[Haskell-cafe] Performance: MD5

[Salvatore Insalaco]

2008/5/18 Andrew Coppin <andrewcoppin at btinternet.com>:
>> unsafeRead/Write
>>
>
> Found in Data.Array.Base, apparently. (I managed to find an example in the
> Gtk2hs "fastdraw" demo.)
>
> Oddly, this seems to make virtually no performance difference. I find this
> highly surprising. But then, I perform 16 write to the array, and 64 reads
> from the ByteString, so maybe that's where I should be worrying about bounds
> checks...

Hi Andrew,
just a profiling suggestion: did you try to use the SCC cost-centre
annotations for profiling?
If you want to know precisely what takes 60% of time, you can try:
        bn = {-# SCC "IntegerConversion" #-} 4 * fromIntegral wn
        b0 = {-# SCC "ByteStringIndexing" #-} B.index bi (bn+0)
        b1 = {-# SCC "ByteStringIndexing" #-} B.index bi (bn+1)
        b2 = {-# SCC "ByteStringIndexing" #-} B.index bi (bn+2)
        b3 = {-# SCC "ByteStringIndexing" #-} B.index bi (bn+3)
        w  = foldl' (\w b -> shiftL w 8 .|. fromIntegral b) 0 [b3,b2,b1,b0]
      in {-# SCC "ArrayWriting" #-} unsafeWrite temp wn w

In profiling the time of all expressions with the same SCC "name" will
be summed.
You can get more information about SCC here:
http://www.haskell.org/ghc/docs/latest/html/users_guide/profiling.html#cost-centres

Obviously, even ultra-optimizing this function to be 60 times faster
(assuming that this is possible) will only make your program go about
2 time faster.

One advice: I've seen various md5sum implementations in C, all using
about the same algorithms and data structures, and they performed even
with 10x time differences between them; md5summing fast is not a
really simple problem. If I were you I would drop the comparison with
ultra-optimized C and concentrate on "does my
high-level-good-looking-super-readable implementation perform
acceptably?".

What "acceptably" means is left as an exercise to the reader.

Salvatore