search

lattice / quda

QUDA is a library for performing calculations in lattice QCD on GPUs.
https://lattice.github.io/quda
Other
279 stars 94 forks source link

Asynchronous QUDA calls? #219

Open mathiaswagner opened 9 years ago

mathiaswagner commented 9 years ago

Just curious whether we could make QUDA calls asynchronous. Might be interesting for heterogenous computing but my intention was from a 'heterogeneous problem': QCD + QED. The QED part is computationally a lot less demanding and thus it might not always be worth offloading it to the GPU. So, it might be the best approach to not worry about getting the QED stuff on the GPU but execute it on the CPU while the GPU does the heavy lifting (QCD).

No need to get that in but maybe we can discuss that at some point.

maddyscientist commented 9 years ago

Also useful for multigrid, where coarse grid is on CPU and fine grid is on the GPU.

alexstrel commented 9 years ago

and for deflation as well (e.g., computing small eigenproblems on the host etc.)

mathiaswagner commented 9 years ago

That's what I also had in my mind but I think this would be asynchronous within QUDA and not necessarily be exposed asynchronous calls in the interface. (Correct me if I am wrong). Anyway, strongly related.