Splitting visibility buffer shaders to reduce register usage:
resolve (async)
binning (async)
velocity (async)
surface sampling (async)
sky (async)
lighting (gfx)
Successfully improved GPU utilization in lighting shader.
Binning is faster, because binning tiles and not pixels
Because tiles are binned, lighting tiles are using coherent tile info for whole thread group
Surface, sky and decal rendering also runs now in async compute.
Velocity is separate pass, improves GPU utilization in resolve, and only runs as optional.
Surface shader occupancy is still bad, and wind computation is still bad per pixel, but now it is only done in surface shader.
Surface shader can now write out proper top layer of clearcoated material, this improves screen space and raytraced reflections.
MSAA not implemented yet for visibility compute shading
Tessellation is still only half way supported in visibility compute shading (tessellation data not available in memory for pixel interpolation, but depth is written in prepass, this causes mismatch)
Turánszki János
54491ad536