I got rid of Point3Endpoint because we have no identified need for it and adding it requires additional code (which was not the case when raw data was transposed). However, for the record, here are some ideas for adding back later (or if nlust or others strongly prefer, on this ticket branch):
- Add a constructor to Point<T, N> that takes a column view of an ndarray. This is easy to use correctly and efficient, but does add another constructor to the API of Point<T, N>.
- Add a "strides" argument to the raw data pointer constructor that defaults to 1. This was jbosch's suggestion. it is very simple to implement and adds a minimum of API clutter, but I worry that it is too easy to get the value of strides wrong.
- Add two new classes: Point3Endpoint and template<N> BasePointEndpoint, to hold code common to Point2Endpoint and Point3Endpoint. This needs no changes to the Point classes but I feel it is a bit too messy.
- Copy the data for each point from the 2D raw array into a C array, then use the C array to construct the Point. krzys suggested this. It is very simple but requires an extra copy for each axis of each point. I worry about the performance hit, though it would surely be small and perhaps negligible.