Merge local summaries #5
Description
Right now, the performance bottleneck is actually putting the rankings into the tree form. This is because the summary is in shared memory and we have to use mutexes to prevent threads from corrupting things.
If each thread had their own summary, and these summaries were merged, there could be as much as a 2x speed increase on 20-core hardware. The speedup would be less pronounced at lower core counts (about 1.33x for 8-core, for instance).