Align with cutlass_profiler GEMM arguments

[youliangh]

Hi, I've been trying to set up the FlashOverlap on my local device, and it's easy to finish the compilation part. When it comes to Generate GEMM configuration, things start to get tricky. The README of this repo is extremely insufficient for me to get over this step.

The script used to generate GEMM instances, ie, "generate_instances.py", contains a group of cutlass arguments. However, these arguments differ from arguments passed to cutlass_profiler, e.g., WarpM (FlashOverlap) against warps_m (cutlass_profiler). The differences can confuse people who are not familiar with cultass_profiler.

In addition, should I keep the candidates in "generate_instances.py" the same with those arguments that passed to cutlass_profiler?

I believe authors could update the README to provide clearer instructions regarding the cutlass_profiler part to set up FlashOverlap.

[hkeee21]

Hi, thanks for the advice.

The execution of "generate_instance.py" should be performed before compiling, and the process is automatic, requiring no manual efforts.

And for the Generate GEMM configuration part, the CUTLASS profiler is utilized to generate the GEMM tiling configs given a specific shape. We use "tune/gen_config.py" to handle the output .csv file generated by the CUTLASS profiler, and the arguments can be referenced at https://github.com/infinigence/FlashOverlap/blob/main/tune/gen_config.py#L15.

I understand the inner arguments (WarpM, ThreadBlockM, ...) are inconsistent with those used by the CUTLASS profiler (warps_m, cta_m, ...). However, the inner arguments in templates follow the style of the CUTLASS examples (e.g., using WarpShape and ThreadBlockShape).

I'll work on the README afterwards. Hope the above explanation helps!