rocBLAS: Add an option to disable tensile

[teonnik]

• Tensile takes a lot of time to compile, add an option to disable it
• Change variant name to be consistent with rocsolver
• Change default variant value to be consistent with rocsolver

Ideally Tensile should be a separate package but that can be handled in a different PR.

[srekolam]

@cgmb ,can you please review it

[cgmb]

Does rocSOLVER build with rocblas~tensile? IIRC, there used to be missing symbols if you tried to do that.

[teonnik]

@cgmb

Does rocSOLVER build with rocblas~tensile? IIRC, there used to be missing symbols if you tried to do that.

The package builds fine. However, I checked for missing symbols and indeed gemms are missing. I thought that rocblas would have those even without tensile, I guess I was wrong. :/

[haampie]

@cgmb do you think rocblas could have a fallback gemm that's quick to build?

[cgmb]

That's been on my personal wishlist for a while... but I have a rather long wishlist. My suggestion would be to open an issue requesting the feature from rocBLAS.

[cgmb]

I haven't actually tested it, but I like this change aside from the default amdgpu_target.

[cgmb]

There will be some minor merge conflicts to resolve if #27436 is merged first, but they should be pretty easy to deal with.

[cgmb]

The defaults on the variants are not appropriate. This should not be merged until they're changed.

By the way, there are build speed improvements coming for Tensile in ROCm 5.0/5.1, so you do have something to look forward to.

[teonnik]

I addressed the comments regarding the default variant and copied the changes for the variant from #28907 for rocblas. This now has the same changes for rocblas as in #28907 except that the default for tensile is set to True. If merging this PR before #28907 is preferable, then this can be merged, if not, then this PR can be closed and only #28907 can be merged (after the default value is changed there).

[cgmb]

This change still needs testing to verify it doesn't break anything, but the code looks good.

[cgmb]

There will be some minor merge conflicts to resolve if #27436 is merged first, but they should be pretty easy to deal with.

It has been merged, so the conflicts will need to be resolved.

[cgmb]

This PR was closed because it was subsumed into #28907.

I talked with the rocBLAS team about this and they were receptive to fleshing out the BUILD_WITH_TENSILE=OFF variant into something more usable by third-parties. The optimized Tensile kernels are important for achieving peak performance with rocblas, but they do introduce build-time and run-time complexity that's not always needed.