I have CUDA host and kernel codes which were developed on a Linux system and work there. I am porting them to a Win 7 system in order to use the Nvidia Nsight debug and analysis tools.
The codes do not run on the Win 7 system; at some point a request for a cudaMalloc or a kernel call returns an error code which translates as: cudaErrorInvalidResourceHandle. (the kernel that fails is the first of the kernel calls; setting a breakpoint at the first line of the kernel with Nsight fails to trigger — the error happens before the kernel actually gets invoked.)
The program creates a single stream and does not create any events so the only resource handle has to be the one associated with the stream but the stream handle does not appear corrupted. I have both a Fermi and a Kepler board in my system so i can check runtimes on each so I have the latest Nvidia driver loaded [9.18.13.448 (6-11-2012) – nvoglv64][304.48]. I can duplicate this problem on a second system.
Has anyone ever seen this error message before ? Any insight into what to check? Normally I would comment out great chunks of the program to isolate the guilty line but this program is wildly complex (can you say spaghetti) and the cuda calls come from all over. so before i start butchering the program I thought i would ask this community if this error has ever come up. I’m also considering that it could be a driver issue but with Nvidia forums closed down i have no access to Nvidia.
thanks for any thoughts.
JM
Problem seemed to be mismatch between nvidia driver and cuda version. dumbed down nvidia driver to cuda version and the problem disappeared.