Use Eigen on CPUs. #1349
Use Eigen on CPUs. #1349
Conversation
|
Logfiles from GitLab pipeline #222085 (:no_entry:) have been uploaded here! Status and direct links: |
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #1349 +/- ##
=======================================
Coverage 86.16% 86.17%
=======================================
Files 179 179
Lines 13665 13659 -6
=======================================
- Hits 11775 11770 -5
+ Misses 1890 1889 -1 ☔ View full report in Codecov by Sentry. |
| nmodl::crout::solveCrout<double>(N, J.data(), F.data(), X_solve.data(), pivot.data()); | ||
| #else | ||
| X_solve = J.partialPivLu().solve(F); | ||
| #endif |
There was a problem hiding this comment.
Could you add a brief description/comment if this is done for performance reasons or something else?
I don't remember this well and hence asking for clarity: do we expect small differences in the results between the two? If there are two code paths, just think from CPU vs GPU debugging or results comparison perspectives. (obviously, not thinking fp differences as they exist on cpu vs gpu. But more from solver/accuracy perspective).
There was a problem hiding this comment.
Yes, without using NMODL is a 25% regression, compared to NOCMODL on the Masoli circuit we've been running recently.
That's a tricky question. The variation we implement ourselves is also "partial pivoting". There's two ways of doing LU decompositions: Dolittle and Crout. They result in almost the same matrices. Mostly I think they just iterate over the matrix differently. For a well-conditioned matrix, I doubt it makes any difference.
No description provided.