My app BarsWF was performing well on variety of different computers with multiple GPUs (up to 8). But in 1 case it constantly crashes with “Invalid device pointer” message right at the beginning of kernel execution.
Hardware: 8800GT and 8400GS (256mb ram). App uses less than 20-30Mb of ram.
According to documentation cudaErrorInvalidDevicePointer has nothing to do with kernel execution itself but rather it’s a memory allocation/copying problem. Are you checking return codes for every cudaMalloc() / cudaMemcpy()?
BTW, I’ve had same problem with your program on 8600GTS + 8800GT. With single 8600GTS it works fine.