That is incorrect. GCN is more suited for general puropose is because scheduling is almost entirely driver independent. Nvidia since Fermi has moved most of the scheduling to their drivers and according to them, scheduling takes up less then 4.5% of a Maxwell SM cluster area.
The reason the particle simulation test runs better on AMD is because it's a probably a port from the CUDA version of that test.
Let's not mix things here, driver issues draw calls, moves data to VRAM and back, and loads compiled shaders. The fact that ASM code of a shader can be optimized by a driver and/or a compiler is irrelevant once it's on GPU in instruction cache. Driver doesn't constantly schedule instructions, it would be ridiculously slow.
There are different kinds of schedulers here, maybe you mean work group scheduling (which threadblock/workgroup goes to which SMM) that can be done in driver. I'm talking specifically about warp instruction scheduler (that 4.5% of SMM)
You are right about GCN scheduler being totally independent ... that doesn't change the fact that you can have a real world problem only solvable by code that can't feed those big vector processors in GCN core optimally, for example something with lot's of scalar instructions with calculations interdependent of each other + lots of branching in the mix ... in that case opencl source or cuda port, it doesn't really matter those frameworks are very similar.
What does matter is optimizing specifically for one architecture over another, and I can't really say if chosen benches are biased that way.
