Bro, you don't even know what you're talkin' about. libsecp256k1 wasn't made for low-level batched optimizations like that. You gotta use JeanLucPons' SECP256K1 if you ever wanna hit 50M keys per second on any CPU.
Actually, JLP is the slow one here. Like, at least 50% slower.
libsecp256k1's primitives can be easily used to implement batched addition, or anything else. I already posted the code to do this a while ago.