Question

I am stumped by what seems to be a problem that should be easy to diagnose and fix. I have a C++ source file that makes use of the CUDA Driver API. When I compile it using nvcc, the executable is produced and works without problems. However, when I try to compile it using g++, the linker complains about undefined references to symbols that are in /usr/lib/libcuda.so:

undefined reference to cuDeviceGet'
... undefined reference tocuDeviceComputeCapability'
... undefined reference to cuDeviceGetName'
... undefined reference tocuDeviceTotalMem_v2'
... undefined reference to cuDeviceGetAttribute'
... undefined reference tocuDeviceGetAttribute'
... undefined reference to cuDeviceGetAttribute'
... undefined reference tocuDeviceGetAttribute'

I tried to ensure that /usr/lib is included in the library path (by supplying the flags -L/usr/lib, -lcuda, and adding /usr/lib to $LIBRARY_PATH). However, the problem still persists. I am sure that /usr/lib/libcuda.so contains the symbols that the linker is complaining about:

$ nm -D /usr/lib/libcuda.so | grep cuDeviceGet
00000000000be410 T cuDeviceGet
00000000000bf120 T cuDeviceGetAttribute
00000000000d2e10 T cuDeviceGetByPCIBusId
00000000000be200 T cuDeviceGetCount
00000000000bdfa0 T cuDeviceGetName
00000000000d2bb0 T cuDeviceGetPCIBusId
00000000000bf380 T cuDeviceGetProperties

I find it strange that nvcc successfully compiles the file, but g++ does not. Do you have any idea as to how I can fix this problem?

Thank you very much for your help!

Answer 1

Ok, the reason I was not able to compile the file was because the -lcuda flag has to come after the filename. Otherwise, it won’t work… a large, prominent warning with the compiler telling me this information would have been much appreciated. Well, that was a waste of two hours.