[JavaCL] Access violation when creating binary kernel

[ochafik]

From @ochafik on September 1, 2011 18:30

Running BinaryKernelTest using 'mvn test' occasionally causes a segfault. The attached log (pid14000) shows that the error occurs when createKernel is called on a binary program. (Sometimes I also get other errors, suggesting memory corruption.)

Revision 1424, Windows, ATI Stream v2.2. I could not reproduce the error on Linux and NVidia.

For testing purposes, I have created a standalone version of this test (maven is slow). When running this version, the binary createProgram variant fails every time with either InvalidBinary or segfault (see pid13980). But insert a System.out.println(device.getName()) and then it hardly fails. Definitely some memory wierdness somewhere.

I have experienced no issues with the BridJ version of JavaCL on the ATI system.

Google Code Info:
Issue #: 39
Author: kazocs...@gmail.com
Created On: 2010-12-06T21:58:29.000Z
Closed On:

Copied from original issue: nativelibs4java/nativelibs4java#41

[ochafik]

Google Code Info:
Author: kazocs...@gmail.com
Created On: 2010-12-06T21:59:24.000Z

[ochafik]

Hi Csaba,

Thanks for your report, it's gonna be a good pretext to fire my Parallels image that's rusting somewhere on my harddrive :-)

About Maven being slow, you can use the offline mode (once Maven has downloaded everything it needs once) and run a single test using 'mvn test -o -Dtest=BinaryKernelTest'.

Cheers

Olivier

Google Code Info:
Author: olivier.chafik
Created On: 2010-12-06T22:19:15.000Z

[ochafik]

Yes, Olivier, I've actually been using the single test execution feature of Maven, but it still takes its sweet time. In this case, over 4 seconds, which is fine in general but not so convenient with these kinds of random, probabilistic memory errors which require several runs to reproduce.

Ah well, the code was simple enough this time to create a separate program. :)

Google Code Info:
Author: kazocs...@gmail.com
Created On: 2010-12-06T23:04:28.000Z

[ochafik]

(hehe, btw, feel free to attach the program's code here ;-))

Google Code Info:
Author: olivier.chafik
Created On: 2010-12-06T23:18:16.000Z

[ochafik]

Ah, yes. Very straightforward rewrites of the test cases (translation: it would be awkward if I messed something up).

Google Code Info:
Author: kazocs...@gmail.com
Created On: 2010-12-06T23:32:24.000Z

[ochafik]

Hi Csaba,

ATI Stream 2.2 and 2.3 won't install on my virtualized Windows 7, and my dual boot is apparently corrupted (was foolish enough to play with rEFIt :'-( ).

Anyway, I've just added some automatic binaries caching support to JavaCL, which might make things even worse... (or not)
Any feedback would be highly welcome :-)

Cheers

Olivier

Google Code Info:
Author: olivier.chafik
Created On: 2010-12-25T21:45:56.000Z

[ochafik]

Hi Olivier,

I've updated to Stream 2.3 myself, and I still get a segfault in the same way. But as a bonus surprise, ATI has apparently removed virtually all of extensions, so now OpenCL4JavaBenchmarkTest throws an exception (but still passes, which surprised me at first). Attaching report, just so you can see what you're missing with your dual boot shenanigans. :)

As to the caching support, I plan to check it out in a couple of weeks, once I have the code for my current project stable. So I'll pop up if there's anything fishy in that area. :)

Thanks,
Csaba

Google Code Info:
Author: kazocs...@gmail.com
Created On: 2010-12-26T10:06:40.000Z

[ochafik]

Hi Csaba,

Hehe, thanks a lot for the bad news ;-)

Well, at least it has the double extension (cl_amd_fp64), which is not that common...
And I wouldn't have expected different OpenCL versions between CPU and GPU, that's funny.

Cheers

Olivier

Google Code Info:
Author: olivier.chafik
Created On: 2010-12-26T11:10:00.000Z

[ochafik]

Does anyone have an idea of how this could be fixed? We happen to be seeing the same error in another project, but can't figure out if we're doing something wrong or if the bug is in JNA itself.

Google Code Info:
Author: trejkaz
Created On: 2011-01-20T22:28:16.000Z

[ochafik]

Hello,

I'd just like to point out: AMD has recently changed the OpenCL platform name from "ATI Stream" to "AMD Accelerated Parallel Processing", which broke the workaround for this bug in CLContext.getCacheBinaries.

Csaba

Google Code Info:
Author: kazocs...@gmail.com
Created On: 2011-04-09T15:56:23.000Z

[ochafik]

Hi Csaba,

Thanks for pointing this out, I'm deploying an updated snapshot to the Maven repo (cf. revision #1875).

Cheers

zOlive

Google Code Info:
Author: olivier.chafik
Created On: 2011-04-09T20:45:27.000Z