common errors while programming CUDA

1. Unknown Error (often doesn't show up unless you call cutilSafeCall)

   • Writing out of bounds: Check that each of your arrays has been initialized to appropriate sizes. ESPECIALLY SHARED MEMORY.

2. Driver crash (Windows 7) or Frozen screen (XP and before)

   • Infinite loop in the code: Make sure all code is guaranteed to terminate.
   • Trying to read from unallocated memory: Make sure every location you access has been allocated appropriately.
   • Too many printfs
   • Deadlocks due to locks / atomics

3. Registers don't retain values, read NaN

   • Conflicting CUDA drivers: Reinstall GPU drivers and use appropriate CUDA SDK. 
     Known problem with recent (Oct 2011) CUDA drivers. Rolling back to 260 driver helped Stanley Tzeng.

4. Kernel Launch Failure

   • Incorrect grid / block size
   • Uninitialized dynamic shared memory or incorrect size (thanks Kerry!)

5. Insufficient Resources Requested

   • Too much shared memory

6. Constant memory reads incorrect values / zeroes

   • Do not use &var in cudaMemcpyToSymbol; just use var.

7. Weird CPU-GPU (PCIE) memory transfer rates (decreasing with data size, too low, etc)

   • Try defaulting to stream ID other than 0, and see if that works (verified by Pinar on a GeForce GT 555M Alienware M14x. Thanks!)

8. nvcc fatal : A single input file is required for a non-link phase when an output file is specified (when compiling using Visual Studio)

   • Remove quotations from the nvcc include directories list
   • Taken from here.