Piko: A Framework for Authoring Programmable Graphics Pipelines

摘要

We present Piko, a framework for designing, optimizing, and retargetingrnimplementations of graphics pipelines on multiple architectures.rnPiko programmers express a graphics pipeline by organizingrnthe computation within each stage into spatial bins and specifyingrna scheduling preference for these bins. Our compiler, Pikoc,rncompiles this input into an optimized implementation targeted to arnmassively-parallel GPU or a multicore CPU.rnPiko manages work granularity in a programmable and flexible manner,rnallowing programmers to build load-balanced parallel pipelinernimplementations, to exploit spatial and producer-consumer localityrnin a pipeline implementation, and to explore tradeoffs between thesernconsiderations. We demonstrate that Piko can implement a widernrange of pipelines, including rasterization, Reyes, ray tracing, rasterization/rnray tracing hybrid, and deferred rendering. Piko allows usrnto implement efficient graphics pipelines with relative ease and tornquickly explore design alternatives by modifying the spatial binningrnconfigurations and scheduling preferences for individual stages, allrnwhile delivering real-time performance that is within a factor six ofrnstate-of-the-art rendering systems.