Skip to content

Latest commit

 

History

History
370 lines (287 loc) · 15.9 KB

optimization-guidelines.md

File metadata and controls

370 lines (287 loc) · 15.9 KB

Guidelines for writing well optimized code

Stream operations are always part of a loop. Usually the loop consists of a stream generation or an unfold operation followed by stream transformation functions (e.g. map) and then a stream elimination operation or a fold operation.

The default or most common stream representation used in streamly is StreamD which is a direct style (compare with CPS representation StreamK) stream representation. All the direct style operations in a loop "fuse" together to form a tight machine loop eliminating any intermediate constructors, therefore, reducing allocations and cpu cost. This elimination of intermediate constructors in a stream loop is known as stream fusion.

Writing high performance code using streamly

This section outlines guidelines for writing code using the available combinators. The guidelines for writing new combinators are provided in the following section.

Note that there is no absolute benchmark for performance, most of the time even without following the guidelines you may get excellent performance for the task at hand. Two important points to keep in mind before you optimize:

  • See if you actually need more performance
  • See if the code you are optimizing is in fast path

Stream Fusion

When your performance requirements are stringent you may want to care about not breaking stream fusion unnecessarily. Certain operations do not fuse and act as a barrier to stream fusion. If these operations are part of a loop, the loop won't fuse completely. In some cases these may be necessary but in others it may be possible to replace these with better fusing operations. These operations include:

  • Avoid unnecessary use of stream append operations in code that should fuse. Append operations in general force CPS style breaking stream fusion. Some direct style append operations can fuse but they won't scale for more than a few appends.
  • Avoid unnecessary use of the stream monad if fusion is important. You may use functor or applicative though.
  • Concurrent stream operations cannot fuse
  • Operations involving exception primitives like catch/throw/mask on elements of the stream cannot fuse.
  • Stream loops involving FFI calls on the elements of the stream

The haddock documentation includes a note when an operation cannot fuse.

Unfolds

  • Use unfolds especially when higher order operations are involved. For example, unfoldMany can fuse completely whereas concatMap would not fuse.
  • Use outerProduct in Unfold module instead of using the monad instance of streams to fuse nested loops where performance matters. See the unfold benchmarks for an example. outerProduct can give you C like nested loop performance.

Inlining

To make sure fusion can occur INLINE any operations that are part of a stream loop but factored out as separate functions.

Strictness

Keep the fold and scan accumulators strict. You could also consider using mutable state in a fold accumulator if the state is large. A large immutable structure as an accumulator may cause optimization issues. See the WordCount example for an example of this case.

Writing streamly combinators

To enable stream fusion in a direct style stream all the operations that are part of a loop must be inlined:

  • Use an explicit INLINE pragma on any combinator that could be part of a loop to ensure they will be inlined.
  • When a higher order function consumes another function then ensure that the higher order function is inlined in the same phase or before the function it is consuming. Use appropriate inline phases to ensure proper inline ordering when required.

Streams and Unfolds

Stream and Unfold State. Direct style stream and unfold combinators use a state data type to represent the internal state of the stream/unfold generator.

  • In some cases we may have to add a FUSE annotation on the state type to ensure that all the internal join points created by GHC are inlined with the help of the fusion-plugin.
  • In rare cases we may need to use a strictness annotation on the state to allow fusion. parseMany and splitOnSeq are such examples.

Step function of a stream or unfold:

  • Do not introduce unnecessary states. If there is only one entry point of a state then perhaps you want to collapse the state into the state from which it is called. More states means more jumps and may affect code locality, and efficiency of low level code (e.g. register allocation) because of independent placement of the code blocks in different states.

    In some cases we may have to go against the above guideline and introduce a separate state even though it is not necessary. See the GroupConsume state in foldMany for an example.

  • Keep minimum possible data in state. More variables in state may lead to poor performance because the state may not fit into registers and the spill may cause allocations on each iteration of the loop. Mutable state may help in such cases.

  • The step function must be annotated such that it gets inlined after the main combinator (INLINE_LATE).

Multiple yield points or single?:

  • A single yield point is usually desirable, however, not always necessary. In some cases multiple yield points may in fact be needed for fusion, see splitOnSeq for an example. Or maybe its fusing because of a direct yield instead of going through an indirect common yielding state.

Recursion in step function:

  • In general, we avoid making the step function recursive. Use the Skip constructor to remove recursion. Recursive step function can introduce optimization barriers that are harder to remove by the GHC simplifier.

    However, in some cases it may be better to have a local recursive loop. A recursive loop can help us avoid threading around some large state values every time. Values that do not change across a loop can be factored in the scope outside the loop (static argument transform), this way we can create a local loop which is more efficient than threading around the state in a larger loop. See the splitOnSeq combinator as an example where we use a local recursive loop, it fuses and is significantly efficient compared to using Skip.

    In a local recursive loop use SPEC and annotate even the rest of the loop arguments as strict where needed. We have observed that when the arguments were not strict the loop does not fuse (splitOnSeq).

Fold and Parser drivers

Recursive loop closing operations:

  • When writing recursive looping combinators using StreamD (e.g. foldlM' in StreamD) use a strict SPEC argument in the recursive loops to ensure that "Spec Constructor" optimization removes boxed arguments and reduces allocations.

Fold and Parser combinators

State of a Fold or Parser:

  • The accumulator of a fold must be a strict data structure. Use the strict data structures provided in Streamly.Internal.Data for this purpose or use explicit bang annotation to make the data strict.
  • In some cases you may need a FUSE annotation on the state type to ensure that any internal join points created by GHC are always inlined with the help of the fusion-plugin.

The step function of a Fold or Parser:

  • Never make the fold step recursive, recursion creates an optimization barrier for the GHC simplifier. Use Partial or Continue constructors to avoid recursion.

  • Sometimes you may need an explicit INLINE on the step function.

NOINLINE, isolating the closed loop

We want the loop iterations to be optimized and the loop stages to be fused to generate a tighter loop. However, it is not necessarily optimal to inline the whole loop itself into a parent function. For example, consider the following function in the FileSystem.Handle benchmarks:

{-# NOINLINE readWriteAfter_Stream #-}
readWriteAfter_Stream :: Handle -> Handle -> IO ()
readWriteAfter_Stream inh devNull =
    let readEx = IP.after_ (hClose inh) (S.unfold FH.read inh)
     in S.fold (FH.write devNull) readEx

If this is inlined into a parent benchmark group list, this leads to many times performance degradation. That's because inlining the loop into a bigger structure interferes with the optimization of the loop body itself and it may not fuse. Whereas it is desirable to INLINE all the stages of a loop, it is often not useful to inline the whole loop itself, in fact we may have to occasionally use a NOINLINE so that the compiler does not inline it.

We need to be careful about NOINLINE on polymorphic functions. When NOINLINE is used, the function cannot be specialized causing huge performance degradation. To avoid that we can either make the function monomorphic or sometimes using NOINLINE [0] can do the trick. NOINLINE [0] blocks inlining in earlier phases but lets it specialize in the last phase, inlining in the last phase is left to the compiler. INLINE [0] may sound the same as NOINLINE [0], however, it behaves differently because we ask the compiler to INLINE it compulsorily and it may not give us the desired results. The Data.Fold.createOfLast benchmark is one such case where NOINLINE [0] provides much better performance than INLINE [0].

NoSpecConstr

It is not always useful to specialize function calls for all constructors, some constructors may just add to code bloat or add overhead of passing unboxed values in a loop. In such cases the NoSpecConstr annotation can be useful. See the parselMx' function in StreamD module.

StreamK operations

StreamK uses foldr/build fusion to a very limited degree. StreamK is not the primary representation in streamly but is used for several operations that cannot scale in StreamD representation. StreamK is relatively immune to compiler optimizations. In some cases you may need an INLINE pragma to improve the performance.

How to debug non-fusing code?

Strip down the code to a minimal version until it starts fusing and then start building it up from there adding more things incrementally. At each stage keep checking if the code is fusing. At some point it won't fuse. See what we added that made the code not fuse. Go through the guidelines above to check if we did something that that is not recommended. Or raise an issue for GHC to be fixed if possible.

GHC Optimizations In Streamly

There are three important levels of optimizations used in streamly:

  • CPS style stream representation (StreamK) to direct style (StreamD) conversion and vice-versa using rewrite rules
  • Stream fusion using case-of-case optimization in StreamD
  • foldr/build fusion using rewrite rules in StreamK

INLINE Phases

Inlining of functions at the right time is crucial for all these optimization to work. A missing inline or inline in an incorrect GHC simplifier phase can adversely impact performance. We use three builtin phases of GHC simplifier for inlining i.e. phase 0, 1 and 2. We have defined them as follows in src/inline.hs:

#define INLINE_EARLY  INLINE [2]
#define INLINE_NORMAL INLINE [1]
#define INLINE_LATE   INLINE [0]

We also use INLINE [3] in some cases.

fromStreamD/toStreamD Fusion

The combinators in Streamly.Prelude are defined in terms of combinators in Streamly.Internal.Data.Stream.StreamD (Direct style streams) or Streamly.Internal.Data.StreamK (CPS style streams). We convert the stream from StreamD to StreamK representation or vice versa in some cases.

Most operations use the StreamD representation, however, stream append operations and monad instances use StreamK representation because StreamD would not perform well in these cases. We use rewrite rules to convert from one representation to another when required. For this reason the combinators in Streamly.Prelude are written using fromStreamD/fromStreamK etc.

In the first inlining phase (INLINE_EARLY or INLINE) we expand the combinators in Streamly.Prelude into fromStreamD/fromStreamK/toStreamD/toStreamK and combinators defined in StreamD or StreamK modules. Once we do that fromStreamD/toStreamD get exposed and we can apply rewrite rules to rewrite transformations like fromStreamK . toStreamK to id. A plain INLINE pragma is usually enough on combinators in Streamly.Prelude.

{-# RULES "fromStreamK/toStreamK fusion"
  forall s. toStreamK (fromStreamK s) = s #-}

Also, we have to prevent fromStreamK and toStreamK themselves from inlining in this phase so that rewrite rules can be applied on them, therefore, we annotate these functions with INLINE_LATE.

Fallback Rules

In some cases, if the operation could not fuse we want to use a fallback rewrite rule in the next phase. For such cases we use the INLINE_EARLY phase for the first rewrite and the INLINE_NORMAL phase for the fallback rules.

The fallback rules make sure that if we could not fuse the direct style operations then better use the CPS style operation, because unfused direct style would have worse performance than the CPS style ops.

{-# INLINE_EARLY unfoldr #-}
unfoldr :: (Monad m, IsStream t) => (b -> Maybe (a, b)) -> b -> t m a
unfoldr step seed = fromStreamS (S.unfoldr step seed)
{-# RULES "unfoldr fallback to StreamK" [1]
     forall a b. S.toStreamK (S.unfoldr a b) = K.unfoldr a b #-}

High Level Operation Fusion

Since each high level combinator in Streamly.Prelude is wrapped in fromStreamD/toStreamD etc. the combinator fusion cannot work unless we have removed those and exposed consecutive operations e.g. a map followed by another map. Assuming that redundant fromStreamK/toStreamK have been removed in the INLINE_EARLY phase, we can then apply the combinator fusion rules in the INLINE_NORMAL phase. For example, we can fuse two map operations into a single map operation. Note that now we have exposed the StreamD/StreamK implementations of combinators and the rules would apply on those.

Inlining Higher Order Functions

Note that partially applied functions cannot be inlined. So if we have a code like this:

concatMap1 src = runStream $ S.concatMap (S.replicate 3) src

We want to ensure that concatMap gets inlined before replicate so that replicate becomes fully applied before it gets inlined. Currently ensuring that both of them are inlined in the same phase (INLINE_NORMAL) seems to be enough to achieve that. In general, we should try to ensure that higher order functions are inlined before or in the same phase as the functions they can consume as arguments. This means StreamD combinators should not be marked as INLINE or INLINE_EARLY, instead they should all be marked as INLINE_NORMAL because higher order functions like concatMap/map/mapM etc are marked as INLINE_NORMAL. StreamD functions in other modules like Streamly.Data.Array.Foreign should also follow the same rules.

Stream Fusion

In StreamD combinators, inlining the inner step or loop functions too early i.e. in the same phase or before the outer function is inlined may block stream fusion opportunities. Therefore, the inner step functions and folding loops are marked as INLINE_LATE.

Specialization

In some cases, the step function in StreamD does not get specialized when inlined unless it is provided with an explicit signature or made a lambda, for example, in the replicate/replicateM combinator we need the type annotation on i to get it specialized:

    {-# INLINE_LATE step #-}
    step _ (i :: Int) =
        if i <= 0
        then return Stop
        else do
                x <- action
                return $ Yield x (i - 1)

-flate-specialise also helps in this case.

Stream and Fold State Data Structures

Since state is an internal data structure threaded around in the loop, it is a good practice to use strict unboxed fields for state data structures where possible. In most cases it is not necessary, but in some cases it may affect fusion and make a difference of 10x performance or more. For example, using non-strict fields can increase the code size for internal join points and functions created during transformations, which can affect the inlining of these code blocks which in turn can affect stream fusion.

See https://gitlab.haskell.org/ghc/ghc/issues/17075 .