While profiling suggests that most operations are cheap and IO-bound rather than CPU-bound, the GEMM for deduplication is pretty slow. As such, use multiple threads for higher throughput.
