1) MLIR Side Effect Modelling - Siddharth Bhat, Jeff Niu
2) Common facilities for ML-Guided Optimizations in LLVM - Mircea Trofin
3) MLIR Dialect for GraphBLAS - Sriram Aananthakrishnan
4) An MPI Dialect for MLIR - Anton Lydike

MLIR Side Effect Modelling - Siddharth Bhat, Jeff Niu
This talk will provide an overview of side-effect modelling in MLIR, point to limitations in the current model and offer tentative suggestions to improve it. We first review the semantics of the memory effects and conditionally speculatable operation interfaces, as defined by the MLIR language reference, and how they are used upstream. We then highlight interesting out-of-tree applications that suggest limitations in the model. For example, there are contentious modelling issues of undefined behaviour, parallelism, and computational divergence. Finally, we hope to open the conversation about side-effecting modelling by proposing several paths for the evolution of MLIR's side-effect API.

Common facilities for ML-Guided Optimizations in LLVM - Mircea Trofin
A comprehensive overview of the facilities for ML-guided optimizations currently available in LLVM.

MLIR Dialect for GraphBLAS - Sriram Aananthakrishnan
Graph analytics is the analysis of graph-based unstructured data, and a wide range of graph algorithms are expressible as sparse matrix and vector operations on an extended algebra of semirings. GraphBLAS defines the standard for creating graph algorithms in the language of linear algebra over semirings. In this work, we present a dialect for GraphBLAS and show our ongoing work of code generation for graph algorithms expressed using GraphBLAS operations.

An MPI Dialect for MLIR - Anton Lydike
This talk will present our work on the MPI dialect, which brings standard MPI calls into the MLIR ecosystem as its own dialect. We want to present the various challenges faced during our exploratory development, and the solutions we came up with. We want to touch on the dialect design and the challenges connected with lowering to a C library in MLIR without a standardised ABI. While this talk focuses on bringing MPI into MLIR, we would like to motivate the addition of the dialect by showing our work that is performing automatic domain decomposition fully in MLIR and targets MPI.