You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We only have two ops in linalgx: pack and unpack. These ops have been in IREE for a while, but we need to make sure they make it to linalg in MLIR, or we'll be stuck with this forever.
We need to create a few examples on pure linalg code and just push those ops into linalg itself as an upstream patch on Phabricator. This will give more people, outside of IREE, visibility on what we're trying to do and help us merge this soon.
I'm hoping @nicolasvasilache can help us here. We're not aiming at perfect, we're just aiming are minimally acceptable, so that we can iterate upstream rather than carrying this with us all the time.
Even if the semantics isn't completely clear, we can just add a bunch of asserts to make sure only the things we can prove are done, everything else fails validation. In time, we can add more functionality upstream once we're clear on what is needed and how to implement it.
Waiting until we have a complete picture won't work, because we need to implement things to have a better picture and right now we can't implement all the things that everyone may one day want, because we don't know what it is, nor have the need for those things.
The text was updated successfully, but these errors were encountered:
We only have two ops in
linalgx
:pack
andunpack
. These ops have been in IREE for a while, but we need to make sure they make it to linalg in MLIR, or we'll be stuck with this forever.We need to create a few examples on pure
linalg
code and just push those ops intolinalg
itself as an upstream patch on Phabricator. This will give more people, outside of IREE, visibility on what we're trying to do and help us merge this soon.I'm hoping @nicolasvasilache can help us here. We're not aiming at perfect, we're just aiming are minimally acceptable, so that we can iterate upstream rather than carrying this with us all the time.
Even if the semantics isn't completely clear, we can just add a bunch of asserts to make sure only the things we can prove are done, everything else fails validation. In time, we can add more functionality upstream once we're clear on what is needed and how to implement it.
Waiting until we have a complete picture won't work, because we need to implement things to have a better picture and right now we can't implement all the things that everyone may one day want, because we don't know what it is, nor have the need for those things.
The text was updated successfully, but these errors were encountered: