You can study what the compiler is doing by studying the SASS code. The cuobjdump tool can be used for that, or even godbolt
I’m of the strong opinion that bank-conflicted access (or not) can be deduced from the source code. The idea that the compiler could/would do something contrary to that is doubtful to me, although I have been wrong before.
I do not think it is strictly possible with 100% accuracy in all cases to determine the presence or absence of shared memory bank conflicts by studying a single metric. At least partly based on this, and partly based on the observations I have made myself from time to time in the past that clearly un-bank-conflicted code shows up as bank conflicts in one metric or another.