deepComp: A CUDA+MPI Library for fast Compositing of "deep" Z-and-alpha Frame Buffers

This library allows a data-parallel application to render, on each individual rank, a "deep" frame buffer with multiple RGBA-z fragments for each pixel, and to then rely on this library to properly z- and alpha-composite all these inputs into a single image.

Some key benefits of this library's approach:

• deepComp is not limited to a single fragment per rank per pixel; each rank can already have multiple framgments for each pixel

• deepComp allows for both alpha- and z-compositing; it automatically sorts all each pixel's (RGBA-z) framgents by z, then performs proper front-to-back alpha-blending

• deepComp does not require a fixed compositing order of all ranks; composting happens on fragments, which ranks can generate in any order

• deepCOmp is not restricted to whole "frame buffers" on each rank; fragments can be written by any CUDA kernel, in any order.

Note:

• deepComp relies on a CUDA-aware MPI implementation that can properly understand device addresses.

Example:

• This here is a screenshot from the (included) testDeepViewer app, running with 4 MPI ranks:

  /home/wald/opt/bin/mpirun -n 4 ./testDeepViewer

In this sample, each rank creates multiple different semi-transparent cubes; with the cude's color encoding the rank it is on. deepComp will properly composite the resulting fragments, even though several pixels have multiple ranks from different ranks, in often "interleaved" order.