Wednesday Sep. 14, 2011
Hamerschlag Hall D-210
Kayvon Fatahalian (CS, CMU)
The modern real-time graphics pipeline is a versatile parallel architecture that accommodates a wide range of graphics techniques. The architecture is implemented by heavily optimized GPUs that employ a mixture of application-programmable and fixed-function processing resources, yet its design lends itself to a simple programming model easily understood by non-expert programmers.
A major goal of future graphics systems is rendering geometrically complex, film-quality scenes at interactive rates. Unfortunately, current GPU implementations not only require additional compute capability to handle high-resolution surfaces represented by pixel-sized “micropolygons”, fundamental system operations such as geometry processing, surface visibility, and shading execute inefficiently under this workload.
In this talk I will describe a proposed evolution of the graphics
pipeline that increases system efficiency under micropolygon
workloads. The redesign includes simultaneous algorithmic, GPU
hardware, and pipeline abstraction changes that integrate high-quality
adaptive tessellation and motion blur, increase rasterization
parallelism, and reduce shading costs by more than a factor of eight.
When making these changes, irregular communication and control-flow
are isolated to non-programmable parts of the system, preserving the
graphics pipeline's simple, implicitly parallel programming model and
the throughput-optimized design of the GPU's programmable cores. This
“graphics-style” approach to system design is interesting to consider
in other compute-intensive domains and in the development of more
general parallel computing systems.
Kayvon Fatahalian is an Assistant Professor in the Computer Science
Department at Carnegie Mellon University. His research focuses on the
design of efficient parallel systems for computer-intensive applications
such as interactive computer graphics.
Back to the seminar page