This is really cool work! I am surprised to see this level of tuning without using cache profiling or other performance counters to identify the bottleneck and quantify the improvement.