Flame On:

Real-Time Fire Simulation for Video Games

Simon Green, NVIDIA
Christopher Horvath, Pixar
Introduction
This talk is about achieving realistic looking
simulations for games / visual effects
Not necessarily physically accurate!

There is a large artistic component

Overview
2D fire simulation using CUDA
Sneak peak: 3D fire simulation using DirectX 11
5 Tips For Good Looking Fluid Sims


A Brief History of
Eulerian Fluids on the GPU


Stable Fluids,
Jos Stam, Siggraph 1999

Mark Harris' 2D fluid solver
(GPU Gems 1, 2004)

3D fluid solver
(GPU Gems 3, Crane, Llamas, Tariq, 2007)

APEX Turbulence
(Cohen, Tariq, 2010)
Interactive Fluid-Particle Simulation using Translating Eulerian Grids
Inspiration
Directable, high-resolution simulation of fire on
the GPU, Horvath, Geiger, SIGGRAPH 2009
Computes high-res 2D slices of
a 3D simulation
Seeded using particle system
GPGPU - used OpenGL
Used in Harry Potter film
Goal Interactive Fire for Video Games
Most video games today use 2D sprites for fire
Procedural, or based on filmed footage
3D simulation probably still too expensive for
real-time use today?


Todays Video Game Fire
Simulated Fire
Advantages
High resolution
Non-repeating animation
Can respond to wind etc.
Less storage (?)

Disadvantages
Computation time
Artist controllability
Implementation
Implemented 2D stable fluids solver in CUDA
Uses pitch-linear textures to store fields
cudaMallocPitch / cudaBindTexture2D
Geometric multi-grid solver
Credit: Nuttapong Chentanez
OpenGL for rendering
Shading done in GLSL pixel shader


Example CUDA Kernel
__global__
void pressureSolveD(float * __restrict__ newPressure,
const float * __restrict__ divergence,
int width, int height,
int pitch)
{
int x = blockIdx.x*blockDim.x + threadIdx.x;
int y = blockIdx.y*blockDim.y + threadIdx.y;
int i = y*pitch+x;
if (x >= width || y >= height) return;

float2 pos = make_float2((float)x + 0.5f, (float)y + 0.5f);
float pL = tex2D(pressureTex, pos.x - 1, pos.y);
float pR = tex2D(pressureTex, pos.x + 1, pos.y);
float pB = tex2D(pressureTex, pos.x, pos.y - 1);
float pT = tex2D(pressureTex, pos.x, pos.y + 1);

float bC = divergence[i];
float pNew = (pL + pR + pB + pT - params.dx2*bC) * 0.25f;

newPressure[i] = pNew;
}
Fire Recipe
Take smoke simulator
Velocity, density
Add new channels
Temperature, Fuel, Noise
Add a simple combustion model
Combustion consumes fuel, generates heat
Heat also generates upwards buoyancy force

Tip 1 Get the Colors Right
Need to map temperature to color
use physically-based black body radiation model
(see later)
Or: just an artist defined color gradient
Dynamic range is important
Fire is very bright!
Can apply curve to density to get sharp flame
edges
Temperature
Color
Tip 2 Use High Quality Advection
Advection determines quality of motion and appearance
detail in velocity and density fields
Bilinear filtering not really good enough
To much blurring over time
Lots of other options:
Higher-order filters (cubic)
Error correction schemes e.g. MacCormack
Particle based - PIC/FLIP
We used Catmull-Rom filter, bounded to neighbourhood
Tip 3 Use a High-Res Density Field
Density field can be much higher resolution than
velocity field
4x or more
Read interpolated velocity field when advecting
density
Need to downsample density to velocity
resolution if simulation is coupled
i.e. buoyancy based on density
Tip 3 Post Processing is Important
Fire is very hot!
Use post-processing to communicate
temperature to viewer
Glow - blur HDR image, add back on top
Heat distortion offset background based on
gradient of temperature
Motion blur
sample image several times along velocity vector
No Glow
With Glow
With Motion Blur
Tip 5 - Embers
Add particles passively advected by velocity field
Shows motion of air even in empty regions
Motion blurred
Drawn as quads stretched between previous and
current position (using geometry shader)
Inherit temperature from simulation
Cool over time
Tip 4 Just Add Noise
Fire is very turbulent and fast moving
Use high levels of vorticity confinement to
preserve vortices
Use procedural (curl) noise to add turbulence
Also advect a 2D noise field
Blend in small amount of noise each frame
Can be used to add detail to other fields
Noise moves with fire
Work in Progress - 3D Simulation
Relatively simple to extend simulation to 3D
Surface writes to 3D textures are now possible
in CUDA 4.x and DirectX 11


3D Performance
Texture performance is great on Kepler
architecture
Sample results:
128 x 128 x 64 (0.5M) voxels for sim, 64 solver steps
2x res density field (8M voxels)
17 msecs per frame, including rendering
(GeForce GTX 680)

Tip 5 Add Light Scattering
Simple scattering approximation
Similar idea to Light propagation volumes
(Discrete Ordinate Method)
Basic algorithm:
Render radiance to 3D texture
Blur radiance in 3 dimensions
Sample blurred radiance (indirect light) in volume
render
Demo
Physically Correct Flame Rendering
Or... How to get the Planck Blackbody Radiation
Function to actually look right.
Overview
Components of Flame Appearance
Blackbody Radiation
Spectral Emission
Tristimulus Response
CIE XYZ
Direct RGB (Human,Camera,Infrared)
Flame Appearance
Blackbody Radiation of Combustion Byproducts
(Soot/Smoke) - The Red/Orange/Yellow part.
Spectral Emission - The Blue/Purple/Green
part.
http://en.wikipedia.org/wiki/Flame
Blackbody Radiation
Plancks Law
Relationship between emitted radiant intensity
at each individual wavelength of light with
temperature
Calculated for a spectrum of wavelengths
within visible range (380nm to 780nm). With
5nm increments, this is a color sample with 81
values.
Spectral Emission
Dependent on type of fuel
Dependent on mixture of oxygen
Also defined as an intensity per wavelength
Butane Spectrum
Stimulus Response
For a given receptor, a Stimulus Response Curve
represents the sensitivity of that receptor to
each individual wavelength of light
The integral of a Stimulus Response Curve with
an Emission Spectrum produces a single scalar
receptor response to a spectrum of radiation
Tristimulus Response
Combination of Stimulus Response Curves for a
triplet of receptor types
Human Color Vision composed of three types of
cells with different spectral sensitivites, called
cones. (L, M, S)
Color Photography created from three types of
color sensitive films or sensors, or alternatively
three different filters (Technicolor)
CIE XYZ
Created in 1931 by International Commission
on Illumination
CIE RGB
RGB curves have negative spectral response
Human Spectral Sensitivity
Significantly Overlapping in Red and Green
Digital Camera Response
Significant Infrared Sensitivity without IR Filter
Digital Camera Response
Significant Infrared Sensitivity without IR Filter
1300K Flame via CIE
Overly Saturated, Computer Generated look.
1300K Flame via cRGB
Properly Balanced Flame Appearance
Questions?
Thanks
Chris Horvath
Mark Harris
Nuttapong Chentanez
References
Jos Stam, "Stable Fluids", In SIGGRAPH 99 Conference Proceedings, Annual
Conference Series, August 1999, 121-128 PDF
Fast Fluid Dynamics Simulation on the GPU, Mark Harris, GPU Gems
Real-Time Simulation and Rendering of 3D Fluids, Keenan Crane, Ignacio
Llamas, Sarah Tariq, GPU Gems 3
Capturing Thin Features in Smoke Simulations, Siggraph Talk 2011, Magnus
Wrenninge, Henrik Falt,Chris Allen, Stephen Marshall PDF