i have just started going down this rabbit hole, but if the MKL team inspects and tests my most recent test case data, they should be able to understand why i'd want this type.
i have noticed that, for "fresh" computations, the eigensolver would occasionally could take up to twice the time than when reading the doubles written to file just before via fwrite.
i then realised it's because there must be some part of the computation that has higher precision when it remains in main memory, and this is lost when written.
then it all made sense with x87, which leads me to believe this precision occasionally 'trickles' into the double, and thus takes almost twice as long.
in future releases, would the team consider adding long double support to MKL? i find it sort of unusual intel hasn't accommodated this type for a precision-focused API like MKL because it's an intel-only advantage and i don't see why its users shouldn't benefit.
it would be cool to have this support.
