Reworking PointSetOperatorBC
Currently the PointSetOperatorBC passes the whole set of training points (including those not in the point set defined in the BC) to the operator, then using a mask to filter out all those not corresponding to values in the point set. This results in far more computation than is necessary for complicated operators and difficulty when defining operators which are only well behaved in a specific point set. I think it's possible to do the mask before passing the arguments to the operator, which would fix the issue. I will open a fork to try this myself but I chose to open an issue too anyways in case there's anything I missed or should be aware about when working on this. Cheers!
Here is on PR https://github.com/lululxvi/deepxde/pull/2038. Not sure if this is what you mean.