Enable MKL Sparse QR solver as alternative to SuiteSparse

[dyollb]

Building ceres on Windows and getting good performance is not very well supported, because there is no (maintained) way to build SuiteSparse.

Please support the MKL Sparse QR, described here: https://www.intel.com/content/www/us/en/developer/articles/technical/intel-mkl-sparse-qr-solver-multifrontal-sparse-qr-factorization-method-for-solving-a-sparse.html

The performance comparisons show that is outperforms SuiteSparseQR on a collection of benchmarks.

This may even be a benefit for other platforms where MKL is available.

[sergiud]

While I support the proposal, I want to also make aware of my SuiteSparse fork with native CMake support which I maintain and continuously test on Visual Studio 2019 and and now 2022 among others.

[dyollb]

Thanks. I was referring to the link provided in the ceres installation instructions : https://github.com/jlblancoc/suitesparse-metis-for-windows. That one hasn't been updated for a while.

Nevertheless, the MKL could be faster, so I still think it would be nice.

[sandwichmaker]

@sergiud do you want to send me a patch to the documentation pointing to your fork of suitesparse?
Also @alexsmac may have some time to look at integrating intelmkl into ceres, so assigning the bug to him.

[sergiud]

Preferably, the integration should be done by employing oneAPI (i.e., oneMKL) to avoid dependency on a specific backend available only for selected platforms. However, it seems that MKL SparseQR is not part of the oneAPI specification, is it?

@sandwichmaker I don't mind it. However, Ceres does not work out-of-the box with my fork and requires a minimal amount of patching to use the provided CMake import targets.

[sandwichmaker]

I have not looked at the oneMKL API yet, its been changing a fair bit over the past two years. I will start another issue to discuss the cmake/suitesparse thing and move the discussion there.