Efficient Matrix-Matrix Products

[tbetcke]

Any efficient linear algebra library requires fast matrix-matrix products. The Rust ndarray project uses the matrixmultiply library https://github.com/bluss/matrixmultiply, which uses the same ideas on which the BLIS project is based. This seems to be a very good start.

In this issue we want to track the use of matrixmultiply and its performance compared to Openblas on different architectures.

[jedbrown]

Is this issue complete in the sense that householder uses matrixmultiply?

[tbetcke]

I leave it open on purpose because of performance evaluation. I'd like to have good comparison metrics against Openblas on different architectures.

[jedbrown]

Yep, or use the BLIS benchmarking infrastructure to test matrixmultiply. https://github.com/flame/blis/blob/master/docs/Performance.md

Or this could be done entirely in Rust, sweeping over problem sizes for each factorization and scalar, but I'd encourage plotting it normalized to match the BLIS figures.