Created attachment 211963 [details]
Python script to demo the bug

The audiofirfilter (and any element which derives from AudioFXBaseFIRFilter) will drop buffers when mutlichannel input is piped in. I believe the cause of this bug is a faulty count of the output samples in the TIME_DOMAIN_CONVOLUTION_BODY macro of audiofxbasefirfilter.c. Specifically, the number of input samples is returned with:

  return input_samples / channels; \

...but this is double counting since input_samples has already been divided by the number of channels previously. The overall effect is that if the kernel is short enough to invoke the time domain convolution algorithm, then the first buffer received will processed correctly and the next (nchannels-1) buffers will be dropped because of the bookkeeping error.

Removing the division by the channels should fix the problem. The attached python script can be used to hear the bug by processing a dual channel audio file (the script takes the audio file as an argument) with the audiofirfilter which should otherwise be an identity operation. The choppiness is the bug in action.