RNN VAD: pitch search optimizations (part 3)
`ComputeSlidingFrameSquareEnergies()` which computes the energy of a
sliding 20 ms frame in the pitch buffer has been switched from backward
to forward.
The benchmark has shown a slight improvement (about +6x).
This change is not bit exact but all the tolerance tests still pass
except for one single case in `RnnVadTest,PitchSearchWithinTolerance`
for which the tolerance has been slightly increased. Note that the pitch
estimation is still bit-exact.
Benchmarked as follows:
```
out/release/modules_unittests \
--gtest_filter=*RnnVadTest.DISABLED_RnnVadPerformance* \
--gtest_also_run_disabled_tests --logs
```
Results:
| baseline | this CL
------+----------------------+------------------------
run 1 | 22.8319 +/- 1.46554 | 22.087 +/- 0.552932
| 389.367x | 402.499x
------+----------------------+------------------------
run 2 | 22.4286 +/- 0.726449 | 22.216 +/- 0.916222
| 396.369x | 400.162x
------+----------------------+------------------------
run 2 | 22.5688 +/- 0.831341 | 22.4902 +/- 1.04881
| 393.906x | 395.283x
Bug: webrtc:10480
Change-Id: I1fd54077a32e25e46196c8e18f003cd0ffd503e1
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/191703
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Karl Wiberg <kwiberg@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32572}
diff --git a/modules/audio_processing/agc2/rnn_vad/pitch_search_internal.cc b/modules/audio_processing/agc2/rnn_vad/pitch_search_internal.cc
index d62cddf..09cb2fb 100644
--- a/modules/audio_processing/agc2/rnn_vad/pitch_search_internal.cc
+++ b/modules/audio_processing/agc2/rnn_vad/pitch_search_internal.cc
@@ -169,7 +169,7 @@
// Auto-correlation energy normalized by frame energy.
const float numerator =
auto_correlation[inverted_lag] * auto_correlation[inverted_lag];
- const float denominator = y_energy[kMaxPitch24kHz - inverted_lag];
+ const float denominator = y_energy[inverted_lag];
// Compare numerator/denominator ratios without using divisions.
if (numerator * best_denominator > best_numerator * denominator) {
best_inverted_lag = inverted_lag;
@@ -253,19 +253,19 @@
void ComputeSlidingFrameSquareEnergies24kHz(
rtc::ArrayView<const float, kBufSize24kHz> pitch_buffer,
- rtc::ArrayView<float, kRefineNumLags24kHz> yy_values) {
- float yy = ComputeAutoCorrelation(kMaxPitch24kHz, pitch_buffer);
- yy_values[0] = yy;
- static_assert(kMaxPitch24kHz - (kRefineNumLags24kHz - 1) >= 0, "");
+ rtc::ArrayView<float, kRefineNumLags24kHz> y_energy) {
+ float yy = std::inner_product(pitch_buffer.begin(),
+ pitch_buffer.begin() + kFrameSize20ms24kHz,
+ pitch_buffer.begin(), 0.f);
+ y_energy[0] = yy;
static_assert(kMaxPitch24kHz - 1 + kFrameSize20ms24kHz < kBufSize24kHz, "");
- for (int lag = 1; lag < kRefineNumLags24kHz; ++lag) {
- const int inverted_lag = kMaxPitch24kHz - lag;
- const float y_old = pitch_buffer[inverted_lag + kFrameSize20ms24kHz];
- const float y_new = pitch_buffer[inverted_lag];
- yy -= y_old * y_old;
- yy += y_new * y_new;
- yy = std::max(0.f, yy);
- yy_values[lag] = yy;
+ static_assert(kMaxPitch24kHz < kRefineNumLags24kHz, "");
+ for (int inverted_lag = 0; inverted_lag < kMaxPitch24kHz; ++inverted_lag) {
+ yy -= pitch_buffer[inverted_lag] * pitch_buffer[inverted_lag];
+ yy += pitch_buffer[inverted_lag + kFrameSize20ms24kHz] *
+ pitch_buffer[inverted_lag + kFrameSize20ms24kHz];
+ yy = std::max(1.f, yy);
+ y_energy[inverted_lag + 1] = yy;
}
}
@@ -375,7 +375,7 @@
float y_energy; // Energy of the sliding frame `y`.
};
- const float x_energy = y_energy[0];
+ const float x_energy = y_energy[kMaxPitch24kHz];
const auto pitch_strength = [x_energy](float xy, float y_energy) {
RTC_DCHECK_GE(x_energy * y_energy, 0.f);
return xy / std::sqrt(1.f + x_energy * y_energy);
@@ -387,7 +387,7 @@
std::min(initial_pitch_period_48kHz / 2, kMaxPitch24kHz - 1);
best_pitch.xy =
ComputeAutoCorrelation(kMaxPitch24kHz - best_pitch.period, pitch_buffer);
- best_pitch.y_energy = y_energy[best_pitch.period];
+ best_pitch.y_energy = y_energy[kMaxPitch24kHz - best_pitch.period];
best_pitch.strength = pitch_strength(best_pitch.xy, best_pitch.y_energy);
// Keep a copy of the initial pitch candidate.
const PitchInfo initial_pitch{best_pitch.period, best_pitch.strength};
@@ -428,8 +428,9 @@
const float xy_secondary_period = ComputeAutoCorrelation(
kMaxPitch24kHz - dual_alternative_period, pitch_buffer);
const float xy = 0.5f * (xy_primary_period + xy_secondary_period);
- const float yy = 0.5f * (y_energy[alternative_pitch.period] +
- y_energy[dual_alternative_period]);
+ const float yy =
+ 0.5f * (y_energy[kMaxPitch24kHz - alternative_pitch.period] +
+ y_energy[kMaxPitch24kHz - dual_alternative_period]);
alternative_pitch.strength = pitch_strength(xy, yy);
// Maybe update best period.
diff --git a/modules/audio_processing/agc2/rnn_vad/pitch_search_internal.h b/modules/audio_processing/agc2/rnn_vad/pitch_search_internal.h
index 693ab9e..0af55f8 100644
--- a/modules/audio_processing/agc2/rnn_vad/pitch_search_internal.h
+++ b/modules/audio_processing/agc2/rnn_vad/pitch_search_internal.h
@@ -62,10 +62,10 @@
// corresponding pitch period.
// Computes the sum of squared samples for every sliding frame `y` in the pitch
-// buffer. The indexes of `yy_values` are lags.
+// buffer. The indexes of `y_energy` are inverted lags.
void ComputeSlidingFrameSquareEnergies24kHz(
rtc::ArrayView<const float, kBufSize24kHz> pitch_buffer,
- rtc::ArrayView<float, kRefineNumLags24kHz> yy_values);
+ rtc::ArrayView<float, kRefineNumLags24kHz> y_energy);
// Top-2 pitch period candidates. Unit: number of samples - i.e., inverted lags.
struct CandidatePitchPeriods {
diff --git a/modules/audio_processing/agc2/rnn_vad/pitch_search_internal_unittest.cc b/modules/audio_processing/agc2/rnn_vad/pitch_search_internal_unittest.cc
index e5826d0..94316a4 100644
--- a/modules/audio_processing/agc2/rnn_vad/pitch_search_internal_unittest.cc
+++ b/modules/audio_processing/agc2/rnn_vad/pitch_search_internal_unittest.cc
@@ -42,7 +42,7 @@
computed_output);
auto square_energies_view = test_data.GetPitchBufSquareEnergiesView();
ExpectNearAbsolute({square_energies_view.data(), square_energies_view.size()},
- computed_output, 3e-2f);
+ computed_output, 1e-3f);
}
// Checks that the estimated pitch period is bit-exact given test input data.
diff --git a/modules/audio_processing/agc2/rnn_vad/pitch_search_unittest.cc b/modules/audio_processing/agc2/rnn_vad/pitch_search_unittest.cc
index c57c8c2..98b791e 100644
--- a/modules/audio_processing/agc2/rnn_vad/pitch_search_unittest.cc
+++ b/modules/audio_processing/agc2/rnn_vad/pitch_search_unittest.cc
@@ -42,7 +42,7 @@
pitch_estimator.Estimate({lp_residual.data(), kBufSize24kHz});
EXPECT_EQ(expected_pitch_period, pitch_period);
EXPECT_NEAR(expected_pitch_strength,
- pitch_estimator.GetLastPitchStrengthForTesting(), 1e-5f);
+ pitch_estimator.GetLastPitchStrengthForTesting(), 15e-6f);
}
}
}
diff --git a/modules/audio_processing/agc2/rnn_vad/test_utils.cc b/modules/audio_processing/agc2/rnn_vad/test_utils.cc
index 74571af..24bbf13 100644
--- a/modules/audio_processing/agc2/rnn_vad/test_utils.cc
+++ b/modules/audio_processing/agc2/rnn_vad/test_utils.cc
@@ -10,6 +10,7 @@
#include "modules/audio_processing/agc2/rnn_vad/test_utils.h"
+#include <algorithm>
#include <memory>
#include "rtc_base/checks.h"
@@ -86,6 +87,10 @@
ResourcePath("audio_processing/agc2/rnn_vad/pitch_search_int", "dat"),
1396);
test_data_reader.ReadChunk(test_data_);
+ // Reverse the order of the squared energy values.
+ // Required after the WebRTC CL 191703 which switched to forward computation.
+ std::reverse(test_data_.begin() + kBufSize24kHz,
+ test_data_.begin() + kBufSize24kHz + kNumPitchBufSquareEnergies);
}
PitchTestData::~PitchTestData() = default;
