`tryMerge` of torch_glow's cusom fuse pass might have some issue

[edimetia3d]

Hi,

I am implementing a custom torchscript jit backend like glow did, and might have found some little issue when custom fuse pass doing tryMerge

It seems that the fuse pass is implemented in a reverse order, but tryMerge tries to fuse producer before consumer. I am wondering if this was intended to be so, or it was some mistake.

The code below shows that glow will not do jit for a supported op aten::mul.

When the aten::mul op is treated as a consumer, the producer will be prim::Param, which will not pass the first canMerge checking, and for this, the consumer aten::mul will not be fused into the FusionGroup, thus will not be JITed.

x = torch.randn(4)
y = torch.randn(4)

@torch.jit.script
def foo(a, b):
    c = a.mul(b)
    return c
print(foo.graph_for(x, y))

Output:

graph(%a.1 : Float(*, requires_grad=0, device=cpu),
      %b.1 : Float(*, requires_grad=0, device=cpu)):
  %c.1 : Float(*, requires_grad=0, device=cpu) = aten::mul(%a.1, %b.1) # /workspace/workspace/torch_glow_dev/glow/torch_glow/examples/basic_example.py:14:8
  return (%c.1)

I have done some modification in this ~~commit~~ PR, it is able to generate a FusionGroup as I expected.

graph(%a.1 : Float(*, requires_grad=0, device=cpu),
      %b.1 : Float(*, requires_grad=0, device=cpu)):
  %c.1 : Float(*, requires_grad=0, device=cpu) = glow::FusionGroup_0(%a.1, %b.1)
  return (%c.1)
with glow::FusionGroup_0 = graph(%a.1 : Float(*, requires_grad=0, device=cpu),
      %b.1 : Float(*, requires_grad=0, device=cpu)):
  %c.1 : Float(*, requires_grad=0, device=cpu) = aten::mul(%a.1, %b.1) # /workspace/workspace/torch_glow_dev/glow/torch_glow/examples/basic_example.py:14:8
  return (%c.1)