Enabling Efficient Sparse Computations Using Linear Algebra Aware Compilers
5 days ago
- #sparse-computations
- #compiler
- #MLIR
- Development of the LAPIS compiler framework based on MLIR for optimizing sparse linear algebra operations.
- Introduction of the Kokkos dialect for lowering high-level code to different architectures and converting MLIR to C++ Kokkos code.
- Creation of a partition dialect for distributed memory architectures to manage sparse tensor distribution and communication patterns.
- Demonstration of MLIR's capability for linear algebra-level optimizations, improving performance on various GPUs.
- Key applications include sparse linear algebra, graph kernels, TenSQL, and subgraph isomorphism/monomorphism kernels.
- Support for productivity, performance, portability, and distributed memory execution in the LAPIS framework.