Advanced compute shaders very often need to synchronize local data within warp. When you program such shaders you have to be aware about SIMD width (32 on NVidia and Intel, 64 on AMD), and design both algorithms and data structures accordingly. Failing to do so often have significant performance costs, I saw up to 10x speedup after implementing cooperative algorithm instead of straightforward one.