This post gives an example how padding and unfold can be applied to create non-overlapping patches as well as how to reshape these patches back to the original input tensor.
Note that for your shapes you would need to pad the input, since the shapes are not factors of the kernel size.