Sameer Agarwal

and actually the slowdown is for two reasons. 1. We cache the local parameterization jacobians so they do not have to be computed over and over again. 2. And generally...

I would recommend the second suggestion before the first, since that would require you to implement an efficent MultibyByJacobian routine, whereas you should be able to make a oneline change...

could you try the two options and let me know which works better? The reason to not do it always is that if you have a lot of residual blocks...

I suspect that suggestion 2 will give you a substantial performance boost.

@dellaert so you have some truly large residual blocks, both in height and in width and they are breaking some of the assumptions we have made. Size 64 parameter blocks...

Let me think about this a bit more. On Mon, Aug 10, 2020, 6:13 PM Frank Dellaert wrote: > I found a *bit* of special structure, but it does not...

Interesting idea. One question for you. What sized problem are you solving? How many rotations and how many pairwise terms ? On Mon, Aug 10, 2020, 11:34 PM Frank Dellaert...

The reason I ask is that got most problem s the linear solver takes way longer than evaluation. Is there a Summary::FullReport that you can share? On Tue, Aug 11,...

If I am reading this right, then the situation is worse for n = 3 than n = 8, where the jacobian evaluation is slower than the linear solve. BTW...

I do not think I quite understand your suggestion @dellaert. Could you elaborate? and what are these plans?