Currently, distance computation kernels are not perfectly load balanced. Most kernels follow this pattern: launch the number of threads on one operand and for each thread loop over points/segments of the correponding pair in the other. This can slow down the operation if the data are unbalanced/skewed.

Instead, the kernels should launch the number of thread that's makes one thread computes one pair of point-point/point-segment/segment-segment, then uses atomic operation to aggregate the result. This avoids slow down if the data is skewed.