Added support for torch arange float module

[Abhishek-TyRnT]

Added Support for float dtype in in torch.arange in TOSA Dialect

This resolves the following issue :-
#2762

The following test cases are passing after this change

1. ArangeDtypeIntModule_basic
2. ArangeFloatModule_basic
3. ArangeNegativeStartFloatModule_basic
4. ArangeStartFloatModule_basic
5. ArangeStartNegativeStepFloatModule_basic
6. ArangeStartOutDtypeModule_basic
7. ArangeStartStepFloatModule_basic

[eric-k256]

Looks good, thanks!

[Abhishek-TyRnT]

CI doesn't seem to pick up jobs... Is there any reason?

[newling]

This isn't going to work for large 64-bit integers, because casting to double will be lossy (mantissa of double is 52 bits, so 2^53 + 1 can't be represented as a double).

[Abhishek-TyRnT]

You are right. I will try to find a way to fix this.

[Abhishek-TyRnT]

I guess there will be no loss if we use long double instead of double. What do you think?

[newling]

Looks fine, just the final comment for the int case.

[newling]

The logic looks good to me. I think the code can be polished a little bit though :-)

[Abhishek-TyRnT]

Hi @newling. The int64 test suddenly started failing cause mlir couldn't typecast float32 or float64 to int64. So I manually typecasted them Torch-to-Tosa pass.

[Abhishek-TyRnT]

The CI failure doesn't seem it to be a issue with TorchToTosa but something in stablehlo unexpectedly passing.

[newling]

Ok. You might need to update the xfail set? #2927
Have you rebased onto main?

[Abhishek-TyRnT]

I just checked, that test has been added there. It should hopefully pass now

[Abhishek-TyRnT]

Now that it has passed. Can you please merge it. @newling

[newling]

Hi @Abhishek-TyRnT I've made a PR against this which should fix the link checks. It also makes some stylistic improvements (IMO) which you can consider pulling in if you like: Abhishek-TyRnT#1