Code hangs...

[eyalhir74]

You can do some pointer arithmetic and work out an offset into the code. But you can also just recompile your code with debugging symbols on and you will get exact line numbers. Going further into the library calls requires disassembly of the cuda runtime library, which can also be done (whether it may be done depends on local copyright laws and how literally you are willing to interpret the EULA you may have agreed to. Or not…).

I’ll try to re-compile with debug when I can - this is supposedly a production system.

From past experience - doing this makes the probelm go away most of the times :)

I’ll also add time stamps and more debug print outs as you’ve suggested.

Many thanks Avid.

Eyal

[James_Sharpe]

I believe that this is a bug in the nvidia driver when it has multiple threads/streams so I’d suggest that you don’t waste too much time trying to debug it - I found disabling L1 Cache helped in some cases but I’ve reproduced it with both cache enabled and disabled; it appears to be prominent when host mapped memory gets loaded directly into local memory.
I submitted a bug report a month ago - bug id 712753.

[James_Sharpe]

I believe that this is a bug in the nvidia driver when it has multiple threads/streams so I’d suggest that you don’t waste too much time trying to debug it - I found disabling L1 Cache helped in some cases but I’ve reproduced it with both cache enabled and disabled; it appears to be prominent when host mapped memory gets loaded directly into local memory.
I submitted a bug report a month ago - bug id 712753.

[eyalhir74]

I believe that this is a bug in the nvidia driver when it has multiple threads/streams so I’d suggest that you don’t waste too much time trying to debug it - I found disabling L1 Cache helped in some cases but I’ve reproduced it with both cache enabled and disabled; it appears to be prominent when host mapped memory gets loaded directly into local memory.

I submitted a bug report a month ago - bug id 712753.

Can you please copy/paste a description of the bug you’ve posted? I cant see it.

thanks

[eyalhir74]

I believe that this is a bug in the nvidia driver when it has multiple threads/streams so I’d suggest that you don’t waste too much time trying to debug it - I found disabling L1 Cache helped in some cases but I’ve reproduced it with both cache enabled and disabled; it appears to be prominent when host mapped memory gets loaded directly into local memory.

I submitted a bug report a month ago - bug id 712753.

Can you please copy/paste a description of the bug you’ve posted? I cant see it.

thanks