support compiling with intel MKL for with LLAMA_MKL=1

[lr1729]

Speeds up prompt processing on cpu by 60% on my i3-1115G4 according to my tests, partially resolves #322

[LostRuins]

I think you might have better results just hijacking the openblas flags instead of adding it as another backend. I can't test MKL myself because I don't have it installed on my system. Have you also compared the difference between Openblas and noblas (which ends up using llamafile)

[lr1729]

Gotcha I'll see if I can figure that out, noblas is similar to openblas speeds for me, sometimes even 10-20% faster

[LostRuins]

Yeah, I can possibly consider adding it as an optional LLAMA_MKL=1 flag that people should set when they do a self-compile (then you can just overwrite the openblas files), but I won't be able to provide prebuilt binaries or support for it, as I don't run it myself.

[lr1729]

That sounds good, it is somewhat of a pain to compile needing the entire 12gb Intel HPC toolkit

[LostRuins]

Exactly, that's why I avoid MKL (also for those on ryzen CPUs it's not useful).

OpenBLAS is portable and can be used with 15mb of files.

But if you have a GPU you should use that, it will beat everything else by miles.