On Sunday, 26 May 2013 at 20:49:36 UTC, Dmitry Olshansky wrote:
It's the kind of thing that is tremendously hard to measure accurately since it depends on the workload, architecture and the time spent is very small. So don't take it by word I'm almost certain that something is amiss (compiler switches and whatnot).
For such cases, I found Agner's benchmarking utilities to be very useful. They print exact CPU statistics, such as numbers of micro-ops, cache misses, mispredicted branches, etc. I've used them very successfully when tuning my appender implementation.
To use them with D, I modified his C++ program to load a DLL and call a function, taking the DLL and function names from the command line.
Original program: http://www.agner.org/optimize/testp.zip My patch (to load a DLL): http://dump.thecybershadow.net/5f55e8be5f8cd38ad60f218957ef24bb/PMCTestB.diff Usage example (sort of): https://github.com/CyberShadow/DAppenderResearch/blob/master/go-dll.bat Hope this helps :)
