This innocent looking 10 lines of code generates more than 600 GPU instructions... oops :) 
It reads 3x3 neighborhood from groupshared memory and combines some bit masks to setup indirect dispatch tile coordinates (later in the shader). Time to refactor it...
HLSLcc GLSL output looks a bit unoptimal. IMHO leaning a bit too much for the IHV compiler to massage the code. My bit extract code does way too much stuff in this case (compare > 0 afterwards). I could SWAR through this mess :)
I have 3 counters packed in one uint (10+12+10 bits). I can binary OR the packed counters directly (SWAR style), because binary OR will only be >0 if any of the combined masks are >0. This of course messes up the counter values, but >0 is the only case I care about.
So much better: gist.github.com/sebbbi/f1a66ae…
Yay. 8x less instruction now for that piece of code. That shader become 0.4 ms faster :)
You could also SWAR add the packed counters, but you risk overflow. Binary OR acts like a carryless per bit adder. Guaranteed to have no overflow. And guaranteed to handle the case of != 0 correctly.
