Note that this is more or less a continuation of the Sparse Representation discussion, but now focused specifically on the progress of MLIR support for sparse tensors.
In a distant past, I pioneered the idea of letting a compiler automatically convert annotated dense linear algebra Fortran to semantically equivalent sparse code, which resulted in the MT1 compiler. But when we started exploring to use a similar approach in MLIR for sparse tensors, I was pleasantly surprised to find the excellent work of Fredrik Kjolstad et al. in the Tensor Algebra Compiler, which formalizes the automatic transformation of annotated tensor algebra to sparse code in an elegant and rigorous manner.
For Q4 this year, we plan to prototype similar ideas in an independent prototype implementation in MLIR. A PDF with a brief description of our plans is posted below, and we welcome your early input. A big kudos to Fredrik and the whole TACO team for clearly explaining their approach in great detail in a series of very well-written papers, which made planning this project for MLIR super easy!
MLIR Support for Sparse Tensors.pdf (229.7 KB)