Optionally use BLAS and MKL to accelerate a few core runtime functions

[jbaayen]

No description provided.

[jgillis]

This needs some more thinking ..

• The binaries are shipped with openblas (which does install cblas.h). Any solution should at least work with that
• In C++, template instantiation does look like the proper solution (note that he runtime functions are sometimes called with SXElem)
• Template instantiations would not be picked up by codegen, we need a solution that works both for codegen and vm without duplication. Modify the codegen parser?

A half-hearted attempt is here https://github.com/casadi/casadi/tree/blas but suffers from SXElem calls

[jbaayen]

To your first point, this should work fine with OpenBLAS. There is an #ifdef switching between mkl_cblas.h and cblas.h, but the function signatures are identical between MKL and OpenBLAS. What alternatives to template overloading do you have in mind? This obviously only works for the VM at the moment, but it does so without introducing any expensive custom run-time type checking. The correct implementation will be dispatched depending on whether SXElem is passed (the template will be instantiated), or a float or double (the type-specific BLAS implementation will be called). The reason I didn't make the changes inside the runtime headers is exactly what you point out: in that approach, calls with SXElem no longer work. Without diving much deeper, C++ overloading is the only mechanism that I could see that would work for all types, SXElem and float/double alike.