I am hoping to load a program which uses Eigen onto a microcontroller with limited memory. I am only using the core, LU, and Cholsky headers, which puts the binary at about 400kb without using any compiler optimization; about 200k with -Os. I don't include any other libs. might anyone have any general advice for further reduction of static code size?

I think that enabling inlining (-O2) is also very important to help the compiler removing boilerplate code.

If your matrices are not very large, you might also compile with:

-D__CUDACC__ -D__host__=" " -D__device__=" "

(using the devel branch). This will disable the heavy matrix product code. The following example:

Code: Select all

#include <Eigen/LU>
#include <Eigen/Cholesky>

using namespace Eigen;
int main() {
  MatrixXf M;
  VectorXf b, x;
  M.setRandom(100,100);
  b.setRandom(100);
  x = PartialPivLU<MatrixXf>(M).solve(b);
  x += LLT<MatrixXf>(M).solve(b);
  return int(x(0));
}


takes 47k with the devel branch and:

Code: Select all

clang++ -Os -O2 -D__CUDACC__ -D__host__=" " -D__device__=" "


and 65k without the "CUDA" trick and Eigen 3.2.