--- Comment #13 from David Simcha <> 2011-05-15 12:15:45 PDT 
(In reply to comment #12)
> Probably we need two more reduce primitives, both of which assume
> associativity. One uses ILP and simple loop unrolling whereas the other uses
> full-blown threads. There are definitely needs for each.

Right.  We already have the full blown threads one in std.parallelism.  What
I'm saying is that, if you're using threads, you're already assuming
associativity.  Therefore, you may as well use ILP and loop unrolling, too. 
std.parallelism currently doesn't do this and once we work out the details and
create an assocReduce in std.algorithm, these techniques should be ported to
std.parallelism.reduce, too.

Configure issuemail:
------- You are receiving this mail because: -------

Reply via email to