md5 :: LazyByteString -> MD5Ctx
In June 2007 I developed a Haskell MD5 implementation intended to be competitive with the 'C' version while still remaining readable. This is a lazy ByteString implementation that makes heavy use of inlining, unboxing and strictness annotations. During this process I [re?]discovered some querks about GHC optimizations and Data.Binary.
The first relatively operational version used unsafePerformIO to extract the Word32s from the ByteString. Minor modifications were made to use Data.Binary - this is, after all, an intended use of Data.Binary.
There are four incarnations. First, there is the manual round unrolling and a foldl' version; second, there is a version that uses Data.Binary to extract the Word32s and another that uses unsafePerformIO + Ptr operations.
I standardized the benchmark as the time it takes to hash a 200MB file (seconds) and show performance relative to standard md5sum.
GHC 6.6.1 results
MD5.unroll.unsafe: 5s 6x MD5.unroll.Binary: 10s 12x MD5.rolled.Binary: 45s 58x MD5.rolled.unsafe: 27s 35x md5sum: 0.78s 1x, by def.
UnsafeIO v. Binary
Sadly, changing the (getNthWord32 :: Int -> ByteString -> Word32) function from unsafePerformIO to using Data.Binary made the system take twice as long. It was my hope that Data.Binary was optimized to the same degree as its unsafePerformIO counterpart.
Manual unrolling? This isn't (insert hated language/compiler)
As you can see, the manual unrolling has saved quite a bit of time. Why did unrolling save computation time? I was optimistic in hoping GHC would eliminate unneeded arrays/accesses when unfolding (among other optimizations).
When profiling (-p -auto-all), the 'Binary' version attributes much memory to 'applyMD5Rounds' while the 'unsafe' version attributes more to the ff,gg,hh,ii functions. I am guessing the 'unsafe' version is correct... is the program profile using Data.Binary wrong?
Dreams of SuperOI am not trying to be Data.Binary/GHC centric. Matter of fact, I look forward to seeing what YHC/supero would do - I just need to be able to get it (and know how to use it).
Observations from core (-ddump-simpl)
I am not patient or knowledgeable enough to know if my understanding of core is correct... that doesn't stop me from interpreting it. It looks like the programs are boxing/unboxing the context between every fold of foldl'! If this is True it explains some of the performance issues. Folding over rounds would cost 64 box/unboxes per block in the rolled version (once for every byte hashed). Folding between each round would cost one box/unbox per block even in the unrolled version (32 million box/unboxings when hashing 200MB). If this is true, it is an obvious place for performance improvement
Minor Alterations With Unexpected Results
Eliminating md5Finalize resulted in a sub 2s run (>2x speed increase, ~3x slower than 'C'). Finalize only runs once (at the end of the hash computation) and is profiled at 1 tick. The only explanation I can see is that md5Finalize use of md5Update complicates the optimization. Inlining md5Update doesn't help and neither does making/using identical copies of md5Update (md5Update') and all sub-functions.
Edit: Benchmark includes nanoMD5 too!
GHC 188.8.131.5270914 Results
Time v. 'C' v. GHC 6.6.1 md5.unroll.unsafe 5s 6x 1x md5.roll.unsafe 17s 21x 0.63x md5.nano 0.9s 1.15x -
So I see a significant improvement in the rolled version thanks to everyone involved in GHC.
Any suggestions on how to close the remaining performance gap would be welcome.
Yes, I know, it looks ugly! I don't feel like cleaning it up much right now. If someone voices that they care, that would be a motivator.
Summary (new from edit)
Footer (for the really curious)
Hardware: 2.2Ghz AMD x86_64 (Athlon 3400+) with 1GB of memory. Software: Linux 2.6.17, GHC (6.6.1 and 6.8.0 as indicated), md5sum 5.97.
flags: -O2 -funfolding-use-threshold66 -funfolding-creation-threshold66 -funfolding-update-in-place -fvia-c -optc -funroll-all-loops -optc-ffast-math
EDIT: Now on darcs:
darcs get http://code.haskell.org/~tommd/pureMD5