I looked for some nvcc BTS but I found not, only a note on this forum, that in case of bug found devs should be post bug reports here. I hope it makes sense.
In my case it is atomic add on shared memory with CC 1.1. It is not supported, but instead of error stating “operation in line 73 is unsupported with this CC. You need at least CC X.Y” I see such error message:
Error 1 error : Ptx assembly aborted due to errors
It is hard not to admit it is pretty meaningless.
What I did, install CUDA 5.0 SDK, install VS, create CUDA project, at the top add such code:
__global__
void foo()
{
__shared__ unsigned int x[1];
atomicAdd(&(x[0]),0);
}
And compile. Observe the error.