Broadcast semantics in GGML?

[tlemo]

I'm trying to understand the broadcasting semantics in GGML, since they seem to differ a bit from the numpy/pytorch semantics:

1. For binary operations, it seems that only src1 can be broadcasted (src0 and dst must match)
2. The broadcast in GGML seems to generalize the repeat broadcasting, in addition to the 1->N broadcasting (ex. numpy), it also supports M->N, when M is less than N, and it's repeated up to N. Does N have to be a strict multiple of M?

Is my interpretation of the current implementation right?