Section: New Results

Portable representation for polyhedral compilation

Participants : Riyadh Baghdadi, Michael Kruse, Chandan Reddy, Tobias Grosser, Sven Verdoolaege, Albert Cohen.

Programming accelerators such as GPUs with low-level APIs and languages such as OpenCL and CUDA is difficult, error prone, and not performance-portable. Automatic parallelization and domain specific languages have been proposed to hide this complexity and to regain some performance portability. We proposed PENCIL, a subset of GNU C99 with specific programming rules. A compiler for a Domain-Specific Language (DSL) may use it as a target language, a domain expert may use it as a portable implementation language facilitating the parallelization of real-world applications, and an optimization expert may use PENCIL to accelerate legacy applications.

The design of PENCIL is simultaneously a key research result and a milestone for parallelizing compiler engineering/design. Aspects of its static-analysis-friendly, formal semantics are highly original, for the language's ability to preserve expressiveness and modularity without jeopardizing a (polyhedral) compiler's ability to perform aggressive transformations. We validated its potential as a front-end to a state-of-the-art polyhedral compiler, extending its applicability to dynamic, data dependent control flow and non-affine array accesses. We illustrated this PENCIL-enabled flow on the generation of highly optimized OpenCL code, considering a set of standard benchmarks (Rodinia and SHOC), image processing kernels, and DSL embedding scenarios for linear algebra (BLAS) and signal processing radar applications (SPEAR-DE). We ran experimental results on a variety of platforms, including an AMD Radeon HD 5670 GPU, an Nvidia GTX470 GPU, and an ARM Mali-T604 GPU.

This work is conducted in collaboration with partners from ARM, RealEyes (a computer vision company) and Imperial College.