[Mlir-commits] [mlir] [mlir][nvgpu] Mark TMA descriptor as MemWriteAt in `tma.async.store` (PR #79427)

[Mehdi Amini]

joker-eph wrote:

I have some questions about the `NVGPU_TensorMapDescriptor` type in MLIR.
Does it map to a `CUtensorMap` object?  Or to a pointer to a `CUtensorMap` object?

In Cuda, the `CUtensorMap` object is setup on the host, and then **copied** to the GPU memory. Inside the kernel the PTX instruction are using a `CUtensorMap *`.

I am afraid that MLIR is sometimes using `NVGPU_TensorMapDescriptor` as a  `CUtensorMap` and other times as a  `CUtensorMap *`

https://github.com/llvm/llvm-project/pull/79427