some aggressive optimizations to vectorize better, mostly switching out repeated inline coordinate->index operations with single int2x4 operation - makes a difference with the loop count we're looking at