Backport edax-reversi-AVX 4.5.1 optimizations those do not affect fingerprint (search node count) onto the v446mod2. Solves endgames 15-40% faster than the v446mod2.
Manually exported from https://github.com/okuhara/edax-reversi-AVX to local on 2018/09/10.
eval.dat.bz2 is from https://github.com/abulmo/edax-reversi/releases/download/v4.4/eval.7z .
Two commits, https://github.com/eukaryo/edax-reversi-AVX-v446mod2/commit/018a4f9b1d79ba441859f7bd8d5ec4ba52816855 and https://github.com/eukaryo/edax-reversi-AVX-v446mod2/commit/9132243769be4d80806243ae2c8904f44d40fadc , are my modifications to the original source code.
Automatically exported from code.google.com/p/okuharaandroid-edax-reversi
Edax is a strong othello program. Its main features are:
fast bitboard based & multithreaded engine.
accurate midgame-evaluation function.
opening book learning capability.
text based rich interface.
multi-protocol support to connect to graphical interfaces or play on Internet (GGS).
multi-OS support to run under MS-Windows, Linux and Mac OS X.
This is SSE/AVX optimized version of Edax 4.4.0. Functionally equivalent to the parent project, provided no bugs are introduced.
64 bit version solves fforum-20-39 40% faster than the original 4.4.0 on my test. Thanks to AVX2, x64-modern build runs 60% faster on Haswell. 32 bit version runs 2 times faster than the original.
Diagonals are SIMD'd using vertical mirroring by bswap.
Athlon -get_moves_sse
problem\fforum-20-39.obf: 111349635 nodes in 0:07.998 (13922185 nodes/s).
mobility: 81.10 < 81.28 +/- 0.17 < 82.03
Athlon +get_moves_sse
problem\fforum-20-39.obf: 111349635 nodes in 0:07.889 (14114544 nodes/s).
mobility: 71.08 < 71.72 +/- 0.34 < 73.53
Core2 -get_moves_sse
problem/fforum-20-39.obf: 111349635 nodes in 0:10.180 (10938078 nodes/s).
mobility: 78.06 < 78.18 +/- 0.08 < 78.41
Core2 +get_moves_sse
problem/fforum-20-39.obf: 111349635 nodes in 0:09.978 (11159514 nodes/s).
mobility: 60.84 < 61.19 +/- 0.13 < 61.47
Now calls SIMD'd get_moves for x86/x64 build.
get_full_lines for horizontal and vertical are simplified. The latter is compiled into rotation instrunction.
The last while loop is rearranged not to call bit_count in case stable == 0.
Athlon -get_stability_sse
stability: 90.10 < 90.28 +/- 0.24 < 91.20
Athlon +get_stability_sse
stability: 81.59 < 81.93 +/- 0.73 < 86.25
Core2 -get_stability_sse
stability: 79.24 < 79.39 +/- 0.15 < 79.93
Core2 +get_stability_sse
stability: 71.80 < 71.85 +/- 0.06 < 72.07
Kindergarten version eliminates bit_count call.
Loop optimization and flip using carry propagation. One time execution but affect total solving time.
Eval feature calculation using SSE2 / AVX2 (now in eval_sse.c) improves midgame by 15-30% and endgame by 8-12%. Restoring eval from backup instead of rewinding. eval_open (one time execution) is also optimized.
I think hash->data.move[0] on line 677 should be hash->data.move[1].
SSE optimization and mirroring reduction. (Not used in solving game)
In many cases AVX2 version is simplest, thanks to variable shift instructions (although they are 3 micro-op instructions).
Benchmarks are on Core i5-4260U (Haswell) 1.4GHz (TB 2.7GHz) single thread.
4.4.0 original x64-modern clang
problem/fforum-20-39.obf: 111349635 nodes in 0:05.726 (19446321 nodes/s).
+optimizations 1-5 above, no-avx2
problem/fforum-20-39.obf: 111349635 nodes in 0:05.342 (20844185 nodes/s).
+get_moves (board_sse.c)
problem/fforum-20-39.obf: 111349635 nodes in 0:05.142 (21654927 nodes/s).
+flip_avx.c
problem/fforum-20-39.obf: 111349635 nodes in 0:04.946 (22513068 nodes/s).
+count_last_flip_sse.c
problem/fforum-20-39.obf: 111349635 nodes in 0:04.906 (22696624 nodes/s).
gcc-old, x86 build should be -m32, not -m64. Some flags and defines added for optimization.