[realy][pass]add split infer shape with convert op layout pass

[chengven027]

add SplitInferCorrectLayout.

[chengven027]

cc @masahi please review when you got time.

[masahi]

@chengven027-intellif I've just hit a bug with SplitInferCorrectLayout. The model I have does split op which splits along C dim in NCHW like this:

  %514 = multiply(%512, %513) /* ty=Tensor[(1, 256, 30, 17), float16] */;
  %515 = split(%514, indices_or_sections=[128], axis=-3) /* ty=(Tensor[(1, 128, 30, 17), float16], Tensor[(1, 128, 30, 17), float16]) */;

This is going to be a problem for x86 or other backend that uses NCHWc layout. Since the original splitting axis is blocked along two axes, SplitInferCorrectLayout needs to change both indices_or_sections and axis parameters. Right now our split op supports only integer axis, so this change could be difficult.

Unless you have a solution, I think we need to revert this PR. Or specially handle NCHWc layout by giving up layout fall-through (insert layout_transform) if the input layout is NCHWc.

You can reproduce this error by running https://gist.github.com/masahi/dce79525a5d0bd9c8a5d04992c060ab0

[chengven027]

ok, thanks. Maybe I didn't consider this situation.

[alter-xp]

This is going to be a problem for x86 or other backend that uses NCHWc layout.

hi @chengven027-intellif @masahi, The same problem occurred to my model. So in addition to disabling this part of the code, is there a good solution to this?

[masahi]

There is no easy solution, we should revert this.