--maxrregcount: sm_10 VS sm_20 If --maxrregcount too low, nvcc aborts for sm_10, continues for sm_20


When using --maxrregcount to limit the number of registers used by a kernel it is possible to select too low a value. When this happens and one is targeting sm_10, sm_11, sm_12, or sm_13 the compiler aborts. However, when sm_20 is the target, the given value is ignored and the minimum number of registers is used.

Will this behaviour be made consistent in the future?

OS: Windows 7 64-bit
nvcc -V: Cuda compilation tools, release 3.0, V0.2.1221.