GPU-friendly truncation implementations #349
Conversation
Codecov Report❌ Patch coverage is
... and 1 file with indirect coverage changes 🚀 New features to boost your workflow:
|
|
I think that implementation is very clean. Not sure I can explain all the errors. Some seem to originate from the eigenvectors of a DiagonalTensorMap also being diagonal, which is being changed, right? But other errors I cannot directly explain without running the code locally. |
|
I will have a look later, I have a local branch to fix the diagonal implementations already, we'll see what remains after. |
kshyatt
force-pushed
the
ld-truncation
branch
2 times, most recently
from
499360d to
77f0ffa
Compare
|
I think this one still needs a little work to get it over the finish line, but then we should be able to really try CUDA with the TNR and PEPS stuff |
kshyatt
force-pushed
the
ld-truncation
branch
from
5e5d87b to
f8d4dd7
Compare
|
Your PR no longer requires formatting changes. Thank you for your contribution! |
kshyatt
force-pushed
the
ld-truncation
branch
from
1b3ae78 to
1190fcd
Compare
|
wait please don't I'm working on this 😆 |
|
I'll stop I just wanted to fix formatting and rebase!!! |
|
Hahah but the rebase screws up the git history, now I gotta see how much git magic I know to fix my local branch |
|
Oh no I'm sorry! |
This reverts commit 77f0ffa.
|
|
||
| function MatrixAlgebraKit.findtruncated( | ||
| values::CuSectorVector, strategy::MatrixAlgebraKit.TruncationByError | ||
| ) |
There was a problem hiding this comment.
This probably needs to be special cased for p == Inf, if we want to support this. That in itself should be quite an easy case.
There was a problem hiding this comment.
I just checked, and at least in MatrixAlgebraKit we are currently not supporting this. While I don't necessarily have any strong opinions in favor or against this, I will for now simply add a check and leave this as TBA?
|
@kshyatt Do you have any clue why the CUDA tests could be giving segfaults? Is this something I caused or are we running into a different issue? Otherwise I think this PR is ready to get merged, would be great to get this done. |
|
CUDA 12 sometimes freaks out due to bad LLVM, I retried the build |
This is an attempt to get rid of the scalar-indexing oriented approach, and instead do more global operations.
The main point is to take all the singular/eigen values at the same time, and sort them in their entirety.