Cutlasss Functionality for SIMT


When I am trying to build my customized kernel in cutlass, I notice that there seems to be some shape options for the block, warp and thread, instead of giving any possible shape configurations.

I wonder if it is possible to change the code to make it work using SIMT with shape where m = 1, or other similar small m. I read through the code and still do not understand which part(computation? data movement?) on which level (warp? thread?)lexactly restrict me to do so.


Please direct your questions to the CUTLASS Github Issues