GDC08_ParticleFluids

7/30/2019 GDC08_ParticleFluids

1/33

-

February 2008


2/33

Overview

Fluid Simulation Techni ues

A Brief History of Interactive Fluid SimulationCUDA particle simulation

Spatial subdivision techniques

Rendering methodsFuture

NVIDIA Corporation 2008


3/33

Fluid Simulation Techniques

Various a roaches:

Grid based (Eulerian)

Stable fluids

Particle level set

SPH (smoothed particle hydrodynamics)

MPS (Moving-Particle Semi-Implicit)

Height fieldFFT (Tessendorf)


. .


4/33

History: 2D Waves

Old demo-scene trick

2D wave equationFirst implemented on GPU

by Greg James, 2003

GeForce 3.

Used register combiners and

texture shader

8-bit fixed point



5/33

2D Fluid Flow

Mark Harris GPU Gems 1

2D Navier-Stokes solverbased on Jos Stams

a e u s

GeForce FX

Cg shaders

Multiple passes to solve

for pressure

Texture interpolation used



6/33

GPU Particle Systems

Buildin a Mill ion Particle S stem

Lutz Latta, GDC 2004Position and velocity

s ore n ex ures

Simple interactions with

terrain hei ht field and

implicit shapes

Emitters done on CPU

Particles rendered usingrender-to-vertex array



7/33

3D Fluids - Box of Smoke Demo

G80 launch demo

Written by Keenan Crane3D Navier-Stokes solver

Used tiled 2D textures

Also tracked free surfaces for waterBFECC advection scheme

Ray cast rendering



8/33

NVIDIA SDK Smoke Demo

3D Navier-Stokes solver

DirectX 10Used render to 3D

texture

Includes interaction



9/33

CUDA N-Body Demo

Computes gravitational attraction between n bodies

Computes all n2

interactionsUses shared memory to reduce memory bandwidth

16K bodies @ 44 FPS

x 20 FLOPS / interaction

x 16K2 interactions /frame

= s

GeForce 8800 GTX



10/33

Particle Systems

Particle S stems: A Techni ue for Modelin a Class

of Fuzzy Objects, Reeves 1983



11/33

Particle-based Fluid Simulation

Advanta es

Conservation of mass is trivialEasy to track free surface

Not necessarily constrained to a finite grid

Easy to parallelize

Disadvantages

ar o ex rac smoo sur ace rom par c esRequires large number of particles for realistic results



12/33

Particle Fluid Simulation Papers

Particle-Based Fluid

Simulation for InteractiveApplications,

. ,

3000 particles, 5fps

Particle-based Viscoelastic Fluid Simulation,

Clavet et al, 2005

1000 particles, 10fps20,000 particles,



13/33

CUDA SDK Particles Demo

Particles with sim le

collisionsUses uniform grid

ase on sor ng

Uses fast CUDA radix

sort

Current performance:

>100 fps for 65Kinteracting particles



14/33

Uniform Grid

Particle interaction requires finding neighbouringparticles

Exhaustive search requires n^2 comparisons

Uniform grid is simplest possible subdivision

Divide world into cubical gr id (cell size = particle size)Put particles in cells

Only have to compare each particle with the particles inneighbouring cells

Building data structures is hard on data parallelmachines like the GPU

possible in OpenGL (using stencil routing technique)


easier using CUDA (fast sorting, scattered writes)


15/33

Uniform Grid using Sorting

Grid is built from scratch each frameFuture work: incremental updates?

Algorithm:

center)

Calculate cell index

Find start of each bucket in sorted list (store in array)

Process coll isions by looking at 3x3x3 = 27 neighbouringrid cells of each article

Advantagessupports unlimited number of particles per grid cell



16/33

Example: Grid using Sorting

0 1 2 3 unsorted list sorted by cell start

4 5 6 7

3

2

1

45

,

0: (9, 0)1: (6, 1)

0: (4, 3)1: (4, 5)

0: -1: -

0

12 13 14 15

02: 6, 23: (4, 3)

4: (6, 4)5: 4 5

2: 6, 13: (6, 2)

4: (6, 4)5: 9 0

2: -3: -

4: 05: -6: 27: -8: -

:10: -...15: -



17/33

Spatial Hashing (Infinite Grid)

For ames we dont want articles to be

constrained to a finite gridSolution: use a fixed number of grid buckets, and

s ore par c es n uc e s ase on as unc on o

grid position

Pro: Allows rid to be effectivel infinite

Con: Hash collisions (multiple positions hashing to

same bucket) causes inefficiency

Choice of hash function can have big impactSee: Optimized Spatial Hashing for Coll ision



18/33

Example Hash Function

__device__ uint calcGridHash(int3 gridPos)

{

const uint p1 = 73856093; // some large primes

const uint p2 = 19349663;

const uint p3 = 83492791;

int n = p1*gridPos.x ^ p2*gridPos.y ^ p3*gridPos.z;

n %= numBuckets;

return n;

}



19/33

Smoothed Particle Hydrodynamics

Particle based fluid simulation techniqueOriginally developed for astrophysics simulations

Interpolates fluid attributes over

For games, we can often get away

with simpler simulationscom ne so par c e co s ons wattractive forces


Image courtesy Matthew Bate


20/33

Fluid Rendering Methods

3D isosurface extraction marchin cubes

2.5D isosurfaces (Ageia screen-space meshes)3D texture ray marching (expensive)

Image-space tricks (blur normals in screen space)



21/33

Marching Cubes

Standard method for extractin

isosurfaces from volume dataCUDA marching cubes uses

scan unc ons rom

library for stream compaction

Up to 8x faster than OpenGLgeometry shader implementation

using marching tetrahedra

But stil l re uires evaluatin f ield

function at every point in spaceE.g. 1283 = 2M points


ery expens ve


22/33

Ray Marching

Volume renderin

techniqueVoxelize particles

n o ex ure

Requires several

passes for thickness

Ray march through

3D texture in pixel

Can shade based

on optical thickness


ery n ens ve


23/33

Density-based Shading

Can calculate er- article densit and normal based

on field functionSPH simulations often already have this data

. .

cells) to get good results expensive

Can use density and normal for point sprite shadingNormal only well defined when particles are close to

each other

. .



24/33

Particle Density



25/33

Particle Normal



26/33

Flat Shaded Point Sprites



27/33

Blended Points Sprites (Splats)

Scale u oint size

so they overlapAdd alpha to points

w auss an a o

Requires sorting

from back to front

Has effect of

interpolating shading

e ween po n sFill-rate intensive,



28/33

Alternative Shading (Lava)

Modifies article

color based ondensity



29/33

Motion Blur

Create uads between revious and current article

positionUsing geometry shader

ry an or en qua owar s v ew rec on

Improves look of rapidly moving fluids (eliminates

a s between articles

p



30/33

Spheres



31/33

Motion Blurred Spheres



32/33

Oriented Discs



33/33

The Future

Practical ame fluids will need to combine article

height field, and grid techniques

GPU performance continues to double every

mon s


Two way coupled SPH and particle level set fluid simulation,Losasso, F., Talton, J., Kwatra, N. and Fedkiw, R

Adaptively Sampled Particle Fluids, Adams 2007

Documents

GDC08_ParticleFluids