Writing kernel modules in Haskell

Have you ever actually looked into the layout of a BLAS implementation? Netlib BLAS consists of hundreds of small files, mostly one file per numerical method, and it's easy to pickup needed files and put them in your code. On the other side, Netlib BLAS is admittedly just a reference implementation, not fast at all (and also actually requiring Fortran 77 compiler to build), and things get more complicated with optimized BLAS implementation. For example, with ATLAS, these files are generated, so either you'd have to generate them up-front for all needed processors, or to include the whole infrastructure for auto-generating these into your code (lots of code there, indeed, but if you want to do it right, you'd have actually to write all of this it yourself). As for vendor libraries (like MKL etc.), you may be in better luck with negotiating with vendors to provide you with some kind of license to use needed routines. But in any case you're better with using an existing solution, than to reimplementing it yourself (especially as numerical codes are notoriously hard to get right, which is why it makes even more sense to stick to re-using existing code).