While fixing RT-4763 the ImageData class that holds data for the inputs of filters was modified to incorporate a transform that must be applied as the data is read.

All of the hardware effect peers were (trivially) modified to handle this transform as they sample their inputs, but some of the hand-tuned software peers were too ingrained with believing that they were dealing with pixels on a 1:1 basis to be so easily modified. This is primarily true of the Java and SSE versions of the Box and Guassian (LinearConvolve) filters. They implement workarounds for the time being (either backing off to a slower, but more generic loop or requesting a transformed version of the input image and then operating on the transformed image), but they should be modified to more optimally operate on transformed (or clipped) inputs.