Hi @flying-sheep I was wondering why the serial (fallback) kernels aren't compiled with nogil=True. This would enable parallelization via joblib and dask schedulers. Is there a reason for it? If not I can open a PR showing the speedup also.
I also want to point to something I learned. Tbb doesn't oversubscribe because it uses a process-wide arena. So, we can suggest tbb in certain OS'es instead of falling back. But setting nogil=True will be enough to be fast enough in itself.
Hi @flying-sheep I was wondering why the serial (fallback) kernels aren't compiled with nogil=True. This would enable parallelization via joblib and dask schedulers. Is there a reason for it? If not I can open a PR showing the speedup also.
I also want to point to something I learned. Tbb doesn't oversubscribe because it uses a process-wide arena. So, we can suggest tbb in certain OS'es instead of falling back. But setting nogil=True will be enough to be fast enough in itself.