Compact legendre polynomials #164
Conversation
|
(convert to draft for the moment ==> it is fnished but we should first merge the other PR, then rebase and review) |
lukasm91
force-pushed
the
compact-legendre-polynomials
branch
from
04194bc to
d0d2349
Compare
lukasm91
force-pushed
the
compact-legendre-polynomials
branch
from
d0d2349 to
2a87188
Compare
|
Beside replacing ext_acc with the proper "copy module", we are only left with this PR from the old GPU branch which I now rebased on top of develop. This PR is compacting the legendre polynomials and with this removing zero padding. This PR is not expected to interfere with any other PR, as it only touches the GEMMs. Since I am touching the CUDA interfaces anyway, I also added const to pointers in the interface.
@samhatfield @wdeconinck Feel free to review when you find time. |
There was a problem hiding this comment.
I think this looks all fine and what a nice saving!
Perhaps @samhatfield can run it in few places to double check all is in order ?
|
If I can find the time this week, I will take a look. It might have to wait until next year. |
Legendre polynomials don't need to be stored zero-padded, we can just concatenate them with proper zero padding.
E.g. on tco2559, this saves up to almost 40GB per rank (used to be 64 GB, now it is 26 GB). Initially, we did this for Leonardo, because legendre coefficients used almost the whole device memory.