Adding ops without e2e support

[henrytwo]

Hi @silvasean, we're now able to trace a Huggingface BERT model with full forward and backwards graph, but we had to add support for a few ops to torch-mlir.

Looking at some existing PRs (e.g. https://github.com/llvm/torch-mlir/pull/570) it seems that adding support for an op usually comes with type refinement, lowering to linalg, and e2e tests (and the e2e tests won't work without the type refinement).

Since we don't need any of this other stuff for LTC, would a PR with only updates to torch_ods_gen.py and GeneratedAtenOps.td be acceptable?

cc: @antoniojkim @ke1337

[silvasean]

Yeah, I guess that would be ok.

[silvasean]

Can you list which ops are missing so we can add them to the e2e todo list?

CC @ramiro050 who is tracking that TODO list.

[henrytwo]

Can you list which ops are missing so we can add them to the e2e todo list?

CC @ramiro050 who is tracking that TODO list.

• aten.native_layer_norm_backward
• aten.embedding_dense_backward
• aten.native_batch_norm_backward
• native_dropout
• native_dropout_backward

[silvasean]

Looks like the ops were added.