This fix comes from @daedalus5 who has also been working on the libmathdx 0.2.1 evaluation in Warp:

in warp/build_dll.py, we need to change Line 396:

OLD:

linkopts.append(f'nvJitLink_static.lib /LIBPATH:"{args.libmathdx_path}/lib" mathdx_static.lib')

NEW:

linkopts.append(f'nvJitLink_static.lib /LIBPATH:"{args.libmathdx_path}/lib/x64" mathdx_static.lib')

While going through the code again, I figured out what caused #768. Turns out that the layout information was not passed along correctly. Since the layout information is only used in the mathdx calls, I took the liberty to directly fix it in this PR (1561b44) since it's directly related. I modified the tests accordingly to test for layout propagation.

Thanks so much for this @RSchwan . Phenomenal effort.

A few small notes:

Could you please add this decorator
@unittest.skipUnless(wp.context.runtime.core.is_mathdx_enabled(), "Warp was not built with MathDx support")

to

test_tile_math_forward_substitution()
test_tile_math_forward_substitution_multiple_rhs()
test_tile_math_back_substitution()
test_tile_math_back_substitution_multiple_rhs()

and this decorator
@unittest.skipUnless(wp.context.runtime.core.cuda_toolkit_version() >= 12600, "CUDA toolkit version is less than 12.6")

to

test_tile_math_cholesky()

in test_tile_mathdx.py? We need these for our CI pipeline.

Done. Let me know if you need other changes.

It looks like we need this:
@unittest.skipUnless(wp.context.runtime.core.cuda_toolkit_version() >= 12600, "CUDA toolkit version is less than 12.6")

over
test_tile_math_matmul() as well.

Apologies, these should both reference version 12060, not 12600. So
wp.context.runtime.core.cuda_toolkit_version() >= 12060

It looks like we need this: @unittest.skipUnless(wp.context.runtime.core.cuda_toolkit_version() >= 12060, "CUDA toolkit version is less than 12.6")

over test_tile_math_matmul() as well.

Just missing this.

Hey @RSchwan, could you also please clean up the commit history with git rebase/squash so it's down to a single commit? We need to merge this into an internal repo first and it's easier if the commits of this pull request are added as-is to the main branch. Rebasing these changes onto the latest main branch would also be nice.

I ended up rebasing the branch on my own, so this pull request did not automatically close when 93f9fef and b7ff683 were added, but you do get credit for the changes in the commit history. Thanks again for this contribution!