What is the exact meaning of `waves per SM`?

[cuda_new_bird]

Thank you. I’ll read it. I’m facing some imbalance issue these days.

Could you please read this post? Thank you!

Is there a way to inspect the time cost of each individual cuda block?

I’m profiling the kernel of flash attention. In this kernel, each cuda block process different workload. Some cuda blocks exit early, but some has to do a lot of work. I want to know how imbalanced the kernel is. Is there a way to inspect this?