Re: [boost] sorting floats by casting to integer

Subject: Re: [boost] sorting floats by casting to integer
From: Steven Ross (spreadsort_at_[hidden])
Date: 2009-07-05 10:41:16

On Sun, Jul 5, 2009 at 1:01 AM, Vladimir Prus <vladimir_at_[hidden]>wrote:

> Steven Ross wrote:
> >
> > If anybody would like to look at my testcase, it is in the boost vault,
> > inside algorithm_sorting.zip.  Inside there, it is in
> > libs/algorithm/sorting/samples/float_as_integer.cpp.
>
> This version fails to compile for me, because it includes
> boost/static_warning.hpp,
> and it does not seem to be present in my SVN HEAD tree. I've changed
> include to
> boost/serialization/static_warning.hpp, but this can't be right. Am I
> missing
> something?

I downloaded boost_1_38_0, and I have boost/static_warning.hpp in my tree.
If it would make it easier for others to compile, I'll shift the include to
serialization.

>
> To bet bigger times that are less prone to noise, I have changed loopCount
> in your test to 10. Then, then time reported without -std is 1 second,
> with very small difference between each run. The time with -std is
> 43 seconds, again with negligible variation.
>
> If I add:
>
>    cxxflags="-march=nocona -mfpmath=sse"
>
> then -std runs in 16 seconds, while the time for your algorithm is the same
> 1
> second. Of course, this is still significant speedup, but smaller than
> number
> you have reported, so probably a deeper look is reasonable.

Thanks for the suggestion Vladimir, that's a substantial difference, and I
will investigate.
I also appreciate the verification that it's not just my system doing this,
and I don't expect a 120X speedup across all tests and systems.

> Incidentally, does your library allow to sort only a collection of
> integers/floats,
> or it also allows to sort an arbitrary collection where each item has
> integer/float
> key? If so, can you point me at example?
>
> Yes.
samples/keyplusdatasample.cpp for integer keys + string data.  I see a 3X
improvement for this test, much more than I get for just sorting integers on
their own.
samples/floatfunctorsort.cpp shows the functors you need to pass to
float_sort, but the test just uses a float; you should be able to just
replace DATATYPE and update the functors.  I'll modify it to use key + data
so there is an example of such usage.
I will note that floatfunctorsort.cpp does not use my new copying approach,
so it only gets around a 7X speedup, and possibly less with the
optimizations you applied.