big matrix multiplication core dump

Question

The following program is trapping.

void main(){  
 fmat A,W,H;  
 W.load("w.csv"); //W is of size 150000x100  
 H.load("h.csv"); //H is of size 300000x100 
 A.set_size(W.n_rows,H.n_rows);  
 A.zeros();  
 A=W*H.t();  
}  

The above program is compiled using g++ 4.8 with fopenmp flag and run on ubuntu 64 bit with 384GB ram. I am using openblas. The W and H are any arbitrary random positive matrix.

The above code is trapping and creating core dump during multiplication. The code is successful till A.zeros(). I checked the ulimit and it is showing unlimited. I also tried to create A as sum of outer products of vectors of W and H. It is also trapping. Also, when W and H are SMALL the code is working and it NOT trapping.

How can I multiple two big matrices? Is there any size restriction?

Answer 1

@mtall's answer in the above comment is the answer. Enable the ARMA_64BIT_WORD in include/armadillo_bits/config.hpp. The include directory can be found in the place where you installed armadillo. For example in my case it was in /usr/local.