SSE4 instructions

[MM]

Hi all,

how many more knodes and how many elo points or how much more speed can sse4 instructions give an engine?

I don't know anything about programming and i wonder if there's a substantial difference in performance.

Thank you

Best Regards

[Adam Hair]

Hi all,

how many more knodes and how many elo points or how much more speed can sse4 instructions give an engine?

I don't know anything about programming and i wonder if there's a substantial difference in performance.

Thank you

Best Regards

I have seen references (mainly in connection to Komodo) that SSE 4.2 compiles are 10% to 15% faster. Also, engines seem to gain 50 to 120 Elo per doubling in speed, depending on the engine and the base time control (probably even less gain as the base time control increases). So a rough guess at the increase in Elo would be as little as 5 to 7 Elo to as much as 25 Elo.

[MM]

Hi all,

how many more knodes and how many elo points or how much more speed can sse4 instructions give an engine?

I don't know anything about programming and i wonder if there's a substantial difference in performance.

Thank you

Best Regards

I have seen references (mainly in connection to Komodo) that SSE 4.2 compiles are 10% to 15% faster. Also, engines seem to gain 50 to 120 Elo per doubling in speed, depending on the engine and the base time control (probably even less gain as the base time control increases). So a rough guess at the increase in Elo would be as little as 5 to 7 Elo to as much as 25 Elo.

Thank you

Regards

[zamar]

It depends how heavy program's evaluation function is...

For program with light evaluation it's probably only 2 elo points.
For program with very heavy evaluation it might be even 10 elo points.

Stockfish gains around 4-5 elo points.

[Engin]

hard to say how much, but its gain of course a little, its depends how often you are using popcount or prefetch functions, on my Tornado its gain a lot with popcount but near nothing if i am using prefetch.

[Engin]

..and i forgot to say that you can only gain if you using 64 bit, its not make any sense if you are using it for 32 bit version.

[Engin]

It depends how heavy program's evaluation function is...

For program with light evaluation it's probably only 2 elo points.
For program with very heavy evaluation it might be even 10 elo points.

Stockfish gains around 4-5 elo points.

how heavy ? light evaluation +2 elo ?

i think its depends on how often you are using the popcount in the evaluation.

[MM]

Thank you all for the contribution

Best Regards

[rbarreira]

..and i forgot to say that you can only gain if you using 64 bit, its not make any sense if you are using it for 32 bit version.

Why? I'm pretty sure that doing "hw_popcount (low_32_bits) + hw_popcount (high_32_bits)" is faster than other methods of doing 64-bit popcnt on a 32-bit architecture.

[bob]

..and i forgot to say that you can only gain if you using 64 bit, its not make any sense if you are using it for 32 bit version.

Why? I'm pretty sure that doing "hw_popcount (low_32_bits) + hw_popcount (high_32_bits)" is faster than other methods of doing 64-bit popcnt on a 32-bit architecture.

There is a long-standing myth that bitboards only work on 64 bit machines. I've been using bitboards since the original pentium PC with good results. David Slate did bitboards on a 60 bit machine that required two instructions to manipulate them for the very same reason, yet chess 4.x seemed to work quite well.

One of the "bitboard tricks of the trade" is to not rely too heavily on popcnt() type operations, and there are some ways to do that for at least mobility which is a heavy popcnt() user...

I am not sure why one would do a pair of 32 bit operations on a machine that is obviously 64 bits (those are the only processors with hardware popcnt)...