[pulseaudio-discuss] resubmit reverted resampler patch

Tanu Kaskinen tanu.kaskinen at linux.intel.com
Fri Nov 8 00:44:05 PST 2013

On Sun, 2013-07-21 at 23:14 +0200, Peter Meerwald wrote:
> Hello, 
> the patch series aims to resubmit a patch to reorder the resampling 
> pipeline; it former patch was dropped due to issues with leftover data 
> handling; patch 11 has more details
> the resampling pipeline is 
> conv -> remap -> resample -> conv
> patch 11 introduces the following, alternative pipeline
> conv -> resample -> remap -> conv
> when the number of input channels is less than the number of output channels
> the resample stage may not consume all data passed; hence the leftover
> data has to be stored; the data is presented again to the resample stage
> together with the next input data
> the strategy for the fix is to allow reinserting leftover data not only
> in the remap stage, but also in the conv(to work format) stage; i.e. the
> stage run before the resampler in case of the alternative pipeline
> patches 1 and 2 introduce helper variables work_channels and w_fz
> representing the number of channels and the byte per frame, respectively,
> in the resample stage
> patch 3 makes the impl_resample() function return the number of leftover
> frames (to make leftover handling more visible in the code)
> patch 4 move leftover handling from impl_resample() to resample()
> patch 5 and 6 is cleanup and introduces a helper function fit_buf() for
> buffer allocation
> patch 7 introduces a pointer to the output buffer of the stage run before
> the resamplers
> patch 8 is cleanup
> patch 9 extends fit_buf() so that it can copy leftover data when allocating
> a new buffer
> patch 10 adds leftover data handling to the convert_to_work_format()
> function
> patch 11 finally enables the two different pipelines; the patch is important
> e.g. when a mono channel is remapped and resampled to quad-channel -- it
> saves resampling 75% of the data!
> I also considered moving the leftover handling into impl_resample(), but
> this leads to more code duplication and buffer copying; new API to find the amount
> of leftover data store in the resamplers might be needed as well
> thanks to the cleanups preceding patch 11, the complexity of the more flexible 
> pipeline keeps managable

I have rebased and reviewed patches 1-5, and all looked good, so pushed.
I'll continue the review later (hopefully today).


More information about the pulseaudio-discuss mailing list