No copy multiplication in Eigen

Question

I'm using Eigen with big matrices and I'm thinking about ways to optimize my code, focusing on reducing dynamic memory allocation.

I'm trying to multiply two matrices. Those matrices change a little bit every now and then, but their sizes stay the same.

I'd like to see the output of the multiplication going to a predefined matrix (that would have memory already allocated for it).

So here's an example of what I'm doing:

Eigen::MatrixXd   left, right,result;
// ... 
result = left * right;
// ... left and right values change a little
result = left * right;

And I'm looking for a solution that would be like that:

  void Multiply(
    Eigen::MatrixXd const&  left,
    Eigen::MatrixXd const&  right,
    Eigen::MatrixXd&        result);

  void Example()
  {
    Eigen::MatrixXd   left, right, result;
    // ... 
    Multiply(left, right, result);
    // ... 
    Multiply(left, right, result);
  }

The aim is basically to reuse the result matrix memory because in theory it should not change dimension. I was thinking about using operator*= but I kind of realize that it still needs an intermediate matrix to do the calculation.

Did you check that result=left*right actually reallocates when result already has the right size? — Marc Glisse

Jitse Niesen Jitse Niesen · Accepted Answer · 2014-06-17T10:01:48

result = left * right allocates a temporary matrix to hold the result of the multiplication, evaluates the product into the temporary matrix, and then copies the result from the temporary matrix to result. This is to deal with statements like A = A * B where a temporary matrix is needed.

If you know that the result is different from the matrices in the product, then you can write result.noalias() = left * right. Eigen will not use a temporary matrix in this case.

More explanation on aliasing in Eigen is at http://eigen.tuxfamily.org/dox/group__TopicAliasing.html

No copy multiplication in Eigen

1 Answers