I am working with OpenGL on Jetson Xavier NX with L4T 32.4.4 and seeing some reliable performance hit on using conditionals with uniform variables. So, there is no chance of warp divergence. Also converting uniform variables to specialization constants in spirv does improve performance, so the code itself does not seem to problem but how it is compiled.
In general, I would expect this hit to rely specifically on driver/shader compiler implementation, so I would like to get qualified confirmation what these observations are not caused by something else and should uniform branching avoided on all NVIDIA devices/drivers/etc.