Fork me on GitHub
We are hiring a PhD student to do research on functional high performance computing, including work on Futhark itself. See here for more details.

Futhark 0.10.1 released

Posted on March 25, 2019 by Troels Henriksen

Futhark 0.10.1 is upon us (full changelog in the usual place). This is not a release with many new features or optimisations. Instead, our focus has been on polishing, fixing bugs, and preparing to enable incremental flattening by default. For this blog post, I am going to focus on one subtle change to the language semantics, but which is simultaneously both a conceptual bug-fix and an enable of further optimisations.

New restrictions on shape polymorphism

One of Futhark’s relatively few unusual language features is the support for size parameters, by which we can express constraints on the shapes of arrays passed to functions. For example, the type of zip is the following:

val zip 'a 'b [n] : [n]a -> [n]b -> [n](a, b)

This signature states that for any types a and b, and any size n, the function zip takes an array of n as and an array of n bs, and produces an array of n pairs of as and bs. If you pass arrays of different sizes to zip, then it will fail at run-time, but we wish to eventually move this checking to compile-time.

Now consider another function, this one for concatenating two arrays:

val concat 'a : []a -> []a -> []a

Here there are no size parameters, which makes sense: two arrays of different sizes should be concatenable. However, there is actually a constraint on the arrays that can be passed to concat, one that is unfortunately not captured in the type. In Futhark, all multidimensional arrays must be regular, meaning that their elements must have the same shape: arrays like [[1,2], [3]] are forbidden. This implies that a concatenation concat [[1,2]] [[3]] is not allowed, as it would produce an irregular array. However, this is not expressible in the type system, because in the type for concat we impose no restrictions on which types can be substituted for a. We could create a version of concatthat makes the restriction apparent for two-dimensional arrays:

val concat_2d 'a [n] : [][n]a -> [][n]a -> [][n]a

However, we would need a function for every possible array rank, of which there are an infinite number (although I think six-dimensional arrays are the most I have ever seen in a real Futhark program, and that was an extreme case). But worse, there is nothing that prevents us from still using the ordinary concat.

To address this, we added a restriction on parametric polymorphism: if an array is passed for a function parameter of a polymorphic type, all arrays passed for parameters of that type must have the same shape. For example, given a function

let pair 'a (x: a) (y: a) = (x, y)

The application pair [1] [2,3] will now fail. Instead, we would have to give each parameter a distinct type:

let pair 'a 'b (x: a) (y: b) = (x, y)

This is also what will be inferred if explicit type annotations are left out.

This sounds draconian, but it turns out not a single program was broken by adding this restriction. It also fits well with one’s intuition about sizes being part of types. Finally, the implementation was natural. As part of its compilation strategy, Futhark performs monomorphisation, where calls to a polymorphic function is replaced with a call to an instantiation of the polymorphic function where all type parameters have been replaced with concrete types. For example, an application concat [[1,2]] [[3]] will be transformed into concat_2d_i32 [[1,2]] [[3]], where

val concat_2d_i32 : [n] : [][n]i32 -> [][n]i32 -> [][n]i32

When we instantiate a type parameter with an array type (here, a is replaced with []i32), we simply invent a new size parameter for each dimension of the type, and use that size parameter in all replacements.

This change is primarily one more step on the long road towards proper dependent types, but it is also a small improvement for the compiler. By specifying size restrictions as early in the program as possible, we are able to move checking out of inner loops, and help some of the transformations that depend on size analysis. However, this is a minor advantage - the compiler was already pretty good at figuring these things out.