Latent Seminar

Mex Coding III

2 min read

MEX PARALLEL

As we talked about, shared memory system will benefit some speedup with parallel computing. MPI is unacceptable for most cases, there are too much overheads.

  • pthread

Low level programming, use extensive mutex for multithreading.

  • OpenMP

Ready to use and easy to integrate. Remember to add -fopenmp during comp

ACCESS DATA

All data exchange has to be done via its internal format mxArray* or mxArray-tag, it holds all information of the data, such as class, size, address, real, complex and so on. Matlab provides almost all get methods which are defined in matrix.h. Once there was a headerdump on fileexchange to find out exact data of object in Matlab, but the format of mxArray changes along versions, it does not work anymore unless modifying it.

PROFILING

tic/toc is not exact, using profiler would be better. Put clock() in code will also work.

MEX VS. LOADLIBRARY

I sometimes experienced hard time with porting codes into MEX domain. It is quite annoying to encounter various problems with internal problem with compiling codes against MEX. However, if I have my ready-to-use codes already built with some library, it is much easier to wrap all things I need to do in pure C/C++ and use loadlibrary as a workaround, porting time could be a loss but also could be negligible in most cases.

Problematic -lmwblas and -lmwlapack

I had a tough time on compiling pardiso interface with Matlab, the manual solution provides me linking options as Matlab uses. However, pardiso has been upgraded to version 5.0.0, it no longer supports the old fashioned codes, and could produce errors like MKL Error.

If compiling pardiso with MEX support, using -lblas and -llapack is fine with compiling process, but would get you crash with Matlab. To overcome the problem:

  • preload lapack before starting Matlab.

is the only solution I have found. Be careful, lapack preloading is the only thing needed, blas is not very important here. The side effect is other functions depending on lapack maybe get lagged. Especially eig, lmwlapack version is ~2x faster than apt-get version of lapack.

Graph View