Profiling a mex-function

Question

I have just rewritten a Matlab program in c++ as a mex-function to speed things up, with fantastic results. This optimization decision was a very very good idea, with up to a factor 20 speed up without threading. It still left me curious about what the mex-function was spending time on and wanting to identify possible bottlenecks.

I'm looking for a way to profile mex-functions. The matlab profiler is not much use, and the other profilers I've downloaded (both free and trial) all want an executable to run. I'm no mex-guru, but as far as I've understood the only way to run a mex is from within Matlab. The mex-function is compiled into a dll, but is called .mex64. So this problem should be similar to profiling a dll. To write the c++ mex-function I used a single-user VS2005 (i.e., not the team version), and am running on a x64-platform.

Does anyone know of a good way to profile a mex-function? What tool should I use and how do I use it when I start from within Matlab? Or is there any other way to profile the c++-code?

Answer 1

the only way i've managed to do this is to separate out the function doing the work and writing a separate wrapper (instead of mexFunction) that loads .mat files with test data and runs as a standalone executable. this can then be profiled using e.g. gprof

Answer 2

Is there a way to run the whole thing under a VC IDE. Like you could say "debug DLL Foo.dll, using Matlab as the startup app".

Then, if the function Bar you're using in Foo.dll runs in < 1 sec, make sure Matlab calls it a lot of times, or add a wrapper function in the DLL to call it a lot of times.

If you can get to that point, you can use the manual call stack sampling technique, that really works in spite of not being popular or requiring installation of a tool.

Answer 3

Two possible solutions in addition to the fine ones already suggested.

1. It's possible to write a script which loads the data, runs your mex function, and then exits. You can then let the profiler call the Matlab executable with that script. This worked for me with nVidia's visual profiler when I was profiling my CUDA mex functions.

2. If you built the mex function in Visual Studio, you should be able to do attach to the Matlab process and then run the function for either debugging or profiling.