[https://linebender.org/blog/towards-fearless-simd/]
The Future of SIMD, with Raph Levien
SIMD = Single Instruction Multiple Data
In GPU we have SIMT i.e. Single Instruction Multiple Threads (32 threads per wrap/group), where each thread can access same or different data
let’s say we have a if statement if A else B then, a SIMT compiler will map that to a predication/predicate mask(a bitmask or a set of boolean values). Each value of predication mask represents a case where condition is true or false according to the boolean value it holds.
true or 1 means operation proceeds
false or 0 means operation is skipped