Had odd performance using asDiagonal at https://github.com/mtao/core/blob/master/include/mtao/eigen/mic0_preconditioner.hpp#L160 so I switched to cwiseProduct. They certainly have deferring performance, especially with and without O2. A sheet looking at the difference is here: https://docs.google.com/spreadsheets/d/1IYLwVsEOQORsDH2eoJXpEP8_oWe8p…

cwiseProduct_vs_asDiagonal

#include <Eigen/Dense>
#include <chrono>
#include <iostream>



template <typename Func>
int run(int size, const std::string& name, Func&& f, bool print=false) {
    Eigen::VectorXd D = Eigen::VectorXd::Random(size);
    Eigen::VectorXd x = Eigen::VectorXd::Random(size);

    using clock_type = std::chrono::steady_clock;

    auto start = clock_type::now();
    for(int i = 0; i < 1e4; ++i) {
        x.noalias() = f(D,x);
    }
    auto end = clock_type::now();
    auto duration = end - start;
    auto dms = std::chrono::duration_cast<std::chrono::milliseconds>(duration);
    if(print) {
        std::cerr << name << "(" << size << "): " << dms.count() << "ms" << std::endl;
    }
    return dms.count();
}

int main() {

    for(int i = 4; i < 40; ++i) {
        int size = std::pow(2,i);
        int ams = run(size, "asDiagonal", [](const Eigen::VectorXd& D, const Eigen::VectorXd& x) {
                return D.asDiagonal() * x;
                }, true);
        int bms = run(size, "cwiseProduct",[](const Eigen::VectorXd& D, const Eigen::VectorXd& x) {
                return D.cwiseProduct(x);
                }, true);
        std::cout << size << ", " << ams << ", " << bms << std::endl;
    }
}