3. Language Overview

The Futhark programming language is a purely functional, call-by-value, mostly first-order language that permits bulk operations on arrays using second-order array combinators (SOACs).

The primary idea behind Futhark is to design a language that has enough expressive power to conveniently express complex programs, yet is also amenable to aggressive optimisation and parallelisation. The tension is that as the expressive power of a language grows, the difficulty of efficient compilation rises likewise. For example, Futhark supports nested parallelism, despite the complexities of efficiently mapping it to the flat parallelism supported by hardware, as many algorithms are awkward to write with just flat parallelism. On the other hand, we do not support non-regular arrays, as they complicate size analysis a great deal. The fact that Futhark is purely functional is intended to give an optimising compiler more leeway in rearranging the code and performing high-level optimisations. It is also the plan to eventually design a rigorous cost model for Futhark, although this work has not yet been completed.

3.1. Lexical Syntax

The syntax of Futhark is derived from Haskell and Standard ML, although somewhat less flexible. Futhark is not whitespace-sensitive, and indentation is only used for readability. An identifier starts with a lowercase letter, followed by any number of letters, digits, apostrophes and underscores. An identifier may also start with an underscore, which is used to indicate that the name will not be used. A structure identifier starts with an uppercase letter. Numeric, string and character literals use the same notation as Haskell (which is very similar to C), including all escape characters. Comments are indicated with -- and span to end of line. Block comments are not presently supported.

3.2. Simple Futhark

An Futhark program consists of a sequence of function definitions, of the following form:

let name params... : return_type = body

A function must declare both its return type and the types of all its parameters. All functions (except for inline anonymous functions; see below) are defined globally. Futhark does not use type inference. As a concrete example, here is the a parallel definition of the factorial function in Futhark:

let fact(n: i32): i32 =
  reduce (*) 1 (map (1+) (iota n))

Indentation has no syntactical significance in Futhark, but recommended for readability.

The syntax for tuple types is a comma-separated list of types or values enclosed in parentheses, so (i32, real) is a pair of an integer and a floating-point number. Single-element and empty tuples are not permitted. Array types are written as the element type preceded by brackets, meaning that []i32 is a one-dimensional array of integers, and [][][](i32, real) is a three-dimensional array of tuples of integers and floats. An array value is written as a sequence of elements enclosed by brackets:

[1, 2, 3]       -- Array of type []i32.
[[1], [2], [3]] -- Array of type [][]i32.

All arrays must be regular (often termed full). This means that, for example, all rows of a two-dimensional array must have the same number of elements:

[[1, 2], [3]]      -- Compile-time error.
[iota(1), iota(2)] -- A run-time error if reached.

The restriction to regular arrays simplifies compilation.

Arrays are indexed using the common row-major notation, e.g., a[i1, i2, i3...]. An indexing is said to be full if the number of given indices is equal to the dimensionality of the array.

A let-expression can be used to refer to the result of a subexpression:

let z = x + y in ...

Recall that Futhark is eagerly evaluated, so the right-hand side of the let is evaluated exactly once, at the time it is first encountered. The in keyword is optional when it precedes another let. This means that instead of writing:

let a = 0 in
let b = 1 in
let c = 2 in
a + b + c

we can write:

let a = 0
let b = 1
let c = 2
in a + b + c

The final in is still necessary.

Two-way if-then-else is the only branching construct in Futhark. Pattern matching is supported in a limited way for taking apart tuples, for example:

let sumpair((x, y): (i32, i32)): i32 = x + y

We can also add the type ascription on the tuple components:

let sumpair(x: i32, y: i32): i32 = x + y

Apart from pattern-matching, the components of a tuple can also be accessed by using the tuple projection syntax: #i, where i indicates the component to extract (0-indexed):

let sumpair(p: (i32, i32)): i32 = #0 p + #1 p

In most cases, pattern matching is better style.

Function calls are written as the function name with the arguments juxtaposed. All function calls must be fully saturated - currying is only permitted in SOACs.

3.2.1. Sequential Loops

Futhark has a built-in syntax for expressing certain tail-recursive functions. Consider the following tail-recursive formulation of a function for computing the Fibonacci numbers:

let fib(n: i32): i32 = fibhelper(1,1,n)

let fibhelper(x: i32, y: i32, n: i32): i32 =
  if n == 1 then x else fibhelper(y, x+y, n-1)

This is not valid Futhark, as Futhark does not presently allow recursive functions. Instead, we can write this using the loop construct:

let fib(n: i32): i32 =
  loop ((x, y) = (1,1)) = for i < n do
                            (y, x+y)
  in x

The semantics of this is precisely as in the tail-recursive function formulation. In general, a loop:

loop (pat = initial) = for i < bound do loopbody
in body

Has the following the semantics:

  1. Bind pat to the initial values given in initial.
  2. While i < bound, evaluate loopbody, rebinding pat to be the value returned by the body. At the end of each iteration, increment i by one.
  3. Evaluate body with pat bound to its final value.

Semantically, a loop expression is completely equivalent to a call to its corresponding tail-recursive function.

For example, denoting by t the type of x, this loop:

loop (x = a) =
  for i < n do
  in body

has the semantics of a call to this tail-recursive function:

let f(i: i32, n: i32, x: t): t =
  if i >= n then x
     else f(i+1, n, g(x))

let x = f(i, n, a)
in body

The purpose of loop is partly to render some sequential computations slightly more convenient, but primarily to express certain very specific forms of recursive functions, specifically those with a fixed iteration count. This property is used for analysis and optimisation by the Futhark compiler.

Apart from the i < n form, which loops from zero, Futhark also supports the v <= i < n form which starts at v. We can also invert the order of iteration by writing n > i or n > i >= v, which loops down from the upper bound to the lower. Due to parser limitations, most non-atomic expressions will have to be parenthesised when used as the left-hand bound.

Apart from for-loops, Futhark also supports while loops. These do not provide as much information to the compiler, but can be used for convergence loops, where the number of iterations cannot be predicted in advance. For example, the following program doubles a given number until it exceeds a given threshold value:

let main(x: i32, bound: i32): i32 =
  loop (x) = while x < bound do x * 2
  in x

In all respects other than termination criteria, while-loops behave identically to for-loops.

For brevity, the initial value expression can be elided, in which case an expression equivalent to the pattern is implied. This is easier to understand with an example. The loop:

let fib(n: i32): i32 =
  let x = 1
  let y = 1
  loop ((x, y) = (x, y)) = for i < n do (y, x+y)
  in x

can also be written:

let fib(n: i32): i32 =
  let x = 1
  let y = 1
  loop ((x, y)) = for i < n do (y, x+y)
  in x

This can sometimes make imperative code look more natural.

3.2.2. Shape Declarations

Optionally, the programmer may put shape declarations in the return type and parameter types of a function declaration. These can be used to express invariants about the shapes of arrays that are accepted or produced by the function, e.g:

let f (a: [n]i32): [n]i32 =
  map (+1) a

The above declaration specifies a function that accepts an array containing n elements and returns an array likewise containing n elements. In general, shape declarations in parameters are fresh names, whilst shape declarations in return types must refer to a name of type i32 in scope. A shape declaration can also be an integer constant (with no suffix).

The same name can be used in several dimensions, or even in several parameters. This can be used to give a natural type to a function for computing dot products:

let dotProduct(a: [n]i32, b: [n]i32): i32 =
  reduce (+) 0 (map (*) a b)

Or matrix multiplication:

let matMult(x: [n][m]i32, y: [m][n]i32): [n][n]i32 =

The dimension names bound in a parameter shape declaration can be used as ordinary variables inside the scope of the parameter.

Shape declarations serve two main purposes:

  1. They document the shape assumptions of the function in an easily understandable form.
  2. More importantly, they help the compiler understand the invariants of the program, which it may otherwise have trouble figuring out.

Note that adding shape declarations is never unsafe - the compiler still inserts dynamic checks, so if an incorrect declaration is made, the result will merely be an abrubt but controlled termination as it collides with reality. Shape declarations matter most when used for the input parameters of the main function and for the return type of functions used to map.

3.2.3. In-Place Updates

In an array programming language, we tend to use bulk operations for most array manipulation. However, sometimes it is useful to directly replace some element. In a pure language, we cannot permit free mutation, but we can permit the creation of a duplicate array, where some elements have been changed. General modification of array elements is done using the let-with construct. In its most general form, it looks as follows:

let dest = src with [indexes] <- (value)
in body

This evaluates body with dest bound to the value of src, except that the element(s) at the position given by indexes take on the new value value. Due to parser limitations, the parenthesis around value are not optional. The given indexes need not be complete, but in that case, value must be an array of the proper size. As an example, here’s how we could replace the third row of an n * 3 array:

let b = a with [2] <- ([1,2,3]) in b

As a convenience, whenever dest and src are the same, we can write:

let dest[indexes] = value in body

as a shortcut. Note that this has no special semantic meaning, but is simply a case of normal name shadowing.

For example, this loop implements the “imperative” version of matrix multiplication of an m * o with an o * n matrix:

let matmult(a: [m][o]f32,  b: [o][n]f32): [m][n]f32 =
  let res = replicate(m, replicate(n,0f32)) in
  loop (res) = for i < m do
      loop (res) = for j < n do
          loop (partsum = 0f32) = for k < o do
            partsum + a[i,k] * b[k,j]
          let res[i,j] = partsum
          in res
      in res
  in res

With the naive implementation based on copying the source array, executing the let-with expression would require memory proportional to the entire source array, rather than proportional to the slice we are changing. This is not ideal. Therefore, the let-with construct has some unusual restrictions to permit in-place modification of the src array, as described in Uniqueness Types. Simply put, we track that src is never used again. The consequence is that we can guarantee that the execution of a let-with expression does not involve any copying of the source array in order to create the newly bound array, and therefore the time required for the update is proportional to the section of the array we are updating, not the entire array. We can think of this as similar to array modification in an imperative language.

3.3. SOACs

The language presented in the previous section is in some sense “sufficient”, in that it is Turing-complete, and can express imperative-style loops in a natural way with do and while-loops. However, Futhark is not intended to be used in this way - bulk operations on arrays should be expressed via one of the second-order array combinators (SOACs) shown below, as this maximises the amount of parallelism that the compiler is able to take advantage of.

e ::=  "map" lambda e
       "filter" lambda e
       "partition" "(" lambda "," ... lambda ")" e
       "reduce" lambda e e
       "scan" lambda e e

A lambda can be an anonymous function, the name of a function (with optional curried arguments), or an operator (possibly with one operand curried):

lambda ::=  "(" \ param... : rettype "->" e ")"
            "(" fname e ... e ")"
            "(" op e ")"
            "(" e op ")"
            "(" op ")"

Parameter- and return type ascriptions are optional in anonymous functions. The semantics of the SOACs is identical to the similarly-named higher-order functions found in many functional languages. For specifics, see Language Reference.

The scan SOAC performs an inclusive prefix scan, and returns an array of the same outer size as the original array. The functions given to reduce and scan must be binary associative operators, and the value given as the initial value of the accumulator must be the neutral element for the function. These properties are not checked by the Futhark compiler, and are the responsibility of the programmer.

3.4. Uniqueness Types

While Futhark is uncompromisingly a pure functional language, it may occasionally prove useful to express certain algorithms in an imperative style. Consider a function for computing the n first Fibonacci numbers:

let fib(n: i32): []i32 =
  -- Create "empty" array.
  let arr = iota(n) in
  -- Fill array with Fibonacci numbers.
  loop (arr) = for i < n-2 do
                 let arr[i+2] = arr[i] + arr[i+1]
                 in arr
  in arr

If the array arr is copied for each iteration of the loop, we are going to put enormous pressure on memory, and spend a lot of time moving around data, even though it is clear in this case that the “old” value of arr will never be used again. Precisely, what should be an algorithm with complexity O(n) becomes (n^2) due to copying the size n array (an O(n) operation) for each of the n iterations of the loop.

To prevent this, we will want to update the array in-place, that is, with a static guarantee that the operation will not require any additional memory allocation, such as copying the entire array. With an in-place modification, a let-with can modify the array in time proportional to the slice being updated (O(1) in the case of the Fibonacci function), rather than time proportional to the size of the final array, as would the case if we performed a full copy. In order to perform the update without violating referential transparency, we need to know that no other references to the array exists, or at least that such references will not be used on any execution path following the in-place update.

In Futhark, this is done through a type system feature called uniqueness types, similar to, although simpler, than the uniqueness types of Clean. Alongside a (relatively) simple aliasing analysis in the type checker, this is sufficient to determine at compile time whether an in-place modification is safe, and signal a compile time error if let-with is used in way where safety cannot be guaranteed.

The simplest way to introduce uniqueness types is through examples. To that end, let us consider the following function definition:

let modify(a: *[]i32, i: i32, x: i32): *[]i32 =
  let a[i] = a[i] + x in

The function call modify(a,i,x) returns a, but where the element at index i has been increased by x. Note the asterisks in the parameter declaration *[]i32 a. This means that the function modify has been given “ownership” of the array a, meaning that the caller of modify will never reference array a after the call. As a consequence, modify can change the element at index i without first copying the array, i.e. modify is free to do an in-place modification. Furthermore, the return value of modify is also unique - this means that the result of the call to modify does not share elements with any other visible variables.

Let us consider a call to modify, which might look as follows:

let b = modify(a, i, x) in

Under which circumstances is this call valid? Two things must hold:

  1. The type of a must be *[]i32, of course.
  2. Neither a or any variable that aliases a may be used on any execution path following the call to modify.

In general, when a value is passed as a unique-typed argument in a function call, we consider that value to be consumed, and neither it nor any of its aliases can be used again. Otherwise, we would break the contract that gives the function liberty to manipulate the argument however it wants. Note that it is the type in the argument declaration that must be unique - it is permissible to pass a unique-typed variable as a non-unique argument (that is, a unique type is a subtype of the corresponding nonunique type).

A variable v aliases a if they may share some elements, i.e. overlap in memory. As the most trivial case, after evaluating the binding let b = a, the variable b will alias a. As another example, if we extract a row from a two-dimensional array, the row will alias its source:

let b = a[0] in
... -- b is aliased to a (assuming a is not one-dimensional)

In Sharing Analysis below, we will cover sharing and sharing analysis in greater detail.

Let us consider the definition of a function returning a unique array:

let f(a: []i32): *[]i32 = body

Note that the argument, a, is non-unique, and hence we cannot modify it. There is another restriction as well: a must not be aliased to our return value, as the uniqueness contract requires us to ensure that there are no other references to the unique return value. This requirement would be violated if we permitted the return value in a unique-returning function to alias its non-unique parameters.

To summarise: values are consumed by being the source in a ``let-with``, or by being passed as a unique parameter in a function call. We can crystallise valid usage in the form of three principal rules:

Uniqueness Rule 1

When a value is passed in the place of a unique parameter in a function call, or used as the source in a let-with expression, neither that value, nor any value that aliases it, may be used on any execution path following the function call. An example violation:

let b = a
let b[i] = 2 in
f(b,a) -- Error: a used after being source in a let-with

Uniqueness Rule 2

If a function definition is declared to return a unique value, the return value (that is, the result of the body of the function) must not share memory with any non-unique arguments to the function. As a consequence, at the time of execution, the result of a call to the function is the only reference to that value. An example violation:

let broken(a: [][]i32, i: i32): *[]i32 =
  a[i] -- Return value aliased with 'a'.

Uniqueness Rule 3

If a function call yields a unique return value, the caller has exclusive access to that value. At the point the call returns, the return value may not share memory with any variable used in any execution path following the function call. This rule is particularly subtle, but can be considered a rephrasing of Uniqueness Rule 2 from the “calling side”.

It is worth emphasising that everything in this chapter is employed as part of a static analysis. All violations of the uniqueness rules will be discovered at compile time during type-checking, thus leaving the code generator and runtime system at liberty to exploit them for low-level optimisation.

3.4.1. Sharing Analysis

Whenever the memory regions for two values overlap, we say that they are aliased, or that sharing is present. As an example, if you have a two-dimensional array a and extract its first row as the one-dimensional array b, we say that a and b are aliased. While the Futhark compiler may do a deep copy if it wishes, it is not required, and this operation thus holds the potential for sharing memory. Sharing analysis is necessarily conservative, and merely imposes an upper bound on the amount of sharing happening at runtime. The sharing analysis in Futhark has been carefully designed to make the bound as tight as possible, but still easily computable.

In Futhark, the only values that can have any sharing are arrays - everything else is considered “primitive”. Tuples are special, in that they are not considered to have any identity beyond their elements. Therefore, when we store sharing information for a tuple-typed expression, we do it for each of its element types, rather than the tuple value as a whole.

Most operations produce arrays without any aliases. You can think of these as producing fresh arrays. The exceptions are split, reshape, transpose, rearrange, zip and unzip, as well as function calls and if expressions (depending on types). You can use copy to “break” sharing by forcing the argument to be manifested freshly in memory.