计算机体系结构论文dynamic warp formation and scheduling for efficient gpu control flow.pdfVIP

Dynamic Warp Formation and Scheduling for Efficient GPU Control Flow Wilson W. L. Fung Ivan Sham George Yuan Tor M. Aamodt Department of Electrical and Computer Engineering University of British Columbia, Vancouver, BC, CANADA {wwlfung,isham,gyuan,aamodt}@ece.ubc.ca Abstract using additional resources to increase throughput by ex- ploiting explicit thread level parallelism in software (forcing Recent advances in graphics processing units (GPUs) software developers to share the responsibility for improv- have resulted in massively parallel hardware that is easily ing performance). programmable and widely available in commodity desktop The modern graphics processing unit (GPU) can be computer systems. GPUs typically use single-instruction, viewed as an example of the latter approach [18, 28, 6]. multiple-data (SIMD) pipelines to achieve high perfor- Earlier generations of GPUs consisted of fixed function 3D mance with minimal overhead incurred by control hard- rendering pipelines. This required new hardware to en- ware. Scalar threads are grouped together into SIMD able new real-time rendering techniques, which impeded batches, sometimes referred to as warps. While SIMD is the adoption of new graphics algorithms and thus motivated ideally suited for simple programs, recent GPUs include the introduction of programmability, long available in tradi- control fl ow instructions in the GPU instruction set archi- tional offline computer animation [35], into GPU hardware tecture and programs using these ins