Cuda-gdb doesn't break and/or step into Kernels

Accelerated Computing CUDA CUDA Programming and Performance
21

gcc4.5 is not supported. That compiler defaults to dwarf3 and nvcc and cuda-gdb doesn’t support dwarf3 yet. The printf causing ULF could be a bug. Please try the latest 4.0RC CUDA release and update with the latest info.

22 
$ gcc --version 

gcc (Ubuntu 4.3.4-10ubuntu1) 4.3.4

$ nvcc --version 

nvcc: NVIDIA (R) Cuda compiler driver

Copyright (c) 2005-2011 NVIDIA Corporation

Built on Fri_Feb_18_18:07:32_PST_2011

Cuda compilation tools, release 4.0, V0.2.1221

The problem manifests in the same way as with cuda-gdb 3.2: it breaks at the kernel call, but does not step into the device code. When trying to invoke any of the cuda-related commands I get the same message as before:

(cuda-gdb) cuda kernel 

Focus not set on any active CUDA kernel.
23

Does it happen even when you turn off X windows on a single GPU? What’s your GPU?

24

I’ve tried with GTX470, 580, and C2050; both with and without X.

25

a similar topic is here ( http://forums.nvidia.com/index.php?showtopic=197360 ) if someone found a way

26

Good that your post reminded me to reply.

I think I’ve got the issue figured out. I am using the Ubuntu PPA repository http://ppa.launchpad.net/ubuntu-x-swat/x-updates/ubuntu to install/update the driver. However, it seems that this driver does not work with debugging. To be honest I haven’t really had time to investigate the issue a lot, but the same code on a CentOS with the official driver on it does work.

@examachine: Are you using this repo? This is not a full explanation of you are seeing the same issue on Mac OS, though.

@CUDA Tools: I think there are quite a number of people who are affected by this issue as on Linux distributions one would obviously prefer a properly packaged version of the driver against a rather distro-agnostic binary installer. This is especially true as the binary driver is infamous of sometimes breaking other packages by removing/overwriting files. Therefor, I think it would be highly beneficial to the community if you could take investigate this issue and if nothing else, suggest the appropriate fixes to the PPA maintainers.

Cheers,

Sz.

27

We have exactly the same problem on our servers:

  • If we use the nvidia-installer to install the driver, there is no problem with the kernel debugging
  • Here, we have RPM’s version of nvidia driver. If we use this version, the debugger doesn’t stop at the correct line.

Our RPM version is just a files set like in nvidia-installer and a few command at the end of the installetion (like ldconfig and depmod).

So, I suppose that we miss a command that the nvidia-installer run during the installation process but I don’t known what.

Thanks for any information

Philippe