Audio Signal Processing Using Graphics Processing Units

摘要

Current graphics processing units (GPUs) are massively parallel computing environments offering remarkable performance boosts in parallelizable tasks. Audio signal processing is a potential application area. Three different cases for GPU implementation are studied: additive synthesis, fast Fourier transforms (FFT), and time-domain convolution. For additive synthesis nearly two million sine waves were computable in real time on the GPU, giving a factor of 250—3000 performance gain over CPU implementation. Similarly, the GPU was able to perform an FFT eight times longer than the CPU version in the same time. For a stereo signal the GPU was able to compute a two-million-point FFT and inverse FFT in real time with an input buffer size of 1024 samples at a sampling rate of 48 kHz with 50% overlap. Finally the GPU could compute approximately 130 times longer FIR filters than the CPU in the same time. For stereo input and output, requiring four filters altogether, the GPU processing was able to implement FIR filters of length 376 000 taps in real time. The latency in all these tasks is tolerable, since the performance is nearly optimal with a hundred-sample buffer, which corresponds to a few milliseconds. In summary the results show that GPUs are highly useful for computationally intensive audio signal processing tasks.