Using a high performance BLAS

[rth]

Performance of numerical calculations is going to be significantly impacted by the BLAS used.

#211 aims to include CLAPACK with the reference BLAS/CBLAS which should have a baseline performance.

Once that is done (hopefully soon), e.g. BLIS could be a possible candidates for a better BLAS. That is probably more achievable than OpenBLAS. BLIS was already previously compiled for emscritpen in https://github.com/Maratyszcza/blis-bench (though targeting Asm.js). The relevant discussion at numpy is numpy/numpy#7372.

[rth]

Some benchmarks for matrix multiplication using the currently included reference BLAS can be found in https://gist.github.com/rth/c71fe792eb56fb271317e35e08576c7a

[rth]

Although as I mentioned to @lesteve, BLIS only includes BLAS, so we would still have to use reference CLAPACK with all the existing hacks. While it we manage to have a generic build of OpenBLAS it would also have optimized LAPACK. Though it's in Fortran, so lfortran is probably our best bet there.