Hi,

is there a way to optimize an expression like this:

Code: Select all

int gradient_cutoff = 15;
v1.noalias() = (m.block(i, j, a, b) * v2.segment(j, b)).array().max(-gradient_cutoff).min(gradient_cutoff).matrix();


NB: I need to clamp the range of the matrixXvector product in my NN training code...

thanks
Stefan

yes, split it in two:

Code: Select all

v1.noalias() = m.block(i, j, a, b) * v2.segment(j, b);
v1 = v1.array().max(-gradient_cutoff).min(gradient_cutoff);


(Eigen 3.3 should be able to do exactly that for you)

ggael wrote:yes, split it in two:

Code: Select all

v1.noalias() = m.block(i, j, a, b) * v2.segment(j, b);
v1 = v1.array().max(-gradient_cutoff).min(gradient_cutoff);


(Eigen 3.3 should be able to do exactly that for you)

Thanks again, ggael, it seems to work. However, this will still force me to use temporaries anyways for expressions like that, right?

Code: Select all

v1.noalias() += m.block(...)*v2.segment(...).array.max(...).min(....);


Is there already a rough schedule for Eigen 3.3?

thanks a lot
Stefan

yes because this will be the most efficient solution. However, if the matrix m is pretty small or if you prefer to sacrifie performance for no temp, you can use A.lazyProduct(B) instead of operator*.