Wrong output shape in `broadcast_in_dim` with empty inputs

[jacobhinkle]

There is an issue with computing the output shape in some cases with broadcast_in_dim:

from nvfuser import DataType, FusionDefinition
import torch
from torch.testing import make_tensor

a = make_tensor((1, 0), device='cuda', dtype=torch.float32)

with FusionDefinition() as fd:
    T0 = fd.define_tensor(symbolic_sizes=[1, -1], contiguity=[None, True], dtype=DataType.Float, is_cpu=False)
    T1 = fd.ops.broadcast_in_dim(T0, output_shape=[0, 0], broadcast_dims=[0, 1])
    fd.add_output(T1)

nvfout, = fd.execute([a])

# This fails because nvfout.shape is (1, 0)
assert nvfout.shape == torch.Size((0, 0))

The Fusion seems to be constructed properly:

Inputs:
  T0_g[ bS0{1}, iS1{i1} ], float
Outputs:
  T2_g[ bS4{0}, iS5{i1} ], float

%kernel_math {
T1_l[ bS2{1}, iS3{i1} ]
   = Set( T0_g[ bS0{1}, iS1{i1} ] )
T2_g[ bS4{0}, iS5{i1} ] = expand( T1_l[ bS2{1}, iS3{i1} ], {0, i1} )
}

Note that changing the first dimension of a to 0 and setting symbolic_sizes=[-1,-1], contiguity=[True,True] gives us the correct behavior. The issue seems to be in expanding to zero size. This might be an issue with the remove empty pass checking extent instead of getMaybeExpandedExtent.

[jacobhinkle]

Confirmed this is an issue with the new RemoveEmptyPass (#543). Using NVFUSER_DUMP=fusion_ir_preseg we see that the replacement is using extent:

Inputs:
  T0_g[ bS0{1}, iS6{0} ], float
Outputs:
  T3_g[ bS9{1}, iS10{0} ], float

%kernel_math {
f10 = (float)(double(0));
T3_g[ bS9{1}, iS10{0} ]
   = full({1, 0}, ( (float)(double(0)) ));
}