Kokkos Kernels - ML/IR

[crtrott]

Authors: @brian-kelley @kliegeois

[brian-kelley]

The MLIR (multi-level intermediate representation) in LLVM can represent operations ranging from high-level (e.g. matrix multiplication) to low-level (e.g. a bitwise OR). It also includes automatic transformations for lowering high-level operations to equivalent sequences of low-level operations. We present an MLIR-based compiler pipeline that can compile PyTorch models and general sparse tensor expressions into parallel Kokkos C++ source code. We also discuss ongoing work involving Kokkos and MLIR.

[brian-kelley]

Kokkos_MLIR.pdf