330 likes | 633 Views
Rohit Nigam, P. J. Narayanan CVIT, IIIT Hyderabad, Hyderabad, India. Hybrid Ray Tracing and Path Tracing of Bezier Surfaces using a mixed hierarchy. Representing a Scene. f>0. f<0. f=0. Triangular Mesh. Implicit Surface. Parametric Surface. Parametric Surface: Motivation.
E N D
Rohit Nigam, P. J. Narayanan CVIT, IIIT Hyderabad, Hyderabad, India Hybrid Ray Tracing and Path Tracing of Bezier Surfaces using a mixed hierarchy
Representing a Scene f>0 f<0 f=0 Triangular Mesh Implicit Surface Parametric Surface
Parametric Surface: Motivation • Provide compact and effective representation. • Remain curved and smooth at arbitrary level of zooming. • Memory efficient, in comparison with triangular mesh.
Bezier Surfaces • Bezier Surfaces are the most basic form of parametric surfaces • A Bezier Surface can be described as: Q(u,v) = [U][M][P][M]T[V]T where [U] = [u3 u2 u 1] and [V] = [v3 v2 v 1], 0 ≤ u,v ≤ 1 [M] is the Bezier Basis Matrix [P] is the set of 16 Control Points defining the patch
Rendering Bezier Surfaces • Tessellation based approches • Eisenacher et al.(2009) : View Dependent Adaptive Subdivision • Direct Ray Tracing • Geimer et al.(2005) : Newton Iteration • Pabst et al.(2006) : Bezier Clipping + Newton Iteration
Ray Tracing Bezier Surface • Constructing an Accelaration Structure • Bounding Volume Hierarchy(BVH)
Ray Tracing Bezier Surface • Ray Traversal through BVH Ray List Outputs Potential Ray-Patch intersections list Initial parameter values
Ray Tracing Bezier Surface • Newton Iteration Picture Courtesy : http://steadyserverpages.com
Geimer, Abert Approach • Based on the flatness criteria, each patch is divided into subpatches. • BVH for original surfaces • Bounding boxes of subpatches at leaf nodes. • For each potential intersection • Generate initial values for Newton Iteration 1 BVH Nodes 3 2 Original Curve Subdivided Linear Curve Patch1 P2 P3 sp2 sp1 sp2 sp1 sp2 Subpatches at Leaf sp1
Limitation of the Model for GPUs • GPU Access time: • High for global memory • Comparatively less for shared memory and registers When subdividing based on flatness criteria, we need to • Store subpatches starting index • Store total number of subpatches • Store initial [u,v] pair for each potential intersection. Thus more global memory operations result in lower throughput. • Need to check every subpatch at leaf node
Our Approach • Create a mixed hierarchy, consisting of two hierarchical structures. • The top level BVH tree is constructed from the bounding boxes of original patches. • Leaf nodes represent the original Bezier Surfaces. • Each Patch is divided into fixed size subpatches, hierarchically, using De Casteljau algorithm. • Make subtree for each patch from bounding boxes of the subdivided patches.
BVH Nodes BVH for Patches Patches 1 2 3 4 Subtree Nodes Subpatch Hierarchy Sub-patches
Mixed Hierarchy Structure • Newton Iteration applied to original patches • No memory required to store subpatches • Fixed depth subtree • Utilize constant degree of bezier surfaces • Utilize shared memory • Apply early termination at subtree level • Leads to tighter bounds • A subdivision depth of 6 was found empirically sufficient.
Mixed Hierarchy Structure • Newton Iteration applied on original patches. • No memory required to store subpatches. • Fixed depth makes it possible utilize shared memory. • A subtree at lower level leads to early termination at this stage, reducing the (Ray, Bounding Box) intersections. • Subdivision also leads to tighter bounds, which further reduces the potential (Ray,Patch) intersections. • A subdivision depth of 6 was found empirically sufficient for our scenes.
GPU Traversal of Mixed Hierarchy Structure • A ‘traverse’ kernel traverses the first level of the BVH. • Lists out Potential (Ray,Patch) intersections. • We make use of atomic operations, to provide scalability. • ‘Recheck’ kernel parallely processes the generated (ray,patch) list. • This leads to further pruning of the list with tighter subpatch bounding boxes. • We make use of ‘t’ values computed here, to not traverse subpatch nodes with higher values. • This leads to reduced computation and in cases of false positive, a little less accurate initial values. • Lists out the reduced potential (Ray,Patch) intersections. • Generates the initial values for each intersection.
Hybrid Ray Tracing Start GPU CPU Preprocessing Ray List rayTraceGPU rayTraceCPU Point and Normal Generate Secondary Rays
Results Teapot Model Fps : 64 Bigguy Model Fps : 28.6 Killeroo Model Fps : 19.2 System Specs GTX 580 + i7 920 1024x1024 2 Killeroos Fps : 10.6 9 Bigguys Fps : 5.2
Path Tracing • We extend our ray tracing approach to Global Illumination effects. • We use Cook’s approach of Monte Carlo based Stochastic Sampling, to sample the image at appropriate non-uniformly spaced points. • Each pixel is sampled for a user defined samples per pixel • We apply our data parallel approach to this massive ray list to generate the desired effects.
Path Tracing Bigguy in a box: 400 spp, 512x512 resolution Rendered in 28.5 minutes
Path Tracing Bigguy in a box: 1000 spp, 512x512 resolution Rendered in 28.5 minutes
Path Tracing Bigguy in a box: 3000 spp, 512x512 resolution Rendered in 28.5 minutes
Path Tracing Bigguy in a box: 5000 spp, 512x512 resolution Rendered in 28.5 minutes
Path Tracing Bigguy in a box: 10000 spp, 512x512 resolution Rendered in 28.5 minutes
Conclusion • A mixed hierarchy model is proposed to speed up Ray Tracing process. • GPU benefits greatly from fixed depth subtree. • A hybrid model is proposed, to fully utilize compute power of CPU and GPU. • We demonstrate the capability of our method by producing Global Illumination effects for Bezier patches.
Hybrid Ray Tracing • Divide the Ray list between CPU and GPU • Ratio decided based on compute capabilities • GPU algorithm comprises of three kernels: • Traverse : Generate Potential Ray-Patch Intersections • Recheck : Further prune intersections and get initial values • Newton : Apply Newton iteration to get hit-point • CPU stage comprises of: • Divide CPU Raylist into 2c threads, where c is number of cores. • Intersection with main BVH • If intersects, further intersection with 2nd level subtree. • Finally, apply Newton iteration and generate hit-point • CPU benefits from early ray termination.
Newton Iteration • We represent a ray as intersection of two planes, (n1,d1) and (n2,d2) The ray patch intersection equation becomes Q(u,v) represents the point on the patch. • We use Newton Iteration to solve for (u,v) Here J is the inverse Jacobian matrix of R.