Michelle <michelle.flanner(a)gmail.com> writes:
I have a muddy situation on my Ubuntu 14.04.
I have the AMDAPP sdk installed. I also have various nvidia drivers (352 is
in use) installed and I have cuda installed.
I'm trying to get pyopencl to use the nvidia gpu.
Running example opencl files from inside cuda works--ie, it finds the nvidia
device and can create a context. I can't get pyopencl to find the device.
My siteconf.py has the directories where the cuda includes and libOpenCL.so
reside. I looked at etc/vendors/nvidia.icd and it has libOpenCL.so but there
are at least two places where it could find libOpenCL.so, inside the AMD sdk
and the cuda, and I'm not sure how to tell it to go to the cuda one. Should I
create a symbolic link and point it to the cuda one, and put the name of the
symbolic link in the .icd? I also have a bunch of libOpenCL.so.1, .1.0,
.1.0.0 in various places, so I'd have to give it a unique name. I tried
putting the full path to the right .so but still get PLATFORM_NOT_FOUND_KHR.
/etc/vendors/nvidia.icd should *not* point at libOpenCL.so. It should
point at the actual OpenCL implementation. On my Nv system, that's
(but that file is probably somewhere different on your machine)