EUROGRAPHICS 2002 / I. Navazo Alvaro and Ph.

Slusallek Short Presentations

(Guest Editors)

A Convolution-Based Algorithm
for Animated Water Waves

J. Loviscach

Fachbereich Elektrotechnik und Informatik, Hochschule Bremen, Bremen, Germany

Abstract

A non-linear partial differential equation solver is too sophisticated for computer graphics applications if they are
only used to render effects like circular waves and ship wakes. We present an approach which simulates waves
through a convolution algorithm. It handles both gravity waves and capillary waves; the latter are often neglected
even though they dominate small-scale behavior. The algorithm can be integrated into a complete solution archi-
tecture: First, standard commercial 3-D software is used to prepare an animated scene with objects traveling on
a water surface. Based on the movements of these objects, waves are calculated and added as bump and displace-
ment maps to the 3-D model.
Categories and Subject Descriptors (according to ACM CCS): I.3.7 [Computer Graphics]: Three-Dimensional
Graphics and Realism—Animation

1. Introduction
Water waves are an everyday phenomenon and thus their
simulation makes for a natural research topic in computer
graphics. Despite all past efforts, both the analytical and nu-
merical treatment of a liquid’s motion still pose consider-
able challenges. Fluid motion is described by the continu-
ity equation plus the Navier-Stokes equations—a system of
three non-linear partial differential equations—in conjunc-
tion with surface conditions, which typically are non-linear
as well. However, a full-blown numerical solution of these
equations is rarely needed to generate realistically-looking
animated films.
Since the motion in the volume of the liquid remains in-
visible in most scenes, it is sufficient to model the surface Figure 1: A raindrop simulation by the proposed algorithm
of the liquid. Additionally, water is virtually incompressible, shows the characteristic multiple rings formed by the inter-
its viscosity can often be neglected, and non-linear effects play of capillary and gravity waves. The results of the simu-
stay small if the waves’ height is small when compared to lation have been used for a bump map and a caustics texture
their length. When using an approximate, linear system of in a rendering created with Cinema 4D XL. Compare with
partial differential equations, one arrives at the conclusion2 figure 3 for the height field.
that small-amplitude surface water waves on an flat infinite
ocean with infinite depth are composed of sinusodial waves
(x,t) 7→ cos(k · x − ω(k)t + φ),
k ∈ R2 the wave vector along the surface, ω(k) the corre-
2
where x ∈ R is the position on the plane of the ocean, sponding circular frequency, and φ ∈ R the phase. The cir-

c The Eurographics Association 2002.

This is an electronic version of an article published in

Eurographics 2002 Short Presentations, pp. 383-389
(c) 2002 The Eurographics Association
 Loviscach / Water Waves

games because of its simple solvability8 . In order to realis-

tically simulate ring waves and Kelvin ship waves for 3-D
animated films without too much mathematical and compu-
tational overhead, we sought a simple solution to the linear
approximation of water surface waves outlined above.
Due to the complicated dispersion relation, there is no
simple finite-difference formulation for this approximation.
However, one can describe the propagation of waves with
help of convolution. Let the elevation of the wave field be
ηt (x) with η = 0 representing the surface at rest. Store the
momentary state of the ocean at time t as a complex-valued
function x 7→ ηt (x) ∈ C and use the real value of this func-
tion for display. Given ηt , one can calculate the elevation at
any time t + ∆t by convolution with a certain kernel function
Ψ∆t (x):
Figure 2: By calculating appropriate convolution kernels for ZZ
each swimming object at each time step, the algorithm is ηt+∆t (x) = d 2 y Ψ∆t (y) ηt (x − y). (2)
R2
able to handle arbitrary ship paths. Here, its results are used
as bump and displacement maps with Cinema 4D XL. The Formally, the kernel Ψ is given by the mathematically ill-
height field can been seen in figure 8. defined expression
d2k
ZZ  
Ψ∆t (x) = 2
exp i(k · x − ω(k)∆t) . (3)
R2 4π

cular frequency only depends on the wave number k = |k| Like in Huygen’s principle, this is the circular wave caused
and is given by by a disturbance at time 0 at the origin, observed after a time
r span of ∆t. In order to use this kernel with a lattice-based
T numerical approximation, it is necessary to introduce cutoffs
ω(k) = gk + k3 , (1)
ρ for long and short waves. These cutoffs have the additional
with gravity acceleration g, surface tension T , and mass den- benefit of eliminating the mathematical problems with this
sity ρ. Under normal circumstances, their values are g = expression.
9.8 kg m/s2 , T = 0.072 kg/s2 , and ρ = 1.0 · 103 kg/m3 . The convolution scheme allows to calculate each frame
The dispersion relation ω(k) causes a number of impor- without intermediate time steps. Sudden disturbances, such
tant visual effects: as a stone dropped into a pool, can be incorporated by alter-
ing ηt at the time t of this incident, for example by adding
• The phase velocity ω(k)/k reaches its minimum of a function to ηt which is zero outside the cross section of
23 cm/s at a wavelength of 17 mm. Shorter waves (cap- the stone. In order to model continuous disturbances such as
illary waves dominated by surface tension) as well as traveling ships, however, one has to integrate over the time-
longer waves (gravity waves) travel faster. Likewise, span ∆t.
the group velocity dω(k)/dk possesses a minimum of
We have developed a prototype system for water wave an-
18 cm/s. A stone thrown into deep water produces com-
imation with this algorithm. The solution is based on Maxon
plex ring waves where the innermost ring expands at this
Cinema 4D XL, but could use virtually any other commer-
speed. In contrast to this behavior of water surface waves,
cial 3-D software package. This program is used to model
waves of light or sound travel at a constant speed, re-
and animate a flat ocean and objects traveling on the wa-
gardless of their frequency. If water behaved this way, a
ter, which still is flat. These pictures are rendered black and
stone thrown into water would yield a single sharp circu-
white in a view form above and then fed into our special
lar wave.
add-on software. It computes the corresponding wave ele-
• A ship traveling at constant speed along a straight line on
vation fields, optionally separates the results into a blurred,
deep water leaves behind a Kelvin wake2 —a wedge with
low-pass version and a corresponding high-pass version. In
semiangle 19◦ . This angle is independent of the ship’s ve-
addition, the software calculates the caustics pattern on the
locity, once more in stark contrast to sound waves: The
ground. These parallel sequences of pictures are fed into the
Mach cone of a supersonic plane gets more pointed as the
3-D software as displacement maps, bump maps, and texture
speed of the aircraft increases.
maps for the final rendering—now including waves and seen
Therefore, an approach using the much simpler wave from an arbitrary point of view. The division into displace-
equation of sound and light often results in unrealistic behav- ment maps and bump maps allows for a relatively coarse
ior. Nonetheless, this equation is favored e. g. in computer tesselation resulting in short rendering times.

c The Eurographics Association 2002.
 Loviscach / Water Waves

The next section outlines previous work done for simu- steps, a damping term has to be introduced. Taking a differ-
lating water waves in 3-D computer graphics. Section 3 de- ent road, Stam16 proposes an intrinsically stable algorithm
scribes our convolution-based algorithm for time evolution. for the full Navier-Stokes equations. It can be simplified us-
Section 4 explains how objects which move on the water sur- ing Fourier transform17 . However, the presented algorithm
face are incorporated into this approach. A discussion of the does not allow free boundaries such as water surfaces. Foster
results follows in section 5. The final section 6 closes with a and Fedkiw4 combine Stam’s solver with marker particles3
summary and an outlook on possible extensions. to trace the fluid surface.
Weimer and Warren19 note that certain multi-grid partial
2. Previous Work differential solvers can be formulated with help of vector
The first two major works about water wave animation subdivision operations. The method they describe assumes
rooted in physics have been published in 1986: Fournier and a slow flow, i. e. viscosity dominates inertial behavior, and is
Reeves5 as well as Peachey14 use wave trains with the shape therefore not well-suited for water simulation.
of a trochoid respectively an approximation to a trochoid
shape. Waves running onto a shore are subjected to refrac- 3. Time Evolution by Convolution
tion and therefore compressed. Gonzato7 carries the idea of
To treat the problem numerically, we discretize the elevation
“wave tracing” further to calculate refraction due to varying
field ηt to a finite lattice of size L × L with N × N points.
water depth and the diffraction at obstacles. Here, capillary
It consists of the (p∆x, q∆x) with p, q = 0, 1, . . . , N − 1 and
waves are introduced, even though they are simply modeled
∆x = L/(N − 1). In our experiments, N = 256 or, in extreme
by fractal noise. In their recent work, Gamito and Musgrave6
cases, 512 proved to be sufficiently large. The real part of
offer a more detailed theory of wave tracing.
the computed ηt will be delivered as a bump map and/or
Thon and Ghazanfarpour18 use wave spectra measured displacement map to the 3-D rendering software. Therefore,
e. g. from photographs to synthesize ocean waves. Using such maps will then consist of e. g. 256 × 256 pixels.
quite a different approach, Schneider and Westermann15
To calculate the elevation field in an animated film from
simulate waves using fractal noise and then render these in
one picture frame to the next, we use the time evolution
real time with light reflection, refraction, and caustics with
given by equation 2 with ∆t being the time between frames.
modern consumer graphics hardware.
The double integral is approximated using the double sum
Goss9 presents an particle-system algorithm tailored to
η̃t+∆t (p∆x, q∆x) = (4)
display ship wakes in real-time. Khan10 superimposes cir-
M−1 M−1
cular waves from points along a ship’s path to simulate its
wake.
(∆x)2 ∑ ∑ Φ∆t (r∆x, s∆x) ηt ((p − r)∆x, (q − s)∆x).
r=−M s=−M
To allow a more complete interaction of the water with Here, ηt is treated as zero outside of [0, L] × [0, L]. M mea-
disturbances as well as changing boundaries, there have been sures the half size of a sliding window which is used for the
several approaches based on partial differential equations. convolution. We use M = 40, so that the window contains
Kaas and Miller11 use a linearized model of a shallow wa- 80 × 80 lattice points. Instead of the ill-defined Ψ∆t of equa-
ter limit, which is equivalent to the wave equation of light tion 3 a regularized version Φ∆t is used. Especially, Φ∆t has
and sound. This is solved by a finite-difference-method. Re- to smoothly approach zero on the boundary of the sliding
cently, Layton and van de Panne12 have improved these ideas window (see below).
by incorporating non-linear advection.
The η̃t+∆t (p∆x, q∆x) resulting from equation (4) cannot
O’Brien and Hodgins13 model the water volume as an ar- be used directly, because the resulting waves would reflect
ray of square columns which exchange fluid. If the content strongly at the borders. This reflection can be suppressed re-
of a column moves upward faster than a threshold veloc- liably by cutting off the elevation field near the borders in
ity, their algorithm generates spray. To achieve real-time re- linear fashion:
sponse, Chen and Lobo1 replace the full three-dimensional
Navier-Stokes equations by the two-dimensionial version. ηt+∆t (p∆x, q∆x) = h(p) h(q) η̃t+∆t (p∆x, q∆x) (5)
The local pressure calculated in this manner is used to steer with
a three-dimensional height field which represents the water 
 p/M if p < M
surface.
h(p) := 1 if M ≤ p < N − M
Foster and Metaxas3 simulate solutions of the full Navier- 
(N − 1 − p)/M if p ≥ N − M.
Stokes equations by using a voxel representation for the vol-
The resulting elevation field ηt+∆t will be used to calculate
ume and marker particles to trace the surface. Thanks to this
the next frame’s field ηt+2∆t , which in turn will serve as base
representation, their algorithm can produce e. g. breaking
for the succeeding one, etc.
waves and overturns. Surface tension is simulated by alter-
ing pressure gradients. To stabilize the method for large time The remaining task is to construct a regularized kernel

c The Eurographics Association 2002.
 Loviscach / Water Waves

sacrifice the fundamental propagation identity of Ψ:

ZZ
Ψs+t (x) = d 2 y Ψs (y) Ψt (x − y). (6)
R2

In order to preserve this property, we have chosen to modify

the dispersion relation given in equation (1) by replacing

ω(k) → ω(k) − ic(k), (7)

where c(k) ≥ 0 is a damping function depending on the wave

number. The well-defined analogue of equation 3 then is a
discrete Fourier transform of a function with fast decay.
To define c(k), one has to identify the waves which are
to be suppressed. Using the stationary phase method2 to ap-
proximate equation (3), one finds that the major contribu-
tions to Ψ∆t (x) arise from wave vectors k at which the ex-
ponent has zero gradient, i. e. x/∆t equals the group velocity
vector
∂ω(k) g + 3T |k|2 /ρ 0
vG (k) := = k . (8)
∂k 2ω(k)
Figure 3: Raindrop simulation by disturbing ηt at pseudo- We want to cutoff contributions to x which lie near the
random places and pseudo-random times. Parameters: L = border of the 2M × 2M window, or, to be specific, where
1 m, ∆t = 0.2 s, N = 512, 2M = 80. See figure 1 for a 3- |x| > 34 M∆x. Hence, the k which follow the following crite-
D rendering. To save calculation time, a quite low M has rion have to be supressed:
been selected. Consequently, Φ∆t is slightly cut off at the
boundary of the sliding window (see figure 4). This causes 3 M∆x
|vG (k)| > .
rectangular artifacts inside the rings. In the 3-D rendering, 4 ∆t
however, these remain nearly invisible. To this end, one can choose c(k) in the modified dispersion
relation (7) as follows:
 
5 3 M∆x
c(|k|) := |vG (k)| − (9)
Φ∆t . Since the time step ∆t is constant throughout the ani- M∆x 4 ∆t +
mation sequence, Φ∆t can be computed in advance, stored in where the pre-factor is somewhat arbitrary (this choice
an 2M × 2M array of complex numbers and used for every proved to be useful in our experiments) and where
picture frame. 
a if a > 0
(a)+ :=
0 else.

To implement the cutoff, we found two other promising

options which appear to merit further investigation: First,
equation (9) could be reformulated using the smoother ex-
pression
 2
100∆t 1 M∆x
|v G (k)| − .
(M∆x)2 2 ∆t +

Figure 4: The convolution kernel Φ∆t used in calculating Secondly, the group velocity could be kept below vmax :=
4 ∆t by approximating the dispersion relation ω(k) of
3 M∆x
figure 3. The left picture shows the real part, the right picture
the imaginary part. For better comprehensibility, they are equation (1) with
enlarged in comparison to figure 3. r
T
gk + k2 a arctan(k/a) + b2 ,
ρ
A necessary property of Φ∆t (p∆x, q∆x) is that it tends to where
zero at the boundary of its domain, i. e. as p or q → −M
or M − 1. This cannot simply be implemented by multiply- 2ρv2max g
a := and b := .
ing with a cutoff function in x-space. Especially, one would πT 2vmax

c The Eurographics Association 2002.
 Loviscach / Water Waves

4. Wave Generation by Swimming Objects The integral over s can be eliminated by partial integration:
Ships and other objects in the water surface generate waves ∆ηt+∆t (x) = (10)
through their motion and/or form changes. Our solution uses ZZ 
(1) (2)

the motion between two picture frames to calculate how the d 2 y ft (x − y)Φ∆t,v (y) + ft+∆t (x − y)Φ∆t,v (y)
R2
elevation field ηt+∆t is to be altered for the following frame.
where
For each swimming object, an additive contribution is com-
d2k
ZZ  
puted. (1)
Φ∆t,v (y) := exp i(k · y − ω(k)∆t)
2
R2 4π
For each object, we render the animation as a picture se- !
quence where the scene is shown from straight above, with e−iφ(k) + iφ(k) − 1
× −1 + iω(k)∆t
the ocean in black and the cross-section formed by the ob- φ(k)2
ject with the water surface in white. Such a rendering can
and
easily be obtained through Boolean operations available in
d2k
ZZ
standard 3-D animation software. (2)
Φ∆t,v (y) := exp(ik · y)
R2 4π2
!
eiφ(k) − iφ(k) − 1
× 1 + iω(k)∆t
φ(k)2

with φ(k) := (k · v − ω(k))∆t.

Figure 5: A picture sequence of the Boolean intersection

between ship and ocean rendered from above in black and
white is used to determine the ship’s momentary velocity vec-
tor and the function ft . For the simulation result see figure 8,
for a 3-D rendering see figure 2.

The function ft (x) describes the greyscale picture ren-

dered this way for time t, see figure 5. From ft and ft+∆t ,
we interpolate a corresponding moving and changing form
for all intermediate times s ∈ [t,t + ∆t]:
Figure 6: For ships moving at constant speed in constant di-
t + ∆t − s rection, the algorithm produces Kelvin wakes, characterized
fs (x) := ft (x − (s − t)v)
∆t by a wedge with a semiangle of 19◦ . Parameters: L = 100 m,
s−t
+ f (x + (t + ∆t − s)v) , ∆t = 4 s, N = 256, 2M = 80.
∆t t+∆t
where v is the object’s momentary velocity vector. It is cal-
culated from the trajectory of the center of gravity of ft . Again, we introduce a regularization by modifying the
dispersion relation as in equation (7). It helps to subtract an
Waves are generated by changes of fs , so we use gs := additional small imaginary offset from ω(k), such as 10−20 i.
∂ fs /∂s as disturbance. In analogy to equations (2) and (3), gs Otherwise, there could occur division by zero errors or over-
acting over the interval s ∈ [t,t + ∆t] leads to the following flow errors when φ(k) becomes zero or nearly zero. This
additive change to the elevation field ηt+∆t (x): happens when the phase velocity is approximately equal to
t+∆t
Z
d2k
ZZ ZZ the velocity of the object. The subtraction of a small imagi-
∆ηt+∆t (x) = ds 2
d2y nary number to avoid theses problems causes no harm: Ac-
t R 4π
2 R2
  tually, the integrands do not possess poles at these values of
× exp i(k · y − ω(k)(t + ∆t − s)) gs (x − y). k since several contributions cancel.

c The Eurographics Association 2002.
 Loviscach / Water Waves

Figure 7: The convolution kernels leading to the simulation

(1) (2)
result of figure 6. From left to right: Φ∆t , Φ∆t,v , and Φ∆t,v .
Upper row: real part, lower row: imaginary part.

The resulting expression for ∆ηt+∆t (x) is discretized in

the same way as in equation (4) and cut-off at the border in Figure 8: The algorithm can treat curved ship paths as well;
analogous form to equation (5). It has to be evaluated for like in nature, the Kelvin wake is distorted. (Simulation pa-
every swimming object and for every picture frame except rameters as in figure 6. For a 3-D rendering see figure 2.)
for the first one.

5. Results and Discussion

Figure 1 displays a simulation of raindrops, rendered with the size of a ring wave ∆t after a point-like disturbance has
Maxon Cinema 4D XL. At pseudo-random moments, the el- occurred. This ring has to fit into the window, but should
evation field will be offset at single, pseudo-randomly placed not be much smaller. The sharply defined inner wave p front
lattice points. Our system reproduces the typical ripples pre- expands radially at a speed set by the minimum 1.1 4 T g/ρ
ceding the relatively large inner ring. of the group velocity. The outer front of the ring wave de-
cays much more softly, so that its velocity is harder to de-
The convolution-based approach produces ship wakes fine. The asymptotic expansion of gravity waves2 contains
which form the correct angle, as can been seen in figure 5. the term cos(g(∆t)2 /4r − π/4), where r is the distance from
By recomputing the kernels, curved paths can be handled as the origin of the disturbance. The largest r for which this
well, see figure 2. Just like in nature, the wedge pattern of expression attains a maximum obeys g(∆t)2 /4r − π/4 = 0.
the Kelvin wake is distorted along the curve. Therefore, an estimate for the outer size of the ring wave can
Up to now, the system does neither reflect nor diffract be given as g(∆t)2 /π. In our experiments, we still got con-
waves at boundaries or swimming objects. Instead, waves vincing results when half of the size of the sliding window
move through obstacles. This remains nearly invisible in was 1/8 of this size.
open-ocean scenes, but can be noticed when shores, is-
This leads to the following conditions on ∆t, ∆x, and M:
lands, or pool walls come into play. In addition, the simula-
tion lacks shallow-water effects, especially the refraction of ! ! !
∆x  1.1 4 T g/ρ ∆t  M∆x, g(∆t)2 /π ≈ 8M∆x.
p
waves hitting a shore. Non-linear effects are not yet incorpo-
rated into the algorithm. Since the water surface is modeled Furthermore, the objects’ velocity moving across the water
by an elevation field, waves cannot break. must not be too large: The waves they generate within ∆t
The discretization to a grid causes little visual defects. Of have to fit into the window used for approximating the k-
(1) (2)
course, the time step ∆t and the grid interval ∆x must be cho- integrals of Φ∆t,v and Φ∆t,v in equation 10.
sen in such a way that the sliding window contains the im-
Because there is only one computational step per pic-
portant features of a ring wave. The size 2M∆x of the win-
ture frame, the stability of the solution algorithm is not a
dow must be large enough to capture the major part of the
very critical issue. In addition, the discretized system obeys
ring wave. On the other hand, the wave must not degenerate
a propagation identity corresponding to equation (6) which
to a few pixels in the middle.
rules out many run-away effects. An open question remains
To adjust the values of ∆t, ∆x, and M, one can estimate whether the suppression of boundary reflections according

c The Eurographics Association 2002.
 Loviscach / Water Waves

to equation (5) can cause instability. However, in our exper- 4. N. Foster and R. Fedkiw. Practical Animation of Liq-
iments such effects were not visible. uids. SIGGRAPH 2001 Conference Proceedings, An-
nual Conference Series, pp. 23–30, 2001. 3
Using the prototype software, a PC equipped with an
Pentium(R) -III processor running at 500 MHz computes 5. A. Fournier and T. Reeves. A Simple Model of Ocean
waves at a speed of 90 s per picture frame (L = 256, 2M = Waves. ACM Computer Graphics (Proc. of SIGGRAPH
80, one swimming object). The most time-intensive part of ’86), 20(4): 75–84, 1986. 3, 7
the algorithm consists in the convolution, even though we
6. M. N. Gamito and F. K. Musgrave. An Accurate Model
limit the evaluation of equation (10) to the neighborhood of
of Wave Refraction Over Shallow Water. Computers &
swimming objects and pre-compute Φ(1) and Φ(2) per object Graphics, in press. 3
for each frame. This pre-computation and the discretized cal-
culation of equation 3 could be implemented through a Fast 7. J. Ch. Gonzato and B. Le Saëc. On Modelling and Ren-
Fourier transform (FFT), but this would not yield an appre- dering Ocean Scenes—Diffraction, Surface Tracking
ciable acceleration. and Illumination. Journal of Visualization and Com-
puter Animation, 11(1):27–37, 2000. 3

6. Summary and Future Work 8. M. Gomez. Interactive Simulation of Water Surfaces.

Game Programming Gems. Edited by M. DeLoura,
We have presented an approach to simulate water surfaces pp. 187–194, Charles River Media, 2000. 2
waves using a simple algorithm. Even though it neglects
9. M. Goss. A Real Time Particle System Display of Ship
nonlinearity and viscosity, the model believably reproduces
Wakes. IEEE Computer Graphics and Applications,
visually important effects such as the distinct behavior of
10(3):30–35, 1990. 3
capillary and gravity waves. It produces realistic ring waves
and ship waves. 10. R. S. Khan A Simple Model of Ship Wakes. Master
thesis, University of British Columbia, 1994. 3
Typically, only a close inspection of the results reveals
shortcomings of this algorithm when compared to more 11. M. Kass and G. Miller. Rapid, Stable Fluid Dynam-
complex approaches. To further improve the simulation, one ics for Computer Graphics. ACM Computer Graphics
can think of several tasks: (Proc. of SIGGRAPH ’90), 24(4): 49–57, 1990. 3
• Simulate ocean waves using a pseudo-random realization 12. A. T. Layton and M. van de Panne. A Numerically
of a spectrum measured in nature18 . Efficient and Stable Algorithm for Animating Water
• Reflect waves at boundaries; diffract waves at obstacles. Waves. The Visual Computer, 18:41–53, 2002. 3
• Generate foam at sharp wave crests. It could be rendered
13. J. F. O’Brien and J. K. Hodgins. Dynamic Simulation of
using a texture map or via a particle system14 .
Splashing Fluids. Proceedings of Computer Animation
• Simulate wakes stirred by the propellers of ships.
’95, pp. 198–205, 1995. 3
• Calculate velocity data to render motion blurred waves.
• Simulate non-linear effects of hydrodynamics—for in- 14. D. R. Peachey. Modeling Waves and Surf. ACM Com-
stance, by including self-advection17 or by deforming the puter Graphics (Proc. of SIGGRAPH ’86), 20(4): 65–
waves in order to produce steeper peaks5 . 74, 1986. 3, 7
After freezing the features of the solution, we want to 15. J. Schneider and R. Westermann. Towards Real-Time
speed up the run-time performance of the algorithm using Visual Simulation of Water Surfaces. Vision, Modeling,
the floating-point SIMD instructions of common micropro- and Visualization 2001, 211–218, 2001. 3
cessors to implement convolution and FFT. 16. J. Stam. Stable Fluids. SIGGRAPH 1999 Conference
Proceedings, Annual Conference Series, pp. 121–128,
References 1999. 3
17. J. Stam. A Simple Fluid Solver Based on the FFT. Jour-
1. J. X. Chen, N. da Vitoria Lobo. Toward Interactive-
nal of Graphics Tools, 6(2):43–52, 2002. 3, 7
Rate Simulation of Fluids with Moving Obstacles Us-
ing Navier-Stokes Equations. Graphical Models and 18. S. Thon and D. Ghazanfarpour. Ocean Waves Synthesis
Image Processing, 57(2):107–116, 1995. 3 and Animation Using Real World Information. Com-
puters & Graphics 26:99–108, 2002. 3, 7
2. L. Debnath. Nonlinear Water Waves. Academic Press,
1994. 1, 2, 4, 6 19. H. Weimer and J. Warren. Subdivision Schemes for
Fluid Flow. SIGGRAPH 1999 Conference Proceedings,
3. N. Foster and D. Metaxas. Controlling Fluid Anima-
Annual Conference Series, pp. 111–120, 1999. 3
tion. Computer Graphics International 97, pp. 178–
188, 1997. 3

c The Eurographics Association 2002.