Parallel general purpose computing across multiple computer graphics devices.

摘要

This thesis proposes that multiple, commodity computer graphics cards in a single system can be used to perform general purpose computation and computer vision many times faster than the CPU alone. Presented is a parallel architecture created by placing multiple graphics cards on a single motherboard. This creates a low cost, commodity, architecture for hardware accelerated general purpose computation. Parallelism, both between the GPU and CPU and between multiple GPUs, however, is new to graphics and current graphics techniques are designed for a single GPU and focus on display, rather than fast communication and synchronization with the CPU. Presented are methods of distributing tasks across multiple graphics cards, synchronizing their execution, and exploiting their fast on-board memories. Additionally, for graphics hardware to provide speedups, methods of mapping software algorithms onto graphics architecture must also be developed. In this thesis, computer vision algorithms are mapped onto the graphics hardware. Novel mappings of the chirplet transform, projective image stitching and parameter estimation, and other common computer vision algorithms are developed in this work. These mappings show how all parts of the graphics pipeline can be used to achieve a wide variety of tasks through novel algorithm mappings. A programming library called "OpenVIDIA" abstracts many of the mapping details allowing programmers to realize graphics hardware acceleration.