Re: Ocaml for Scientific computing

Re: Ocaml for Scientific computing

[Mike Lin]

On Sep 25, 7:11 am, Alex Mikhalev <a.mikha...@cranfield.ac.uk> wrote:
> Dear all,
> I am wondering is anyone using Ocaml for scientific computing? I didn't
> mean parsing, but for number crunching applications, like signal/image
> analysis. Is it suitable for this kind of tasks in general? I would like
> to hear from someone practically using it, not just theoretical
> possibility.

I use it for a lot of genome sequence processing, comparative genomics
and phylogenetic modeling. Algorithmically this involves conditional
random fields (for which I have my own library) and some linear
algebra and numerical optimization (for which I use Lacaml and
ocamlgsl). I had a lot of frustrations at first, but it's been better
since we got exception stack traces and ocaml+twt.

For numerical computing, I wish ocamlopt would do at least basic loop
optimizations, like hoisting invariant values -- this type of stuff is
easily done manually, but often at the expense of code readability. I
can see how this may be a bottomless pit for the dev team though,
since I would probably always feel like we are one crucial
optimization short of not having to rewrite that tight loop in C.

In any case though, I work in a group where everyone else uses Python,
so I'm already beating their pants off :o)

Mike

Re: [Caml-list] Re: Ocaml for Scientific computing

[skaller]

On Tue, 2007-09-25 at 13:27 -0400, Mike Lin wrote:
> On Sep 25, 7:11 am, Alex Mikhalev <a.mikha...@cranfield.ac.uk> wrote:

> For numerical computing, I wish ocamlopt would do at least basic loop
> optimizations, like hoisting invariant values -- this type of stuff is
> easily done manually, but often at the expense of code readability.

Gcc does this quite well I think. However in a higher level 
procedural language (like Ocaml and Felix) it is very hard
to get right and potentially very expensive.


-- 
John Skaller <skaller at users dot sf dot net>
Felix, successor to C++: http://felix.sf.net

Re: [Caml-list] Re: Ocaml for Scientific computing

[Jon Harrop]

On Tuesday 25 September 2007 19:46:34 skaller wrote:
> On Tue, 2007-09-25 at 13:27 -0400, Mike Lin wrote:
> > On Sep 25, 7:11 am, Alex Mikhalev <a.mikha...@cranfield.ac.uk> wrote:
> >
> > For numerical computing, I wish ocamlopt would do at least basic loop
> > optimizations, like hoisting invariant values -- this type of stuff is
> > easily done manually, but often at the expense of code readability.
>
> Gcc does this quite well I think. However in a higher level
> procedural language (like Ocaml and Felix) it is very hard
> to get right and potentially very expensive.

I'm not sure that it is conceptually more difficult to do similar things for 
OCaml but my vote goes to hoisting bounds checks. I don't like having to 
write unsafe code by hand in OCaml and F# does a great job of improving 
performance by hoisting bounds checks.

I discussed some of the benchmarks where F# is faster than OCaml recently on 
one of our blogs:

  http://ocamlnews.blogspot.com/2007/09/spectral-norm.html

For some reason, the current implementation of spectral-norm on the shootout 
leaves out some important optimizations in OCaml. F# doesn't require these 
optimizations to be done by hand.

-- 
Dr Jon D Harrop, Flying Frog Consultancy Ltd.
OCaml for Scientists
http://www.ffconsultancy.com/products/ocaml_for_scientists/?e

Re: [Caml-list] Re: Ocaml for Scientific computing

[skaller]

On Tue, 2007-09-25 at 21:38 +0100, Jon Harrop wrote:

> I'm not sure that it is conceptually more difficult to do similar things for 
> OCaml but my vote goes to hoisting bounds checks. I don't like having to 
> write unsafe code by hand in OCaml and F# does a great job of improving 
> performance by hoisting bounds checks.

I guess that this is easier. How much speed is gained eliding
bounds checks?

-- 
John Skaller <skaller at users dot sf dot net>
Felix, successor to C++: http://felix.sf.net

Re: [Caml-list] Re: Ocaml for Scientific computing

[Florian Hars]

skaller schrieb:
> I guess that this is easier. How much speed is gained eliding
> bounds checks?

It may even make the code slower:

http://caml.inria.fr/pub/ml-archives/caml-list/2004/01/679987f7b8b70e4cd9f088ac4e6c8097.en.html

Yours, Florian.

Re: [Caml-list] Re: Ocaml for Scientific computing

[skaller]

On Wed, 2007-09-26 at 10:27 +0200, Florian Hars wrote:
> skaller schrieb:
> > I guess that this is easier. How much speed is gained eliding
> > bounds checks?
> 
> It may even make the code slower:
> 
> http://caml.inria.fr/pub/ml-archives/caml-list/2004/01/679987f7b8b70e4cd9f088ac4e6c8097.en.html

Yes, I tend to agree with Xavier's analysis.
This may also explain why C code generated by Felix is 
often faster than hand written C code: on AMD64 Felix
generated C code for Ackermann's function is almost the same
as the hand written C, but runs 2x faster, and 2.5x faster
than Ocamlopt generated code.

At this point with gcc as the compiler and using stack
programming (no heap), I'm now expecting Felix to beat
C on all programs (and if it doesn't it's a bug 
in the compiler that needs fixing).

I am guessing this is because GCC has good SSA analysis,
and it works a lot better on goto spaghetti with a fixed
stack frame of the kind Felix tends to generate, than
block structured C code using for loops, etc.

So I'd add to Xaviers comments on processor behaviour
that small differences in source code structure can make
a big difference to compiler technology.


-- 
John Skaller <skaller at users dot sf dot net>
Felix, successor to C++: http://felix.sf.net