On Thu, Feb 07, 2008 at 11:01:54AM +0000, Richard Jones wrote:
> On Wed, Feb 06, 2008 at 11:34:02PM +0000, Jon Harrop wrote:
> > In this case, most of the speed loss can be regained by simply
> > aggressively inlining everything, which is exactly what you (ab)used
> > C macros for in the C code.
> 
> I don't understand this.  In 'blend2.ml' (which I was responsible for)
> C macros are used to inline all the OCaml functions the same as in the
> C version.  Yet it's still 70% slower than the C version.
> 
> My suspicion was that it was to do with his use of a string as a byte
> array.

I get a 30% speedup by unrolling the BLEND loop and performing some additional
CSE (just getting rid of that extra (j_+3), actually). 

I think the major culprit is 31-bit integer arithmetic: there are lots of 
"orl $1, ...", "sarl $1, ...", decl and incl in the generated code.

If Bigarray.Array1 had unsafe_get and unsafe_set, operating on int32 values
could be faster (ocamlopt is often smart enough to do without boxed int32s).
Currently, bigarray bound checking is performed even with -unsafe, however.

-- 
Mauricio Fernandez  -   http://eigenclass.org