| Programmability of graphics pipelines has proven to be an extremely valuable tool in the hands of shader authors and game programmers, and is one of the driving forces behind the ubiquity of high-quality visual content across the computing ecosystem. This dissertation presents the design, implementation, and evaluation of an abstraction that extends the programmability of graphics systems from individual shaders to entire pipelines, and addresses the performance, programmability, and flexibility of the resulting implementations.;To this end, we propose Piko, a high-level abstraction for designing flexible and efficient pipelines for computer graphics applications. Piko extends traditional pipelines by allowing each stage to consist of three phases, and uses 2D spatial tiling to help express information about not just the algorithm, but the granularity and schedule of computation as well. Having these different programs address different implementation concerns helps optimize a Piko pipeline for multiple scenarios, such as varying application inputs and target platforms.;The focus of this dissertation is twofold. First, we motivate the importance of efficient programmable pipelines, and survey existing graphics systems to explore implementation practices that help improve their efficiency. Specifically, we identify concurrency, locality, and fine-grained synchronization as important goals, and explore design practices that help achieve them. Second, we propose an abstraction to embody these practices as first-class constructs. We present a detailed study of how pipelines expressed using Piko could translate to two different architectures, study the behavior of resulting implementations, and compare it to the state of the art in high-performance graphics. |
