If you want it to be faster you have to change the high level parameters of perception and adjust for the specific device characteristics. VanitySearch is more like a port from CPU to CUDA rather than a parallel-thought problem. So all clones follow the same philosophy more or less.

Micro-optimizations with PTX may end up producing the same final SASS as plain C. You might lose a year to optimize some inverse only to find it works slower than initially on some random new GPU, and faster on another. I have 3 versions (SafeGCD, BinGCD and the one by RC), each of them runs better or worse depending on whether I change a single line of code in a totally different part of the kernel source. So it's more a game of luck to have a perfect faster kernel, depending on whether the compiler decides or not to maybe not spill an extra register just because you swapped two lines that are logically non-dependent