Performance model and memory problems on GPU
Dear developers,
first of all, thank you very much developing such a great solver.
In my current project I am using PastiX' hybrid GPU mode through PaRSEC. It works great and the speedup is really showing. However, there are currently two problems I am facing:
-
A default performance model and cost function is provided for CPU and GPU (common/model.c) but I wonder how to reasonably determine coefficients for my target hardware? I found the CMake-option "PASTIX_GENERATE_MODEL", but can't figure out whether i can turn that data into coefficients.
-
When factorizing some of my larger matrices and choosing a larger block size at the same time, PaRSEC returns illegal memory access errors during factorization. I assume this is happens because more memory than available is requested from the GPU. Do you possibly know a solution for this problem or are you aware of this problem?
Best regards,
Peter