STACK OVERFLOW 50

[mcmingchang]

The code used compute_80 and sm_80 did not encounter any issues, but the above error occurred under default parameters.
The problem lies in calling the function：

template<typename CandidateList,
typename data_t,
typename data_traits>
inline device
float knn(CandidateList &result,
const SpatialKDTree<data_t,data_traits> &tree,
typename data_traits::point_t queryPoint)

[mcmingchang]

The reason is that the construction of the KD tree failed，tree.numPrims=10000，but tree.numNodes=2

[ingowald]

Leaving for vacation right now, can't look at that til weekend. Can u send me google drive link with file containing these 10k points to reproduce?

[ingowald]

Hey - any change i could get some more info and/or data to test with?
What's happening here is pretty clear - it cannot find any way of splitting your input points into two separate halves, and then creates a single huge leaf - but the question is why: I currently see these options:
a) your input data is 10k copies of the same identical point, in which case the bounding box collases to a single point, and the builder can't find any plane to split.
b) maybe your input data contains some coordinates that are messing up my code (NANs or INFs?) that are messing up the way i compute the bounding box,
or
c) something's wrong with how cuda treats the atomic min/max'es in the bounding box computation.

given that you explicitly mention "compute_80" vs "default parameters" i kind of assume it's the latter, but some more info would be very useful. Also, please provide the cmdline flags you used to compile.

[ingowald]

BTW: i'd have to double-check, but I think that starting with either compute_75 or compute_80 architectures you can no longer run code that was explicitly compiled for earlier archs; in particular, not if your code was explicitly compiled for compute_50, which unfortunately is what some older cmake archs will default to if arch isn't set.