Finally it is the time for looking at audio resampling more seriously. I'm starting for a simplest case - a decimator. Decimation is dividing the sampling frequency by an integer number M. It is done with a lowpass FIR filter with cutoff frequency at pi/M, calculating only every M-th output sample.

I've researched FIR-s optimization already, but this one is a special case. The data reusability (which is critical for SIMD performance) is lower, because of sample skipping. A filter coefficients table is shifted over the input vector not one sample at a time, but M samples.

I've implemented naive reference code today, and did some benchmarking. The reference code is slow, Pegasos 2 achieves 175 to 195 Mtaps/s depending on filter length and decimation factor M. Efika does 55 to 65 Mtaps/s. It needs optimization badly :-).