Blender's compositor always uses 4 floats to store data of a pixel. Even if the pixel is a value (only one floats needed) or vector (only three floats needed). This patch will allocate only the memory needed reducing memory. (normally around 50%).
This memory model also suits better with other area's in blender where we don't need that much of conversions between memory models This will have some (but little) positive impact on performance.
The free memory can be used eventually to keep buffers alive etc, buffering and other features.
more information can be found
- overall project page: http://wiki.blender.org/index.php/Dev:Ref/Proposals/Compositor2014
- technical design this project: http://wiki.blender.org/index.php/Dev:Ref/Proposals/Compositor2014_p1.1_TD