Characterizing and Enhancing the Performance of Sound Synthesis Applications on Pentium III

Computationally intensive applications such as sound synthesis, animation, scientific computation, can take advantage of table-lookup interpolation to reduce computation and execution time. When designing table-lookup interpolation, application writers face various choices such as table size and interpolation mechanism (linear, quadratic, etc). The better the interpolation, the more intensive the computation, and as a result, often computationally least expensive mechanism is chosen, compromising quality for performance. In our project, we characterize the performance a particular kind of streaming application: sound synthesis oscillator. We determine the existing performance bottlenecks in this application, and investigate the effectiveness of using Streaming SIMD Extension (SSE) instruction set in this application. We demonstrate that the SSE-enhanced code can deliver performance improvement when ratio of computation to memory operations is favorable. In particular, in case of quadratic interpolation, we show a 27% performance increase as compared to the best alternative version that does not use SSE. Further, we demonstrate that the additional penalty of using quadratic interpolation is reduced to 11% as compared to linear interpolation, when using SSE. This compares favorably to 50% penalty in case when SSE is not used. This provides a much more favorable trade-off in the application design space. Introduction Table-lookup interpolation is an old but still important technique in many computation-intensive programs like sound synthesis [Roads], key-frame animation [Burtnyk+], and scientific computations [Ueberhuber]. The technique is especially effective in replacing the computation of expressions where there is limited range of input values and such expressions are costly to evaluate directly. The most common usage of table-lookup interpolation is to restore the curve of the continuous signal from limited amount of sampled data. However, there are a lot of factors can affect the performance of this straightforward method. As normally it is impossible to restore the exact signal, table-lookup interpolation introduces the error which results in noise. Error can be reduced by increasing the table size and/or by increasing the quality of the interpolation. But both of these also affect the signal computation cost. In addition, general purpose processors like Pentium series are not particularly good at either random access to main memory or interpolation, but these operations are at the core of tablelookup interpolation.