[OnnxToTorch] Casting float to integer should round to nearest for pow with int result type

[catswe]

Fixes #4091. I assume this will also need to be fixed for AtenPowScalarOp and AtenPowTensorScalarOp as well. I'm putting up a PR to ensure the initial approach is correct (new contributor :D ) before I put up another fix for AtenPowScalarOp and AtenPowTensorScalarOp.

[catswe]

It'd be great if you could review this @zjgarvey!

[zjgarvey]

Hey, this generally looks like good work, but I have a few notes:

1. The issue #4091 really pertains to the behavior of the TorchOnnxToTorch lowering.
2. Note the comment here , which indicates that the torch op only has an integer result type when both the base and exponent dtypes are integer types. This means that if we properly generate IR for AtenPowTensorTensorOp, we will never be in the situation covered by your current changes.
3. A proper resolution to the issue is likely to edit

   torch-mlir/lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp

   Lines 3074 to 3076 in 866786c

   	rewriter.replaceOpWithNewOp<Torch::AtenToDtypeOp>(
   	binder.op, resultType, pow, outTyConst, cstFalse, cstFalse, none);

   to also include something like:

pow = rewriter.create<AtenRoundOp>(loc, pow.getType(), pow);

(Note: AtenRoundOp lowers to an elementwise application of math::RoundEvenOp).

[catswe]

Thank you so much for the review! I put up changes. Embarrassing haha. I totally misunderstood where the issue was coming from 😅. The newly added test, test_pow_i32_f32_to_i32, is the same as test_pow_i32 but with mixed operand types to test for the original issue.

[zjgarvey]

Thank you so much for the review! I put up changes. Embarrassing haha. I totally misunderstood where the issue was coming from 😅. The newly added test, test_pow_i32_f32_to_i32, is the same as test_pow_i32 but with mixed operand types to test for the original issue.

All good! Thanks for the quick work. I'm running the CI now and if it turns green, I'll stamp and merge.

[catswe]

Just a quick ping so the PR doesn't get stale @zjgarvey

[zjgarvey]

Sorry for the delay! Thanks for the ping.