I also recall seeing that message in the release notes a while back, but since it is not mentioned in the CUDA 4.1 or 4.2 release notes (I have not checked 4.0), I assumed this was no longer an issue. Could it be that since the new CUBLAS interface passes the scaling values (alpha in the case of ZGEMM for example) as a call by reference parameter, that this is the reason that it is not mentioned. The legacy API still uses call by value parameters.
One thing that makes me wonder if the launch failures I am experiencing (see viewtopic.php?f=2&t=536
) is not simply due to compiler incompatibility is that they occur on Windows as well. Am I right in assuming that the Windows CUDA binaries are not built using GCC?
I will do a test or two here and see if this is indeed a problem.