What does “NVRM: Xid (0001:00): 13, 0002 00000000 000050c0 00000368 00000000 00000100” appeared after running a cuda program mean? A illegal address access? Is there any other kind of error which can cause such an error? I checked the accessed addresses in emurelease mode, which seemed no access beyond bound, but this error message always came out when i ran my program in release mode.
this error info appeared several times half a hour before, not for now. there is no any distinction between the code for emu mode and that for release, except that i add several ‘printf’ statements in code for emu. now my program can run without any error message coming out, but the output is wrong.
the CUDA module should be loaded, although i run a CUDA program in console, cause i am sure that X windows is initialized when the machine boots.
Note: i tried to add a script according to CUDA_Toolkit_Release_Notes for running CUDA program without X window when i installed CUDA toolkit, didn’t succeed.
there is no a stable means to reproduce this error message, and what i installed is CUDA_1.1. after having rectified a logic fault my program can run without any error info coming out, and result in a correct output in emu mode now, but the same error (NVRM: Xid:13) always turns out in release mode. there may still be another falut which i don’t find out, let me try to figure out, i will do what you suggest if i can’t solve this problem.
i found out what’s wrong with my program. a memory region shouldn’t be accessed, may be is private for some CUDA kernels, just like region on main memory begongs to some system kernels is read by my kernel. since data is read from an array in a block of 256 ints iteratively, region next to this array which is forbidden to be accessed is read with a ‘NVRM Xid: 13’ error coming outwhen the last block is access. as i know, there is no bound checking in c language on cpu, and i just read it then don’t do anything with this memory region, so i didn’t notice this may cause a illegal access.
anyway, there is nothing wrong with CUDA, just my fault, thank you netllama