*   Improved DirectSession::Run() overhead and error checking. Feeding a value
    of the wrong type will now synchronously raise an INVALID_ARGUMENT error
    instead of asynchronously raising an INTERNAL error. Code that depends on
    the (undefined) behavior when feeding a tensor of the wrong type may need to
    be updated.
*   Added unreduced NONE, and reduced MEAN options for losses. Removed

R5 #define KEYP R6 #define ROUNDS R7 #define IVP R8 #define ENC R9 #define INOUT V2 #define TMP V3 #define IVEC V4 // Load the crypt key into VSRs. // // The expanded key is stored and loaded using // STXVD2X/LXVD2X. The in-memory byte ordering // depends on the endianness of the machine. The // expanded keys are generated by expandKeyAsm above. // // Rkeyp holds the key pointer. It is clobbered. Once // the expanded keys are loaded, it is not needed. // // R12,R14-R21 are scratch registers. // For...