Proceedings of the COST G-6 Conference on Digital Audio Effects (DAFX-00), Verona, Italy, December 7-9, 2000

A PARALLEL 3D DIGITAL WAVEGUIDE MESH MODEL WITH

TETRAHEDRAL TOPOLOGY FOR ROOM ACOUSTIC SIMULATION

Guilherme Campos and David Howard

Music Technology Research Group

Department of Electronics - University of York
agrc101@york.ac.uk

ABSTRACT 2. DIGITAL WAVEGUIDE MODELLING

Following a summary of the basic principles of 3D waveguide

By discretising time and space, the travelling wave solution to the
mesh modelling and the context of its application to room
1-D wave equation for either flow or pressure can be
acoustic simulation, this paper presents a detailed analysis of the
implemented digitally with a bi-directional pair of delay lines.
tetrahedral mesh topology and describes its implementation on a
Such a structure is called a digital waveguide [3].
parallel computer model. Its structural characteristics are
analysed, with particular emphasis on how they influence The point of intersection of n digital waveguides is called a
execution speed. Performance deterioration due to scattering junction or node. Using scattering junctions, multi-
communication overhead in the parallelised model is discussed. dimensional digital waveguide meshes can be constructed [3]. In
Theoretical predictions are compared with data from performance this paper, superscripted + and denote wave components
tests carried out on different computer platforms and both are travelling respectively to and from a junction.
contrasted with the corresponding results from the rectilinear Assuming lossless transmission (i.e. neglecting energy
model, in order to assess the practical efficiency of the model. absorption by the propagation medium), the sound pressure ( at
Objective validation tests are reported and discussed. a scattering junction can be expressed as a function of incoming
sound pressure travelling waves ( i+) [4]:
1. INTRODUCTION n
2 pi+ Z i
(1)
The applications of room acoustic modelling range from auditoria p= i =1
n
design to the development of musical equipment and the creation 1 Zi
of tools for sound synthesis and manipulation. Although varying i =1

degrees of accuracy may be required by different applications,

where i is the acoustic impedance along each line of
auralisation can be regarded as the ultimate objective.
propagation.
The problem of auralisation amounts to solving the sound
The sound pressures ( i = i+ + i-) in all crossing waveguides
wave equation for a particular room under analysis.
are equal at the junction [4]:
Unfortunately, analytical solutions can be obtained only for very
simple geometries and idealised boundary conditions [1]. p1 = p 2 = ... = p n = p (2)
Models based on numerical methods, known as physical
models, have to be used. Excessive computational loading has In a mesh, all waveguide segments between nodes have
been their main drawback for practical application in room the same length, being called . Wave components take
acoustic modelling. one to travel the distance through a delay unit from a
Simplified approaches to the analysis of sound propagation node to the one [4]:
have been proposed to circumvent this problem, namely
geometric models inspired by optics [1]. However, their results pi+ [nT ] = pi ,opp [(n 1)T ] (n N) (3)
are far from satisfactory when a high level of perceptual accuracy
Strings and air columns can be modelled using one-
is required [2], as is the case in the acoustical reconstruction of
dimensional . Higher-dimensional models,
ancient buildings or structures for musical or musicological
known as , can be formed by digital
purposes.
waveguides interconnected in regular arrangements [3].
Finding ways of making physical models practical for room
The application of two-dimensional models to the simulation
acoustic simulation is therefore a research priority. The work
of acoustic membranes and percussion instruments has been
reported in this paper is based on digital waveguide modelling, a
particularly successful. Very rich and natural-sounding timbres
Finite Difference Time Domain (FDTD) method developed for
can be obtained [5]. This adds to the idea that 3D meshes have
musical applications by Van Duyne and Smith [3]. Being a
the potential to provide accurate room simulation.
physical modelling technique, it automatically accounts for all
wave propagation phenomena, including diffraction. The main handicap of the model, common to any FDTD
method, is . This means that wave propagation
speed is variable with frequency and direction of propagation.

DAFX-1
Proceedings of the COST G-6 Conference on Digital Audio Effects (DAFX-00), Verona, Italy, December 7-9, 2000

Higher-frequency components generally lag behind, causing

direction-dependent signal distortion [3].
This error, as well as those caused by finite spatial resolution x=y=z 3. 4.
in the representation of boundaries, can be reduced by increasing
the mesh density, but at the expense of computation time Figure 2 shows there are 16 bi-directional delay units per unit
Several 2D mesh topologies have been studied, particularly cell or, equivalently, 2 per node. Figure 2 highlights the 7 bonds
rectilinear, hexagonal and triangular, the aim being to find the between nodes within the same unit cell.
best compromise between dispersion error, computation speed z
and ease of implementation [5].
Similar efforts are being applied to the 3D case. This paper
explores the tetrahedral structure proposed by the originators of
the waveguide-mesh modelling technique [6]. They show that
G
tetrahedral models compute a valid finite difference
approximation to the 3D wave equation and point out several
F D
potential advantages which could lead to practical application in
room acoustic simulation.
C
B
3. THE TETRAHEDRAL MESH STRUCTURE
A H x
Tetrahedral meshes replicate the molecular structure of the
diamond crystal, with nodes corresponding to carbon nuclei and E
the tetrahedrally-spaced bi-directional delay units around each
node corresponding to the chemical bonds with neighbouring
nuclei [6]. Figure 1 presents the geometry of the tetrahedral (a)
arrangement: the four neighbours of a node are vertices of a cube y
whose geometrical centre is the given node. The angle between s/4
bonds is 2arctan 2 109.47 . The inter-nodal distance is (b)
Front view Side view
d= 3 s/2, s being the edge length of the tetrahedral site [7].
R Plan view
C F D G D G C F
P R u
Q B H B H
s
d A E A E
s
P 2*s Figure 2. [9]
2
Q S A mesh is formed by unit cells placed adjacent to one
another, as figure 3 illustrates.
S
Figure 1. A tetrahedral site z
An important concept in the analysis of crystalline structures 0
is that of the unit cell the smallest repeating unit which shows
the full symmetry of the crystal structure ysize-1
cubic zsize-1
a x

0
y 0 xsize-1

Figure 3.
| | s

Considering only an infinitesimal cell boundary contour

_
| | s displacement, 0 ( , , ), fig. 2 holds and the node
b coordinates of a generic unit cell ( , , ) in the model of fig. 3 can
be worked out as:

DAFX-2
Proceedings of the COST G-6 Conference on Digital Audio Effects (DAFX-00), Verona, Italy, December 7-9, 2000

A (i, j, k) s(i, j, k)+(0, 0, 0); mesh, because of the eight different node positions in a unit cell.
B (i, j, k) s(i, j, k)+ (s/2, s/2, s/2); Table 2 identifies the neighbours corresponding to each position.
C (i, j, k) s(i, j, k)+(0, s, s);
Nodes in cell (i, j, k)
D (i, j, k) s(i, j, k)+ (s, 0, s);
A B C D E F G H
E (i, j, k) s(i, j, k)+ (s, s, 0);
i, j, k B A/E/D/C F/B G/B H/B C D E
F (i, j, k) s(i, j, k)+ (s/2, 3s/2, 3s/2); i, j, k-1 F/G
G (i, j, k) s(i, j, k)+ (3s/2, s/2, 3s/2); i, j, k+1 E E
H (i, j, k) s(i, j, k)+ (3s/2, 3s/2, s/2); i, j-1, k F/H
i, j+1, k D D
where 0 i xsize-1, 0 j ysize-1, 0 k zsize-1 and 2s = u is G/H
i-1, j, k
the edge of the unit cell. C C
i+1, j, k
i, j-1, k-1 F
4. THE TETRAHEDRAL MODEL ALGORITHM i-1, j, k-1 G
i-1, j-1, k H
i, j+1,k+1 A
Each tetrahedral mesh node is represented by 10 fields, as i+1, j, k+1 A
detailed in table 1. Eight of them implement bi-directional i+1, j+1, k A
communication ports with the 4 neighbouring nodes.
Table 2. Map of node interconnections.
Field Symbol Data type
There are links not only to the 6 unit cells with faces adjacent
Node pressure (air nodes) / to it (as was the case in the rectilinear model) but also to 6 other
p / 4-byte float
/ Reflection factor (boundary nodes)
Output port p1+ 4-byte float
Propagation axis 1
Input port p1 4-byte float
+ 4-byte float
Output port p2
Propagation axis 2 i+1, j, k+1
Input port p2 4-byte float
(right/up)
Output port p3+ 4-byte float i, j+1,k+1
Propagation axis 3 i-1, j-1, k (front/up)
Input port p3 4-byte float
(left/back)
Output port p 4+ 4-byte float
Propagation axis 4 i, j, k
Input port p4 4-byte float
i+1, j+1, k
Configuration node_config char (right/front)
Table 1. Node structure
i-1, j, k-1
As in the 3D rectilinear model [10], nodes can be configured (left/down) i, j-1, k-1
(node_config) as air nodes or boundary nodes. The field that (back/down)
holds the wave variable p in the former case, is used to store the
reflection factor of the corresponding surface in the latter. Air Relative position of the six neighbouring unit
nodes can be assigned special functions, namely mesh excitation cells exchanging data with a generic unit cell (i, j, k),
(modelling sound sources) and output (modeling sound through a common edge
receivers). This allows the acoustic characteristics of any room
with various surface coverings and different source and receiver
locations to be modelled. 5. PARALLEL IMPLEMENTATION
The modelling algorithm is the iterative two-pass
computation for lossless propagation [3] applied before [10] to
the 3D rectilinear topology.
The scattering pass calculations for boundary nodes remain
exactly the same, but for air nodes equations 1 and 2 yield:
1 +
p= ( p1 + p 2+ + p 3+ + p 4+ ) (4)
2 model blocks

pi = p pi+ i {1 ,2 , 3, 4} (5)

The delay pass implements equation 3 for every delay unit in

the mesh. Data transfer is more complex than in the rectilinear

DAFX-3
Proceedings of the COST G-6 Conference on Digital Audio Effects (DAFX-00), Verona, Italy, December 7-9, 2000

[xsize.ysize.zsize] node arrays, corresponding to designations A 3c

through H; this was the implementation approach adopted. d=
fu
Based on a map of their positions on a 3D grid, a master
program spawns one slave task per block and enables it to c fu T update
communicate directly with the tasks modelling neighbouring frequency T
blocks, for data transfer between surface nodes. This requires that fu
model blocks be associated with appropriate sets of fs
communication buffers. Dispersion error
In order to help improve performance, the computation
sequence in each iteration cycle (fig. 5) is organised so that data
required by neighbours is made available as early as possible and
data required from them can be received as late as possible [10].
T 4d / 3
Computation stage 1: Scattering pass for surface nodes.
First stage of delay pass: data transfer from external output
ports of surface nodes to out-buffers

Data Send: Packing and sending of data from out-buffers

to corresponding neighbours
3
Computation stage 2: Scattering pass for interior nodes 2d
tetrahedral = s3 =
Computation stage 3: Second stage of delay pass: data 3
transfer to and from interior nodes

Data Receive: Waiting stage if necessary. Receiving and

unpacking of data to in-buffers from corresponding rectilinea r = d3
neighbours
Computation stage 4: Last stage of delay pass: data
transfer from in-buffers to external input ports of surface 1/ v
nodes
3
3
(1 / v )tetrahedra l
3
Final 2d 3
= = 65%
iteration? (1 / v )rectilinea r 1
3
2
d
Figure 5. Sequence of main computation stages in one rx ry rz
cycle of parallelised operation.[10]
tx ty tz
The main difference to the rectilinear topology in terms of
data transfer is that it occurs not only between any two blocks
with a common face but also between blocks having a common ri d =
4d
ti ri =
4
ti i {x, y, z}
3 3
edge on certain positions. The symmetry properties exhibited at
ri ti
equivalent
unit cell(s) model block(s)

3
44100
TM = .V . RT 60 . f Sn4 .t n
2c

V m3 RT60
s tn=44100t
6. COMPUTATION TIME s
fn

1/ 3 fS
fn =
d 44100 Hz

DAFX-4
Proceedings of the COST G-6 Conference on Digital Audio Effects (DAFX-00), Verona, Italy, December 7-9, 2000

The only difference between the two computation time 0.2 tn (s)
formulae is the numerical factor, 65% smaller in the tetrahedral Test 2

mesh as a direct result of its lower density. 15.3% 12.3% 10.1% 8.3%
0.15
In addition, tn should be lower:
43.1%
1) The scattering pass (equations 4 and 5) involves only 7 0.1
27.3% 26.9% 22.6%
algebraic additions and one division by 2, as opposed to 11 Test 1
additions and a division by 3 with the rectilinear model 6-port
0.05
junctions. The division by 2 cannot be implemented here through
bit-shifting as calculations are all floating-point. Block’s edge (m/fn)
0
2) As there are only 2 delay units per node (3 in the 0.25 0.5 0.75 1 1.25
rectilinear mesh), the number of delay pass operations also
decreases by 1/3.
Figure 6. Comparative communication overhead tests.
Thus, with tn reduced by approximately 1/3 (see figure 5), TM The percentage values indicate performance deterioration
would be 57.1% lower than in the rectilinear mesh. in test 2 (worst-case inter-block communication) relative
In parallelised operation, the computation time is affected by to the tn values from test 1 (no inter-block communication)
inter-block data transfer. The ratio S/V between the number of
surface delay units (connecting to exterior nodes) and the total As predicted, after a steep decrease in the initial part of the
number of nodes gives a clear indication of communication graph, the communication overhead ratio between equivalent
overhead. This ratio, a decreasing function of block size, is tetrahedral and rectilinear blocks stabilises at a value only slightly
minimised for cubic shapes in both the rectilinear and tetrahedral greater than that given by equation 13. The higher percentages
topologies. For equivalent cubic blocks, it can be shown that: seen for the tetrahedral mesh are due mainly to communication
overhead being compared to much lower values of tn.
( S / V )tetrahedral 2 1 2
= 1 (13) Normal parallel model operation was tested on a much faster
( S / V )rectilinear 3 3n 3 SGI Origin 2000

where n is the edge of the rectilinear block, in number of nodes.

This indicates a higher communication overhead in the
tetrahedral mesh. The larger number of active neighbours and
communication buffers (respectively 12 and 60 against only 6 and
12 in the rectilinear case) is likely to aggravate the situation,
especially for small blocks.
55
tn (ms)
50
7. PERFORMANCE TESTS
45

40
In order to assess the efficiency of the model and the effect of
35
communication overhead, the same technique as used in [10] for
30
the rectilinear model was adopted. Two simple tests were carried 1 processor
25
out on identical SGI O2 workstations connected by a network
(100 Mbit/s Ethernet). In the first test, only one cubic block (no 20
2 processors
neighbours) was tested under PVM on one of the machines, with 15

all send-receive operations disabled. In the second test, 2 tasks 10 4 processors

were spawned, one on each workstation, both modelling blocks 5
identical to the previous one. Blocks were made to interact as if 0
both had the maximum possible number of active neighbours 0.2 0.6 1 1.4 1.8 2.2 2.6 3 3.4 3.8 4.2 4.6
Edge of equivalent cubic model (m/fn)
(12), by sending all the corresponding buffers to each other,
therefore simulating a worst-case scenario in terms of
communication overhead. Multiprocessor comparative performance of the
tetrahedral mesh (solid lines) and rectilinear mesh
To allow a comparative analysis, the two tests were applied to (dashed lines).
sets of equivalent tetrahedral and rectilinear blocks with
increasing size. The results are presented in figure 6.
The abscissa unit is scaled by fn (see equation 12) to make the
graph applicable to any sampling frequency (this also applies to
figure 7).
Test 1 reveals a performance improvement in excess of the
theoretical prediction of 33%, caused by delay pass calculations
being faster than expected. This may be related to more efficient
memory access, resulting from data being transferred between
interlaced small-size arrays rather than within a single large array.

DAFX-5
 Proceedings of the COST G-6 Conference on Digital Audio Effects (DAFX-00), Verona, Italy, December 7-9, 2000

communication overhead is much more apparent than in the

rectilinear mesh, increasing with number of blocks, as expected.
Performance gains relative to the rectilinear model are
variable with block size, reaching more than 100% for the largest
sizes tested.

8. MODEL VALIDATION

Using a model parallelised over 6 processors on the computer 10. REFERENCES

tested before, a 2s impulse response at 44.1kHz was obtained for
a rectangular room with dimensions as indicated in figure 8. The Room Acoustics
source and the receiver were placed at opposite corners of the
room model.
Figure 8 shows the initial portion of the frequency spectrum
obtained by applying FFT to the last 1.8s of the impulse response, Proceedings of the Institute of Acoustics
using uniform windowing for best frequency resolution.
There is an excelent correspondence between the modal
frequencies calculated analytically and the spectral peaks, ICMC
providing objective validation of the model. The slight Proceedings
differences observed, more apparent for higher frequencies, can
be attributed to dispersion error [12].

9 4 .9 12 0 .3 12 7.5
oblique , Espoo,
tangential
6 5.2 8 1.1 8 4 .7 9 8 .6 10 7.3 10 9 .8 12 0 .2 13 0 .4 13 7.3
Finland, 24-27 Sept., pp 455-458).
axial
4 2 .7 4 9 .2 6 8 .9 8 5.5 9 8 .4 12 8 .2 13 7.8
[5]

, vol. 84, pp 529-542).

[6]

f (Hz)
ICMC Proceedings,
20.0 40.0 60.0 80.0 100.0 120.0 140.0

. John Wiley &

Figure 8. Comparison between the spectral modes
Sons.
calculated analytically for a room with W=4.025m,
D=3.495m; H=2.496m and the frequency response of the [8] West, A. R. (1988) . John Wiley
corresponding 3D tetrahedral model & Sons.
[9] Smart, L. and Moore, E. (1995)
Preliminary listening tests with anechoically recorded sounds
. Chapman & Hall.
suggest that the model is behaving appropriately. In addition,
binaural room responses obtained with a coarse model of a [10]

Proceedings of the 26th Euromicro

Conference
9. CONCLUSIONS AND FURTHER WORK

PVM: Parallel Virtual

. Cambridge, Massachussets, MIT

Press.
[12]

Proceedings of the 15th ICA (ICA 95)

DAFX-6