Transpilation Utilizing Language-agnostic IR and Interactivity for Parallelization

Abstract

Migrating codes between architectures is difficult because different execution models require different types of parallelism for optimal performance. Previous approaches like libraries or source-level tools generate correct and natural-looking syntax for the new parallel model, but require the programmer to entirely bear the burden of performance engineering. Recent approaches in the compiler intermediate representation (IR) level can automate some performance engineering at the expense of readability of the generated parallel code. Both approaches leave performance on the table. Even with manual rewriting, source-level tools will miss out on performance-enabling compiler optimizations and pure compiler-level tools will neglect expert-level tuning and parallelization techniques. This paper introduces Tulip, a framework that combines the best of compiler and programmer-level program rewriting. Tulip first transpiles parallel programs in a compiler IR, enabling it to automatically optimize and retarget parallelism to match the target platform. It will then generate natural source code in a high-level parallel programming language (OpenMP) that can be interactively optimized and tuned with programmer intervention. For 19 Polybench benchmarks, Tulip-generated OpenMP-CUDA offloading programs perform 14% faster than the original sources on NVIDIA GPUs. Moreover, transpilation to CPU leads to 2.9× speedup over best state-of-the-art results.