Version 1 (modified by chak, 7 years ago) (diff)


Vectorisation: The lifted case example

This example is taken from

Flattening sums

We represent arrays of type

data Either a b = Left a | Right b

essentially as

instance (ArrayElem a, ArrayElem b) => ArrayElem (Either a b) where
  type [:Either a b:] = ([:Bool:], [:a:], [:b:])

For instance, [:Left 5, Right 4, Left 2:] will be represented by ([:True, False, True:], [:5,2:], [:4:]). The boolean array (the selector) indicates whether the corresponding element is a Left or a Right one; the actual data is stored in two separate arrays.

Case distinction (take 1)

For case distinction on Either, we have

either :: (a -> c) -> (b -> c) -> Either a b -> c
either f g (Left  x) = f x
either f g (Right y) = g y

The interesting question, of course, is how this works in a data-parallel context, i.e., what semantics something like mapP (either f g) xs has. For the moment, we assume that [:a -> b:] = [:a:] -> [:b:].

Sometimes it works …

Let us first consider

inc :: Int -> Int
inc n = n+1

good :: Either Int Int -> Int
good x = either inc id x

The term mapP good [:Left 5, Right 4, Left 2:] translates to good^ (EitherArr [:True, False, True:] [:5,2:] [:4:]); to see how this is evaluated, we have to derive the lifted version of good^. This is easy (we ignore the tupling of functions here):

good^ :: [:Either Int Int:] -> [:Int:]
good^ xs = either^ inc^ id^ xs

We need the lifted versions of either, inc and id. The latter is trivial:

id :: a -> a
id x = x

id^ :: [:a:] -> [:a:]
id^ xs = xs

We will omit the definition of inc^ - it just increments every element in an array. The definition of either^ is easily derived from its type:

either^ :: ([:a -> c:])     -> ([:b -> c:])     -> [:Either a b:]           -> [:c:]
        :: ([:a:] -> [:c:]) -> ([:b:] -> [:c:]) -> ([:Bool:], [:a:], [:b:]) -> [:c:]   -- assuming [:a -> b:] = [:a:] -> [:b:]
either^ f g (bs, xs, ys) = combineP bs (f xs) (g ys)

Here, we apply f and g to the respective data arrays and the combine the results according to the selector. For instance, we have

combineP [:True, False, True:] [:6,3:] [:4:] = [:6,4,3:]

and therefore

  mapP good [:Left 5, Right 4, Left 2:]
= good^ ([:True, False, True:], [:5,2:], [:4:])
= either^ inc^ id^ ([:True, False, True:], [:5,2:], [:4:])
= combineP [:True, False, True:] (inc^ [:5,2:]) (id^ [:4:])
= combineP [:True, False, True:] [:6,3:] [:4:]
= [:6,4,3:]

This is precisely the expected result.

... but sometimes it doesn't

Now, let us consider a slightly more complex function:

bad :: (Either Int Int, Int) -> Int
bad x = either ((+) (snd x)) id (fst x)

Again, we derive the lifted version of bad:

bad^ :: [:(Either Int Int, Int):] -> [:Int:]
bad^ xs = either^ ((+^) (snd^ x)) id^ (fst^ x)

So, how is mapP bad [:(Left 5, 1), (Right 4, 3), (Left 2, 7):] evaluated? We have

  mapP bad [:(Left 5, 1), (Right 4, 3), (Left 2, 7):]
= bad^ [:(Left 5, 1), (Right 4, 3), (Left 2, 7):]
= either^ ((+^) [:1,3,7:]) id^ [:Left 5, Right 4, Left 2:]
= either^ ((+^) [:1,3,7:]) id^ ([:True,False,True:], [:5,2:], [:4:])
= combineP [:True,False,True:] ((+^) [:1,3,7:] [:5,2:]) (id^ [:4:])

This program diverges when trying to evaluate (+^) [:1,3,7:] [:5,2:] since (+^) only works with arrays of the same length. Note that it is not enough to just throw away the last element of the first array like Haskell's zipWith would do; we really want to evaluate (+^) [:1,7:] [:5,2:], i.e. throw away the 3.

Case distinction (take 2)

So why doesn't it work? Simply because our definition of either^ is wrong. In either^ ((+^) [:1,3,7:]) id^ ([:True,False,True:], [:5,2:], [:4:]) two function arguments each contain one computation for each element of the Either array; it is the job of either^ (and of case distinction in general) to select those which really should be applied. In particular, either^ should filter the function arguments according to the selector. Without assuming [:a -> b:] = [:a:] -> [:b:], it has the type

either^ :: [:a -> c:] -> [:b -> c:] -> [:Either a b:] -> [:c:]

The real type is actually [:a -> c:] -> [:(b -> c) -> Either a b -> c:] but this is not important at the moment. So, either^ has three array parameters; and since it is a lifted function, these arrays must have the same length. The correct definition is

either^ fs gs (bs, xs, ys) = let fs' = packP bs fs
                                 gs' = packP (not^ bs) gs
                                 xs' = fs' $^ xs
                                 ys' = gs' $^ ys
                             combineP bs xs' ys'

Here, $^ denotes the elementwise application of an array of functions to an array of arguments; packP filters an array according to a selector.

Let us see how this works out for our example. Now, the type of (+^) is

(+^) :: [:Int:] -> [:Int -> Int:]

and we will denote the result of f^ [:x1, ..., xn:] by [:f x1, ..., f xn:]. Moreover, bad^ should really be defined as

bad^ :: [:(Either Int Int, Int):] -> [:Int:]
bad^ xs = either^ ((+^) (snd^ x)) (replicateP (lengthP xs) id) (fst^ x)

Note that where we previously lifted id to id^, we now replicate it to get a function array. Then, we have

  mapP bad [:(Left 5, 1), (Right 4, 3), (Left 2, 7):]
= bad^ [:(Left 5, 1), (Right 4, 3), (Left 2, 7):]
= either^ ((+^) [:1,3,7:]) (replicateP 3 id) [:Left 5, Right 4, Left 2:]
= either^ [:(+) 1, (+) 3, (+) 7:] [:id, id, id:] ([:True,False,True:], [:5,2:], [:4:])
= let fs' = packP [:True,False,True:] [:(+) 1, (+) 3, (+) 7:]
      gs' = packP [:False,True,False:] [:id,id,id:]
      xs' = fs' $^ [:5,2:]
      ys' = gs' $^ [:4:]
  combineP [:True,False,True:] xs' ys'
= let fs' = [:(+) 1, (+) 7:]
      gs' = [:id:]
      xs' = fs' $^ [:5,2:]
      ys' = gs' $^ [:4:]
  combineP [:True,False,True:] xs' ys'
= let xs' = [:(+) 1, (+) 7:] $^ [:5,2:]
      ys' = [:id:] $^ [:4:]
  combineP [:True,False,True:] xs' ys'
= combineP [:True,False,True:] [:6,9:] [:4:]
= [:6,4,9:]

This gives us the correct result.

Arrays of functions

The crucial mistake in our first approach was to assume that arrays of functions can be represented by functions over arrays. This does not work because the latter do not (and cannot) support array operations. Unfortunately, it is not enough to simply forbid [:a -> b:] as the transformation introduces such types and sometimes needs to manipulate these arrays (as in packP above - either^ is just an example of how we would translate case expressions).

So why can't we just use boxed arrays of functions? First, this kills performance. The main reason, however, is that this would break the semantics of the data-parallel model. The flattening transformation ensures that in the parallel case, all processes essentially execute the same code at the same time. This gives us very strong semantic guarantees; in particular, we can use collective operations for communication and synchronisation. This would not be the case if we could run unlifted code in a lifted context, which is necessarily the case with boxed function arrays. Explicit closures seem to be the only solution to this problem.