1
Cyril Crassin David Luebke Michael Mara Morgan McGuire Brent Oster Peter Shirley Peter-Pike Sloan Chris Wyman Main references: Interactive indirect illumination using voxel cone tracing. Computer Graphics Forum 2011. CRASSIN, C., NEYRET, F., SAINZ, M., GREEN, S., AND EISEMANN, E. Toward practical real-time photon mapping: Efficient gpu density estima- tion. ACM I3D 2013. MARA, M., LUEBKE, D., AND MCGUIRE, M. Shading in Valve’s source engine. ACM SIGGRAPH Courses 2006. MITCHELL, J., MCTAGGART, G., AND GREEN, C. Model courtesy of 3D-Coat/Rick Results We demonstrate scaling up to 50 users per server node, allowing better quality global illumination for similar per-user costs. For each thechniques, we analyse the server-side overhead depend- ing on the number of clients, as well as the effective latency of indirect lighting. We show that only coarse synchronization between direct and indirect light is necessary, and even high latencies can be accept- able for indirect lighting. For voxel, the common overhead becomes quickly insignificant with increasing number of clients. Each K5000 GPU can simultaneously serve at least 5 clients with a consistent 30 FPS. For irradiance map, per-user latency is essentially constant. Due to larger bandwidth requirements, photon mapping does not scale as well. We developed three different pipelines within CloudLight to solve this for different network and client scenarios: Voxel cone-tracing: Clients only decode final frames computed on the server and streamed as a video. This allows the most amortization of server resources, but also requires a very low latency network. Irradiance maps: Clients render direct light and combine it with indirect light streamed asynchronously from server. This is good for medium-power clients with low latency that are limited by low bandwidth and is robust to network instabilities. Photon mapping: Clients reconstruct indirect light from photons generated on the server. This is gives the most reactive lighting through progressive update for a PC or console-class client. Indirect Lighting Pipelines Scene model courtesy of Guillermo M. Leal Llaguno Motivation Games on mobile phones, tablets, and laptops have graphics process- ing limited by heat dissipation in their small form factors. CloudLight is a system for real-time game rendering that computes lighting on distributed servers separate from the processors that compute the final images. This allows even better graphics than today's PC and console games on low-power mobile platforms. Cloud graphics also offers virtualization, load balancing, lower production costs, and piracy protection for developers. CloudLight offers three different pipelines, targeting different client devices. CloudLight maps the traditional graphics pipeline onto a distributed system which separates direct (view-dependent) and indirect (view- independent) lighting computations so that they can be computed on different machines connected by a network and compressed data streaming. A distributed graphics pipeline has several differences from a single- machine one. It must: Leverage the asymmetric resource distribution between stages of the pipeline. Tolerate high latency and low bandwidth between certain pipeline stages. Amortize computation across users. Distributed System ^ The CloudLight pipeline for three global illuminaon algorithms: voxels, irradiance maps, and photons. Shiſt from cloud (blue) to local computaon (black), suggests use of different algorithms depending on the targeted user device. ^ Our CloudLight framework demonstrates feasibility of Cloud indirect lighng for game-scale assets ^ Our experiments run in the Cloud on a GeForce TITAN + 2 secondary Quadro K5000 GPU for voxel (render frames and perform H.264 encoding). Photon reconstrucon use a GeForce 670. ^ Images of the scenes used for tesng, as seen on the client with real-me dynamic indirect light. Our system is based over commercially-deployed systems like OpX, hardware video encoders and NVIDIA ^ Our pipeline splits lighng between “direct” and “indirect”. View-independent “indirect” can be shared and amorzed in the Cloud over mulple clients CloudLight : Interactive Indirect Illumination in the Cloud H.264 NETWORK Lossless NETWORK View-Independent Indirect Lighting Final Frame Rendering Image Display LOCAL DEVICE CLOUD Photon Splat [Mara 13] c) Photon OptiX Photon Trace 30 hz HDMI Display Direct + Post 30 fps 30 fps b) Irradiance Map Reconstruct Irradiance Maps OptiX Ray Trace 10 hz HDMI Display 30 fps [Mitchell 06] + Direct + Post 30 fps H.264 NETWORK Voxel Injection Voxel Interpolate + Direct + Post PCI Voxel Cone Trace [Crassin 11] a) Voxel 15 hz 30 fps Proprietary 30 fps CONTACT NAME Cyril Crassin: [email protected] POSTER P4166 CATEGORY: REMOTE GRAPHICS & CLOUD-BASED GRAPHICS - RC02

CloudLight : Interactive Indirect Illumination in the Cloud · for different network and client scenarios: Voxel cone-tracing: Clients only decode final frames computed on the server

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

Page 1: CloudLight : Interactive Indirect Illumination in the Cloud · for different network and client scenarios: Voxel cone-tracing: Clients only decode final frames computed on the server

Cyril Crassin David Luebke Michael Mara Morgan McGuire Brent Oster Peter Shirley Peter-Pike Sloan Chris Wyman

Main references: • Interactive indirect illumination using voxel cone tracing. Computer Graphics Forum 2011. CRASSIN, C., NEYRET, F., SAINZ, M., GREEN, S., AND EISEMANN, E.• Toward practical real-time photon mapping: Efficient gpu density estima-tion. ACM I3D 2013. MARA, M., LUEBKE, D., AND MCGUIRE, M. • Shading in Valve’s source engine. ACM SIGGRAPH Courses 2006. MITCHELL, J., MCTAGGART, G., AND GREEN, C.

Model courtesy of 3D-Coat/Rick

Results We demonstrate scaling up to 50 users per server node, allowing better quality global illumination for similar per-user costs. For each thechniques, we analyse the server-side overhead depend-ing on the number of clients, as well as the effective latency of indirect lighting. We show that only coarse synchronization between direct and indirect light is necessary, and even high latencies can be accept-able for indirect lighting. For voxel, the common overhead becomes quickly insignificant with increasing number of clients. Each K5000 GPU can simultaneously serve at least 5 clients with a consistent 30 FPS. For irradiance map, per-user latency is essentially constant. Due to larger bandwidth requirements, photon mapping does not scale as well.

We developed three different pipelines within CloudLight to solve this for different network and client scenarios:

• Voxel cone-tracing: Clients only decode final frames computed on the server and streamed as a video. This allows the most amortization of server resources, but also requires a very low latency network.

• Irradiance maps: Clients render direct light and combine it with indirect light streamed asynchronously from server. This is good for medium-power clients with low latency that are limited by low bandwidth and is robust to network instabilities.

• Photon mapping: Clients reconstruct indirect light from photons generated on the server. This is gives the most reactive lighting through progressive update for a PC or console-class client.

Indirect Lighting Pipelines

Scen

e m

odel

cou

rtes

y of

Gui

llerm

o M

. Lea

l Lla

guno

Motivation Games on mobile phones, tablets, and laptops have graphics process-ing limited by heat dissipation in their small form factors. CloudLight is a system for real-time game rendering that computes lighting on distributed servers separate from the processors that compute the final images. This allows even better graphics than today's PC and console games on low-power mobile platforms. Cloud graphics also offers virtualization, load balancing, lower production costs, and piracy protection for developers. CloudLight offers three different pipelines, targeting different client devices.

CloudLight maps the traditional graphics pipeline onto a distributed system which separates direct (view-dependent) and indirect (view- independent) lighting computations so that they can be computed on different machines connected by a network and compressed data streaming.

A distributed graphics pipeline has several differences from a single-machine one. It must:

• Leverage the asymmetric resource distribution between stages of the pipeline.

• Tolerate high latency and low bandwidth between certain pipeline stages.

• Amortize computation across users.

Distributed System

^ The CloudLight pipeline for three global illumination algorithms: voxels, irradiance maps, and photons. Shift from cloud (blue) to local computation (black), suggests use

of different algorithms depending on the targeted user device.

^ Our CloudLight framework demonstrates feasibility of Cloud indirect lighting for game-scale assets

^ Our experiments run in the Cloud on a GeForce TITAN + 2 secondary Quadro K5000 GPU for voxel (render frames and perform H.264 encoding). Photon reconstruction use a GeForce 670.

^ Images of the scenes used for testing, as seen on the client with real-time dynamic indirect light. Our system is based over commercially-deployed systems like OptiX, hardware video encoders and NVIDIA

^ Our pipeline splits lighting between “direct” and “indirect”. View-independent “indirect” can be shared and amortized in

the Cloud over multiple clients

CloudLight : Interactive Indirect Illumination in the Cloud

H.264

NETWORK

Lossless

NETWORK

View-Independent Indirect Lighting

Final FrameRendering

ImageDisplay

LO

CA

LD

EV

ICE

CLO

UD

Photon Splat[Mara 13]

c) Photon

OptiXPhoton Trace

30 hz

HDMI

DisplayDirect + Post

30 fps 30 fps

b) Irradiance MapReconstruct Irradiance

Maps

OptiX Ray Trace

10 hz

HDMI

Display

30 fps

[Mitchell 06] + Direct + Post

30 fps

H.264

NETWORK

Voxel InjectionVoxel

Interpolate + Direct + Post

PCI

Voxel Cone Trace

[Crassin 11]

a) Voxel

15 hz 30 fps

Proprietary

30 fps

CloudLight

contact name

Cyril Crassin: [email protected]

P4166

category: Remote GRaPhiCs & Cloud-Based GRaPhiCs - RC02