I have two GPUs, one is AMD (fglrx-updates) and one nVidia (nvidia-current), drivers from x-swat PPA. The AMD card is used byt xorg (without problems). Both devices show up in lspci|grep VGA:
01:00.0 VGA compatible controller: ATI Technologies Inc Device 6718
02:00.0 VGA compatible controller: nVidia Corporation Device 1200 (rev a1)
The nvidia kernel module is loaded (dmesg|grep nvidia):
[ 16.033488] nvidia 0000:02:00.0: enabling device (0000 -> 0003)
[ 16.033501] nvidia 0000:02:00.0: PCI INT A -> GSI 40 (level, low) -> IRQ 40
[ 16.033507] nvidia 0000:02:00.0: setting latency timer to 64
OpenCL problem
I created /etc/OpenCL/vendors/nvidia.icd containing
/usr/lib/nvidia-current/libcuda.so
The device is detected, platform NVIDIA CUDA, device GeForce GTX 560 Ti. However, no code gets compiled for the device, clBuildProgram always returns CL_BUILD_RPOGRAM_FAILURE and querying clGetProgramBuildInfo crashes with uninformative backtrace
#0 0x00007ffff7442bef in ?? () from /lib/x86_64-linux-gnu/libc.so.6
#1 0x00007ffff22c0e67 in ?? () from /usr/lib/nvidia-current/libcuda.so
The program is build with cl.h from Intel's SDK; that is OK since everything was fine until I installed the AMD card alongside.
Running /usr/share/doc/python-pyopencl/examples/benchmark.py crashes just the same (in pyopencl::program::get_build_info).
Any hint on what is wrong?
The implementation probably can't dlopen one of the libraries it uses, try running your app with
LD_LIBRARY_PATH=/usr/lib/nvidia-current:$LD_LIBRARY_PATH ./yourapp, and if that doesn't help run it with strace, and look out for libraries it's unable to load and adjust the path so it can find them.