Closed TeachRaccooon closed 1 year ago
Not sure what's causing it to segfault in CI. Will need
geqrt
in tester, but omitting that shouldn't cause segfaults, it seems.
I was thinking that that is due to the wrong setup of the line in run_tests.py, but seems like something else is the issue. Also, I noticed that there are no tests for gemqr.
Just tested gemqrt
in RandLAPACK and the routine itself works fine.
Looks like something is off with the way I've set up tests.
Is there a bug in LAPACK with the workspace query?
Remember, the LAPACK++ wrapper code is written by a Python script reading the LAPACK documentation. It's entirely possible that something goes amiss in that parsing and translation process. The generator is meant to get most of the code, but likely needs tweaks, especially if the documentation doesn't exactly follow LAPACK's (unwritten) style.
Is there a bug in LAPACK with the workspace query?
Remember, the LAPACK++ wrapper code is written by a Python script reading the LAPACK documentation. It's entirely possible that something goes amiss in that parsing and translation process. The generator is meant to get most of the code, but likely needs tweaks, especially if the documentation doesn't exactly follow LAPACK's (unwritten) style.
So the generated code did not actually perform the worspace query, I just decided to add it because it was done in gemqr
(any my querying strategy was wrong).
It looks like gemqrt
has a very well-defined workspace size in the first place.
Did the manual mapping of transposition types, resolved the last issue.
Hi @mgates3 I see that there is apending change request, but I think I addressed all the issues that were there (unless github is not showing me something for whatever reason). So just wanted to check if we can merge this PR.
@TeachRaccooon That just means I haven't approved the changes. I'll look at it probably today or tomorrow.
I noticed your fork is specific to this feature (TeachRaccooon/lapackpp-add_gemqrt
) and uses the master
branch in that fork. That works, but it's usually cleaner to create one fork (say, TeachRaccooon/lapackpp
) and create a branch for each specific feature (say, branch gemqrt
).
Can you add an empty commit to trigger the CI actions again? The "re-run jobs" button will re-run with the old master, not with the latest master, which seems odd to me.
git commit --allow-empty -m "Empty-Commit"
I'm going to squash all this into one commit when merging, so the empty commit will disappear.
Adding a gemqrt() routine, needed in RandLAPACK. Code is initially generated via
lapackpp/tools/wrapper_gen.py
.