[attachment=14139]
Wave Field Synthesis for 3D Audio: ArchitecturalProspectives
Motivation (I)
Wave Field Synthesis (WFS) 3D sound technology is utilized in many audio applications (e.g. virtual reality, cinemas)
Previous audio systems based on WFS are implemented on mainstream PCs
Processing bottlenecks
Excessive power consumption
Exploration of other platforms
Motivation (II)
There is huge amount of parallelism in WFS
Candidate platforms with potential for parallel processing
Graphics Processor Units (GPUs)
Field Programmable Gate Arrays (FPGAs)
Outline
The WFS algorithm
GPU implementation
FPGA implementation
Experimental results
Conclusions
Future work
The WFS algorithm (I)
Christiaan Huygens’ principle
The WFS algorithm (II)
High number of loudspeakers
Sweet spot is the entire listening area
Sound sources can be rendered in front and behind the speaker array
The WFS algorithm (III)
Rayleigh 2.5D operator
Dz constant
z inner product between and
θ angle between and
The WFS algorithm (IV)
The WFS algorithm (V)
The WFS algorithm (VI)
GPU implementation (I)
Generalized GPU architecture
GPU implementation (II)
Thread hierarchy
Grid of thread blocks
A multiprocessor maps each thread onto a scalar processor core
Single Instruction Multiple Thread (SIMT) unit
Manages groups of 32 threads, called warps
GPU implementation (II)
WFS thread hierarchy
GPU implementation (III)
WFS kernel implemented in GPU
GPU implementation (IV)
Constant memory
Loudspeakers coordinates
Sound source coordinates
Shared memory
Intermediate data during computation
FPGA implementation (I)
Design
infrastructure
FPGA implementation (II)
Fabric Coprocessor Module for WFS
FPGA implementation (III)
Preprocessor
Amplitude decay
Unit Distance
FPGA implementation (IV)
WFS Engine
Delayed sample
Source distance
Experimental results (I)
GPU setup
NVidia GeForce 8600 GT, 256 MB DDR3
CUDA 2.1 toolkit
FPGA setup
Xilinx EDK 9.1 and ISE 9.1
Xilinx ML410 board with Virtex4 FX60 FPGA
GPP setup
Pentium D @ 3.4GHz
Windows XP
Experimental results (II)
FPGA resource utilization
Experimental results (III)
GPU time breakdown
Experimental results (IV)
Kernel speedup
Experimental results (V)
Experimental results (V)
# of real-time sources
Conclusions (I)
WFS can be considerably accelerated when mapped onto platforms with parallel processing capabilities
Contemporary GPUs and FPGAs could be used to WFS based 3D-Audio systems
New GPUs should boost performance even more
Conclusions (II)
Future work (I)
Profile other 3D audio algorithms
Beamforming
High Order Ambisonics
Binaurals
Explore architectural prospectives
Experiments with GPUs and FPGAs