I have the following unexpected 2X slow down, when I am using Eigen's column expression instead of manual looping through each element of the column. I am using clang on Mac with compile flags -O3 -DNDEBUG -march=native. I am using the latest trunk version of Eigen.

MethodA: 1.89s
MethodB: 0.83s

Code: Select all

Eigen::MatrixXf compute(const Eigen::MatrixXf& in, const Eigen::MatrixXi& offset, const Eigen::MatrixXf& weight) {
  assert(offset.size() == weight.size());
  Eigen::MatrixXf values(in.rows(), 2000);
  for (int j = 0; j < in.cols(); ++j) {
   
    for (int i = 0; i < offset.rows(); ++i) {
      int o = offset(i, j);
      float w = weight(i, j);

       // MethodA: Eigen Col Expression
       values.col(o) += w * in.col(j) ;

      // MethodB: Raw Loops
      //for (int k = 0; k < in.rows(); ++k)
      //   values(k, o) += w * in(k, j);
    }
  }
  return values;
}

The full code to reproduce the problem is https://gist.github.com/anonymous/29560512fb748bcbd706

I guess methodA is slower due to the vector resizing.

tienhung wrote:I guess methodA is slower due to the vector resizing.

Hi Tienhung, what do you mean by resizing?

hijokpayne wrote:

tienhung wrote:I guess methodA is slower due to the vector resizing.

Hi Tienhung, what do you mean by resizing?

in.col(j) and values.col(o) return column vectors, so their size must be verified before the addition and assignment. Eigen resizes values.col(o) on the left hand side.
You can refer the link below for more information.
http://eigen.tuxfamily.org/dox/group__TutorialMatrixClass.html

tienhung wrote: in.col(j) and values.col(o) return column vectors, so their size must be verified before the addition and assignment. Eigen resizes values.col(o) on the left hand side.
You can refer the link below for more information.
http://eigen.tuxfamily.org/dox/group__TutorialMatrixClass.html

I still don't understand why there should be resizing? Both values and in has same number of rows. Also since I am using non-debug mode, I expect no overhead of size check assertions.

Hi,

Did you figure out why method 1 is more time consuming, and how do you optimize the code when assign one matrix to another?

Thanks.

When I compiled the same with gcc, the differences go away. So the problem is on Clang (I am on Mac OSX).