Crafty-22.2 is available

[Dann Corbit]

Be aware that it's going to crash on Windows if you use more than one thread.
There is a patched version here, but it only contains a 64 bit binary:
http://cap.connx.com/chess-engines/new- ... ty22-3.zip

However, anyone can recompile the code.

Hi Dann,

I'm using Crafty 22.3 32-bit and only intend doing single-cpu testing at this stage.

Regards, Graham.

Then there is no problem in using that version.

[Graham Banks]

Be aware that it's going to crash on Windows if you use more than one thread.
There is a patched version here, but it only contains a 64 bit binary:
http://cap.connx.com/chess-engines/new- ... ty22-3.zip

However, anyone can recompile the code.

Hi Dann,

I'm using Crafty 22.3 32-bit and only intend doing single-cpu testing at this stage.

Regards, Graham.

Then there is no problem in using that version.

I should have posted in the Crafty 22.3 thread. My bad for creating confusion. Sorry.

Cheers, Graham.

[Peter Skinner]

Be aware that it's going to crash on Windows if you use more than one thread.
There is a patched version here, but it only contains a 64 bit binary:
http://cap.connx.com/chess-engines/new- ... ty22-3.zip

However, anyone can recompile the code.

This is the original code:

Code: Select all

for (i = MAX_BLOCKS_PER_CPU * ((int) tid) + 1; n; i++, n--) {
    block[i] =
        (TREE *) ((~(size_t) 127) & (127 + (size_t) WinMalloc(sizeof(TREE) +
                127, tid)));

Here is your patched code:

Code: Select all

for (i = MAX_BLOCKS_PER_CPU * ((int) tid) + 1; n; i++, n--) {
    block[i] =
        (TREE *) (WinMalloc(sizeof(TREE) + 2047, tid)));

Your patched code has an extra ) in it, and should be as follows:

Code: Select all

for (i = MAX_BLOCKS_PER_CPU * ((int) tid) + 1; n; i++, n--) {
    block[i] =
        (TREE *) (WinMalloc(sizeof(TREE) + 2047, tid));

Other than that one error, it compiles cleanly, and doesn't crash when using two threads. I will update my website with new binaries.

Peter

[bob]

How's this look Peter? I want to use closest to 128mb hash, single-cpu, ponder off, no learning, generic book.

smpnice=0
egtb
tbpath=c:\Tablebases
hash=96M
hashp=64M
cache=32M
ponder off
learn 0
book off
log=off
exit

Seems to work okay.

Cheers, Graham.

Got confirmation via messenger that it's ok. Thanks Peter.
Doesn't matter whether smpnice=0 or 1.

Actually it does. With default smpnice=1, if you use multiple processors, the program terminates processes when it is waiting and you have ponder=off. With smpnice=0, when waiting the processes will spin in a tight busy-wait loop and burn N-1 CPUs at 100%. Ponder=on would make that not very important, but with ponder=off and mt=anything other than 1, smpnice=1 is more system-friendly (and is the default as I mentioned).

[bob]

Be aware that it's going to crash on Windows if you use more than one thread.
There is a patched version here, but it only contains a 64 bit binary:
http://cap.connx.com/chess-engines/new- ... ty22-3.zip

However, anyone can recompile the code.

Why would it crash? windows and linux both now use threads and the old exit(0) problem has been removed. I'm not aware of anything that breaks on windows with 22.2 and beyond. 22.1 had a problem for sure.

[bob]

Be aware that it's going to crash on Windows if you use more than one thread.
There is a patched version here, but it only contains a 64 bit binary:
http://cap.connx.com/chess-engines/new- ... ty22-3.zip

However, anyone can recompile the code.

This is the original code:

Code: Select all

for (i = MAX_BLOCKS_PER_CPU * ((int) tid) + 1; n; i++, n--) {
    block[i] =
        (TREE *) ((~(size_t) 127) & (127 + (size_t) WinMalloc(sizeof(TREE) +
                127, tid)));

Here is your patched code:

Code: Select all

for (i = MAX_BLOCKS_PER_CPU * ((int) tid) + 1; n; i++, n--) {
    block[i] =
        (TREE *) (WinMalloc(sizeof(TREE) + 2047, tid)));

Your patched code has an extra ) in it, and should be as follows:

Code: Select all

for (i = MAX_BLOCKS_PER_CPU * ((int) tid) + 1; n; i++, n--) {
    block[i] =
        (TREE *) (WinMalloc(sizeof(TREE) + 2047, tid));

Other than that one error, it compiles cleanly, and doesn't crash when using two threads. I will update my website with new binaries.

Peter

What is this supposed to fix? The above "fix" actually breaks cache usage. The original code forces the split blocks to begin on a 128-byte cache boundary, while the "fixed code" accepts whatever winmalloc returns which is guaranteed only to be on an 8-byte boundary.

What exactly is the problem???

[Graham Banks]

How's this look Peter? I want to use closest to 128mb hash, single-cpu, ponder off, no learning, generic book.

smpnice=0
egtb
tbpath=c:\Tablebases
hash=96M
hashp=64M
cache=32M
ponder off
learn 0
book off
log=off
exit

Seems to work okay.

Cheers, Graham.

Got confirmation via messenger that it's ok. Thanks Peter.
Doesn't matter whether smpnice=0 or 1.

Actually it does. With default smpnice=1, if you use multiple processors, the program terminates processes when it is waiting and you have ponder=off. With smpnice=0, when waiting the processes will spin in a tight busy-wait loop and burn N-1 CPUs at 100%. Ponder=on would make that not very important, but with ponder=off and mt=anything other than 1, smpnice=1 is more system-friendly (and is the default as I mentioned).

Thanks Bob.

[Peter Skinner]

Be aware that it's going to crash on Windows if you use more than one thread.
There is a patched version here, but it only contains a 64 bit binary:
http://cap.connx.com/chess-engines/new- ... ty22-3.zip

However, anyone can recompile the code.

This is the original code:

Code: Select all

for (i = MAX_BLOCKS_PER_CPU * ((int) tid) + 1; n; i++, n--) {
    block[i] =
        (TREE *) ((~(size_t) 127) & (127 + (size_t) WinMalloc(sizeof(TREE) +
                127, tid)));

Here is your patched code:

Code: Select all

for (i = MAX_BLOCKS_PER_CPU * ((int) tid) + 1; n; i++, n--) {
    block[i] =
        (TREE *) (WinMalloc(sizeof(TREE) + 2047, tid)));

Your patched code has an extra ) in it, and should be as follows:

Code: Select all

for (i = MAX_BLOCKS_PER_CPU * ((int) tid) + 1; n; i++, n--) {
    block[i] =
        (TREE *) (WinMalloc(sizeof(TREE) + 2047, tid));

Other than that one error, it compiles cleanly, and doesn't crash when using two threads. I will update my website with new binaries.

Peter

What is this supposed to fix? The above "fix" actually breaks cache usage. The original code forces the split blocks to begin on a 128-byte cache boundary, while the "fixed code" accepts whatever winmalloc returns which is guaranteed only to be on an 8-byte boundary.

What exactly is the problem???

I did email about this patch by Dan, and hadn't updated the website with new compiles until I heard back from you.

Since it doesn't fix things and actually breaks things, then I won't update it with his "patch".

The problem seems to be that Windows users, using more than one thread are seeing crashing problems. While I can't seem to reproduce this with my compiles on any of my machines, I don't know what the real issue is.

If I start Crafty with mt=2, and select "machine white" in Winboard, it works perfectly. No issues. Crafty also played 30+ games via ICS last night with no issues.

So again, I don't know where the issue is...

[gerold]

Crafty 22.3 32 is now working in Arena. The download from
your page.

Thanks Peter,

Gerold.

[Dann Corbit]

Access to null pointers and/or unintialized memory.
I sent you a patch via email last night.