[Mlir-commits] [mlir] [mlir][gpu] Skip address space checks for memrefs between launchFuncOp and kernel func (PR #102925)

[Fabian Mora]

fabianmcg wrote:

> but, as there are examples of address space mismatches between host and device in LLVM, I would reconsider the approach proposed by @kurapov-peter here.

Those mismatches actually have a `TODO` attached see: https://github.com/llvm/llvm-project/blob/main/mlir/lib/Dialect/GPU/IR/GPUDialect.cpp#L389-L394 
We shouldn't aggravate the situation.


> That's interesting. Probably not in gpu too? Like, for all the devices on a platform?

What I'm saying, is, if we add a 4th GPU address space: `unified`, that could fit your needs for the SPIR-V-like model.

https://github.com/llvm/llvm-project/pull/102925