Conversation
…ements don't work for the pthreads variant
|
✅ Linter reported no issues All Python files are correctly formatted with Black. |
|
✅ Linter reported no issues All C/C++ files are correctly formatted with clang-format. |
|
You only changed the conda environment for the package build, but we are not using that anymore. In fact, perhaps we should just delete that folder entirely. If you open a PR there, you will trigger a rebuild of the conda-forge package. If you are creating a new env yourself, either using the CF package or building from source, you can add the openmp variants to your env file and it should work. |
|
Ah- I somehow assumed the systems in the two repos were linked up and didn't even think to check how builds were managed. Thanks for pointing me in the right place! I always manually add the openmp variant myself, but I figured it'd be nice for openmp to be a default so a user doesn't have to discover this on their own (even if we do put it in our documentation). |
|
@RaulPPelaez I converted this to remove devtools and the associated CI files. Merge it if you'd like that to go through, otherwise feel free to close. I'll open a PR in the feedstock since I think this is a default worth having- I've found DPStokes to be nearly unusable on large domains without your magnificent speedup. |
I keep running into this issue when I create a new environment where I set up a DPStokes simulation on a large domain and it never initializes. Each time, I eventually recall it's because the pthreads variant of libopenblas is really slow compared to the openmp variant which incorporates your speedup. I'm proposing to prefer the openmp version in the conda build.