Programs don’t magically become faster when they are run on GPUs. E.g. Linear Algebra algorithms work really well on CPUs, but if ported 1 to 1 ( as this would ) to a GPU their performance is just abysmal. Usually one needs to use a specially designed algorithm that can actually use the massive parallelism of a GPU and not get stuck e.g. trying to synchronize or doing other kinds of communication. GPUs really like doing the same operation on independent data, which is basically what happens when rendering an image, they are not really designed to have operations that need information of all other data, or neighbouring data in a grid…. . Just because something works on a GPU does not mean its efficient, thus the performance could be much worse using a GPU .

Also balancing CPU and GPU usage is even harder ( maybe impossible ? ) as you cannot predict what kind of System you will run your software on, thus usually these days the CPU feeds the GPU with data ( with the current Tesla cards only 1 core per GPU, this changes in the Kepler version to 32 ) and does some processing that can’t be done on the GPU, but do not share any kind of workloads.

I don’t know how the h.264 codec is structured or if it is possible to have performance gains on encoding. However I really doubt that x.264 can be just ported as they rely heavily on CPU specific features ( SSE etc ) which is quite different to the much higher level bytecode that Java would produce.

via Rootbeer GPU Compiler Lets Almost Any Java Code Run On the GPU – Slashdot.