nanoflann orders of magnitude slower when using large samples about nanoflann

[Nelthirion]

And also when I looked inside the constructor for mat_index I saw that we have a call to index->buildIndex(); so why do we have to explicitly call this again in the next line:

my_kd_tree_t   mat_index(dim /*dim*/, samples, 10 /* max leaf */ ); // This already calls `index->buildIndex()`
mat_index.index->buildIndex();

[jlblancoc]

But When using 1e7 samples which is 10 times larger, the tree takes a LOT more time to build and also to index.

No, I think you didn't do anything wrong. But you might have hit a performance limit related to the size of your L3 cache memory.
However, I would not recommend the vector_of_vector container for very large datasets, since it involves N calls to the memory allocator (once per vector). Instead, use the Matrix adaptor or even better write your own wrapper to avoid data duplication, directly accessing to data in whatever form it's originally kept in your program.

And also when I looked inside the constructor for mat_index I saw that we
have a call to index->buildIndex(); so why do we have to explicitly call this again in the next line:

Good point. Please, feel free of opening a pull request to fix the code/example. I guess that keeping the call inside the ctor is the expected behavior by the user, so the second one is the one that needs to be removed.

[jlblancoc]

PS: An efficient kind of Matrix for this case is not Eigen::MatrixXd, but a matrix with one dimension (e.g. number of cols) fixed in compile-time to 128 or whatever is the feature length.