Steven Watanabe wrote:
AMDG
David Abrahams wrote:
I think the optimizer can remove some redundant tests if the functions are inlined. But, that said, I am becoming more and more convinced as this discussion goes on that compressing stacked iterator adaptors is barking up the wrong tree if you care about efficiency. The difference between the code generated in the abstracted case and the code you'd write by hand is simply too great.
No, I don't have a picture of a better solution yet. But it's an interesting problem.
Here are some measurements for my current version of filter_iterator
In brief:
4 function objects containing two integers each: 3.39 seconds 4 stateless function objects: 2.813 seconds 4 stateless function objects combined using phoenix's && rather than stacked iterators: 2.563 seconds 1 stateless function object 2.015 seconds raw for loop using vector iterators 2.516 seconds raw for loop using pointers: 1.703 seconds
These were all doing the same calculation over the same underlying vector.
I ran your attached test on VC9 SP1 default console app release mode with _SECURE_SCL=0 on my Core 2 Quad Q6600 2.4Ghz with 4GB of RAM 50.702 seconds 19.984 seconds 48.453 seconds 13.343 seconds 4.313 seconds 2.953 seconds Press any key to continue . . . Are these all truly doing the same work? Thanks, Michael Marcin