tcgen05.mma.ws

[jeromeku]

What is your question?

What are the use cases for tcgen05.mma.ws (ptx and cccl bindings)?

How does this differ from conventional tcgen05.mma which is what Cutlass uses for sm100+ mma?

[ahendriksen]

.ws stands for weight-stationary. It allows reuse of the A operand. It's not typically used by GEMMs as the weight is not stationary in that case (both A and B matrices are iterated through).

[jeromeku]

I understand that ws stands for weight-stationary as is clearly stated in the PTX documentation.

My question is more along the lines of use cases for this specific instruction type. I.e., are there any documented / undocumented uses of this instruction type that people have seen in the wild or in any NVIDIA library?

[ahendriksen]

I cannot think of one of the top of my head. Any use case that requires repeatedly multiplying the same matrix A with a varying matrix B would benefit from this instruction.