search

syzygy1 / Cfish

C port of Stockfish
GNU General Public License v3.0
137 stars 59 forks source link

Just AVX2 test #116

Open JavaMast opened 4 years ago

JavaMast commented 4 years ago

CPU Ryzen 3900X

Screenshot_651

syzygy1 commented 4 years ago

Thanks! So avx2 move generation doesn't seem to beat regular magic move generation on Zen, and compiling specifically for Zen does seem to help a bit (with -march=znver1 seemingly outperforming -march=znver2).

JavaMast commented 4 years ago

Just for information. BMI2_FANCY is much slower than BMI2_PLAIN on Ryzen 3900X

Screenshot_90 Screenshot_91 Screenshot_94

JavaMast commented 4 years ago

Looks like AVX2_BITBOARD is a bit faster with NNUE=Pure

Screenshot_95 Screenshot_96 Screenshot_97

syzygy1 commented 4 years ago

That AVX2 move generation is doing well in pure NNUE mode is very interesting. I should look into this to try to understand why.

That BMI2_FANCY does poorly on Ryzen is not so strange, as it uses the pdep/pext instructions even more.