Automatic Data Layout at Multiple Levels for CUDA

Abstract

Trasgo is a source-to-source compiler system that translates simple high-level specifications of parallel algorithms to lower-level native programs, with data partition and communication details generated automatically. Hitmap is the run-time library used by the backends of Trasgo for hierarchical tiling and mapping of arrays, currently built on top of the MPI message-passing interface. Hitmap includes a plug-in system for automatic data-layouts. In this paper we extend Hitmap with a new type of data-layout techniques suitable for the CUDA parallel programming model. The combination with the previous type of data-layout techniques allow to generate data distributions, at multiple levels of parallelism, for GPU clusters. The new Hitmap version hides to the programmer the details about the machine structure and thread management, allowing to easily generate programs with multiple levels of parallelism in heterogeneous systems. This work opens the road to develop a new back-end for the Trasgo compiler system to automatically generate CUDA programs.