Stockfish 1.6.3 JA update available

[Jim Ablett]

Stockfish 1.6.3 JA (uci)
by Tord Romstad, Marco Costalba & Joona Kiiski.

Windows/Linux 64/32 Intel/Gcc p.g.o compiles.

Whats new:

This is a maintenance release of Stockfish 1.6.x series.

It is 100% functionally and speed equivalent to 1.6.2 so there is NO ELO DIFFERENCE, so, as
usual, testers do not need to redo anything.

The aim of this release is to fix all the reported bugs that have been accumulated in these weeks.


So here are the fixes:

- Extend maximum hash size to 8GB (it was 2 GB)

- Fix a possible crash when using many threads (thanks to Bruno Causse for the fix)

- Fix en-passant parsing from fen string

- Fix various UCI interface issues so that now connection with GUI should be reliable with all the
GUIs

- Fix compatibility with gcc 4.4

- Fixes a Chess960 bug when playing with more than one search thread

- Add hardware POPCNT support for gcc: use 'make gcc-popcnt' to enable hardware POPCNT support
if you have an i7 / i5 CPU

- Fix sending of best move during an infinite search

- Optimized "Minimum Split Depth" for 8 cores and removed auto-limitation to 7 cores (thanks to
Louis Zulli)

Download:
http://www.mediafire.com/?ywtum2tmwi5
Alt Download:
http://good.net/_7cYkXC2PF

Jim.

[M ANSARI]

Great! Thanks a milllion to all the Stockfish team. Can we have a run down of changes, improvements etc ...

[M ANSARI]

Well first thing I noticed is that it seems to have a much more improved MP implementation. On 8 cores it seems in task manager that all cores are at 100% while before that was jumping all over the place.

[mcostalba]

Thanks Jim,

Greatly appreciated, as usual.

[swami]

Very good improvement.

Stockfish 1.6.3 JA
by Marco Costalba, Tord Romstad, Joona Kiiski, Europe.

Strategic Test Suite Conditions:

Core2Quad 32 bits, Q6600, 2 GB RAM, 2.4GHZ
10 seconds per position
900 positions
Engine uses 155 Mb Hash.
Single CPU
Arena GUI


Overall Performance:

• Total Score: 695/900 [.....] Average : 77.22% [.....] Grade: A [.....] Total Rated Time: 41.37/150 minutes [2482 Seconds/9000 Seconds]

Subject-wise Scores:

STS (v1.0) - Undermining:
82/100, Grade: A+

STS (v2.1) - Open Files and Diagonals:
81/100, Grade: A+

STS (v3.0) - Knight Outposts/Centralization/Repositioning:
81/100, Grade: A+

STS (v4.1) - Square Vacancy:
85/100, Grade: S

STS (v5.0) - Bishop vs Knight:
77/100, Grade: A

STS (v6.0) - Re-Capturing:
77/100, Grade: A

STS (v7.0) - Offer of Simplification:
74/100, Grade: A-

STS (v8.1) - Advancement of f/g/h Pawns:
63 /100, Grade: B

STS (v9.0) - Advancement of a/b/c Pawns:
75/100, Grade: A

Best Wishes,
Swami

[Martin Thoresen]

Very good improvement.

Do you have the test results for 1.6.2 somewhere?

[swami]

Very good improvement.

Do you have the test results for 1.6.2 somewhere?

http://www.talkchess.com/forum/viewtopic.php?t=31363

I note the improvement in Square Vacancy and Advancement of f/g/h pawns.

Also, this version does so much better than many engines in a/b/c pawns test which is supposedly the most difficult of all tests.

[yanquis1972]

out of curiousity swami (or anyone else), what are the leaders in the f/g/h pawn category? my guesses would be naum & shredder 12. it would be interesting to see results for junior 2010 as well.

i tested the latest ivanhoe at 5s/move (but on 4 cores) and it's % scores for all tests were S (superduper?) with a couple As at the end. this despite some bug in the engine/gui/which caused it to skip about 10 positions or so. very impressive nonetheless, something like 87% best move score. i see now looking at the stats it scored 'only' 76 on the f/g/h one, with 869 points (i'm pretty sure no positions were skipped on that one since i watched it as it was the one i was most interested in). that was definitely it's lowest best move %.

other bugs in the sts stat report -- it was awarded only 30 points in sts5 despite scoring an 88 in that suite, & all of sts6 apparently failed to save or translate. i have no idea if it is the engine/gui/stat program thats responsible for these errors but thought i'd let you know.

edit -- sorry that got so offtopic, i thought i was posting in one of swami's sts thrreads

[mcostalba]

Very good improvement.

Hi Swami,

I have verified 1.6.3 JA is 100% equivalent with 1.6.2 JA, if you run:

Code: Select all

stockfish bench 128 1 12 default depth

node count at the end is 22308138 as for 1.6.2, even the smallest difference would yield to a completely different node count.


So test result should be the same. If it is not then there is some source of randomness in the test itself, here following are the first ideas that come to my mind:

1) Frequency throttling: assuming you run all the tests with the same hardware could happen (and do happen) that the OS changes the CPU frequency during the test to prevent overheating. In modern CPU /OS this is transparent to the user and very dynamic in nature, some time it goes down, after a while it comes up again, and so on. Because frequency, that is strictly related to CPU speed, changes in random ways test could be not reproducible. To fix that you should use an external software tool that fixes the frequency and set it to a lower value then maximum so to prevent overheating and run all the tests at the same fixed frequency.


2) Multi core test. If you run the test using more then one core then result is _not_ reproducible because different threads are subject to random timing issues that make the number and the type of nodes searched different from run to run. For instance if instead of the above you run:

Code: Select all

stockfish bench 128 2 12 default depth

with '2' threads instead of '1' you will get different node count at every run.

So I suggest to run all your tests at _single core_ to have reproducible results. This is not related to Stockfish, but is the same for all the MP engines.

[swami]

Hi Marco,

I don't think that it's either of the points you made:

You forgot to raise one important:

Arena has a minor bug that introduces new move of its own and awards points for the engine that guessed the newly introduced move.

It happens when STS Epd files have partial credit moves. Arena doesn't understand partial credit moves at all.

See this thread:
http://www.talkchess.com/forum/viewtopic.php?t=32022

Stockfish 1.6.2 was tested when STS suites had a partial credit moves and Arena erroneously awarded points for certain moves. I wasn't aware of this bug until Wesley pointed it out in a thread later on.

Stockfish 1.6.3 is tested with no-partial scoring-STS suites which consists of only best moves.

Best Wishes,
Swami

Very good improvement.

Hi Swami,

I have verified 1.6.3 JA is 100% equivalent with 1.6.2 JA, if you run:

Code: Select all

stockfish bench 128 1 12 default depth

node count at the end is 22308138 as for 1.6.2, even the smallest difference would yield to a completely different node count.


So test result should be the same. If it is not then there is some source of randomness in the test itself, here following are the first ideas that come to my mind:

1) Frequency throttling: assuming you run all the tests with the same hardware could happen (and do happen) that the OS changes the CPU frequency during the test to prevent overheating. In modern CPU /OS this is transparent to the user and very dynamic in nature, some time it goes down, after a while it comes up again, and so on. Because frequency, that is strictly related to CPU speed, changes in random ways test could be not reproducible. To fix that you should use an external software tool that fixes the frequency and set it to a lower value then maximum so to prevent overheating and run all the tests at the same fixed frequency.


2) Multi core test. If you run the test using more then one core then result is _not_ reproducible because different threads are subject to random timing issues that make the number and the type of nodes searched different from run to run. For instance if instead of the above you run:

Code: Select all

stockfish bench 128 2 12 default depth

with '2' threads instead of '1' you will get different node count at every run.

So I suggest to run all your tests at _single core_ to have reproducible results. This is not related to Stockfish, but is the same for all the MP engines.