Designing an analysis friendly Stockfish?

[Prima]

gaard,Jeremy, Marcos Costalba etc,

thanks for the patch builds. What version of Preserve Analysis (F - J ?) is this latest patch from Marcos Costalba based on?

It's Stockfish 2.0.1 + the Stockfish team's own code for PA on top of it.

Thanks for the reply Jeremy. I know the patch is from SF team (through Marcos Costalba) but what version of PA is this patch applied to? Currently, version PA_G is considered safe and stable.

Was this latest patch from the SF team (Marcos Costalba) applied to Stockfish 2.01PA_F/G, H, I, or J?

[gaard]

gaard,Jeremy, Marcos Costalba etc,

thanks for the patch builds. What version of Preserve Analysis (F - J ?) is this latest patch from Marcos Costalba based on?

It's Stockfish 2.0.1 + the Stockfish team's own code for PA on top of it.

Thanks for the reply Jeremy. I know the patch is from SF team (through Marcos Costalba) but what version of PA is this patch applied to? Currently, version PA_G is considered safe and stable.

Was this latest patch from the SF team (Marcos Costalba) applied to Stockfish 2.01PA_F/G, H, I, or J?

The patch is applied to the original source code for 2.0.1. It only includes the latest patch suggested by mcostalba. It does what versions G did and it's probably more thoroughly tested if it is going to be part of the next Stockfish release, therefore, I suggest using pamc or pamc with the addition changes by other forum members.

[Prima]

gaard,Jeremy, Marcos Costalba etc,

thanks for the patch builds. What version of Preserve Analysis (F - J ?) is this latest patch from Marcos Costalba based on?

It's Stockfish 2.0.1 + the Stockfish team's own code for PA on top of it.

Thanks for the reply Jeremy. I know the patch is from SF team (through Marcos Costalba) but what version of PA is this patch applied to? Currently, version PA_G is considered safe and stable.

Was this latest patch from the SF team (Marcos Costalba) applied to Stockfish 2.01PA_F/G, H, I, or J?

The patch is applied to the original source code for 2.0.1. It only includes the latest patch suggested by mcostalba. It does what versions G did and it's probably more thoroughly tested if it is going to be part of the next Stockfish release, therefore, I suggest using pamc or pamc with the addition changes by other forum members.

Okay I get it now. Thanks gaard.

[BB+]

These incorporate proper root node probing, Peter's other changes, the Marco changes, the Decembrist TT changes, and so on

Think it's not yet counting tb_hits after splitpoints. Here are some possible fixes (gleaned from ComStock):

Line 1509:

Code: Select all

    if (SpNode)
    {
        // Here we have the lock still grabbed                                  
        sp->slaves[threadID] = 0;
        sp->nodes += pos.nodes_searched();
        sp->tbhits += pos.tb_hits (); // add this
        lock_release(&(sp->lock));
    }

Line 2590:

Code: Select all

    splitPoint.nodes = 0;
    splitPoint.tbhits = 0; // add this

Line 2639:

Code: Select all

    // We have returned from the idle loop, which means that all threads are    
    // finished. Update alpha and bestValue, and return.                        
    lock_grab(&mpLock);

    *alpha = splitPoint.alpha;
    *bestValue = splitPoint.bestValue;
    masterThread.activeSplitPoints--;
    masterThread.splitPoint = splitPoint.parent;
    pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes);
    pos.set_tb_hits (pos.tb_hits () + splitPoint.tbhits); // add this
    lock_release(&mpLock);

and in thread.h line 66:

Code: Select all

  // Shared data
  Lock lock;
  volatile int64_t nodes;
  volatile int64_t tbhits; // add this
  volatile Value alpha;
  volatile Value bestValue;
  volatile int moveCount;
  volatile bool betaCutoff;
  volatile int slaves[MAX_THREADS];

[Jeremy Bernstein]

Thanks. Fixes incorporated. I moved a few things around and I'm pretty confident that it's working properly now. My only concern now is the use of hard probing at the root. If we don't, root values will be missed under low time conditions. However, it certainly adds some overhead. Any thoughts about that? I'll post new builds when I get off my train...

[Jeremy Bernstein]

Here new builds. Source is up-to-date at https://github.com/jeremybernstein/Stockfish_PA_GTB (anyone who wants write access, please drop me a PM).

[Peter C]

Thanks. Fixes incorporated. I moved a few things around and I'm pretty confident that it's working properly now. My only concern now is the use of hard probing at the root. If we don't, root values will be missed under low time conditions. However, it certainly adds some overhead. Any thoughts about that? I'll post new builds when I get off my train...

I think we shouldn't hard probe at root. The positions where that would matter are less important than the speed we would lose.

Peter

[Jeremy Bernstein]

Thanks. Fixes incorporated. I moved a few things around and I'm pretty confident that it's working properly now. My only concern now is the use of hard probing at the root. If we don't, root values will be missed under low time conditions. However, it certainly adds some overhead. Any thoughts about that? I'll post new builds when I get off my train...

I think we shouldn't hard probe at root. The positions where that would matter are less important than the speed we would lose.

Peter

We should do some testing first, to see if there is any practical speed loss. I've demonstrated already that there are very practical situations in which Stockfish can lose on the basis of missed tb values. And that was just one random game.

Another question: is there any point to adding the root tb probe return value to the hash? Any hit found will end the search, anyway. Jb

[Jeremy Bernstein]

I feel like the MC settings for PA are slightly disadvantageous for the engine. I'm running a 3/1 tournament right now. Results so far:

Engine                     Score              St             St             St             St             St    S-B
Stockfish-201-64-ja        30,5/52 ·············· =1==101=1=10=  =0=10====1=11  11==1010=0===  1==1=11=0====   775,00
Stockfish_201_x64          28,5/53 =0==010=0=01=  ·············· ==1=0=01====== ==========1=1  =====1=11=11=   722,75
Stockfish_201_64_pamc      25,0/53 =1=01====0=00  ==0=1=10====== ·············· =10=0=1==1==0  ==0=1=0=0==1=   667,50
Stockfish_201_popcnt_pamc  24,5/53 00==0101=1===  ==========0=0  =01=1=0==0==1  ·············· =1=0=1=00=====  655,00
Stockfish_201_PAMC_GTB_x64 23,5/53 0==0=00=1====  =====0=00=00=  ==1=0=1=1==0=  =0=1=0=11===== ··············  625,25

132 of 500 games played
Level: Blitz 3/1
Hardware: Intel(R) Xeon(R) CPU  W3520 @ 2.67GHz 2659 MHz with 1,0 GB Memory
Operating system: Windows 7 Professional Professional (Build 7600) 64 bit

201_x64 is my compile of the standard stockfish code, 64_pamc and popcnt_pamc are gaard's builds, PAMC_GTB_x64 is my build with tablebases. There aren't enough games to draw any real conclusions yet, but standard stockfish seems to slightly outplay the pa builds.

Jeremy

[Uly]

What about the PA_G and PA_I builds?