Lc0 evaluation of startpos is -2147483648

[zullil]

Anyone else seeing stuff like this from Lc0?

info depth 18 seldepth 53 time 196336 nodes 2568924 score cp -2147483648 wdl -2147483648 1000 -2147483648 nps 13283 tbhits 0 pv e2e4 e7e5 g1f3 b8c6 f1b5 a7a6 b5a4 g8f6 e1g1 f6e4 d2d4 b7b5 a4b3 d7d5 d4e5 c8e6 c1e3 f8c5 d1d3 e8g8 c2c3 d8e7 b1d2 c5e3 d3e3 e4d2 e3d2 c6a5 f1e1 a5c4 b3c4 b5c4 b2b4 c4b3 a2b3 c7c5 b3b4 c5b4 c3b4 f8b8 e1b1 e6f5

I already reported this issue at https://github.com/LeelaChessZero/lc0/issues

Seems independent of network used, compiler used in building, and has apparently been present for awhile.

Maybe it's my fault??

[Dann Corbit]

The great question has finally been answered. Chess is a loss for white.

[zullil]

The great question has finally been answered. Chess is a loss for white.

But the WDL score says the probability of a draw is 1000/1000.

[Ajedrecista]

Hello Louis:

Anyone else seeing stuff like this from Lc0?

info depth 18 seldepth 53 time 196336 nodes 2568924 score cp -2147483648 wdl -2147483648 1000 -2147483648 nps 13283 tbhits 0 pv e2e4 e7e5 g1f3 b8c6 f1b5 a7a6 b5a4 g8f6 e1g1 f6e4 d2d4 b7b5 a4b3 d7d5 d4e5 c8e6 c1e3 f8c5 d1d3 e8g8 c2c3 d8e7 b1d2 c5e3 d3e3 e4d2 e3d2 c6a5 f1e1 a5c4 b3c4 b5c4 b2b4 c4b3 a2b3 c7c5 b3b4 c5b4 c3b4 f8b8 e1b1 e6f5

I already reported this issue at https://github.com/LeelaChessZero/lc0/issues

Seems independent of network used, compiler used in building, and has apparently been present for awhile.

Maybe it's my fault??

-2147483648 = -2^31 (32-bit integer range is usually from -2^31 to 2^31 - 1, this is, from -2147483648 to 2147483647). Could it be a kind of overflow or more likely an incorrect hadling of 0?

Regards from Spain.

Ajedrecista.

[zullil]

Hello Louis:

Anyone else seeing stuff like this from Lc0?

info depth 18 seldepth 53 time 196336 nodes 2568924 score cp -2147483648 wdl -2147483648 1000 -2147483648 nps 13283 tbhits 0 pv e2e4 e7e5 g1f3 b8c6 f1b5 a7a6 b5a4 g8f6 e1g1 f6e4 d2d4 b7b5 a4b3 d7d5 d4e5 c8e6 c1e3 f8c5 d1d3 e8g8 c2c3 d8e7 b1d2 c5e3 d3e3 e4d2 e3d2 c6a5 f1e1 a5c4 b3c4 b5c4 b2b4 c4b3 a2b3 c7c5 b3b4 c5b4 c3b4 f8b8 e1b1 e6f5

I already reported this issue at https://github.com/LeelaChessZero/lc0/issues

Seems independent of network used, compiler used in building, and has apparently been present for awhile.

Maybe it's my fault??

-2147483648 = -2^31 (32-bit integer range is usually from -2^31 to 2^31 - 1, this is, from -2147483648 to 2147483647). Could it be a kind of overflow or more likely an incorrect hadling of 0?

Regards from Spain.

Ajedrecista.

Thanks for the reply. Yes, almost certainly an incorrect handling of a near-zero number.

Just wondering if anyone else has experienced this, or can reproduce it.

[zullil]

Hello Louis:

Anyone else seeing stuff like this from Lc0?

info depth 18 seldepth 53 time 196336 nodes 2568924 score cp -2147483648 wdl -2147483648 1000 -2147483648 nps 13283 tbhits 0 pv e2e4 e7e5 g1f3 b8c6 f1b5 a7a6 b5a4 g8f6 e1g1 f6e4 d2d4 b7b5 a4b3 d7d5 d4e5 c8e6 c1e3 f8c5 d1d3 e8g8 c2c3 d8e7 b1d2 c5e3 d3e3 e4d2 e3d2 c6a5 f1e1 a5c4 b3c4 b5c4 b2b4 c4b3 a2b3 c7c5 b3b4 c5b4 c3b4 f8b8 e1b1 e6f5

I already reported this issue at https://github.com/LeelaChessZero/lc0/issues

Seems independent of network used, compiler used in building, and has apparently been present for awhile.

Maybe it's my fault??

-2147483648 = -2^31 (32-bit integer range is usually from -2^31 to 2^31 - 1, this is, from -2147483648 to 2147483647). Could it be a kind of overflow or more likely an incorrect hadling of 0?

Regards from Spain.

Ajedrecista.

Thanks for the reply. Yes, almost certainly an incorrect handling of a near-zero number.

Just wondering if anyone else has experienced this, or can reproduce it.

Further investigation suggests this is my fault. The odd behavior seems to occur only when I apply an offset of +200 Mhz or more to my GPU clock. Up until now, I was able to apply a +200 MHz offset with no apparent ill effects. Looks like I'll need to reduce that.