Conversation
Presumably, this was done to guarantee a certain minimum version of Thrust, but it fails with current (2.x) versions of Thrust. Let's assume that any reasonably recent version of Thrust will work, and don't try to enforce a specific version.
Without it, the config wasn't included, and `USE_CUDA` wasn not defined.
nvcc was complaining about "half" being redefined. It didn't say anything about where the original definition was, maybe it's a conflict with CUDA's `half` type. In any case, renaming the function to `one_half` seems to fix the issue.
This is kinda hacky, since MparticlesCuda hasn't been fully adapted, but it compiles, at least.
basically, reviving the hacky way it was handled...
There was a problem hiding this comment.
Was this related to the problem that MparticlesCuda has to be converted before being dumped to HDF5?
I don't understand why this would do anything, unless the operator() were actually being called somewhere other than in perform_diagnostic and actually got some other type of Mparticles. I trust that it compiles, I'm just confused.
(Referring to 09bdf21, gpu: fix hdf5 output with MparticlesCuda.)
|
Yes, I have had a FIXME comment indicating that I took the easy way out by just converting It was an easy fix, though obviously is probably even more opaque now. In the end, this probably should use some accessor rather than conversion anyway, which might improve the situation, as in, eliminates the need for special casing. Or, it could use host mirroring. Whatever, one thing is pretty clear: Having three different ways of achieving the same purpose is not KISS... |
2 participants
This fixes build errors when CUDA is enabled.