high performance computing on graphics processing units: hgpu.org

We propose Tangram, a general-purpose high-level language that achieves high performance across architectures. In Tangram, a program is written by synthesizing elemental pieces of code snippets, called codelets. A codelet can have multiple semantic-preserving implementations to enable automated algorithm and implementation selection. An implementation of a codelet can be written with tunable knobs to allow architecture-specific parameterization. The Tangram compiler produces highly optimized code by choosing and composing architecture-friendly codelets, and then tuning the knobs for the target architecture. We demonstrate that Tangram’s synthesized programs are comparable in performance to existing well-optimized codes on both CPUs and GPUs. The language is defined in a concise and maintainable way to improve debuggability and to enable progressive improvement. This strategy allows users to extend their applications and achieves higher performance on existing architectures and new architectures.