conv about fuse_scale

[disanda]

I found there is one conv of each block who added fuse_scale, the trick implemented by transform_kernel.
(see kernel_weight in lreq.py line 142):
'''
if self.transform_kernel:
w = F.pad(w, (1, 1, 1, 1), mode='constant')
w = w[:, :, 1:, 1:] + w[:, :, :-1, 1:] + w[:, :, 1:, :-1] + w[:, :, :-1, :-1]
'''

I wonder what the trick uses. can anyone explain it?