Any suggestion for setting params for float?

[whtqh]

I am testing the time cost for proxqp, and I wonder if the time could be less using T = float compared to T = double.
However, using the same setting will make the qp unstable and unsolved.
I guess the default setting params in proxqp/setting.hpp is tuned for double.
Is there any suggestion for tuning and the most important params for time cost?

btw, I have adjusted the eps_abs and eps_rel

// float -> 0.0035
problem_.settings.eps_abs = sqrt(std::numeric_limits<T>::epsilon()) * 10;
problem_.settings.eps_rel = sqrt(std::numeric_limits<T>::epsilon()) * 10;

[whtqh]

https://inria.hal.science/hal-03683733/file/Yet_another_QP_solver_for_robotics_and_beyond.pdf

I found the paper very informative, and I'm eager to learn more tuning techniques. Could you please provide me with additional tips and strategies?

[whtqh]

I use the params below, and the abnormal time cost will not appear in my own practice.

    // problem_.settings.mu_min_eq = 1e-4;
    // problem_.settings.mu_min_in = 1e-3;
    // problem_.settings.mu_max_eq_inv = 1e4;
    // problem_.settings.mu_max_in_inv = 1e3;
    // problem_.settings.eps_refact = 1e-3;

    // problem_.settings.max_iter = 500;
    // problem_.settings.max_iter_in = 100;
    // problem_.settings.default_rho = 1e-2;

However, the time cost of float does not decrease compared to double.
I think the float type should have some adv on Eigen SSE speed-up?