MDH Lowering: From High-Level to Low-Level

The MDH’s Lowering Process fully automatically lowers a data-parallel computation expressed in its high-level program representation to an architecture- and data-optimized instance in its low-level representation. For this, we have designed MDH’s formal lowering process as generic in performance-critical parameters, such as the size of tiles and number of threads – different values of these parameters lead to semantically equal instances of our low-level representation, but the instances are differently optimized. Currently, MDH uses the Auto-Tuning Framework (ATF) to fully automatically determine architecture- and data-optimized values of tuning parameters.

Overview of MDH Tuning Parameters: