Array math broadcastability #74
Description
In the DAWN implementation, it was possible to do array maths with arrays of non-equal dimensions, where arrays would be padded with extra dimensions on the right. For example: if we do A * B, with A shape (16, 1) and B of shape (16, 1, 1030, 1064), A would be padded to (16,1,1,1) before operation.
According to @timsnow, the broadcasting rules were module-dependent, usually more complex, and caused some difficult-to-trace issues. We're therefore proposing to not pad left or right, but instead demand from the user / data source that the arrays on which maths is supposed to be done are broadcastable in numpy operations. That means sanitising input rather than quietly adapting shapes.
Opinions?