Avoid vector copy in NTT call
Perform in-place NTT in SwitchFormat()
, yielding ~1% slowdown in SwitchFormat()
.
Running:
OMP_NUM_THREADS=1 ./bin/benchmark/poly-benchmark-16k --benchmark_filter="DCRT_ntt" --benchmark_min_time=10
yields
Before:
DCRT_ntt/towers:1 1375 us 1374 us 10186
DCRT_ntt/towers:2 2761 us 2748 us 5096
DCRT_ntt/towers:4 5521 us 5509 us 2540
DCRT_ntt/towers:8 11043 us 11031 us 1268
After:
DCRT_ntt/towers:1 1394 us 1391 us 10063
DCRT_ntt/towers:2 2782 us 2780 us 5036
DCRT_ntt/towers:4 5593 us 5568 us 2515
DCRT_ntt/towers:8 11181 us 11149 us 1255
Edited by Fabian Boemer