Laser 1.6 Release

[CMCanavessi]

Well that explains it then, Ive found that Ginkgo is similar to laser in that the bmi2 version does not run slow on ryzens.

[CMCanavessi]

Here are some benchmarks using "go depth 30" with a 4gb hash size:

Laser-1_6-NOPOPCNT.exe
info depth 30 seldepth 54 score cp 25 time 3464369 nodes 2118013920 nps 611370 tbhits 0 hashfull 970 pv e2e4 e7e6 d2d4 d7d5 b1d2 c7c5 g1f3 c5d4 e4d5 d8d5 f1c4 d5d6 e1g1 g8f6 d2b3 b8c6 b3d4 c6d4 d1d4 d6d4 f3d4 c8d7 c1f4 a8c8 c4d3 f8c5 d4f3 d7c6 f3e5 c6d5 c2c4 d5e4 d3e4 f6e4

Laser-1_6.exe
info depth 30 seldepth 54 score cp 25 time 2933501 nodes 2118013920 nps 722008 tbhits 0 hashfull 970 pv e2e4 e7e6 d2d4 d7d5 b1d2 c7c5 g1f3 c5d4 e4d5 d8d5 f1c4 d5d6 e1g1 g8f6 d2b3 b8c6 b3d4 c6d4 d1d4 d6d4 f3d4 c8d7 c1f4 a8c8 c4d3 f8c5 d4f3 d7c6 f3e5 c6d5 c2c4 d5e4 d3e4 f6e4

Laser-1_6-BMI2.exe
info depth 30 seldepth 54 score cp 25 time 2927108 nodes 2118013920 nps 723585 tbhits 0 hashfull 970 pv e2e4 e7e6 d2d4 d7d5 b1d2 c7c5 g1f3 c5d4 e4d5 d8d5 f1c4 d5d6 e1g1 g8f6 d2b3 b8c6 b3d4 c6d4 d1d4 d6d4 f3d4 c8d7 c1f4 a8c8 c4d3 f8c5 d4f3 d7c6 f3e5 c6d5 c2c4 d5e4 d3e4 f6e4


You can see there's no difference at all between the POPCNT and the BMI2 versions, and that both of them are about 18% faster than the NOPOPCNT version. Now compare that to Ethereal 10.00:


Ethereal10.00-x64-nopopcnt.exe
info depth 30 seldepth 44 score cp 32 time 240015 nodes 377673587 nps 1573000 tbhits 0 hashfull 161 pv d2d4 g8f6 c2c4 e7e6 g1f3 d7d5 b1c3 f8b4 e2e3 e8g8 c1d2 c7c5 a2a3 b4c3 d2c3 c5d4 d1d4 b8c6 d4d3 d5c4 d3c4 f6d5 a1d1 c8d7 f1d3 a8c8 c4b3 d5c3 b2c3 c6a5 b3c2 f7f5 e1g1 a5c4 d3c4 c8c4

Ethereal10.00-x64-popcnt.exe
info depth 30 seldepth 44 score cp 32 time 199344 nodes 377673587 nps 1894000 tbhits 0 hashfull 161 pv d2d4 g8f6 c2c4 e7e6 g1f3 d7d5 b1c3 f8b4 e2e3 e8g8 c1d2 c7c5 a2a3 b4c3 d2c3 c5d4 d1d4 b8c6 d4d3 d5c4 d3c4 f6d5 a1d1 c8d7 f1d3 a8c8 c4b3 d5c3 b2c3 c6a5 b3c2 f7f5 e1g1 a5c4 d3c4 c8c4

Ethereal10.00-x64-pext.exe
info depth 30 seldepth 44 score cp 32 time 320344 nodes 377673587 nps 1178000 tbhits 0 hashfull 161 pv d2d4 g8f6 c2c4 e7e6 g1f3 d7d5 b1c3 f8b4 e2e3 e8g8 c1d2 c7c5 a2a3 b4c3 d2c3 c5d4 d1d4 b8c6 d4d3 d5c4 d3c4 f6d5 a1d1 c8d7 f1d3 a8c8 c4b3 d5c3 b2c3 c6a5 b3c2 f7f5 e1g1 a5c4 d3c4 c8c4

Here you can see the huge drop in performance from the BMI2 version. While the POPCNT version is 20% faster than the NOPOPCNT one (similar to Laser's 18%), the BMI2 version takes a huge hit (on Ryzen) being 25% slower than the NOPOPCNT one, and a whooping 38% slower than the POPCNT one. Stockfish and the rest of the engines show a similar trend, with a couple of exceptions.

[AndrewGrant]

Here are some benchmarks using "go depth 30" with a 4gb hash size:

Laser-1_6-NOPOPCNT.exe
info depth 30 seldepth 54 score cp 25 time 3464369 nodes 2118013920 nps 611370 tbhits 0 hashfull 970 pv e2e4 e7e6 d2d4 d7d5 b1d2 c7c5 g1f3 c5d4 e4d5 d8d5 f1c4 d5d6 e1g1 g8f6 d2b3 b8c6 b3d4 c6d4 d1d4 d6d4 f3d4 c8d7 c1f4 a8c8 c4d3 f8c5 d4f3 d7c6 f3e5 c6d5 c2c4 d5e4 d3e4 f6e4

Laser-1_6.exe
info depth 30 seldepth 54 score cp 25 time 2933501 nodes 2118013920 nps 722008 tbhits 0 hashfull 970 pv e2e4 e7e6 d2d4 d7d5 b1d2 c7c5 g1f3 c5d4 e4d5 d8d5 f1c4 d5d6 e1g1 g8f6 d2b3 b8c6 b3d4 c6d4 d1d4 d6d4 f3d4 c8d7 c1f4 a8c8 c4d3 f8c5 d4f3 d7c6 f3e5 c6d5 c2c4 d5e4 d3e4 f6e4

Laser-1_6-BMI2.exe
info depth 30 seldepth 54 score cp 25 time 2927108 nodes 2118013920 nps 723585 tbhits 0 hashfull 970 pv e2e4 e7e6 d2d4 d7d5 b1d2 c7c5 g1f3 c5d4 e4d5 d8d5 f1c4 d5d6 e1g1 g8f6 d2b3 b8c6 b3d4 c6d4 d1d4 d6d4 f3d4 c8d7 c1f4 a8c8 c4d3 f8c5 d4f3 d7c6 f3e5 c6d5 c2c4 d5e4 d3e4 f6e4


You can see there's no difference at all between the POPCNT and the BMI2 versions, and that both of them are about 18% faster than the NOPOPCNT version. Now compare that to Ethereal 10.00:


Ethereal10.00-x64-nopopcnt.exe
info depth 30 seldepth 44 score cp 32 time 240015 nodes 377673587 nps 1573000 tbhits 0 hashfull 161 pv d2d4 g8f6 c2c4 e7e6 g1f3 d7d5 b1c3 f8b4 e2e3 e8g8 c1d2 c7c5 a2a3 b4c3 d2c3 c5d4 d1d4 b8c6 d4d3 d5c4 d3c4 f6d5 a1d1 c8d7 f1d3 a8c8 c4b3 d5c3 b2c3 c6a5 b3c2 f7f5 e1g1 a5c4 d3c4 c8c4

Ethereal10.00-x64-popcnt.exe
info depth 30 seldepth 44 score cp 32 time 199344 nodes 377673587 nps 1894000 tbhits 0 hashfull 161 pv d2d4 g8f6 c2c4 e7e6 g1f3 d7d5 b1c3 f8b4 e2e3 e8g8 c1d2 c7c5 a2a3 b4c3 d2c3 c5d4 d1d4 b8c6 d4d3 d5c4 d3c4 f6d5 a1d1 c8d7 f1d3 a8c8 c4b3 d5c3 b2c3 c6a5 b3c2 f7f5 e1g1 a5c4 d3c4 c8c4

Ethereal10.00-x64-pext.exe
info depth 30 seldepth 44 score cp 32 time 320344 nodes 377673587 nps 1178000 tbhits 0 hashfull 161 pv d2d4 g8f6 c2c4 e7e6 g1f3 d7d5 b1c3 f8b4 e2e3 e8g8 c1d2 c7c5 a2a3 b4c3 d2c3 c5d4 d1d4 b8c6 d4d3 d5c4 d3c4 f6d5 a1d1 c8d7 f1d3 a8c8 c4b3 d5c3 b2c3 c6a5 b3c2 f7f5 e1g1 a5c4 d3c4 c8c4

Here you can see the huge drop in performance from the BMI2 version. While the POPCNT version is 20% faster than the NOPOPCNT one (similar to Laser's 18%), the BMI2 version takes a huge hit (on Ryzen) being 25% slower than the NOPOPCNT one, and a whooping 38% slower than the POPCNT one. Stockfish and the rest of the engines show a similar trend, with a couple of exceptions.

Thanks for confirming with some good numbers.

[CMCanavessi]

Thanks for confirming with some good numbers.

Having a Ryzen processor, which Laser binary would you choose? Considering there's almost no difference at all between the BMI and POPCNT ones...

[AndrewGrant]

Thanks for confirming with some good numbers.

Having a Ryzen processor, which Laser binary would you choose? Considering there's almost no difference at all between the BMI and POPCNT ones...

His "BMI2" version is built with -march=haswell on. I am VERY surprised that it runs for you on a Ryzen. I would recommend not using the BMI2 version, as I wonder if there is any hidden flaw put in place.