MaheshRavishankar

+1 to having constant group count. IIUC you might need to do reshapes to expand dimensions, and `tensor.expand_shape` does not allow you to expand a single dynamic dimension into multiple...

I am not sure I follow (a) fully. If the case is 3D LHS, 2D RHS, I would expect it to be lowered as 1) Broadcast the 2D RHS to...

Hey @chelini thanks for starting on this. I didn't get around to review it today. I'll take a look tonight or tomorrow morning for sure.

> > Thanks @chelini . Have a few comments. I am happy to work off of this. Let me know. > > Hi @MaheshRavishankar if you don't mind I will...

> > I do suspect that we only need the tensor version of the operation eventually, but we can make that determination later. > > Yes, I think eventually we...

This is actually wrong. Ill fix it.

@antiagainst the pattern to detach the `outs` from named ops should probably be dropped. I looked at the original issue that led to this work. That should have been fixed...

It would be preferable not to do that. AFAICS, this is an issue with RISC-V only. Having such escape hatches for individual cases will result in a death by thousand...

To ask another way, lets say we dont fuse, there is still a gather op which is not vectorized... that is still going to be almost as expensive as the...

> > AFAICS, this is an issue with RISC-V only. > > This is not a RISC-V specific issue. This is a limitation in the vectorizer that impacts all the...