memory bank conflict check in new cuda versions

In previous versions of cuda, there was a CUT_BANK_CHECKER for memory conflicts [1]. I don’t see much details about that in the programming guides which means it is deprecated.

On the other hand, I searched for memory bank conflict in newer versions of cuda, but didn’t find a function or macro for that [2]. What does that mean? Does that mean it no longer necessary to do that check? or the name has been changed in newer versions?

If somebody knows, please leave a reply.

[1] https://devtalk.nvidia.com/default/topic/377455/what-is-cut_bank_checker
[2] https://stackoverflow.com/questions/3841877

As I recall, CUT was an unsupported collection of helper functions in early versions of CUDA, i.e. this was never part of CUDA proper. It has long since disappeared. A quick read of the linked thread seems to indicate that the bank checker was tied to a basic host-based CUDA simulator that shipped as a debugging aid prior to the introduction of a proper debugger around 2008 or 2009.

These days, the CUDA profiler can tell you about shared memory bank conflicts, so I suggest you give that a try if you haven’t already.

From your statement and what I found (although little) on the web is that, the bank conflict check is now used for tuning the performance. Therefore, removing the CUT_BANK_CHECKER lines in the old codes won’t affect the correctness of the execution. Am I right?