Steven Ross wrote:
On Thu, Jul 3, 2008 at 9:36 AM, Phil Endecott < spam_from_boost_dev@chezphil.org> wrote:
I'll have a look at that. What do we think about "magic tuning parameters" in algorithms? My feeling is that things dependent on e.g. cache sizes (or number of processors etc) really need to be determined at run time to be useful to most people.
The values I picked are pretty good across most systems. If there is a standard, quick way to find the size of the processor cache, then I could consider using it. By far the most important parameter MAX_SPLITS is dependent on the page size, and 9 corresponds to the number of bins that take up 4096 bytes on a 32-bit system, which is a common page size. introsort looks like it has a tuning parameter of 16 in SGI's implementation.
It would be really good to get an ATLAS-like tuning framework into Boost so people could configure these things properly at installation time. -- Dave Abrahams BoostPro Computing http://www.boostpro.com