1. Wei Hong, Feng Qiu, Arie Kaufman Center for Visual Computing and Department of Computer Science, Stony Brook University http://www.cs.sunysb.edu/~vislab/projects/gpgpu Hybrid Volumetric Ray-Casting Overview Hybrid volumetric ray-casting algorithm using both CPU and GPU: Hybrid volumetric ray-casting algorithm using both CPU and GPU: CPU: ray traversal and empty space skipping CPU: ray traversal and empty space skipping GPU: local illumination and ray integration GPU: local illumination and ray integration Exploit the parallelism and trade-off between CPU and GPU to obtain further acceleration. Exploit the parallelism and trade-off between CPU and GPU to obtain further acceleration.Overview Hybrid volumetric ray-casting algorithm using both CPU and GPU: Hybrid volumetric ray-casting algorithm using both CPU and GPU: CPU: ray traversal and empty space skipping CPU: ray traversal and empty space skipping GPU: local illumination and ray integration GPU: local illumination and ray integration Exploit the parallelism and trade-off between CPU and GPU to obtain further acceleration. Exploit the parallelism and trade-off between CPU and GPU to obtain further acceleration. Algorithm 1. Ray Determination (CPU): Compute normalized ray direction and first intersection point - Figure 1(a) 2. Slab Rendering (GPU): Make each ray travel N steps forward - Figure 1(b) 3. Hole Predicting (CPU): Find the non- terminated rays 4. Hole Filling (GPU): Finish the non- terminated rays - Figure 1(c) Algorithm 1. Ray Determination (CPU): Compute normalized ray direction and first intersection point - Figure 1(a) 2. Slab Rendering (GPU): Make each ray travel N steps forward - Figure 1(b) 3. Hole Predicting (CPU): Find the non- terminated rays 4. Hole Filling (GPU): Finish the non- terminated rays - Figure 1(c) Figure 1: Overview of the algorithm Figure 2: Volume rendering with different transfer functions. (a) Depth Image(b) Intermediate Image(c) Final Image Figure 3: Two images inside the human colon. ResultsPlatform: 2.4 GHz Pentium 4 and 1G RAM2.4 GHz Pentium 4 and 1G RAM Nvidia Geforce FX5950 with 256M RAMNvidia Geforce FX5950 with 256M RAM Average rendering (frames/sec): Rendering (frames/sec) for human colon: Conclusion: Our algorithm provides high quality and good performance for versatile direct volume rendering applications.Our algorithm provides high quality and good performance for versatile direct volume rendering applications.ResultsPlatform: 2.4 GHz Pentium 4 and 1G RAM2.4 GHz Pentium 4 and 1G RAM Nvidia Geforce FX5950 with 256M RAMNvidia Geforce FX5950 with 256M RAM Average rendering (frames/sec): Rendering (frames/sec) for human colon: Conclusion: Our algorithm provides high quality and good performance for versatile direct volume rendering applications.Our algorithm provides high quality and good performance for versatile direct volume rendering applications. Discussion Key innovations: Exploit the parallelism of CPU and GPUExploit the parallelism of CPU and GPU Adjust the workload of CPU and GPU by controllingAdjust the workload of CPU and GPU by controlling  Number of steps processed in slab rendering  Number of steps processed in fragment program Discussion Key innovations: Exploit the parallelism of CPU and GPUExploit the parallelism of CPU and GPU Adjust the workload of CPU and GPU by controllingAdjust the workload of CPU and GPU by controlling  Number of steps processed in slab rendering  Number of steps processed in fragment program DatasetSizeOpaqueTranslucent Engine256×256×12823.39.1 Foot152×256×22020.610.2 DatasetSizeFig. 3(a)Fig. 3(b) Colon CT512 × 512 × 36127.320.9 (a)(b) CT EngineVisible Human Foot Acknowledgement: NSF CCR-03064