I think that getting decent speedup on this kind of non-obvious GPU problem is an impressive result. Their presentation here is somewhat lacking in academic rig

I think that getting decent speedup on this kind of non-obvious GPU problem is an impressive result. Their presentation here is somewhat lacking in academic rigor (I for one would have liked to see a description of how they got their timing results, as well as some tests with a smaller source file, or possibly many small source files), but given that they've provided the source code their results should be easy enough to verify.