Wrench distribution QP computation times

[stephane-caron]

We've been comparing the performance of LSSOL with QuadProg in the topic/NoLSSOL branch for the wrench distribution quadratic program (WDQP) of the stabilizer. Currently QuadProg solves it much faster:

On average QuadProg runs around 0.02 ms, vers 025 ms for LSSOL.

[stephane-caron]

There is a short walking phase (warmup plan) in the plot above, but it is not so clear. Here is a walking plan (forward_30cm_steps):

The difference between DSP and SSP problem solving times is quite clear.

[stephane-caron]

@aescande Computation times are the same with and without custom QR decomposition for QuadProg:

At the last DSP of this example, the "no QR" (i.e. Q = A.transpose() * A) solution is roughly 2x closer (than the "QR" one) to the one found by LSSOL:

[aescande]

In the comparison QuadProgQR vs QuadProdNoQR, are you measuring the time spent to form Q for the NoQR case ?

How come that the earlier timing you did between the controllers based on LSSOL and QuadProgNoQR were very close ?

[stephane-caron]

In the comparison QuadProgQR vs QuadProdNoQR, are you measuring the time spent to form Q for the NoQR case ?

Yes, you can check the instructions between startTime and endTime in both case the QR and NoQR cases. Actually, the NoQR case has a slight disadvantage here as it also includes the computation of c = -A^T b 😉

How come that the earlier timing you did between the controllers based on LSSOL and QuadProgNoQR were very close ?

I didn't commit this intermediate version so I can't tell you for sure. In the commit history timings start after we switched to the QR decomposition.

[stephane-caron]

Actually, the NoQR case has a slight disadvantage here as it also includes the computation of c = -A^T b

Removing this bias the NoQR version is actually faster:

[stephane-caron]

We investigated the performance of LSSOL deeper with @aescande, using settings from Tasks to improve its solve times, notably:

  lssol_.warm(true);
  lssol_.persistence(true);
  lssol_.forceMinSize(false);

This brought the number of iterations from 10-40 to 1-4, so that the average solve time drops from 200-400 µs to 50 µs. Incidentally, these are roughly the same solve times than Tasks' IK. Still, much slower than QuadProg:

Roughly, QuadProg is 6x faster than LSSOL on the double-support QP and 20x faster on the single-support QP.