Sort-First Parallelism for Image-Based Rendering
Voicu Popescu, Anselmo Lastra, John Eyles
ABSTRACT
Rendering from images with depth (Image-Based Rendering by Warping, IBRW)
is an appealing 3D-computer-graphics technique since it alleviates the
modeling bottleneck and promises photorealism at interactive rates. Current,
and probably near future silicon technology cannot produce single-chip
IBRW renderers that are powerful enough. Thus a high-level parallelism
scheme needs to be employed that divides the rendering task among several
nodes.
We show that sort-first is the technique best suited for parallel IBRW
since it takes advantage of the locality characteristic of depth images.
One can easily determine the output-image region to which a contiguous
set of depth samples warps. Thus the nodes get the samples pertinent to
their region which they fully render independently, eliminating the need
of very high bandwidth node-to-node communication paths for sorting the
warped samples.
We previously presented the WarpEngine, a sort-first architecture for
IBRW. In this paper we will measure the performance of sort-first by analyzing
the efficiency of the parallelization scheme, the communication costs,
and the load balancing for various system configurations and output-image
resolutions.
Read
paper