Add flag to APM to force multichannel even with AEC3
Currently, APM fakes multichannel in two ways:
- With injected AECs, capture processing is only performed on the left
channel. The result is copied into the other channels.
- With multichannel render audio, all channels are mixed into one
before analysing.
This CL adds a flag to disable these behaviors, ensuring proper
multichannel processing happens throughout the APM pipeline.
Adds killswitches to separately disable render / capture multichannel.
Additionally - AEC3 currently crashes when running with multichannel.
This CL adds the missing pieces to at least have it run without
triggering any DCHECKS, including making the high pass filter properly
handle multichannel.
Bug: webrtc:10913, webrtc:10907
Change-Id: I38795bf8f312b959fcc816a056fba2c68d4e424d
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/152483
Commit-Queue: Sam Zackrisson <saza@webrtc.org>
Reviewed-by: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#29248}
diff --git a/modules/audio_processing/audio_processing_impl.cc b/modules/audio_processing/audio_processing_impl.cc
index cdc37c6..b1187fa 100644
--- a/modules/audio_processing/audio_processing_impl.cc
+++ b/modules/audio_processing/audio_processing_impl.cc
@@ -48,6 +48,7 @@
#include "rtc_base/ref_counted_object.h"
#include "rtc_base/time_utils.h"
#include "rtc_base/trace_event.h"
+#include "system_wrappers/include/field_trial.h"
#include "system_wrappers/include/metrics.h"
#define RETURN_ON_ERR(expr) \
@@ -348,8 +349,12 @@
}
AudioProcessingImpl::AudioProcessingImpl(const webrtc::Config& config)
- : AudioProcessingImpl(config, nullptr, nullptr, nullptr, nullptr, nullptr) {
-}
+ : AudioProcessingImpl(config,
+ /*capture_post_processor=*/nullptr,
+ /*render_pre_processor=*/nullptr,
+ /*echo_control_factory=*/nullptr,
+ /*echo_detector=*/nullptr,
+ /*capture_analyzer=*/nullptr) {}
int AudioProcessingImpl::instance_count_ = 0;
@@ -382,13 +387,17 @@
/* enabled= */ false,
/* enabled_agc2_level_estimator= */ false,
/* digital_adaptive_disabled= */ false,
- /* analyze_before_aec= */ false),
+ /* analyze_before_aec= */ false,
#else
config.Get<ExperimentalAgc>().enabled,
config.Get<ExperimentalAgc>().enabled_agc2_level_estimator,
config.Get<ExperimentalAgc>().digital_adaptive_disabled,
- config.Get<ExperimentalAgc>().analyze_before_aec),
+ config.Get<ExperimentalAgc>().analyze_before_aec,
#endif
+ !field_trial::IsEnabled(
+ "WebRTC-ApmExperimentalMultiChannelRenderKillSwitch"),
+ !field_trial::IsEnabled(
+ "WebRTC-ApmExperimentalMultiChannelCaptureKillSwitch")),
#if defined(WEBRTC_ANDROID) || defined(WEBRTC_IOS)
capture_(false),
#else
@@ -630,10 +639,18 @@
RTC_DCHECK_NE(8000, render_processing_rate);
- // Always downmix the render stream to mono for analysis. This has been
- // demonstrated to work well for AEC in most practical scenarios.
if (submodule_states_.RenderMultiBandSubModulesActive()) {
- formats_.render_processing_format = StreamConfig(render_processing_rate, 1);
+ // By default, downmix the render stream to mono for analysis. This has been
+ // demonstrated to work well for AEC in most practical scenarios.
+ const bool experimental_multi_channel_render =
+ config_.pipeline.experimental_multi_channel &&
+ constants_.experimental_multi_channel_render_support;
+ int render_processing_num_channels =
+ experimental_multi_channel_render
+ ? formats_.api_format.reverse_input_stream().num_channels()
+ : 1;
+ formats_.render_processing_format =
+ StreamConfig(render_processing_rate, render_processing_num_channels);
} else {
formats_.render_processing_format = StreamConfig(
formats_.api_format.reverse_input_stream().sample_rate_hz(),
@@ -658,6 +675,10 @@
rtc::CritScope cs_render(&crit_render_);
rtc::CritScope cs_capture(&crit_capture_);
+ const bool pipeline_config_changed =
+ config_.pipeline.experimental_multi_channel !=
+ config.pipeline.experimental_multi_channel;
+
const bool aec_config_changed =
config_.echo_canceller.enabled != config.echo_canceller.enabled ||
config_.echo_canceller.use_legacy_aec !=
@@ -733,6 +754,12 @@
private_submodules_->voice_detector->Initialize(
proc_split_sample_rate_hz());
}
+
+ // Reinitialization must happen after all submodule configuration to avoid
+ // additional reinitializations on the next capture / render processing call.
+ if (pipeline_config_changed) {
+ InitializeLocked(formats_.api_format);
+ }
}
void AudioProcessingImpl::ApplyAgc1Config(
@@ -809,7 +836,14 @@
size_t AudioProcessingImpl::num_proc_channels() const {
// Used as callback from submodules, hence locking is not allowed.
- return capture_nonlocked_.echo_controller_enabled ? 1 : num_output_channels();
+ const bool experimental_multi_channel_capture =
+ config_.pipeline.experimental_multi_channel &&
+ constants_.experimental_multi_channel_capture_support;
+ if (capture_nonlocked_.echo_controller_enabled &&
+ !experimental_multi_channel_capture) {
+ return 1;
+ }
+ return num_output_channels();
}
size_t AudioProcessingImpl::num_output_channels() const {
@@ -1338,7 +1372,11 @@
capture_buffer->SplitIntoFrequencyBands();
}
- if (private_submodules_->echo_controller) {
+ const bool experimental_multi_channel_capture =
+ config_.pipeline.experimental_multi_channel &&
+ constants_.experimental_multi_channel_capture_support;
+ if (private_submodules_->echo_controller &&
+ !experimental_multi_channel_capture) {
// Force down-mixing of the number of channels after the detection of
// capture signal saturation.
// TODO(peah): Look into ensuring that this kind of tampering with the
@@ -1846,8 +1884,8 @@
echo_control_factory_->Create(proc_sample_rate_hz());
} else {
private_submodules_->echo_controller = std::make_unique<EchoCanceller3>(
- EchoCanceller3Config(), proc_sample_rate_hz(),
- /*num_render_channels=*/1, /*num_capture_channels=*/1);
+ EchoCanceller3Config(), proc_sample_rate_hz(), num_reverse_channels(),
+ num_proc_channels());
}
capture_nonlocked_.echo_controller_enabled = true;