Hi there,
OpenCL already is supported, via JOCL, a Java wrapper for OCL, and there's a tutorial here...
http://jogamp.org/wiki/index.php/JOCL_Tutorial . I've spent some time making my own Association Rule operators which use the GPU; the speedup on frequent itemset generation is more than 100, so that justifies for me the very considerable effort spent.
It's not only that you need good low-level programming skills to harness the power, but you also need to recast the problem in a way that the GPU likes, a parallel implementation is not just a sequential algorithm running on lots of processors, sadly.
However, having a GPU turbo for RM makes a rather cool mix, great interface to limitless grunt...
