Julia’s type system for linear algebra is quite sophisticated. I have a case where, after some
transpose operations, I obtain two matrices
A,B for which for some reason,
A*B falls back to the slow
generic_matmul instead of using BLAS.
Given two arrays
A, B (which are matrices or vectors), what determines if the product
A*B will be carried out by BLAS, instead of the fallback
generic_matmul? Is there a way to have something like a predicate
hitsblas(A, B)? That would be very useful debugging this situation.