audioconvert: passthrough conversion causing high CPU load

Submitted by Petr Kulhavy

Link to original bug (#772872)

Description

While measuring the audioconvert plugin performance I found that converting S24LE to S24LE format (i.e. no change of the format at all) causes superfluous quantization. This results in extra CPU load.

The reason for this peculiar behaviour is that the first and mandatory step in an audio conversion is to unpack the data to either 16 or 32 bits. So 24-bit data is internally expanded to 32-bits. Then in the following if statement the chain_quantize() function logic decides that a quantization is needed:

/* we still want to run the quantization step when reducing bits to get

• the rounding correct */
  if (out_int && out_depth < 32
  && convert->current_format == GST_AUDIO_FORMAT_S32)

Obviously, if 24-bits are expanded to 32-bits and then again compacted to 24-bits, no quantization is needed.

Not sure how exactly the if statement should be altered, but shouldn't it also take the in_depth into account (like: ... && in_depth != out_depth)?

My test pipeline:

gst-launch-1.0 audiotestsrc samplesperbuffer=48 ! "audio/x-raw,format=(string)S24LE,rate=(int)48000,channels=(int)2" ! audioconvert ! "audio/x-raw,format=(string)S24LE,rate=(int)48000,channels=(int)2" ! alsasink can-activate-pull=true provide-clock=false buffer-time=8000 slave-method=none blocksize=192

Version: 1.8.3