Alfio Lazzaro

A bit more, I have some screenshots that I can share... this is low priority...

We have some lapack dependencies in utils/dbcsr_blas_operations.F, which are used for testings (generation of random matrices). Probably we can refactor and remove the dependency...

I'm reopening this issue. Ideally, we need LAPACK only for the testing, the library itself doesn't need this dependency.

Thanks for reporting this problem. For any future problem related to DBCSR (anything file inside the exts/dbcsr), please post on the DBCSR page (https://github.com/cp2k/dbcsr/issues)

Comment those tests for the time being (suggested by @hfp !) See https://github.com/cp2k/dbcsr/pull/595/commits/dec083a1d1432c8ca946500d56b8f21383a1eda4

Could you check by compiling with `-DUSE_CUBLAS=OFF`? It seems an error due to https://github.com/cp2k/dbcsr/issues/229...

OK, thanks for the clarification, we are trying to reproduce the problem. Concerning this part: > It looks like dbcsr is trying to allocate memory in device 1 but launch...

@vitesselin Unfortunately, we don't have a multi-gpu system to try. We would like to ask you if you could try the following patch for replacing the lines https://github.com/cp2k/dbcsr/blob/f35f901e4460980aa06757294463a1e6308f8dc9/src/acc/dbcsr_acc_devmem.F#L435-L438 with ```...

Few clarifications: 1. we don't test multi-gpu configurations (see point 3. too) 2. I don't think #205 is related to this bug in the first place. #205 is when you...

I think I got the answer to the problem. Sorry for my ignorance in CUDA threading: https://devblogs.nvidia.com/cuda-pro-tip-always-set-current-device-avoid-multithreading-bugs/ @vitesselin I assume that 1 thread the code should work fine?