I hope this has been measured with the binary optimized for the CPU architecture in question, in oder to sqeeze out reasonable gains from likely more portability-focused opus code, particularly regarding vectorization opportunities.