qpOASES zeroes an array when it's not necessary; wastes time about acados HOT 6 CLOSED

Next suspects for improvement are the many d_free and d_zeros calls in ocp_qp_qpoases_condensing. Rewrite after condensing interface is completed.

from acados.

Hey Robin. I would be reluctant to change qpoases without a deeper understanding of the code. Skipping Initialization to zeros might work for some operating systems and/or problem instances but not for others. Perhaps you can ping Joachim about this. Of course all the mallocs from the acados side should be gone once the new interface is in place.

from acados.

from acados.

Alright! By the way, I suspect that malloc is much slower in mac than in linux. At least on my system. Ps: you can have a shot at Greg's server for the benchmarking.

from acados.

The connection with Greg's box is too slow to do that from here. Could someone just try this for me? It's basically

• commenting out that line,
• building qpOASES and
• running chen_nmpc_qpoases_example again and see if it changes anything (put number of repetitions 10000).

If it also occurs on Linux, I will then write to Joachim.

from acados.

I can confirm, roughly the same speedup (~25%) can be seen on a Linux box. Note that this probably occurs because NVMAX and NCMAX are their default values (i.e. way bigger than the actual size of the problem). For smaller maximum dimensions, this would probably not occur, as other routines will become relatively more CPU-intensive.

Still, these lines are unnecessary, it seems. I will contact Joachim about this, see what his opinion is.

from acados.