more precise version of clock() function

[stanlvw]

I have tried using clock() [in time.h] to time function execution time, but getting 0 out of it b/c it may not be precise enough. Is there a clock function that has a higher precision than clock()? Thanks.

[dwks]

With standard C/C++, there's no real way.

But if you're on Windows: http://cboard.cprogramming.com/showp...7&postcount=27

And there are ways to do this on other platforms as well. Try googling for more information.

FWIW, you can often just run your code in a loop say 1000 times, and then it will take long enough for clock() to measure its runtime.

[stanlvw]

I saw in the link that you have previously written. I don't want to purposely add delays to the functions in order to find out the performance. I'm also investigating the use of a profiler. Would it be compiler specific in terms of that information (I'm writing mex-file generating C++ code that is used on Matlab)?

[dwks]

I'm not quite sure what you're asking. Profiling information is definitely compiler-specific, and run-specific as well. But it's still very useful nevertheless; profiling can tell you the bottlenecks in your code.

If you're compiling with gcc, you can use its profiler, gprof. cpwiki.sf.net/gprof

What exactly are you trying to do? Figure out what is slowing down your program so that you can improve its perfomance?

[stanlvw]

I'm compiling with Microsoft Visual Studio 2005 compiler in Matlab environment. I'm trying to figure out which function I have in a project that is running the slowest so that I can start improving the performance by spending the effort optimizing that function.

[dwks]

Then a profiler is exactly what you need.

Unfortunately I've never done any profiling with Visual Studio, so I can't really help you there. I can only suggest google, and this site, which looks like it might be useful.

http://blogs.interknowlogy.com/adamc.../08/22470.aspx

[stanlvw]

The thing is I'm not sure whether I should be using a profiler in Visual Studio 2005 or Matlab. I have trouble finding profiler information that are helpful with google before I asked the question on this forum already. Anyhow, I guess I can keep searching and asking for help on finding the information until I find what I'm looking for. Thanks anyway.

[matsp]

A sampling profiler (oprofile for Linux, or VTune(Intel)/CodeAnalyst(AMD) for Windows) is often better than gprof that instrument the code - the reason being that the instrumentation adds to the run-time, and doesn't necessarily show the actual time spent in each function.

The sampling profiler will sample the program counter (where in the code you are) on each timer-tick, which usually gives a good picture of where your code is spending most of it's time.

Another option is of course to use clock() over a bigger function. But instrumenting the code to run a 1000 (or 1000000) loops of something you want to see how fast it is, is often a good plan that works well - and then clock() is good enough.

There are also other higher precision timing methods - TSC is one of the most precise ones, as it gives you a clock-cycle accurate time [2GHz clock on a 2GHz processor, for example]. http://en.wikipedia.org/wiki/Time_Stamp_Counter
However, there are some small problems, particularly if your machine has multiple processors - you may want to force your thread to run on a single processor, e.g. SetThreadAffinityMask.

--
Mats