Komodo...wonderful

[fern]

Hi Don:
happy to know you are still finding ways to improve even more Komodo, but as you could imagine, for me is kind of a disappointing to see you engaged in those technicalities to enhance an already incredible strong program just because, and I understand that, you cannot simply stay quiescent looking any other program getting some 5 elo pints more than Komodo.
I still hope once day your effort will aim to the matter we have talked before, GUI capabilities, friendly approach to middle level players as me, etc.

One day....

fern

[Don]

Why not just communicate trough PM's then. You will not have a problem with avoiding people, mistrusted by you.

I just want to get to the truth, I don't want to argue about which compiler is better. I did not mean to imply you were being dishonest - there are many different reasons your statements might appear inconsistent and I was hoping you could clarify this.

You have to admit that it's strange that you say on either platform Intel will be 5-10% faster than gcc and yet they will both be the same speed - that means the windows binary has to be 15-25% faster (that the gcc version) to reconcile what I am seeing. I am basically seeing 10% difference on gcc.

I think it's almost a proven fact that you cannot get the same speed on windows due to the register issue discussed on the other thread, the compiler simply has less registers to work with on Windows. Linux has an advantage that cannot be compiled away - so if you are getting the same numbers on the Linux Intel compiler I have to suspect that you are running your tests incorrectly. It could also be that on Intel the Linux version is purposely crippled which really seems insane, but it does not reconcile with your statement that Intel is much faster on Linux than GCC.

So I apologize for challenging you, but I want to plow through this and find out the relevant details and this has to be reconciled or it remains a blocker to understanding what is going on.

[Dragan]

I did some performance testing of my engine using MSVC10 Express vs MingW with the latest GCC.

The MSVC version without PGO was 10% faster then the GCC version with the PGO.
Test was done under Windows 7 on an Intel machine.

[Don]

Hi Don:
happy to know you are still finding ways to improve even more Komodo, but as you could imagine, for me is kind of a disappointing to see you engaged in those technicalities to enhance an already incredible strong program just because, and I understand that, you cannot simply stay quiescent looking any other program getting some 5 elo pints more than Komodo.
I still hope once day your effort will aim to the matter we have talked before, GUI capabilities, friendly approach to middle level players as me, etc.

One day....

fern

I am seriously considering it, but if I do it I want it to be really different, not just a copy of what is already available. I do have ideas ....

[Don]

I did some performance testing of my engine using MSVC10 Express vs MingW with the latest GCC.

The MSVC version without PGO was 10% faster then the GCC version with the PGO.
Test was done under Windows 7 on an Intel machine.

Richard Vida, the Critter programmer reports that he uses MSVC on his windows compile and gcc on his Linux compiler and he gets about 6% difference in favor of Linux and he used PGO with both, so this is another serious discrepancy.

This is discrepancy of a huge magnitue. For you the MSVC is 10% faster on Windows, For Richard the MSVC version is 6% slower on Windows! That is hard to explain.

What do you think is going on?

[Vinvin]

After MP we will probably gain a significant amount of ELO because I have never put any energy into optimizations. I did an experiment a few months back to see how much I could get in a day with micro-optimizations and was able to produce something like a 5-10 percent speedup without trying very hard by doing a number of pretty simple and obvious things. As you can see Komodo is far slower in NPS than any other top playing program as I have always focused on chess heuristics. So I will go into a serious optimization effort to increase our performance. If we did the same NPS as Stockfish, Houding or Critter we would be untouchable. I don't expect to be able to achieve that (as I believe part of our success is due to our "heavy" evaluation function), but we would be untouchable even if I made up half the difference and I don't think this will be a problem - it will just be a lot of work for me.

Larry and I may also restructure the evaluation function which I think will bring more performance and enable even more evaluation without bogging down the search even more. It's actually been a long time since I have put serious work into the real guts of the program - we have made huge progress despite this but Komodo is like an old car that needs a complete overhaul! It's going to blow your socks off when we get that done!
...
Don

It's good to see you know where it's possible to improve your engine ... I bet a lot of top programmers don't even know where to improve their engine !
We look forward Komodo 6 MP !!!

[Werewolf]

Do you still have plans for a Cluster Komodo?

[fern]

I look forward for that day and remember that i have since some years a lot of ideas and perhaps, one or two of them could be useful for you. I am also thinking not just in another GUI more or less similar to existent, but a radically new approach.
If ever you begin with that, please let me know. Maybe I can be of some utility.

fern

[Dragan]

I didn't try compiling on Linux. I used MingW GCC to create Windows executable and this executable is much slower then MSVC one.
The difference with Critter could be related to the use of popcnt. My engine uses rotated bitboards, but I don't use popcnt much.
GCC has intrinsic for popcnt that performs faster then my popcnt MSVC code.

Dragan

[Dragan]

Forgot to mention very important fact.
I tested 32-bit compiles only.