[Mlir-commits] [mlir] [mlir][nvgpu] update commit group and wait async ops (PR #130482)
lonely eagle
llvmlistbot at llvm.org
Fri Mar 14 12:01:55 PDT 2025
linuxlonelyeagle wrote:
> > > When you break SSA semantic like following, is reordering the OPs allowed?
> >
> >
> > >
> >
> >
> > >
> >
> >
> > > nvgpu.device_async_create_group %0
> >
> >
> > > nvgpu.device_async_wait { numGroups = 1 : i32 }
> >
> >
> > >
> >
> >
> > It cannot be reordered, `nvgpu.device_async_wait` should be similar to gpu.barrier.
>
> They should be reordered yes. But what stops the compiler to reorder them? When there was SSA semantic, the reordering wasn't possible.
I think they are not Pure Ops, like constant Op, it will be Pure Op, so by running -canonicalize, they will be reordered.
https://github.com/llvm/llvm-project/pull/130482
More information about the Mlir-commits
mailing list