Hi,

I'm thinking of porting some of my research code to Eigen. There are several parts of my program that need to perform element-by-element operations over matrices, but I'm anxious about using myMatrix(rowIndex,colIndex) because that's presumably recomputing positions on each invocation. I was thinking of defining EIGEN_DEFAULT_TO_ROW_MAJOR and then just getting pointers to the beginning of each row, but it seems that row() returns a copy rather than a reference. Can I do something like:

Code: Select all

MatrixXf mat(m, n);
for (int i = 0; i < m; i++) {
  RowReference theRow = mat.row(i);
  for (int j = 0; j < n; j++) {
    // do some processing on theRow
  }
}

What type should "RowReference" be? I want the changes to mirror back into "mat".

Cheers,
Alex

You can find this sort of information in the API documentation if you know where to look. The row() function is documented at http://eigen.tuxfamily.org/dox/classEig ... 8665c28813 . You'll see that it returns something of type RowXpr. This is a type member of MatrixXf, so in your code you would write the third line as "MatrixXf::RowXpr theRow = mat.row(i);" A RowXpr is basically a reference to the row, but once you assign it to a vector or a matrix it makes a copy.

However, depending on what you want to do, other solutions may be better. It sounds like the unaryExpr() function would be useful for you. Have a look at the example at http://eigen.tuxfamily.org/dox/classEig ... b0525ddd65 .

I'd go even farther by saying that in general you even do not need custom functors (unaryExpr/binaryExpr), but instead use Eigen's API to build complex and optimized expressions, e.g.:

(std::sin(m.row(i).array() + M_PI)).maxCoeff()

compiles to something like:

double* base = &m(i,0);
double res = sin(base[0] + M_PI);
for(int j=1; j<m.row(i).size();++j)
res = std::max(res, sin(base[j] + M_PI));

and in some cases this expression can even be unrolled and/or vectorized by Eigen for you!