Digital signal processing (DSP) applications commonly transform input data before performing an FFT, or transform output data afterwards. For example, if the input data is supplied as low-resolution samples from an 8-bit analog-to-digital (A/D) converter, the samples may first have to be expanded into 32-bit floating point numbers before the FFT and the rest of the processing pipeline can start.
The cuFFT library included with CUDA 6.5 introduces device callbacks to improve performance of this sort of transforms. Callback routines are user-supplied device functions that cuFFT calls when loading or storing data. You can use callbacks to implement many pre- or post-processing operations that required launching separate CUDA kernels before CUDA 6.5.
Example DSP Pipeline
In this blog post we will implement the first stages of a typical DSP pipeline as depicted in Figure 1. We will first discuss a solution without callbacks using multiple custom kernels which we then use as a stepping stone towards a solution based on cuFFT device callbacks. The source code for both versions is available on github.
Batches of 8-bit fixed-point samples are input to the DSP pipline from an A/D converter. Each sample consists of 1024 data points. For more efficient processing, we group samples into batches of 1000 samples each. Therefore, you can think of this input as a 1000×1024 matrix of 8-bit fixed-point values. Continue reading