dh-core icon indicating copy to clipboard operation
dh-core copied to clipboard
verified publisher icon DataHaskell

BLAS layer

Open ocramz opened this issue 7 years ago • 4 comments

Unify dense and sparse lin.alg. , for a given underlying vector type, under one same interface

Blocked by #1 and #3

ocramz avatar Oct 16 '18 11:10 ocramz

Now this one is difficult, especially if efficiency is important (it is). We will probably end up with different backends with every single once having different constrains on elements which is possible to put into matrices. It's:

  • Double, Float, and Complex Double/Float for BLAS
  • (Storable a)/(Unbox a) if we're we store data in vectors
  • Anything else really. It's only bounded by craziness of backend implementor

Unifying all of these going to be difficult

Shimuuar avatar Oct 16 '18 16:10 Shimuuar

We've got to try at least! I have a couple ideas to put on the table for this (based on closed type families).

ocramz avatar Oct 17 '18 07:10 ocramz

@Shimuuar I'm wondering where to put any typeclasses that must be shared between dense-linear-algebra and the main package. In DLA itself, as a new module that will get re-exported ?

ocramz avatar Oct 18 '18 07:10 ocramz

We'll probably split type classes into separate package in the end. But it sounds fine at the moment

BTW I have some BLAS bindings lying around. I should find them

Shimuuar avatar Oct 18 '18 07:10 Shimuuar