www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.announce - amoeba, a chess engine written in D

reply Richard Delorme <abulmo club-internet.fr> writes:
I am pleased to announce the release of a chess engine written in 
D:
https://github.com/abulmo/amoeba

I am not aware of any other chess engine written with the D 
language.

The source can be compiled with dmd, ldc or gdc, but the best 
performance are obtained with the latter (almost twice faster).

This is my first program using the D language (I am a former C 
programmer), and I enjoyed it a lot.
May 20 2016
next sibling parent reply extrawurst <stephan extrawurst.org> writes:
On Friday, 20 May 2016 at 23:16:01 UTC, Richard Delorme wrote:

 I am pleased to announce the release of a chess engine written 
 in D:
 https://github.com/abulmo/amoeba

 I am not aware of any other chess engine written with the D 
 language.

 The source can be compiled with dmd, ldc or gdc, but the best 
 performance are obtained with the latter (almost twice faster).

 This is my first program using the D language (I am a former C 
 programmer), and I enjoyed it a lot.


That is awesome!

Two questions:
1) If it is using UCI protocol I guess there are other engines 
that you can compare amoeba with when it comes to performance and 
other aspects, did you ?

2) It would be awesome if you could write down how it was to use 
D for that project as someone freshly coming from C. What were 
the reasons you chose D, what did you like, what did u hate ? 
That would help you being read on reddit by a hundred blood 
thursty devs (street cred and all) and us for knowing what we can 
improve and also get the PR :)

Cheers,
Stephan
May 20 2016
parent reply Richard Delorme <abulmo club-internet.fr> writes:
On Saturday, 21 May 2016 at 00:29:13 UTC, extrawurst wrote:

 Two questions:
 1) If it is using UCI protocol I guess there are other engines 
 that you can compare amoeba with when it comes to performance 
 and other aspects, did you ?


Yes, It is a strong program, but far from the top programs yet.
In the ccrl scale: http://www.computerchess.org.uk/ccrl/4040/
I guess its rating is close to 2700.
The move generator is pretty fast, though:
$ amoeba-linux-x64-sse4.2 perft -d 7
perft  7 :      3195901860 leaves in     17.920 s    178344094 
leaves/s

In comparison, stockfish 7, one of the strongest chess program in 
C++:
$ stockfish-7
perft 7
[...]
Total time (ms) : 16669
Nodes searched  : 3195901860
Nodes/second    : 191727269

So my implementation of a different algorithm in D is only 7% 
slower at generating moves, which is negligible in the overall 
program.


 2) It would be awesome if you could write down how it was to 
 use D for that project as someone freshly coming from C. What 
 were the reasons you chose D, what did you like, what did u 
 hate ? That would help you being read on reddit by a hundred 
 blood thursty devs (street cred and all) and us for knowing 
 what we can improve and also get the PR :)


Good idea, I will try to write something clever and constructive 
within the next days.

--
Richard
May 21 2016
parent reply Abdulhaq <alynch4047 gmail.com> writes:
On Saturday, 21 May 2016 at 10:10:21 UTC, Richard Delorme wrote:

 On Saturday, 21 May 2016 at 00:29:13 UTC, extrawurst wrote:

 [...]


 Yes, It is a strong program, but far from the top programs yet.
 In the ccrl scale: http://www.computerchess.org.uk/ccrl/4040/
 I guess its rating is close to 2700.
 The move generator is pretty fast, though:
 $ amoeba-linux-x64-sse4.2 perft -d 7
 perft  7 :      3195901860 leaves in     17.920 s    178344094 
 leaves/s

 [...]


Congratulations  2700 is great, vastly better than I managed many 
years ago.
Do you plan to take it further?
May 22 2016
parent Richard Delorme <abulmo club-internet.fr> writes:
On Sunday, 22 May 2016 at 15:48:13 UTC, Abdulhaq wrote:


 Do you plan to take it further?


Yes, but I do not expect to make it as strong as stockfish.
100 elo above the present level by the end of this year will be 
great.
I have ideas for improving the evaluation function, many things 
to tune, test or retest in the search.

--
Richard
May 22 2016
prev sibling next sibling parent reply Johan Engelen <j j.nl> writes:
On Friday, 20 May 2016 at 23:16:01 UTC, Richard Delorme wrote:

 The source can be compiled with dmd, ldc or gdc, but the best 
 performance are obtained with the latter (almost twice faster).


Can you give cmdline details and compiler version data for this? 
(how much faster is gdc relative to ldc?)

Thanks,
   Johan
May 21 2016
next sibling parent Iain Buclaw via Digitalmars-d-announce writes:
On 21 May 2016 at 10:23, Johan Engelen via Digitalmars-d-announce
<digitalmars-d-announce puremagic.com> wrote:

 On Friday, 20 May 2016 at 23:16:01 UTC, Richard Delorme wrote:

 The source can be compiled with dmd, ldc or gdc, but the best performance
 are obtained with the latter (almost twice faster).



 Can you give cmdline details and compiler version data for this? (how much
 faster is gdc relative to ldc?)


I think he meant that ldc or gdc is twice as fast.
May 21 2016
prev sibling parent reply Richard Delorme <abulmo club-internet.fr> writes:
On Saturday, 21 May 2016 at 08:23:59 UTC, Johan Engelen wrote:

 On Friday, 20 May 2016 at 23:16:01 UTC, Richard Delorme wrote:

 The source can be compiled with dmd, ldc or gdc, but the best 
 performance are obtained with the latter (almost twice faster).


 Can you give cmdline details and compiler version data for 
 this? (how much faster is gdc relative to ldc?)


I used the following compiler versions,

dmd: DMD64 D Compiler v2.071.0
ldc: 0.17.0-beta2 (DMD v2.068.2, LLVM 3.7.1)
gdc: gcc version 5.2.0 (crosstool-NG 203be35 - 
20160205-2.066.1-e95a735b97)

to compile, you can use the following commands:
$ DC=ldc2 make amoeba (to compile with ldc)
$ DC=dmd make amoeba (to compile with dmd)
$ DC=gdc make amoeba  (to compile with gdc)
$ make amoeba (use gdc by default)
$ make pgo (for profile guided optimization),
$ make pgo BUILD=popcount (to use the CPU popcnt instruction)
With the above versions & my poor knowledge of the compilers, pgo 
& BUILD=popcount only worked with gdc. I will try to upgrade to 
more recent versions to see if things changed.

to test the speed , you can do:
amoeba bench -f bk.epd -d 14

Here is what I get on my computer (an old i7 2600k at 4Ghz):
     dmd: 24 positions  29981105 nodes in 20.000s : 1499024 nps, 
depth = 13.71
    ldc2: 24 positions  29981105 nodes in 18.340s : 1634711 nps, 
depth = 13.71
     gdc: 24 positions  29981105 nodes in 12.005s : 2497388 nps, 
depth = 13.71
gdc+pgo: 24 positions  29981105 nodes in 10.224s : 2932328 nps, 
depth = 13.71
* gdc used the BUILD=popcount options. without, it is slower, 
doing the test in 15.4. & 13.6 s., respectively for gdc & gdc + 
pgo. So gdc benefits a lot from both pgo and using the popcnt 
instruction.

--
Richard
May 21 2016
parent Johan Engelen <j j.nl> writes:
On Saturday, 21 May 2016 at 16:25:27 UTC, Richard Delorme wrote:

 With the above versions & my poor knowledge of the compilers, 
 pgo & BUILD=popcount only worked with gdc. I will try to 
 upgrade to more recent versions to see if things changed.


With LDC, you can pass "-mattr=+popcnt" to use the popcount 
instruction. Or "-mattr=+sse42" to enable SSE4.2 (like you seem 
to do in the Makefile).
May 22 2016
prev sibling next sibling parent Johan Engelen <j j.nl> writes:
On Friday, 20 May 2016 at 23:16:01 UTC, Richard Delorme wrote:

 I am pleased to announce the release of a chess engine written 
 in D:
 https://github.com/abulmo/amoeba


In the makefile, I see you tried LDC's PGO, awesome! I hope to 
get (part of) it into LDC master very soon. All feedback is very 
welcome.

(small typo: better use ldc-profdata instead of llvm-profdata)

cheers,
   Johan
May 21 2016
prev sibling next sibling parent reply John Colvin <john.loughran.colvin gmail.com> writes:
On Friday, 20 May 2016 at 23:16:01 UTC, Richard Delorme wrote:

 I am pleased to announce the release of a chess engine written 
 in D:
 https://github.com/abulmo/amoeba

 I am not aware of any other chess engine written with the D 
 language.

 The source can be compiled with dmd, ldc or gdc, but the best 
 performance are obtained with the latter (almost twice faster).

 This is my first program using the D language (I am a former C 
 programmer), and I enjoyed it a lot.


LDC might benefit from copying the _popcnt source from ldc's 
druntime in to your code as it has a problem inlining it from 
druntime. You might also see a benefit from the -single-obj flag 
(enabled by default in ldmd).
May 22 2016
parent reply Richard Delorme <abulmo club-internet.fr> writes:
On Sunday, 22 May 2016 at 11:20:44 UTC, John Colvin wrote:

 LDC might benefit from copying the _popcnt source from ldc's 
 druntime in to your code as it has a problem inlining it from 
 druntime. You might also see a benefit from the -single-obj 
 flag (enabled by default in ldmd).


Thank you for those advices.
I succeeded to have popcnt in the LCD build, (using llvm_ctpop). 
-singleobj looks like a good idea too.

for the above test:
original: 18.7 s
with popcnt & singleobj: 11.1s.
LDC is now close to GDC in performance.
A question: why singleobj is not activated by a -Ox options?

--
Richard
May 22 2016
parent David Nadlinger <code klickverbot.at> writes:
On Sunday, 22 May 2016 at 21:22:30 UTC, Richard Delorme wrote:

 A question: why singleobj is not activated by a -Ox options?


It changes compiler behaviour – only a single object file is 
produced. Historically, the default behaviour was used by some 
people/build systems for incremental compilation.

However, since incremental compilation only works when exactly 
the same (sub)sets of commands are invoked every time these days, 
-singleobj should arguably be on by default, at least when 
producing an executable.

This is actually what ldmd2 does already, but we should probably 
take the plunge and break backwards compatibility to enable it 
for the main driver too. I can't think of a scenario where you 
wouldn't want to be using it.

  — David
May 22 2016
prev sibling next sibling parent David Nadlinger <code klickverbot.at> writes:
On Friday, 20 May 2016 at 23:16:01 UTC, Richard Delorme wrote:

 The source can be compiled with dmd, ldc or gdc, but the best 
 performance are obtained with the latter (almost twice faster).


Allowing LDC to do cross-module optimisations (by adding the 
-singleobj flag) and make use of popcnt increased the performance 
by 47% for one specific benchmark, which should bring it into the 
same ballpark as GDC. See https://github.com/abulmo/amoeba/pull/2 
for more details.

PGO might still give the latter a bit of an edge, though, as 
LDC's implementation is still experimental and there is a lot of 
unused potential for making use of the profile data in LLVM's 
optimiser.

  — David
May 22 2016
prev sibling parent Andrei Alexandrescu <SeeWebsiteForEmail erdani.org> writes:
On 05/20/2016 07:16 PM, Richard Delorme wrote:

 I am pleased to announce the release of a chess engine written in D:
 https://github.com/abulmo/amoeba

 I am not aware of any other chess engine written with the D language.

 The source can be compiled with dmd, ldc or gdc, but the best
 performance are obtained with the latter (almost twice faster).

 This is my first program using the D language (I am a former C
 programmer), and I enjoyed it a lot.


https://www.reddit.com/r/programming/comments/4koi8q/amoeba_a_chess_engine_written_in_d/

https://www.facebook.com/dlang.org/posts/1302718359741842

https://twitter.com/D_Programming/status/734769561518510081


Andrei
May 23 2016