Possible optimization?

[Nielsbishere]

Hi, this was a good read and interesting to see. One thing that might be interesting reading about is https://www.khronos.org/blog/vulkan-subgroup-tutorial, it basically allows to get rid of the atomicOr and barrier. Since you're executing 1 warp (32 threads), you can exchange it using subgroupOr (GL_KHR_shader_subgroup_arithmetic). Basically uint x = subgroupOr(result << gl_LocalInvocationIndex) in https://github.com/diharaw/HybridRendering/blob/master/src/shaders/ao/ao_ray_trace.comp#L120 and also the shadow pass from what I've read. I'm not sure if this is already done by the driver or compiler, but it might be interesting to check out anyways (if this wasn't done for reducing the extensions required to run the sample).