[AIROCMLIR-612] Lower migraphx.convert into linalg.generic

[Mr-Anyone]

Motivation

There is a nice helper in arith/Utils/Utils.h that converts from type A to type B. In the case of integer, we always assume it is signed. See AIROCMLIR-595

Technical Details

Added a linalg.generic loop elementwise converter that converts type type A to type B. The emitted IR looks like the following:

Starting from

func.func @func_convert_i32_to_f32(%arg0: !migraphx.shaped<4x8xi32, 8x1>) -> !migraphx.shaped<4x8xf32, 8x1> {
  %0 = migraphx.convert %arg0 : <4x8xi32, 8x1> to <4x8xf32, 8x1>
  return %0 : !migraphx.shaped<4x8xf32, 8x1>

to

    %1 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel", "parallel"]} ins(%expanded : tensor<4x8xi32>) outs(%0 : tensor<4x8xf32>) {
    ^bb0(%in: i32, %out: f32):
      %2 = arith.sitofp %in : i32 to f32
      linalg.yield %2 : f32
    } -> tensor<4x8xf32>

Test Plan

Lit test

Test Result

Passed lit test

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[dhernandez0]

sorry I'm late to the review, We have always rewritten migraphx.convert to tosa.cast, so migraphx expects tosa's way of casting. We are having some issues on rocmlirTriton (https://github.com/ROCm/rocmlirTriton/pull/125) related to this. We use tosa spec for converting (tosa.cast), so the logic they follow is not as simple as arith::FPToUIOp. https://github.com/ROCm/rocmlirTriton/pull/125#discussion_r3043866753

[dhernandez0]

we probably want to sync with migraphx team if we really want to follow tosa's spec for casting. My intuition is that it's better to follow what ONNX/pytorch do instead (which might be the same as tosa or not).