its one of those things... micro optimisation becomes the big one once you do enough algorithm optimisation. even the humble memory copy becomes faster from prefetching and clever use of registers... and actually making the algorithm theoretically worse. :P
also, note that it is possible to remove the magic from your magic number with some thought about latency, the size and alignment of your data etc. fetching a cache line takes a fairly predictable amount of cycles.
opposite to the prefetch is the non temporal read/writes where you don't pollute cache to prevent big one off copies from damaging the performance of other code...
also, note that it is possible to remove the magic from your magic number with some thought about latency, the size and alignment of your data etc. fetching a cache line takes a fairly predictable amount of cycles.
opposite to the prefetch is the non temporal read/writes where you don't pollute cache to prevent big one off copies from damaging the performance of other code...