There are 2 optimizations I’m after:
- How to send the
C
function a pointer tocolptr
,rowval
andnzval
. - How to make my whole Julia setup work with
32 Bit
indices for the sparse matrix (See Convert a Current Installation of Julia to Use `BlasInt = `Int32`).
In (2) the problem I have is that most non trivial algorithms on Sparse Matrix (Decomposition) will gain from having indices in 32 Bit
yet I don’t know about the other eco system (MKL Sparse
, Suite Sparse
, etc…). If I set my matrix to be 32 Bit
, does it mean any call to a function of those libraries will create a conversion step?
I hope we have / we’ll have a way to configure all to be coherent. At the moment I don’t see a way to set MKL
, MKL Sparse
, Pardiso
and other Intel based libraries to use LP64
mode (Some has but it will go away with Julia 1.7). I have no idea about SuiteSparse
. It seems some of the algorithms support 32 Bit
indices but it is not system wide.