NVIDIA supports CL1.2 and it's intentionally slow so people will use CUDA. Plain CL2.0 shouldn't bring performance improvements unless the kernels used by Blender perform things like work_group functions. Also CL2.0 will require people to specify if a WG is uniform, otherwise they must assume it's not uniform and that's taxing in some cases. SPIR-V on the other hand should be interesting to see.