> One thing I’ve learned is that spending time building, benchmarking and profiling basic blocks may seem like it can delay the overall plan; however, in this case, exploring different designs for the basic parts made it clear which features of the basic blocks were varying, and which were necessary, yielding a more modular design in the end, which was key to the performance improvements I got.

This. I try point this out so often and am challenged with the 'premature optimisation...' mantra but in my view this is in no way a premature optimisation but an essential part of the design process.

Thanks to the author for an enlightening article.