Relax dynamic restriction on ScatterND stablehlo conversion

[srcarroll]

The previous implementation asserted that all dims of the indices operand were static. First of all, assertion isn't appropriate here as the definition of the op allows dynamic. This is a condition for the pattern match and thus should fail the match with notifyMatchFailure. Second, the implementation only relies on the last dim being static, so we can relax the match condition and only fail if last dim is dynamic.

[jenkins-droid]

Can one of the admins verify this patch?

[srcarroll]

I'll be honest. I don't completely understand the lowering of ScatterND so I'm not sure if the result is correct. But I just noticed that the assertion was too restrictive based on the fact that only the last dim of indices needed to be static. So I'm assuming everything else is correct when relaxing the restriction. Please confirm.

[AlexandreEichenberger]

@jenkins-droid test this please

[jenkins-droid]

Can one of the admins verify this patch?

[AlexandreEichenberger]

@jenkins-droid test this please

[AlexandreEichenberger]

LGTM, feel free to get a second opinion from an HLO expert.

[AlexandreEichenberger]

@srcarroll let me know when you are ready for me to push the code.

[Connor-XY]

I'll be honest. I don't completely understand the lowering of ScatterND so I'm not sure if the result is correct. But I just noticed that the assertion was too restrictive based on the fact that only the last dim of indices needed to be static. So I'm assuming everything else is correct when relaxing the restriction. Please confirm.

The assertion is restrictive since I would like to provide a basic implementation for static-shaped tensors first, and fail dynamic ones temporarily. I believe that you are now supporting dynamic cases and we could remove some restrictions gradually. As for the ScatterND op, it indeed only requires the last dim of indices to be static, and your test case should work as I don't see any problem with it.

[srcarroll]

I'll be honest. I don't completely understand the lowering of ScatterND so I'm not sure if the result is correct. But I just noticed that the assertion was too restrictive based on the fact that only the last dim of indices needed to be static. So I'm assuming everything else is correct when relaxing the restriction. Please confirm.

The assertion is restrictive since I would like to provide a basic implementation for static-shaped tensors first, and fail dynamic ones temporarily. I believe that you are now supporting dynamic cases and we could remove some restrictions gradually. As for the ScatterND op, it indeed only requires the last dim of indices to be static, and your test case should work as I don't see any problem with it.

@Connor-XY Thanks for the confirmation! @AlexandreEichenberger you can merge. I'm rebasing now just in case

[jenkins-droid]

Can one of the admins verify this patch?

[AlexandreEichenberger]

@jenkins-droid test this please

[jenkins-droid]

Jenkins Linux s390x Build #14626 [push] Relax dynamic restrictio... started at 11:37

[jenkins-droid]

Jenkins Linux amd64 Build #14596 [push] Relax dynamic restrictio... started at 10:37

[jenkins-droid]

Jenkins Linux ppc64le Build #13621 [push] Relax dynamic restrictio... started at 11:47

[jenkins-droid]

Jenkins Linux amd64 Build #14596 [push] Relax dynamic restrictio... passed after 1 hr 3 min

[jenkins-droid]

Jenkins Linux s390x Build #14626 [push] Relax dynamic restrictio... passed after 1 hr 24 min

[jenkins-droid]

Jenkins Linux ppc64le Build #13621 [push] Relax dynamic restrictio... passed after 1 hr 59 min