Advanced Real-time Post-Processing using GPGPU techniques. Presentation overview

Advanced Real-time Post-Processing using GPGPU techniques

Presentation overview Problem description and objectives Depth of field Methods GPGPU programming Results Conclusion Questions

Problem description and objectives Post processing filters Different depth of field algorithms Visual quality Implement using HLSL and CUDA Performance Usability

Depth of field Depth cue Focus plane Focus in area in front of and beyond Different blurriness

Depth of field Thin lens camera model Circle of confusion

Depth of field Calculate Circle of confusion Depth value and lins parameters Depth map COC map

Methods Poisson disc blur Multi-passed diffusion Separable diffusion Summed-area table

Methods – Poisson disc blur Distribution function COC defines scale Downscaled image

Methods – Poisson disc blur Calculate values and interpolate depending on COC

Methods – Multi-passed diffusion Every pixel gets new value depending on the COC gradient

Methods – Separable diffusion Use a tridiagonal system to represent the heat conductivity Cyclic reduction can solve the matrices for each row

Methods – Separable diffusion Each row is solved independently In each step a reduced tridiagonal matrix is calculated (and output value) until the system is solved

GPGPU programming General Better flexibility Potential advantages CUDA Extension of C Large community

GPGPU programming Executes in chunks of threads User specified blocks Several memory types Global Texture Shared Constant More choices and possibilities Hardware specific limits Great potential

GPGPU programming Gaussian blur timings

GPGPU programming Implementation impact using CUDA + Easy to get started (C) Memory indexing (no more floating point texture indices) Good support for timing on the GPU Good control over computations (threads and memory) - A lot of ”rules” (amount of threads, occupancy, etc) Hard to optimize Beta problems (lack of interop, slow operations)

Results HLSL and CUDA for most methods Exceptions Poisson disc (HLSL only) Summed Area-Table (CUDA only) Timings in runs of 100 on recent hardware

Results Poisson disc timings Separable simluated diffusion timings Multi-passed diffusion timings

Results Artifacts Color leaking Sharp edges

Results Input data

Results Poisson disc Multi-passed diffusion Separable simulated diffusion

Poisson disc Results Multi-passed diffusion Separable simulated diffusion

Lens parameter settings Results

Conclusions Current depth of field filters are good enough Not really, but better is too expensive Cut scenes do get time for more computations GPGPU techniques have great potential Not mature enough (hardware support etc.) Maybe better for other things than image processing Future work Diffusion based approach offers best visual quality Compute shaders anyone?

Videos

End