Use cusolverMpSygvd - a generalized eigenvalue solver #3787
Conversation
sliwowitz
force-pushed
the
cusolvermp
branch
3 times, most recently
from
c900392 to
73864ca
Compare
sliwowitz
force-pushed
the
cusolvermp
branch
from
73864ca to
a6e6004
Compare
|
The automatic checks passed, but I suppose none of test are actually testing a build + run with cuSOLVERMp enabled, right? For the record, using the fixes in PR #3770, I'm building CP2K with the Nvidia HPC SDK like |
That's correct! There were a couple of reasons:
In all likelihood things have improve over the last year. So, maybe you want to give it another try? |
2 participants
In
cp_fm_geeig, instead of reducing the generalized eigenvalue problem into standard from on the CPU and then solving the standard form on the GPU usingcusolverMpSyevd, solve the whole generalized problem on the GPU usingcusolverMpSygvd. This gives a nice speedup for all problems which are large enough to benefit from using cuSOLVERMp in the original code.The original API for solving only the standard form on the GPU has been kept in the code, should there be a need to use it from some other routine.