Currently there is no concurrent code when KIJ order is required with block multiplication due to cross thread interaction. This can be fixed by accumulating the output matrix, but to keep it bit equivlanet with single thread that needs to be done carefully.
Currently there is no concurrent code when KIJ order is required with block multiplication due to cross thread interaction. This can be fixed by accumulating the output matrix, but to keep it bit equivlanet with single thread that needs to be done carefully.