Support for fast GPU matrix assembly
Merge request reports
Activity
mentioned in issue #16 (closed)
mentioned in issue #15 (closed)
Hi @szampini - I'll need a bit of time to review in the next few days. I was discussing aspects with @sbna in December. As you know, he has a similar chunk of code for a direct coupling to amgx. It could possibly be useful to treat the transcription as a form of "algorihthm" with template parameters for the data types.
BTW: of the 7 commits, are they all related to this merge request? Or have some slipped in from pre-2012?
Each commit is self-contained. The COO support is one of them. I'm aware they have such code, but this addition does not preclude others. @mark We should chat next week.
Edited by stefano zampini@mark Any specific reason to hold up this?
-
crossed-wires. I thought that the three of us (incl. @sbna) wanted to discuss this along with amgx refactoring before proceeding.
-
Hi @szampini - need to pick up on this again. Its faster GPU assembly makes sense, but I'd probably use OpenFOAM structures to memory management. However, not sure how many of the other commits on the same branch are still relevant.
You can edit commit !8 (415ff47c) to use OpenFOAM structures. All the other commits contain small fixes here and there, and I guess they can be merged.
added 7 commits
- 218abc71 - Preliminary support for geometric multigrid from PC_ML
- 11693a1b - Skip allreduce calls in MatAssembly
- 679da174 - Implementing fast assembly with COO support
- d1bec4b2 - Report L1 final norm if KSP_NORM_NONE has been used without foam convergence check
- 53efa810 - Need to log setup of block solvers into pc stage
- 90134581 - Add example dictionary for CUDA solvers
- b61bab96 - Support older versions of PETSc
Toggle commit list-
Hi @szampini - made a few very minor cosmetic changes. Can fixup again in the future.
Thanks - and sorry about it having taken so long. Wanted to coordinate with the bits that Simone had been working on, but will have to revisit in the future.
mentioned in commit cc029f24
@mmartineau has recently update the foam2scr APIs and made some optimizations in AMGx (private branch). We faced some issues in having a stack software running with cuda 11 on M100.
