On arm64 and at least one other architecture, the error says: -3.2862601528904633e-16 != 0
It looks as though the test is computing (1.2 * 3.4 - 3.4 * 1.2). Now, the log to base 2 of (1.2 * 3.4) divided by 3.286e-16 is about 53.5. There are 52 bits in the mantissa of a 64-bit float, or 53 including the implicit "1". So I would guess that the error comes from multiply and subtract being combined into a single operation, so that the error in computing 1.2 * 3.4 is revealed. If this is a real problem, perhaps "cross" could be implemented in a way that prevents multiply-subtract from being used. If it is not a real problem, then the test should be made less sensitive, perhaps by requiring the absolute value of the result to be less than 1e-12 or some other arbitrary epsilon.
