[QST] What is the general process of developing and tuning GEMM kernels with CUTLASS?

[tiandi111]

Hi CUTLASS Community,

I'm a newbie to CUTLASS. I'm trying to use CUTLASS to develop multiple GEMM kernel templates and then tune them on different shapes to select the best one.

However, I'm confused with the APIs CUTLASS provided:

1. Which version of API should I use? I noticed that there're 2.x and 3.x API, is there any standard to choose the version?
2. Is the API dependent on Hardware architectures? I only saw Hopper examples with 3.x API, is it only applicable on Hopper?
3. The APIs from different layers, e.g. device and kernel, shares many similarities in terms of warp shape, instruction shape and etc. What's the difference between them?

Look forward to any reply, Thank you very much!