Structured Matrix / Tensors Detection

Open avik-pal opened this issue 6 months ago • 1 comments

Listing out structured tensor that we can detect (ideally put these in a separate cc file and not with the HLO patterns):

[ ] UpperTriangular
[ ] UnitUpperTriangular
[ ] LowerTriangular
[ ] UnitLowerTriangular
[ ] Banded
[ ] Diagonal
[ ] Bidiagonal
[ ] Tridiagonal
[ ] Symmetric
[ ] Hermitian
[ ] BlockSparse
- (usually, BlockDiagonal. ideally, block sparsity pattern is known at comptime)
[ ] OneHot

(Feel free to add more here)

These enable us to lower to faster factorization routines, matmul, etc.

Jun 20 '25 01:06 avik-pal

mmm what's your idea for how to implement this? like adding a MLIR type per structured tensor type? or maybe adding an attribute for marking ops that support a specific structure?

Jun 20 '25 09:06 mofeing