Oh wow, hard numbers, thanks! That's very impressive, but strangely misleading in the graph towards the bottom that appears to clock TAIL->Futhark->OpenCL and Futhark->OpenCL exactly the same in more than half the cases (if you look at the raw numbers if appears that the second approach blows the first away in most).
That's because the speedup graph cuts away at a 100x speedup, but some of these (somewhat silly) benchmarks are compute-bound and hence gets these crazy 1000x speedups on the GPU.