使用Intel C ++编译器的Armadillo错误

Question

I have been using Armadillo 4.2 version (from linux mint package manager) for quite a some time. That version had no problem when compiled with icc, and codes compiled with icc ran very fast (about 10%-15% fast compared to g++).

Recently I have upgraded Armadillo to recent 6.1 version (from the official website). But in this case, all the codes compiled correctly with icc, but gave runtime error.

error: eig_sym(): failed to converge

terminate called after throwing an instance of 'std::runtime_error'
what():  eig_sym(): failed to converge
Aborted

Codes compiled correctly with g++ ran perfectly. I have also tried previous 5.6 version, but the same problem. Any solutions?? I had to revert back to old 4.2 version to get that extra speed from icc...

Answer 1

The Intel C++ compiler is known to contain of bugs when dealing with heavily templated libraries like Armadillo (see for example here and here ).

You can increase the performance of gcc compiled code by enabling auto-vectorization (using -O3 ) and using the native machine instruction set (using -march=native ). For example:

g++ code.cpp -o code -O3 -march=native -larmadillo

Also make sure that the matrix you're passing to eig_sym() is actually symmetric.