Fix infercorrect layout in Layoutrewrite and improve naming.

[lazycal]

This model

def @main(%x: Tensor[(1, 1, 1, 1), float32] /* ty=Tensor[(1, 1, 1, 1), float32] */) -> Tensor[(1, 1, 1), float32] {
  %0 = image.resize2d(%x, size=[2, 4], roi=[0f, 0f, 0f, 0f], rounding_method="") /* ty=Tensor[(1, 1, 2, 4), float32] */;
  %1 = mean(%0, axis=[3]) /* ty=Tensor[(1, 1, 2), float32] */;
  image.resize1d(%1, size=[1], roi=[0f, 0f], rounding_method="") /* ty=Tensor[(1, 1, 1), float32] */
}

fails with error Check failed: (tir::BijectiveLayout(new_src_layout, dst_layout).defined()) is false: Cannot insert layout transform because there are inconvertible layouts: NCW v.s. NCH. This should be a bug introduced in #10118. This line 7805d6f#diff-c3bb97b7e813244d6dae39925cc10bb9d50e8b7a280556b423068b3705f9f114R363 fixes it.

In short, there should be no layout rewriting here since no conv etc. are used. But when the pass is checking resize1d, the previous code do not pass consistent layouts for the argument old_in_layouts for the two calls of ResizeInferCorrectLayout (the first time NCH from old_in

tvm/src/relay/transforms/transform_layout.h

Line 341 in ef08c36

InferCorrectLayouts(ref_call, Array<Layout>(nullptr), old_in, types);

, second time NCW from old_in_2

tvm/src/relay/transforms/transform_layout.h

Line 363 in ef08c36

std::tie(infer_out, success) = InferCorrectLayouts(new_call, new_in_tmp, old_in2, types);

), misleading the optimizer into thinking that the layout has changed and hence an ill-formed layout_transform is inserted.

This PR also improves the naming based on suggestions from @masahi in #10118, as well as adds several comments and logging statements to help people understand. (Hope this helps you understand that change better too @masahi :-)).

[lazycal]

To reproduce the problem, run the following code:

import tvm
from tvm import relay
from tvm.relay.transform import InferType


@tvm.instrument.pass_instrument
class PrintIR:
    """Print the name of the pass, the IR, only before passes execute."""

    def __init__(self, print_mod=True, show_meta_data=False) -> None:
        self.pass_cnt = 0
        self.print_mod = print_mod
        self.show_meta_data = show_meta_data

    def run_before_pass(self, mod, info):
        with tvm.transform.PassContext(instruments=[]):
            global prev_mod
            if self.print_mod:
                print(relay.transform.InferType()(mod))
            print('>' * 40, f'Running Pass#{self.pass_cnt}:', info)

        self.pass_cnt += 1


x = relay.var("x", shape=(1, 1, 1, 1))
y = relay.image.resize2d(x, (2, 4))
z = relay.mean(y, axis=3)
a = relay.image.resize1d(z, (1,))
func = relay.Function((x,), a)
mod = tvm.IRModule.from_expr(func)
print(InferType()(mod))

with tvm.transform.PassContext(opt_level=4, instruments=[PrintIR()]):
    relay.create_executor("graph", mod).evaluate()
print('pass')