Boost logo

Boost :

From: Sam Schetterer (samthecppman_at_[hidden])
Date: 2007-03-19 20:19:25

Lewis Hyatt wrote:

> Are you sure about the quicksort speed difference? That seems like a huge
> difference. Did you have inlining enabled? Also, the current versions are
> not the final forms. I am implementing random-access iterators and
> user-provided functions.

Yes, this was on cygwin compiled with g++ -O3. One thing to keep in mind is
your routines are competing with the standard library, which can be highly
optimized for whatever target platform the compiler was written for. Unless
have a dramatically improved algorithm, it is unlikely that your sorting
routines will beat std::sort in general, even if they do on some platforms.

As an example, suppose your goal was to write an equivalent of std::bitset
beat the standard implementation. It is unlikely that you could exceed the
performance of std::bitset::count(), for instance, because most
use native assembly instructions to do this whenever possible. (e.g., gcc
its built-in popcount function on platforms where it is available.) On the
hand, if you hand-coded a version with the best algorithm you could find, it
might beat a few poor implementations, but in general it would be inferior.
only way to make it better in general would be to implement all these
optimizations yourself for each target platform, but at that point, it's
not worth the time.

I don't think it's true that std::sort, std::stable_sort, and
are the end of the story, but I think any work in a sorting library should
on needs which are not already met. So far, only radix sort appears to fall
that category, but you would also need to demonstrate a case in which radix
is preferable to std::sort.


I think that you are right that each compiler's version of the stl has the
advantage of assembly language. Since I know both that speed and features
are important to developers, what would people think if I uploaded to vault
x86 assembly language extensions to the algorithm for users to download. The
whole algorithm will not be coded in assembly language(only the tight inner
loop), but it could only be used on built in types. Would you use the
extensions if they were supplied?

Boost list run by bdawes at, gregod at, cpdaniel at, john at