Hi

In my particular problem i am working with some very sparse constant 6x6 matrices, which i am multiplying with a vector at runtime - as fast as possible. The fastest approach would be to manually unroll the multiplication and manually remove the multiplications by zero, however i am also trying to achieve readability.

My question is, is there a way to encourage compilers to unroll the such an expression and remove the multiplication of zero constants? Or are my only options manual unrolling or some meta programming approach?. So far i haven't had any success with any compiler optimizations (ie the density of the constant matrix makes no difference to speed).

Thanks for any assistance.

Honestly, if your 6x6 matrices are really very sparse, maybe they have such a special form that you can reinterpret the matrix products in a completly different way. An example is permutation matrices...

Otherwise, if you really think you have to do it, you can do the following. First, play with EIGEN_UNROLLING_LIMIT to make sure that your operations are completely unrolled. For matrix-vector product, the default should be OK.
Then write your own "very sparse matrix expression". An expression is a class inheriting Eigen::MatrixBase. For an example, you can have a look at the CwiseNullaryOp expression. The crucial part will be the coeff() method. For example, for a matrix that is zero everywhere except that m(2,3)=1, your coeff() method would be:

Code: Select all

EIGEN_STRONG_INLINE float MyExpression::coeff(int row,int col) const
{
  if(row == 2 && col == 3) return 1;
  else return 0;
}


(the EIGEN_STRONG_INLINE here is because if you want this to evaluate at compile-time, you've got to make sure that this function is inlined)

The idea is that if everything gets completely unrolled, then coeff(row,col) will only get called with values of row and col that are known at compile-time. Then if it is inlined, the compiler will be able to resolve the if's at compile-time. You'll need a good compiler here.

Thank you for your help bjacob, i will give that a try.