Alfio Lazzaro
Alfio Lazzaro
Given the fact that it is suggested to set the GPU even per thread (my understanding was that each thread should keep the same GPU of the rank, it turns...
> @alazzaro > Is there any plan of DBCSR to support the Multi-GPU in one node in the neer future? @vitesselin Assuming now other issues && Assuming I can run...
@hfp Yes, this is the idea... I must say, we don't need to save per stream, we need to save per rank and apply the setdevice any time we call...
Still, each stream within a rank cannot only access to a single device, so all stream will have the same value... Does the ACC implementation allow to have streams with...
Concerning the streams, what you did (@oschuett ) is still OK. NIVIDA didn't change that much since then... There are still 2 copy engines/compute with PASCAL, probably you have more...
@oschuett I'm far to be an expert, I'm just reporting what it is written here https://devblogs.nvidia.com/cuda-pro-tip-always-set-current-device-avoid-multithreading-bugs/ Quoting the bottom line: `To save yourself from a variety of multithreading bugs, remember:...
No idea about CP2K... Probably @oschuett knows better? In principle (never tried), you can drop GPU DBCSR by removing `-D__DBCSR_ACC` in the arch file.
Is this still relevant? Please close otherwise....
Good spot, I will fix it. Thanks!
Related to https://github.com/cp2k/dbcsr/issues/261