[Mlir-commits] [mlir] [mlir][gpu] Adding the ELF section to the NVVMTargetAttr, to propagate to the gpu.binary Op (PR #119440)

[Fabian Mora]

fabianmcg wrote:

> The `SelectObjectAttr` does everything we need. Having our own implementation of `OffloadingLLVMTranslationAttrInterface` would lead to a lot of cut-and-paste.

The reason I mention a new attribute, is because a new attribute would be able to handle cuda binary registration for using CUDA RT.

> Would this mean that I would have to replace the gpu.binary op with a new one?

No, you would only need to update the attributes.

> The blobs attached to them can be huge. Are we confident we will not be duplicating memory passing one to the next?

It shouldn't `#gpu.object` stores a string attribute to the binary precisely to avoid duplicating the blob https://github.com/llvm/llvm-project/blob/main/mlir/include/mlir/Dialect/GPU/IR/CompilationAttrs.td#L226


https://github.com/llvm/llvm-project/pull/119440