Fork me on GitHub

Futhark 0.18.2 released

Posted on November 7, 2020 by Troels Henriksen

The previous release 0.18.1, had two large changes: the new multicore backend, and a switch to 64-bit sizes for arrays. Due to the scope of these changes, I expected that programmers would rapidly find many critical bugs, and that we would have to issue many small bugfix releases, like when 0.15.1 added size-dependent typing. Surprisingly, this didn’t happen - while programmers did find several bugs, they were the usual fare of obscure edge cases, often in old code and unrelated to the big changes in 0.18.1. The one exception was a few places in the multicore backend, which did not properly use 64-bit sizes when dealing with large arrays.

Still, the changes in the just released 0.18.2 show that it is still mostly a bugfix release. Indeed, the only improvement to the optimiser is a loosening of certain restrictions in the implementation of loop tiling. The programmer-facing effect of this improvement is that code that resembles matrix-vector multiplication in its dataflow will now be tiled. This can provide a substantial speedup, often in excess in 50%, but it is probably not many programs that are bottlenecked by this operation.

My favourite fix in this release is one that addresses a bug in our coverage checking for pattern matches. In Futhark, inexhaustive matching (where the patterns do not cover all possible values) is not allowed, so it was an actual soundness bug that the type checker was unable to properly detect incomplete matches. Fixing this required a fundamental redesign of how we check for exhaustivity. Fortunately, Futhark is fundamentally a rather conventional dialect of ML, so I was able to implement a pre-existing algorithm presented by Luc Maranget in the paper Warnings for pattern matching.

Oh, and to all US citizens reading this: congratulations on your own major bugfix. Be careful to avoid future regressions.