Pull out the PostFilter to its own NonlinearBeamformer API

webrtc/modules/audio_processing/splitting_filter.cc

Index: webrtc/modules/audio_processing/splitting_filter.cc
diff --git a/webrtc/modules/audio_processing/splitting_filter.cc b/webrtc/modules/audio_processing/splitting_filter.cc
index e87a7f86218b9fd50c8fe109f55b128f8af7d04b..b6f642c09d4db448b320748538d485bd731d6449 100644
--- a/webrtc/modules/audio_processing/splitting_filter.cc
+++ b/webrtc/modules/audio_processing/splitting_filter.cc
@@ -72,8 +72,8 @@ void SplittingFilter::TwoBandsAnalysis(const IFChannelBuffer* data,
 
 void SplittingFilter::TwoBandsSynthesis(const IFChannelBuffer* bands,
                                         IFChannelBuffer* data) {
-  RTC_DCHECK_EQ(two_bands_states_.size(), data->num_channels());
-  for (size_t i = 0; i < two_bands_states_.size(); ++i) {
+  RTC_DCHECK_LE(data->num_channels(), two_bands_states_.size());

[peah-webrtc, 2016/06/01 14:51:01]

Could you please explain why this change is needed?

[aluebs-webrtc, 2016/06/01 22:13:20]

When there is a change in number of channels between splitting and merging, the
size and channels in the ChannelBuffer are not the same, but it only needs to be
lower or equal than the former.

[peah-webrtc, 2016/06/08 12:04:55]

Do you mean that before this CL, the size in the channels was changed by the
beamformer, but that the number of channels needed to be changed inside APM by
setting it to 1?

And with the changes in this CL, which no longer applies downmixing inside the
beamformer, the downmixing instead needs to be done as part of the synthesis?

[aluebs-webrtc, 2016/06/09 02:11:46]

No. This was a bug. Before this CL the number of channels of the AudioBuffer was
changed, but this was not reflected in the number of channels of the
ChannelBuffers, so the synthesis filter was merging the original number of
channels, which were later discarded in CopyTo (or InterleaveTo). So the bug had
no effects except adding complexity. Now that the number of channels of
AudioBuffer is also reflected in the ChannelBuffer, it triggered this overly
strict CHECK. So I fixed it.

[peah-webrtc, 2016/06/09 09:36:30]

Great find!! It would, however, be good to have that bugfix as a separate CL as
it
1) Is not at all related to the beamformer.
2) Will also affect all calls where the beamformer is not active.
3) Will change the complexity (reduce).

Would it be possible to extract that part into another CL, and land it before
this CL is landed so that the changes in the first CL can be rebased into the
basis for this CL?

[aluebs-webrtc, 2016/06/09 19:21:41]

1) It is related to the Beamformer, since...
2) It is only triggered when there is a num_channels change, which only happens
when Beamformer is enabled.

But I agree it is better to break it out in a different CL. Done here:
https://codereview.webrtc.org/2053773002/

+  for (size_t i = 0; i < data->num_channels(); ++i) {

[peah-webrtc, 2016/06/01 14:51:01]

Could you please explain why this change is needed?

[aluebs-webrtc, 2016/06/01 22:13:20]

Please see above.

     WebRtcSpl_SynthesisQMF(bands->ibuf_const()->channels(0)[i],
                            bands->ibuf_const()->channels(1)[i],
                            bands->num_frames_per_band(),
@@ -95,8 +95,8 @@ void SplittingFilter::ThreeBandsAnalysis(const IFChannelBuffer* data,
 
 void SplittingFilter::ThreeBandsSynthesis(const IFChannelBuffer* bands,
                                           IFChannelBuffer* data) {
-  RTC_DCHECK_EQ(three_band_filter_banks_.size(), data->num_channels());
-  for (size_t i = 0; i < three_band_filter_banks_.size(); ++i) {
+  RTC_DCHECK_LE(data->num_channels(), three_band_filter_banks_.size());

[peah-webrtc, 2016/06/01 14:51:01]

Could you please explain why this change is needed?

[aluebs-webrtc, 2016/06/01 22:13:20]

Please see above.

+  for (size_t i = 0; i < data->num_channels(); ++i) {
     three_band_filter_banks_[i]->Synthesis(bands->fbuf_const()->bands(i),
                                            bands->num_frames_per_band(),
                                            data->fbuf()->channels()[i]);