nbody: Control barrier in non-uniform control flow

[msiglreith]

The N-body example uses a control barrier which might not be reached by all executed: https://github.com/SaschaWillems/Vulkan/blob/master/data/shaders/computenbody/particle_calculate.comp#L60

The early return here will result in possible divergence further on with potential stalls.

[oscarbg]

might be worth examining SPIR-V glslang generated code versus disassembled DXBC in case the shader compiles fine to DXBC(blob) with D3DCompiler API.. anyway just a question: do you know if D3D12 spec more strict in that regard vs Vulkan spec regarding control barriers in non unif. control flow? I mean this sample works ok on a lot of Vulkan drivers (implementations)..

[msiglreith]

d3d compiler is more strict here compared to SPIR-V which allows it. In practice, no divergence will occur as the example is crafted in such a way that the number of particles is a multiple of the group size. Therefore, it's soley a restriction by my d3d12 wrapper. A possible workaround would be to move the discard statement down after the control barrier.