Yes. I have watched it. Though what I’m talking is mainly for the Sparse Matrices.
Namely, what I’m asking is if both paths exist at once and if Julia chose the LP
path for Sparse Matrices built with Int32
.
The code on MKLSparse
currently use BlasInt
to chose the code path. It doesn’t even look at the matrix itself.
This is what I’m talking about (Though I’d say the 3-4 last comments should appear on Convert a Current Installation of Julia to Use `BlasInt = `Int32`).
The solution has 2 stages:
- Allow, at the same time, accessing BLAS / LAPCAK / Sparse BLAS / LAPACK libraries with
Int32
andInt64
API. - For Sparse Matrices chose the path based on the type of the indices. Go for
LP64
for Sparse Matrices withInt32
/UInt32
indices. Use theILP64
path for matrices built withInt64
/UInt64
. One could even say that the path can have AUTO mode for matrices with less than2 ^ 31 - 1
elements (Be Sparse or Dense).
It will bring performance and memory optimization for the eco system.