Open kovacs-levent opened 3 years ago
I think the implementation definitely needs SIMD integrated to have good performance. At this point for some reason the multi-threaded version gets slow on large inputs, but it's definitely faster than the single-threaded, the bottleneck seems to be the recursive part, where we could integrate SIMD.
The prototype at this point is bare-boned and doesn't use any implementation tricks to decrease runtime.
Things to check: