Open ReubenHill opened 3 years ago
An alternative would be to use libxsmm
:
To quote Paul Kelly:
You give libxsmm the dense matrix (with zeroes), and it JITs a matrix multiply implementation for you for that specific matrix, which you can reuse. It does full unrolling so this only works for matrices not much bigger than 100x100
I think both of these probably potentially have a place. AIUI, for libxsmm to work, the dense matrix needs to exist first, which if it is very large is bad news. libxsmm doesn't really work above 100x100
A sparse version of a
gem.Literal
would be very useful, e.g. for FInAT dual evaluation where the dual bases are, in general, sparse tensors.