GPU Architecture

Good

• 2 first warps would do branch a
• the third one would do branch b

if (ID < 64)
    a
else
    b

Bad

• only one instruction can run on a warp (32 of threads), meaning branch a would run only on 1 thread wasting 31 threads branch b would be run on 31 threads and wasting 1 thread

if (ID == 0)
    a
else
    b

Worst

• worst case, wasting half of threads in multiple warps!!

if (ID%2==0)
    a
else
    b