Source-to-source compilation of loop programs for manycore processors

摘要

It is widely accepted today that the end of microprocessor performance growthudbased on increasing clock speeds and instruction-level parallelism (ILP)uddemands new ways of exploiting transistor densities.udManycore processors (most commonly known as udGPGPUs or simply GPUs) provide a viable solution to this performance udscaling bottleneck through large numbers of lightweight compute cores udand memory hierarchies that rely primarily on software for their udefficient utilization. The widespread proliferation of this class of udarchitectures today is a clear indication that exposing and managing udparallelism on a large scale as well as efficiently orchestrating udon-chip data movement is becoming an increasingly critical concern for udhigh-performance software development. In such a computing landscape udperformance portability -- the ability to exploit the power of a variety udof manycore chips while minimizing the impact on software development udand productivity -- is perhaps one of the most important and challenging udobjectives for our research community. ududThis thesis is about udperformance portability for manycore processors and how source-to-source udcompilation can help us achieve it. In particular, we show that for anudimportant set of loop-programs, performance portability is udattainable at low cost through compile-time polyhedral analysis and optimizationudand parametric tiling for run-time performance udtuning. In other words, we propose and evaluate a source-to-source udcompilation path that takes affine loop-programs as input and udproduces parametrically tiled parallel code amenable to run-time tuning udacross different manycore platforms and devices -- a very useful udand powerful property if we seek performance portability because it uddecouples the compiler from the performance tuning process. The produced udcode relies on a platform-independent run-time environment, called Avelas,udthat allows us to formulate a robust and portable code generation algorithm.udOur experimental evaluation shows that Avelas induces low run-time overheadudand even substantial speed-ups for wavefront-parallel programs compared to a state-of-the-artudcompile-time scheme with no run-time support. We also claim that the low overhead of Avelas is a strongudindication that it can also be effective as a general-purpose programming modeludfor manycore processors as we demonstrate for a set of ParBoil benchmarks.