CuTe and Tensor Core

[osayamenja]

This may be a naive question, so please feel free to direct me to documentation or code that resolves it.

TL;DR

Does CuTe gemm with TiledMMA leverage the tensor core?

MMA Instructions

Per the CUDA C++ programming guide, wmma targets the tensor core via warp level MMA.

Also, the PTX ISA mentions both wmma and mma under warp-level MMA but does not explicitly say that the latter instruction targets tensor cores.

CUTLASS uses wmma and mma for the cutlass namespace (see here) and only mma for cute (see here), leading to the question I asked above.

Also, tensor core examples available use cutlass only and there are none using cute, so my understanding is that maybe this is the case because there is no support?

Please help me clarify.

[thakkarV]

Yes. Using the mma ptx targets tensor cores.

[osayamenja]

Thank you! @thakkarV