(De/Re)-Composition of Array Computations + An MDH-Based DSL for Array Computations
summary
MDH provides a formal high-level representation plus a low-level tiling/placement IR and a fully auto-tuned lowering pipeline that emits CUDA/OpenCL code; a Python MDH DSL and an MLR front end are implemented and produce kernels whose ATF autotuner runs CUDA kernels as cost functions. Experiments report wins over state-of-practice tools (sometimes >2.5x vs vendor GPU libraries, up to ~170x on tiny dot-product cases vs TVM) and higher portability by the Pennycook metric; search handles constrained, multi-layer tile/parallelism parameters via a tree→coordinate mapping.
AI-written summary · about