haskell / vector

An efficient implementation of Int-indexed arrays (both mutable and immutable), with a powerful loop optimisation framework .
Other
366 stars 139 forks source link

Optimize Mutable.nextPermutation and add {next/prev}permutation(By) #498

Closed gksato closed 3 months ago

gksato commented 4 months ago

This implements some optimization of nextPermutation from Data.Vector.Generic.Mutable, and supercedes #497. The main content of this re-implementation is the following two points:

  1. Wrapping the whole implementation in stToPrim. This allows the compiler to optimize the code better.
  2. When finding the rightmost increasing pair v[k]<v[k+1], we now search from the right, instead of from the left. This allows us to abort the search as soon as we find such a pair, giving average-case constant performance, instead of best-case linear in the previous implementation.

Also, this adds some clarification on doc comments; nextPermutation does not update the source vector when the given vector is the last permutation.


Update

This also adds the following API:

Data.Vector.*.Mutable.nextPermutationBy :: Constraint => (e -> e -> Ordering) -> v (PrimState m) a -> m Bool
Data.Vector.*.Mutable.prevPermutation :: (Constraint, Ord e) => v (PrimState m) a -> m Bool
Data.Vector.*.Mutable.prevPermutationBy :: Constraint => (e -> e -> Ordering) -> v (PrimState m) a -> m Bool
Shimuuar commented 4 months ago

Thanks!

I'll look into this tomorrow but at a glance nextPermutationByLt miss INLINE pragma which means it won't be specialized. Is this intentional?

gksato commented 4 months ago

I just aligned it with the original implementation; I assumed it was intentional. Is it better to add it before your review?

Shimuuar commented 4 months ago

I just aligned it with the original implementation

I see. I missed that somehow.

Is it better to add it before your review?

No need. I'll benchmark it and going to tweak code a bit anyway

Shimuuar commented 4 months ago

Yes. This is an optimization. I used very simple benchmark

permute :: Int -> Int -> IO (Vector Int)
permute sz n_it = do
  vec <- MV.generate sz id
  replicateM_ n_it $ MV.nextPermutation vec
  V.unsafeFreeze vec

With vector size 20 and n_it varying in range 1000-10000. See Shimuuar/vector branch bench-next-permutation for very quick and dirty code.

Benchmarks

I've run five benchmarks:

  1. [00.Base] Baseline from current master
  2. [01.strict] Added strictness. First change from #497
  3. [02.stToPrim] Added stToPrim
  4. [03.INLINE] Added INLINE pragma
  5. [04.gksato] Version from this PR

image

So for unboxed vector of ints stToPrim gives ~4× runtime speedup and your optimization another 3×. Of course inlined and specialized version beats them all.

Do you plan to include functions proposed in #499 in this PR?

gksato commented 4 months ago

Do you plan to include functions proposed in https://github.com/haskell/vector/issues/499 in this PR?

I was not sure if it was appropriate to put optimization and API addition in one PR, but I'm happy to do it if you propose so! Should CHANGELOG be added now?

gksato commented 4 months ago

Since I forgot to tweak the comments in Data.Vector.*.Mutablein 4ac750f, I'm going to force-push an updated history.

gksato commented 4 months ago

I added INLINE to the internal routine nextPermutationByLt, because cabal v2-bench with a line of bench "permute" $ whnfIO 20 useSize proved it 10x faster with INLINE (I haven't tweaked vector-bench-papi, because I'm on MacOS).

gksato commented 4 months ago

Oh, and I added:

Data.Vector.*.Mutable.nextPermutationBy :: Constraint => (e -> e -> Ordering) -> v (PrimState m) a -> m Bool
Data.Vector.*.Mutable.prevPermutation :: (Constraint, Ord e) => v (PrimState m) a -> m Bool
Data.Vector.*.Mutable.prevPermutationBy :: Constraint => (e -> e -> Ordering) -> v (PrimState m) a -> m Bool

The TODOs we have now are... tests, benchmarks, and changelog? Are they all?

gksato commented 4 months ago

I've written some tests on a separate branch (gksato:optimize-nextperm-tidyup), since I don't know if it is wanted. I'd fast-forward the branch (gksato:optimize-nextperm) of this pull request to include that commit whenever any of you request it!

Shimuuar commented 3 months ago

I finally found time to review PR. Please add tests and this PR is good to go. It would be nice to add changelog entry as well and whether you want to add benchmarks or not is up to you.

gksato commented 3 months ago

@Shimuuar OK, I'll do that soon. I've been writing some benchmark, so I might add it if implementation doesn't take much time. For the purpose of the changelog and @since pragma, what should be the version? I assume 0.13.2.0, but I'd like to be sure that it'll not be 0.13.1.1?

Shimuuar commented 3 months ago

Yes 0.13.2.0. Adding new definitions to module requires at least minor bump according to PVP. Patch releases are for bugfixes and documentation

gksato commented 3 months ago

Added tests, a changelog entry, @since annotations, and benchmarks. Force-pushed some commit message rewording because @since without quotes refers to someone's name on a GitHub!

Note that vector-bench-papi hasn't been updated, because they say it doesn't go nicely on MacOS (I haven't tried it myself). However, I assume that applying the same diff as vector/benchmarks/Main.hs to vector-bench-papi/benchmarks/Main.hs should work.

Shimuuar commented 3 months ago

That's quite formidable battery of tests all right! It's more thorough than rest of benchmark suite :)

I'm unhappy with inlining of worker function. It's too large. But we're running into GHC's limitations since we don't way to ask GHC to generate specializations for it and its callers automatically.

Excellent pull request. Thank you!

Note that vector-bench-papi hasn't been updated

Yes ..-papi package should modified same way. And that's the point when you start to want to use backpack in order to be able to swap bencmarking engine. I implemented poor man's version in math-functions. Sadly backpack never caught on

P.S. @since. That poor guy. He must hate haskell.

gksato commented 3 months ago

I appreciate your work on review and merge, thank you!

Oh, yes. I haven't noticed this would of course be a nice use case for backpack. I imagined what if we try to comply DRY without backpack... urgh, it's a mess!

Sadly backpack never caught on

I truly agree with you.

gksato commented 1 month ago

I've noticed that the new version is on its way, which I'm glad to see. I appreciate your work. However, I would like to poke you about a small, minute detail:

Note that vector-bench-papi hasn't been updated

Yes ..-papi package should modified same way.

Can we leave it as it is? I was just too lazy to build a Docker container to run vector-bench-papi on my MacOS and I didn't modify it in this PR.

Shimuuar commented 1 month ago

Thanks for reminding me. I totally forgot. I'll fix it after making release. It's fine since vector-bench-papi is not published on hackage anyway

gksato commented 1 month ago

I see, I leave it to you! Thanks for your good maintaining work!