This should be a TypeError, I think

Does it actually have to be a C order array? I think we can just keep those error details to the np.dot function.

I think that checking only shape should be enough. np.dot will check other details.

The danger is that reshape may create a copy, so F contiguous arrays may be accepted by dot, but not necessarily work here (at least not without a copy back into out). In any case, just make sure we have tests for these things, then it is easy to try around to see how many checks are needed.

This will end up returning a view of out when out is specified (since it is a reshaped version of the reshaped version). I am not sure I mind that though.

I think we promise that return(func, out=out) is out for functions that have an out kwarg, no?

@mattip: Indeed, although that promise is hard to keep unless transpose does not return a view (something subclasses can change, and I think to a certain extent ma.array does). I think the logic would be:

• (somehow) check if transpose returns a view
• If it does, then return the original out array
• If it does not, then do a copy from the transposed array back to the original array.

Can we rewrite this function to use matmul instead of dot, using the new axes argument to matmul? Then the out would be handled for us without any reshaping.

One small issue with such rewriting is that it gives a different speed tradeoff currently. I would not actually mind the lazy way of doing type(out) is np.ndarray saying that subclasses may be allowed in the future.
The whole __array_wrap__ dance is not fun, also the function currently uses np.asarray on the inputs in any case.