packed_or_broadcasted: gpu::mlir_op: Shapes are not packed nor broadcasted

[shivadbhavsar]

Torch benchmarks fail. Affected models: TIMM: xcit_large_24_p8_224 HF: MobileBertForQuestionAnswering, MobileBertForMaskedLM

[CharlieL7]

~Issue is with find_reduce_reduce matcher. Sometimes it gives the mentioned error, sometime it's a seg fault.~ Investigating what is the cause of the issue. Error is actually in gpu::compile_ops. Something goes wrong when merging reduce -> reduce. into a seg fault.

[pfultz2]

Does #3039 fix this issue?

[CharlieL7]

#3039 does get around the packed_or_broadcasted error. Now has a seg. fault in gpu::compile_ops

[CharlieL7]

#3070 fixes the seg fault