Direct to GPU Video Transfers

Modern graphics processing units (GPUs) are high-performance parallel floating point processors ideal for the image and video processing operations required for broadcast and digital film applications. Historically, image data transfer to the GPU required that the data pass through a staging buffer prior to upload to or download from the GPU as the buffers utilized by the video I/O device could not be used directly by the GPU. This memory copy operation created a performance bottleneck. Direct to GPU Video Transfer enables the sharing of a lockable system memory buffer between a video I/O device and the GPU eliminating the latency incurred by the memory copy. Results demonstrate transfer times for SD, HD, 2K and 4K frames approaching realized PCI Express 2.0 limits. Overall system latency can be further reduced by overlapping sub-frame transfers with the video device capture and scanout.

Published

2011-10

Content type

Original Research

Keywords

GPU, Video I/O

DOI

10.5594/M001104

ISBN

978-1-61482-940-9