You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, only 2D and 3D transforms were supported on the GPU. During the original implementation, the thinking was that there wouldn't be much advantage to doing 1D on the GPU (IIRC).
However, in nifty-ls, we found that 1D was substantially faster on the GPU once you get to sufficiently large problems (e.g. N_data ~ 10^3, N_freq ~ 10^4). And besides, there may be cases where users have a multi-step workflow where they want to stay on the GPU the whole time.
I imagine this isn't that hard, it just needs to be implemented and tested.
The text was updated successfully, but these errors were encountered:
Currently, only 2D and 3D transforms were supported on the GPU. During the original implementation, the thinking was that there wouldn't be much advantage to doing 1D on the GPU (IIRC).
However, in nifty-ls, we found that 1D was substantially faster on the GPU once you get to sufficiently large problems (e.g.
N_data ~ 10^3
,N_freq ~ 10^4
). And besides, there may be cases where users have a multi-step workflow where they want to stay on the GPU the whole time.I imagine this isn't that hard, it just needs to be implemented and tested.
The text was updated successfully, but these errors were encountered: