Real-time Auralisation of the Lower Frequency Sound Field using Numerical Methods on the GPU

Master defence by Nikolas Borrel-Jensen

Abstract:

We present a software system for simulating the lower frequency sound field in real-time using the FDTD method. The FDTD method is a numerical method for solving the wave equation by approximating the time and space derivatives by finite-differences. With the goal of achieving real-time performance, the FDTD method was implemented using CUDA on the GPU.

A family of 3-D non-staggered compact explicit FDTD schemes have been implemented incorporating a frequency-dependent boundary model that is consistent with locally reacting surfaces, taking the full 3-D wave field into account. The boundaries are modelled using digital impedance filters (DIFs), realised as IIR filters. Three GPU versions have been implemented using CUDA, with differences in the number of kernels used.

The physical correctness of the FDTD method has been compared with the FEM method for cubic rooms using seven FDTD schemes with frequency-dependent and -independent boundaries. In general, the results showed good correspondence between the sound fields simulated with the FEM and FDTD methods, when less than 2% of the dispersion errors are allowed, of which the Standard Leapfrog method introduces fewest errors (less than 0.3 dB) and the Interpolated Digital Waveguide Mesh method introducing the most errors (less than 0.7 dB). The precision is also studied when oblique boundaries are present, and when discretisation errors due to meshing the scenes are introduced.

A performance test for three selected schemes were performed in different environments, and it was shown that real-time simulations below the Schröder frequency is possible.

Time: Thursday, April 12, 2012 17.15-18.30.

Place: Njalsgade 126, 24.5.62 (building 24, 5th floor, room 62)

Supervisors:

Frank Wefers (RWTH Aachen)

Jakob Grue Simonsen (University of Copenhagen)