This Thursday, January 26th (9am California Time, 17:00 UTC), @ftynse will discuss the transform dialect and related infrastructure for precise control of compiler transformations, particularly useful for code generation. Numerous high-performance code generators rely on explicit or implicit schedules that describe how the program is to be run separately from the program itself, including Halide, TVM and Rise. The transform dialect approach brings the notion of schedules to the MLIR ecosystem and connects it to code generation for structured ops, bufferization and vectorizaton. It allows for code generation heuristics to be externalized (e.g., stored and shipped as IR or bytecode) or generated by third-party tools. It also inspired the extension of the core infrastructure, in particular dialect extensions, and more variants of pattern rewriting. This talk will discuss the current state of the transform dialect infrastructure, its evolution since the original RFC, and the expected longer-term effects on MLIR-based compiler construction.
Note: this is a re-run of the MLIR developer summit talk from October, which wasn’t recorded, with slight additions.
Meeting ID: 851 5109 0498