VAES/AVX2 OCB implementation suffer from same performance oddity with Intel CPUs that was previously seen with AES-NI implementation (e924ce456d5728a81c148de4a6eb23373cb70ca0). Performance of encryption would start to drop with large buffer sizes (beyond 16KiB). Same drop is not seen with decryption. Here's graph from Intel Tigerlake:
Same fix as was done with AES-NI implementation should work with VAES/AVX2: inline checksumming with OCB encryption.