[Inductor] Fusion of Tiled Point-Wise and Reduction Operators about pytorch HOT 1 OPEN

I tried manually splitting the reduction in python:

def test(x):
    y = torch.sum(x, dim=1)
    z = (x + x.t()) / 10.0 # x need to be a squared tensor here
    return y.sum(), z

And indeed it does allow the fusion to happen and the fused kernels are faster.

from pytorch.