所有CUDA设备均用于显示：无法在桌面环境中调试我的CUDA代码

Question

since last week I got a big problem with my CUDA-development setup. I have an integrated GPU which I attached my monitors too and an extra NVIDIA Card for running my CUDA kernels on. However, i can not debug my code anymore, because it says:

fatal:  All CUDA devices are used for display and cannot be used while debugging. (error code = CUDBG_ERROR_ALL_DEVICES_WATCHDOGGED(0x18)

Somehow it seems that my X-Server is blocking my NVIDIA GPU because if I switch to another virtual console (CTRL+ALT+F1) I am able to run my code using cuda-gdb. No monitor cable is plugged into the NVIDIA-card...

"lsof /dev/nvidia*" does not give any output. I am using Xubuntu 14.04.

Does anyone have an idea how to solve this problem?

Answer 1

In devices with compute capability of at least SM35, we can apparently get around this by setting the environment variable

CUDA_DEBUGGER_SOFTWARE_PREEMPTION=1

We can see it at the cuda-gdb documentation page: http://docs.nvidia.com/cuda/cuda-gdb/#axzz4BrMPoaoW

Here's test. I am running on a Maxwell Quadro GPU:

nvidia-smi
Fri Jun 17 10:59:47 2016       
+------------------------------------------------------+                       
| NVIDIA-SMI 352.63     Driver Version: 352.63         |                       
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro M4000M       Off  | 0000:01:00.0      On |                  N/A |
| N/A   37C    P8     9W / 100W |    158MiB /  4087MiB |      1%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|    0      2981    G   /usr/bin/X                                      57MiB |
|    0      9186    G   ...ves-passed-by-fd --v8-snapshot-passed-by-    85MiB |
+-----------------------------------------------------------------------------+

Build and run the application

 nvcc -g -G foo.cu
 cuda-gdb ./a.out 

...
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
fatal:  All CUDA devices are used for display and cannot be used while debugging. (error code = CUDBG_ERROR_ALL_DEVICES_WATCHDOGGED(0x18)

Now set the environment variable.

export CUDA_DEBUGGER_SOFTWARE_PREEMPTION=1
cuda-gdb ./a.out
(cuda-gdb) r
...
warning: Cuda API error detected: cudaMemcpy returned (0xb)

warning: Cuda API error detected: cudaFree returned (0x11)

[Thread 0x7fffed3ff700 (LWP 10302) exited]
[Thread 0x7ffff7fc6780 (LWP 10293) exited]

Answer 2

For me it helped to change in the block

Section "ServerLayout"
  Identifier "layout"
  Screen 0 "intel"
EndSection

the Identifier from 'nvidia' to 'intel'.