realizing multi-hit ray tracing in embree & ospray · •ospray permits easy integration...

Realizing Multi-Hit Ray Tracingin Embree & OSPRay

Christiaan GribbleSURVICE Engineering

Intel HPC Developer Conference

12 November 2016

Take-home messages

• Multi-hit ray traversal

Enables a new class of ray-based rendering &simulation applications

• Embree permits efficient implementation

• OSPRay permits easy integration

2 C. Gribble, Realizing Multi-Hit Ray Tracing in Embree & OSPRay

Take-home messages



Intersection filter functions enable user-level implementationof state-of-the-art multi-hit ray traversal techniques



Take-home messages




Supports scalable, high performance visual analysis toolsacross optical & non-optical domains


Acknowledgements

• SURVICE• Joe Rosenthal

• Mark Butkiewicz

• Intel• Jeff Amstutz

• Ingo Wald

• Jim Jeffers


Overview

Interval computation

Optical rendering



Optical rendering



Optical rendering Non-optical rendering


Non-optical rendering



Non-optical rendering Interval computation




• Difficult or impossible• Epsilon hacks

• Missed/repeated intersections

• Performance impacts



• Difficult or impossible

• Performance impacts• Traversal restart

• Operational overhead



• Difficult or impossible

• Performance impacts

Is there a better solution?


Multi-hit ray traversal

• Which primitives are intersected?• One or more, & possibly all

• Ordered by t-value along ray

• Possible applications

• GNK14

• AGGW15

• Gri16, GWA16



• Which primitives are intersected?

• Possible applications• Transparent rendering

• Alpha textures

• GNK14

• AGGW15

• Gri16, GWA16





• GNK14• Spatial partitioning

• Two algorithms: naive, buffered

• AGGW15

• Gri16, GWA16





• GNK14

• AGGW15• Object partitioning

• User-level implementation mechanisms

• Gri16, GWA16





• GNK14

• AGGW15

• Gri16, GWA16• Enable early-exit in BVH

• Implement in Embree & OSPRay


Implementation

Mechanisms

• Direct implementation• Kernels specific to multi-hit

• Runs counter to our goal

• Intersection callbacks

• Traversal callbacks

• Reference implementation


Mechanisms

• Direct implementation

• Intersection callbacks• Invoked on valid ray/primitive intersection

• User accepts/rejects hit




Mechanisms



• Traversal callbacks• Invoked on ray/node interaction

• Two variants: every-node, leaf-node



Mechanisms




• Reference implementation• Supports callback mechanisms

• Opts for clarity & simplicity


http://www.rtvtk.org/~cgribble/research/mhBVH/

Embree implementation

• Intersection filters• Compatible with mainline developments

• v2.10.0+ – enables node-culling

• Assumptions

• Full source code available



• Intersection filters

• Assumptions• Nquery known a priori

• Preallocated hit data buffer

• Initial value beyond range

• Full source code available



• Intersection filters

• Assumptions

• Full source code available• Apache License, v2.0

• Public git repository


http://www.rtvtk.org/~cgribble/research/ospMultiHit/

Scalar implementation

static void collectIntersectionsFilter(void* /* unused */,

RTCRay& _ray)

{

0 // Find index at which to store candidate intersection

1 uint idx = Nquery;

2 while (idx > 0 && ray.tfar < hits[idx-1].tval)

3 {

4 hits[idx] = hits[idx-1];

5 --idx;

6 }



7 // Store intersection, possibly beyond index of the

8 // N ≤ Nquery closest intersections (i.e., at

9 // idx = Nquery)

10 HitData& hit = hits[idx];

11 hit.geomID = ray.geomID;

12 hit.primID = ray.primID;

13 hit.tval = ray.tfar;

14 hit.Ng = ray.Ng;

15

16 // Update number of intersections identified so far

17 ray.nhits += (ray.nhits < Nquery ? 1 : 0);



17 if (ray.nhits < Nquery)

18 {

19 // Reject intersection to continue traversal with

20 // incoming ray interval, as in previous work

21 // [Amstutz et al. 2015]

22 ray.geomID = RTC_INVALID_GEOMETRY_ID;

23 return;

24 }



25 // Induce node culling

26 // Trick: set ray.tfar to farthest value among the

27 // N = Nquery intersections identified so far

28 // and (implicitly) accept intersection with

29 // modified ray interval

30 ray.tfar = hits[Nquery-1].tval;

31 }


Results

Performance – tests

• Find-first-intersection• First-hit v. multi-hit variants

• Isolates multi-hit overhead

• Find-all-intersections

• Find-some-intersections



• Find-first-intersection

• Find-all-intersections• Naive v. node culling multi-hit

• Bounds performance expectations

• Find-some-intersections



• Find-first-intersection

• Find-all-intersections

• Find-some-intersections• Naive v. node culling multi-hit

• Demonstrates in situ performance


Performance – scenes

sibe

80K tris

fair

174K tris

conf

282K tris

truck

426K tris

tank

1.0M tris

hball

2.8M tris

sanm

10.5M tris

pplant

12.7M tris


Performance – truck scene

0.0

20.0

40.0

60.0

80.0

100.0

120.0

140.0

10%1 30% 70% all

Mhps

+4.60x

+1.64x

+1.08x 0.96x 0.97x

naive culling



0.0

20.0

40.0

60.0

80.0

100.0

120.0

140.0

10%1 30% 70% all

Mhps

+4.60x

+1.64x

+1.08x 0.96x 0.97x

naive culling


Limitations

• Number of hits specified a priori• Too few – incorrect

• Too many – wasteful

• C++ templates?

• Ordered BVH traversal [WAB16]

• Spatial partitioning [GNK14]


Limitations

• Number of hits specified a priori

• Ordered BVH traversal [WAB16]• Enables early-exit

• No dynamic allocation

• Requires new traversal kernels

• Spatial partitioning [GNK14]


Limitations

• Number of hits specified a priori

• Ordered BVH traversal [WAB16]

• Spatial partitioning [GNK14]• Enables early-exit

• No dynamic allocation

• Requires user-defined geometry


Demonstrations

Engineering CAD visualization Vulnerability analysis


Wrap-up

Take-home messages


Enables a new class of ray-based rendering &simulation applications




Take-home messages



Intersection filter functions enable user-level implementationof state-of-the-art multi-hit ray traversal techniques



Take-home messages




Supports scalable, high performance visual analysis toolsacross optical & non-optical applications


Key references

[AGGW15] Amstutz, J., Gribble, C., Gunther, J., & Wald, I. (2015) An evaluation of multi-hitray traversal in a BVH using existing first-hit/any-hit kernels. Journalof Computer Graphics Techniques, 4(4):72—90.

[GNK14] Gribble, C., Naveros, A., & Kerzner, E. (2014) Multi-hit ray traversal. Journal ofComputer Graphics Techniques, 3(1):1—17.

[Gri16] Gribble, C. (2016) Node culling multi-hit BVH traversal. Eurographics Symposiumon Rendering. doi:10.2312/sre.20161213

[GWA16] Gribble, C., Wald, I., & Amstutz, J. (2016) Implementing node culling multi-hitBVH traversal in Embree. Journal of Computer Graphics Techniques,5(4):29—35.


Contact information

AddressApplied Technology Operation

SURVICE Engineering

4695 Millennium Drive

Belcamp, MD 21017

[email protected]

Webhttp://www.survice.com/employees/~cgribble/


realizing multi-hit ray tracing in embree & ospray · •ospray permits easy integration...

Documents