Why not add contiguous check for tensor in cuda kernel?

[otakuxiang]

I just implement a InstanceNorm layer in SparseConv and it using permute and unsqueeze function.This will change the stride and make tensor no more contiguous. The convolution after the norm function will give a wrong answer because the stride changes. I see the source code and I find it always assume the input tensor is contiguous because it use stride(0) multiply row id to get the value.
So why not add a contiguous check to make it more robustly?