Transpose scheduler fails with Resize ops

[jacobhinkle]

The following fails in TransposeScheduler::canScheduleRuntime using the #397 branch:

import torch
from nvfuser import FusionDefinition, DataType

def nvfuser_fusion_id0(fd : FusionDefinition) -> None :
    T0 = fd.define_tensor(symbolic_sizes=[-1, -1, -1], contiguity=[True, True, True], dtype=DataType.Float, is_cpu=False)
    T1 = fd.define_tensor(symbolic_sizes=[-1, -1, -1, -1], contiguity=[True, True, True, True], dtype=DataType.Float, is_cpu=False)
    T2 = fd.ops.permute(T1, dims=[0, 2, 1, 3])
    T3 = fd.ops.slice(T0, start_indices=[0, 0, 0], end_indices=[1024, 64, 2], strides=[1, 1, 1])
    T4 = fd.ops.reshape(T3, original_shape=[1024, 64, 2], new_shape=[1, 1024, 1, 64, 2])
    T5 = fd.ops.slice(T4, start_indices=[0, 0, 0, 0, 0], end_indices=[1, 1024, 1, 64, 1], strides=[1, 1, 1, 1, 1])
    T6 = fd.ops.reshape(T2, original_shape=[8, 1024, 32, 128], new_shape=[8, 1024, 32, 64, 2])
    T7 = fd.ops.slice(T6, start_indices=[0, 0, 0, 0, 0], end_indices=[8, 1024, 32, 64, 1], strides=[1, 1, 1, 1, 1])
    T8 = fd.ops.squeeze(T7, original_shape=[8, 1024, 32, 64, 1], dims=[4])
    T9 = fd.ops.squeeze(T5, original_shape=[1, 1024, 1, 64, 1], dims=[4])
    T10 = fd.ops.broadcast_in_dim(T9, output_shape=[8, 1024, 32, 64], broadcast_dims=[0, 1, 2, 3])
    T11 = fd.ops.mul(T8, T10)
    T12 = fd.ops.slice(T6, start_indices=[0, 0, 0, 0, 1], end_indices=[8, 1024, 32, 64, 2], strides=[1, 1, 1, 1, 1])
    T13 = fd.ops.squeeze(T12, original_shape=[8, 1024, 32, 64, 1], dims=[4])
    T14 = fd.ops.slice(T4, start_indices=[0, 0, 0, 0, 1], end_indices=[1, 1024, 1, 64, 2], strides=[1, 1, 1, 1, 1])
    T15 = fd.ops.squeeze(T14, original_shape=[1, 1024, 1, 64, 1], dims=[4])
    T16 = fd.ops.broadcast_in_dim(T15, output_shape=[8, 1024, 32, 64], broadcast_dims=[0, 1, 2, 3])
    T17 = fd.ops.mul(T13, T16)
    T18 = fd.ops.sub(T11, T17)
    T19 = fd.ops.slice(T6, start_indices=[0, 0, 0, 0, 1], end_indices=[8, 1024, 32, 64, 2], strides=[1, 1, 1, 1, 1])
    T20 = fd.ops.squeeze(T19, original_shape=[8, 1024, 32, 64, 1], dims=[4])
    T21 = fd.ops.slice(T4, start_indices=[0, 0, 0, 0, 0], end_indices=[1, 1024, 1, 64, 1], strides=[1, 1, 1, 1, 1])
    T22 = fd.ops.squeeze(T21, original_shape=[1, 1024, 1, 64, 1], dims=[4])
    T23 = fd.ops.broadcast_in_dim(T22, output_shape=[8, 1024, 32, 64], broadcast_dims=[0, 1, 2, 3])
    T24 = fd.ops.mul(T20, T23)
    T25 = fd.ops.slice(T6, start_indices=[0, 0, 0, 0, 0], end_indices=[8, 1024, 32, 64, 1], strides=[1, 1, 1, 1, 1])
    T26 = fd.ops.squeeze(T25, original_shape=[8, 1024, 32, 64, 1], dims=[4])
    T27 = fd.ops.slice(T4, start_indices=[0, 0, 0, 0, 1], end_indices=[1, 1024, 1, 64, 2], strides=[1, 1, 1, 1, 1])
    T28 = fd.ops.squeeze(T27, original_shape=[1, 1024, 1, 64, 1], dims=[4])
    T29 = fd.ops.broadcast_in_dim(T28, output_shape=[8, 1024, 32, 64], broadcast_dims=[0, 1, 2, 3])
    T30 = fd.ops.mul(T26, T29)
    T31 = fd.ops.add(T24, T30)
    fd.add_output(T18)
    fd.add_output(T31)

with FusionDefinition() as fd:
    nvfuser_fusion_id0(fd)

inputs = [
    torch.randn((4096, 64, 2), dtype=torch.float32, device='cuda:0').as_strided((4096, 64, 2), (128, 2, 1)),
    torch.randn((8, 32, 1024, 128), dtype=torch.float32, device='cuda:0').as_strided((8, 32, 1024, 128), (4194304, 131072, 128, 1)),
]
fd.execute(inputs)

The error we encounter is

RuntimeError: downcast_ptr != nullptr INTERNAL ASSERT FAILED at "/opt/pytorch/nvfuser/csrc/utils.h":196, please report a bug to PyTorch.
This occurs in DomainMap::getInnerLeafDim which assumes that any transforms between root and leaf domains are either Split or Merge. We should handle the Resize operation here, and we should place a better error message here for unhandled expressions since we may introduce more in the future.

[jacobhinkle]

Sorry the repro is not exactly minimal, but even reordering these will sometimes cause the transpose scheduler not to fire. I am working on a smaller C++ repro.

[jacobhinkle]

This no longer fails on main (checked with 8226e61).