add aten.fill converter

[shivadbhavsar]

Using masked_fill migraphx op for now. Can be updated to use a new fill op if migraphx supports it in the future.

[CharlieL7]

I think we actually have gpu fill in migraphx after Paul's PR here: https://github.com/ROCm/AMDMIGraphX/pull/2908

[shivadbhavsar]

I think we actually have gpu fill in migraphx after Paul's PR here: ROCm/AMDMIGraphX#2908

So it looks like hip::fill takes value as an attribute. Since the fill value here can be a tensor (ie. output of some other non constant node) it wouldnt work for that case I think. It also seems to take a buffer allocation instruction as input. Not sure we want to add that sort of low level instructions to a front end parser. It also means the parsed program cant be compiled using ref

[bpickrel]

Is there a functional difference between filling an array with a where op under the hood vs. creating a literal with a single value and then using multibroadcast and then contiguous to allocate and copy it, as we do in various other places in torch_migraphx?

[shivadbhavsar]

Is there a functional difference between filling an array with a where op under the hood vs. creating a literal with a single value and then using multibroadcast and then contiguous to allocate and copy it, as we do in various other places in torch_migraphx?

I am assuming you're talking about the add_literal call right? The difference is that the input value to the where op can be non-constant (ie. the single value is not known at compile time) and so it cannot be added as a literal. Here the fill op can have a constant value, and in that case migraphx will just const fold the inserted where op anyways, and so we can just insert it this way without worrying about whether value is constant or not.