Timing output interpretation

User Questions
1

Hi guys,

first of all thank you very much for sharing such a useful tool!

I wanted to ask you how I should interpret the timing output. On one example I got:
time_tot: 0.00452284 # total cpu time previous call
time_lin: 0.000916956 # cpu time for linearization
time_sim: 0.0007608240000000002 # cpu time for integrator
time_sim_ad: 0.0007382460000000001 # cpu time for integrator contribution of external function calls
time_sim_la: 0.0 # cpu time for integrator contribution of linear algebra
time_qp: 0.003321902 # cpu time qp solution
time_qp_solver_call: 0.003229635 # cpu time inside qp solver (without converting the QP)
time_qp_xcond: 8.934899999999999e-05
time_reg: 2.93e-07 # cpu time regularization

Since many of these times are not even approximately the sum of the others, I am not sure what they refer to. I also reported some comments I found in the python interface. However, what is the difference between the integrator and linearization time?

Does linearization mean integrator + system dynamics linearization + cost linearization + constraint linearization + Lagrangian Hessian computation?

Does time_sim_ad mean integrator minus the system dynamics linearization (and, eventually, Hessian computation)?

2

Hi Mario,

Linearization includes linearization of dynamics (discretized through integrator), cost and constraint.

Yes, all operations to compute the QP matrices.

time_sim_ad is the time that is spent in external (typically CasADi generated) functions (AD - algorithmic differentiation).
time_sim_la is the time spent in Blasfeo linear algebra within the integrator.
It is 0 in your example, probably because you use ERK.

I hope that helps,
Best,
Jonathan

3

Awesome! Thanks a lot Jonathan!

One last thing, I have time_reg >0 (though as you see it’s very small) even though I select the option
ocp.solver_options.regularize_method = ‘NO_REGULARIZE’
is that expected?

4

Great!

I just had a look and the time_reg is taken around the call to the regularization module and not inside the module.
So, the ~1e-7 should be the overhead of the empty function + the overhead of the timer.