One potential way to further improve the speed of your new and interesting approach would be to use SWIPE as a replacement of SWPS3 for your pairwise sequence comparisons. I quote Torbjørn Rognes' paper: "SWIPE was about six times faster than SWPS3 and Farrar's own implementation for a typical length query, but the factor varied between 2 and 18 depending on query length and number of threads used."

SWIPE is also a Smith-Waterman implementation, and should provide you with exactly the same results, only faster (if of course you run your analyses on x86/SSE3 processors, and if SWIPE can use your scoring matrix).