On Tue, Mar 25, 2025 at 9:34 AM Shasang Thummar via NumPy-Discussion < numpy-discussion@python.org> wrote:
> *Why Should This Be Allowed?* > > If *a larger dimension is an exact multiple of a smaller one*, then > logically, the smaller array can be *repeated* along that axis *without > physically copying data*—just like NumPy does when broadcasting (1,M) to > (N,M). > > In the above example, > > - A has shape (2,3), and B has shape (4,3). > - Since 4 is *a multiple of* 2 (4 % 2 == 0), A could be *logically > repeated 2 times* to become (4,3). > - NumPy *already does* a similar expansion when a dimension is 1, so > why not extend the same logic? > > It's not a particularly similar expansion. The reason broadcasting a 1-dimension works is that we just set the stride for the broadcasted dimension to be 0, so the pointer never advances along that dimension and just keeps pointing at the one value along that axis. You can't do the same thing with tiling n>1. It would have to advance within the tile and then go back. So, unlike current broadcasting, I don't think we can just adjust the strides on each operand separately to match the broadcasted shape, which is something the current machinery relies upon. You could reshape `B` to be (2, 2, 3), broadcast as normal, operate, then reshape the result back. I'm not sure if it's guaranteed in all cases that the reshaping back could be done without copying, and I have no clear idea of how this works in the ternary+ operator case. But even so, this is not a straightforward extension of the broadcasting functionality and would not just plug in to the rest of the broadcasting machinery. I could be wrong; haven't put too much effort into it, but I think the first step would be to walk through what the stride tricks actually would be to get this operation to work. Then we can talk about whether or not it's desirable to plug it in. -- Robert Kern
_______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
