Broken matrix operations

[devreal]

https://github.com/nileshchat/MRA-TTG/actions/runs/17109772322/job/48528016804?pr=97 suggests that some of the mxm variants are broken. We need a test case t hat compares the results of our implementation against blaspp. The handwritten GEMMs are still important as a fallback, in case we don't find anything else useful.

[nileshchat]

in case we don't find anything else useful

Can you please elaborate?

[devreal]

We talked about this: unless we find a way to generate efficient GEMMs inside the thread block across all architectures we need a fallback. HIP is currently one candidate where we don't have any other option. The cublasDx integration is experimental and needs a lot of work still, so we rely on the custom implementations too for the most part.

[nileshchat]

The conversation I recall is us using in-house GEMM from MADNESS only for HIP. For host, we use blaspp and cublasDx for CUDA. But I do agree, in the long run, we should have customized GEMM implementation.