA backend from this pr.
We first load the
cuSparseandcuSolverlibs via theDynamicLoaderclass in this pr.A new class,
CUDADriverBaseis introduced as the base class for the driver classes of these libs, which might also be helpful for the potential libs that we can use in the future.