mutable

Documentation and Walkthrough

Beautiful Mutable Values

Mutability can be awesome!

Take back the power of mutable objects with all the safety and explicit state of Haskell. Associate and generate "piecewise-mutable" versions for your composite data types in a composable and automatic way. Think of it like a "generalized MVector for all ADTs". It also leverages GHC Generics to make working with piecewise mutability as simple as possible.

Making piecewise updates on your giant composite data types (like artificial neural networks or game states in your game loop) got you down because they require re-allocating the entire value? Tired of requiring a full deep copy every time you make a small change, and want to be able to build mutable versions of your types automatically in composable ways? This is the package for you.

Useful for a situation where you have a record with many fields (or many nested records) that you want to use for efficient mutable in-place algorithms. This library lets you do efficient "piecewise" mutations (operations that only edit one field), and also efficient entire-datatype copies/updates, as well, in many cases.

Check out the documentation home page, haddock reference, or read below for motivation and high-level descriptions.

Motivation

Piecewise-Mutable

For a simple motivating example where in-place piecewise mutations might be better, consider a large vector.

Let's say you only want to edit the first item in a vector, multiple times. This is extremely inefficient with a pure vector:

addFirst :: Vector Double -> Vector Double
addFirst xs = iterate incr xs !! 1000000
  where
    incr v = v V.// [(0, (v V.! 0) + 1)]

That's because addFirst will copy over the entire vector for every step --- every single item, even if not modified, will be copied one million times. It is O(n*l) in memory updates --- it is very bad for long vectors or large matrices.

However, this is extremely efficient with a mutable vector:

addFirst :: Vector Double -> Vector Double
addFirst xs = runST $ do
    v <- V.thaw xs
    replicateM_ 1000000 $ do
        MV.modify v 0 (+ 1)
    V.freeze v

This is because all of the other items in the vector are kept the same and not copied-over over the course of one million updates. It is O(n+l) in memory updates. It is very good even for long vectors or large matrices.

(Of course, this situation is somewhat contrived, but it isolates a problem that many programs face. A more common situation might be that you have two functions that each modify different items in a vector in sequence, and you want to run them many times interleaved, or one after the other.)

Composite Datatype

That all works for MVector, but let's say you have a simple composite data type that is two vectors:

data TwoVec = TV { tv1 :: Vector Double
                 , tv2 :: Vector Double
                 }
  deriving Generic

Is there a nice "piecewise-mutable" version of this? You could break up TwoVec manually into its pieces and treat each piece independently, but that method isn't composable. If only there was some equivalent of MVector for TwoVec...and some equivalent of MV.modify.

That's where this library comes in.

instance PrimMonad m => Mutable m TwoVec where
    type Ref m TwoVec = GRef m TwoVec

This gives us thawRef :: TwoVec -> m (GRef m TwoVec), where GRef m TwoVec is a mutable version of TwoVec, like how MVector s Double is a mutable version of Vector Double. It stores each field tv1 and tv2 as a seaprate MVector in memory that can be modified independently.

Now we can write:

addFirst :: TwoVec -> TwoVec
addFirst xs = runST $ do
    v <- thawRef xs
    replicateM_ 1000000 $ do
      withField #tv1 v $ \u ->
        MV.modify u 0 (+ 1)
    freezeRef v

This will in-place edit only the first item in the tv1 field one million times, without ever needing to copy over the contents tv2. Basically, it gives you a version of TwoVec that you can modify in-place piecewise. You can compose two functions that each work piecewise on TwoVec:

mut1 :: PrimMonad m => Ref m TwoVec -> m ()
mut1 v = do
    withField #tv1 v $ \u ->
      MV.modify u 0 (+ 1)
      MV.modify u 1 (+ 2)
    withField #tv2 v $ \u ->
      MV.modify u 2 (+ 3)
      MV.modify u 3 (+ 4)

mut2 :: PrimMonad m => Ref m TwoVec -> m ()
mut2 v = do
    withField #tv1 v $ \u ->
      MV.modify u 4 (+ 1)
      MV.modify u 5 (+ 2)
    withField #tv2 v $ \u ->
      MV.modify u 6 (+ 3)
      MV.modify u 7 (+ 4)

doAMillion :: TwoVec -> TwoVec
doAMillion xs = runST $ do
    v <- thawRef xs
    replicateM_ 1000000 $ do
      mut1 v
      mut2 v
    freezeRef v

This is a type of composition and interleaving that cannot be achieved by simply breaking down TwoVec and running functions that work purely on each of the two vectors individually.

Show me the numbers

Here are some benchmark cases --- only bars of the same color are comparable, and shorter bars are better (performance-wise).

There are four situations here, compared and contrasted between pure and mutable versions

1. A large ADT with 256 fields, generated by repeated nestings of data V4 a = V4 !a !a !a !a

   1. Updating only a single part (one field out of 256)
   2. Updating the entire ADT (all 256 fields)

2. A composite data type of four Vectors of 500k elements each, so 2 million elements total.

   1. Updating only a single part (one item out of 2 million)
   2. Updating all elements of all four vectors (all 2 million items)

We can see four conclusions:

1. For a large ADT, updating a single field (or multiple fields, interleaved) is going to be faster with mutable. This speedup is between x4 and x5, suggesting it is a speedup arising from the fact that the top-level type has four fields.
2. For a large ADT, updating the whole ADT (so just replacing the entire thing, no actual copies) is faster just as a pure value by a large factor (which is a big testament to GHC).
3. For a small ADT with huge vectors, updating a single field is much faster with mutable.
4. For a small ADT with huge vectors, updating the entire value (so, the entire vectors and entire ADT) is actually faster with mutable as well.

Interestingly, the "update entire structure" case (which should be the worst-case for mutable and the best-case for pure values) actually becomes faster with mutable when you get to the region of many values... somewhere between 256 and 2 million, apparently. However, this may just be from the efficiency of modifying vectors sequentially.