Enabling Efficient Sparse Computations using Linear Algebra Aware Compilers
Summary
The report introduces the LAPIS compiler framework built on MLIR to optimize sparse linear algebra with performance portability across architectures. It centers on the Kokkos dialect for lowering to diverse targets and a partition dialect for distributed memory execution, enabling efficient sparse tensor operations and integration with SciML workflows. Applications include sparse graphs, GraphBLAS-based TenSQL, and kernels for subgraph isomorphism, demonstrating productivity, performance, portability, and scalable computation.