Optimizations and refactoring of the APM 3-band split filter
This CL refactors and optimizes the 3-band split-filter in APM, which
is a very computationally complex component.
Beyond optimizing the code, the filter coefficients are also quantized
to avoid denormals.
The changes reduces the complexity of the split filter by about 30-50%.
The CL has been tested for bitexactness on a number of aecdump
recordings.
(the CL also removes the now unused code for the sparse_fir_filter)
Bug: webrtc:6181
Change-Id: If45f8d1f189c6812ccb03721156c77eb68181211
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/168189
Reviewed-by: Sam Zackrisson <saza@webrtc.org>
Reviewed-by: Karl Wiberg <kwiberg@webrtc.org>
Commit-Queue: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#30592}
diff --git a/common_audio/BUILD.gn b/common_audio/BUILD.gn
index 48bd906..6c89bf2 100644
--- a/common_audio/BUILD.gn
+++ b/common_audio/BUILD.gn
@@ -32,8 +32,6 @@
"resampler/sinc_resampler.cc",
"smoothing_filter.cc",
"smoothing_filter.h",
- "sparse_fir_filter.cc",
- "sparse_fir_filter.h",
"vad/include/vad.h",
"vad/vad.cc",
"wav_file.cc",
@@ -47,6 +45,7 @@
deps = [
":common_audio_c",
":sinc_resampler",
+ "../api:array_view",
"../rtc_base:checks",
"../rtc_base:gtest_prod",
"../rtc_base:rtc_base_approved",
@@ -331,7 +330,6 @@
"signal_processing/real_fft_unittest.cc",
"signal_processing/signal_processing_unittest.cc",
"smoothing_filter_unittest.cc",
- "sparse_fir_filter_unittest.cc",
"vad/vad_core_unittest.cc",
"vad/vad_filterbank_unittest.cc",
"vad/vad_gmm_unittest.cc",
diff --git a/common_audio/channel_buffer.h b/common_audio/channel_buffer.h
index dc44369..f027080 100644
--- a/common_audio/channel_buffer.h
+++ b/common_audio/channel_buffer.h
@@ -14,7 +14,9 @@
#include <string.h>
#include <memory>
+#include <vector>
+#include "api/array_view.h"
#include "common_audio/include/audio_util.h"
#include "rtc_base/checks.h"
#include "rtc_base/gtest_prod_util.h"
@@ -48,40 +50,60 @@
num_frames_per_band_(num_frames / num_bands),
num_allocated_channels_(num_channels),
num_channels_(num_channels),
- num_bands_(num_bands) {
- for (size_t i = 0; i < num_allocated_channels_; ++i) {
- for (size_t j = 0; j < num_bands_; ++j) {
- channels_[j * num_allocated_channels_ + i] =
- &data_[i * num_frames_ + j * num_frames_per_band_];
- bands_[i * num_bands_ + j] = channels_[j * num_allocated_channels_ + i];
+ num_bands_(num_bands),
+ bands_view_(num_allocated_channels_,
+ std::vector<rtc::ArrayView<T>>(num_bands_)),
+ channels_view_(
+ num_bands_,
+ std::vector<rtc::ArrayView<T>>(num_allocated_channels_)) {
+ // Temporarily cast away const_ness to allow populating the array views.
+ auto* bands_view =
+ const_cast<std::vector<std::vector<rtc::ArrayView<T>>>*>(&bands_view_);
+ auto* channels_view =
+ const_cast<std::vector<std::vector<rtc::ArrayView<T>>>*>(
+ &channels_view_);
+
+ for (size_t ch = 0; ch < num_allocated_channels_; ++ch) {
+ for (size_t band = 0; band < num_bands_; ++band) {
+ (*channels_view)[band][ch] = rtc::ArrayView<T>(
+ &data_[ch * num_frames_ + band * num_frames_per_band_],
+ num_frames_per_band_);
+ (*bands_view)[ch][band] = channels_view_[band][ch];
+ channels_[band * num_allocated_channels_ + ch] =
+ channels_view_[band][ch].data();
+ bands_[ch * num_bands_ + band] =
+ channels_[band * num_allocated_channels_ + ch];
}
}
}
- // Returns a pointer array to the full-band channels (or lower band channels).
- // Usage:
- // channels()[channel][sample].
- // Where:
- // 0 <= channel < |num_allocated_channels_|
- // 0 <= sample < |num_frames_|
- T* const* channels() { return channels(0); }
- const T* const* channels() const { return channels(0); }
-
- // Returns a pointer array to the channels for a specific band.
- // Usage:
- // channels(band)[channel][sample].
+ // Returns a pointer array to the channels.
+ // If band is explicitly specificed, the channels for a specific band are
+ // returned and the usage becomes: channels(band)[channel][sample].
// Where:
// 0 <= band < |num_bands_|
// 0 <= channel < |num_allocated_channels_|
// 0 <= sample < |num_frames_per_band_|
- const T* const* channels(size_t band) const {
+
+ // If band is not explicitly specified, the full-band channels (or lower band
+ // channels) are returned and the usage becomes: channels()[channel][sample].
+ // Where:
+ // 0 <= channel < |num_allocated_channels_|
+ // 0 <= sample < |num_frames_|
+ const T* const* channels(size_t band = 0) const {
RTC_DCHECK_LT(band, num_bands_);
return &channels_[band * num_allocated_channels_];
}
- T* const* channels(size_t band) {
+ T* const* channels(size_t band = 0) {
const ChannelBuffer<T>* t = this;
return const_cast<T* const*>(t->channels(band));
}
+ rtc::ArrayView<const rtc::ArrayView<T>> channels_view(size_t band = 0) {
+ return channels_view_[band];
+ }
+ rtc::ArrayView<const rtc::ArrayView<T>> channels_view(size_t band = 0) const {
+ return channels_view_[band];
+ }
// Returns a pointer array to the bands for a specific channel.
// Usage:
@@ -100,6 +122,13 @@
return const_cast<T* const*>(t->bands(channel));
}
+ rtc::ArrayView<const rtc::ArrayView<T>> bands_view(size_t channel) {
+ return bands_view_[channel];
+ }
+ rtc::ArrayView<const rtc::ArrayView<T>> bands_view(size_t channel) const {
+ return bands_view_[channel];
+ }
+
// Sets the |slice| pointers to the |start_frame| position for each channel.
// Returns |slice| for convenience.
const T* const* Slice(T** slice, size_t start_frame) const {
@@ -140,6 +169,8 @@
// Number of channels the user sees.
size_t num_channels_;
const size_t num_bands_;
+ const std::vector<std::vector<rtc::ArrayView<T>>> bands_view_;
+ const std::vector<std::vector<rtc::ArrayView<T>>> channels_view_;
};
// One int16_t and one float ChannelBuffer that are kept in sync. The sync is
diff --git a/common_audio/sparse_fir_filter.cc b/common_audio/sparse_fir_filter.cc
deleted file mode 100644
index 772eb82..0000000
--- a/common_audio/sparse_fir_filter.cc
+++ /dev/null
@@ -1,60 +0,0 @@
-/*
- * Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "common_audio/sparse_fir_filter.h"
-
-#include "rtc_base/checks.h"
-
-namespace webrtc {
-
-SparseFIRFilter::SparseFIRFilter(const float* nonzero_coeffs,
- size_t num_nonzero_coeffs,
- size_t sparsity,
- size_t offset)
- : sparsity_(sparsity),
- offset_(offset),
- nonzero_coeffs_(nonzero_coeffs, nonzero_coeffs + num_nonzero_coeffs),
- state_(sparsity_ * (num_nonzero_coeffs - 1) + offset_, 0.f) {
- RTC_CHECK_GE(num_nonzero_coeffs, 1);
- RTC_CHECK_GE(sparsity, 1);
-}
-
-SparseFIRFilter::~SparseFIRFilter() = default;
-
-void SparseFIRFilter::Filter(const float* in, size_t length, float* out) {
- // Convolves the input signal |in| with the filter kernel |nonzero_coeffs_|
- // taking into account the previous state.
- for (size_t i = 0; i < length; ++i) {
- out[i] = 0.f;
- size_t j;
- for (j = 0; i >= j * sparsity_ + offset_ && j < nonzero_coeffs_.size();
- ++j) {
- out[i] += in[i - j * sparsity_ - offset_] * nonzero_coeffs_[j];
- }
- for (; j < nonzero_coeffs_.size(); ++j) {
- out[i] += state_[i + (nonzero_coeffs_.size() - j - 1) * sparsity_] *
- nonzero_coeffs_[j];
- }
- }
-
- // Update current state.
- if (!state_.empty()) {
- if (length >= state_.size()) {
- std::memcpy(&state_[0], &in[length - state_.size()],
- state_.size() * sizeof(*in));
- } else {
- std::memmove(&state_[0], &state_[length],
- (state_.size() - length) * sizeof(state_[0]));
- std::memcpy(&state_[state_.size() - length], in, length * sizeof(*in));
- }
- }
-}
-
-} // namespace webrtc
diff --git a/common_audio/sparse_fir_filter.h b/common_audio/sparse_fir_filter.h
deleted file mode 100644
index 5197a8e..0000000
--- a/common_audio/sparse_fir_filter.h
+++ /dev/null
@@ -1,53 +0,0 @@
-/*
- * Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#ifndef COMMON_AUDIO_SPARSE_FIR_FILTER_H_
-#define COMMON_AUDIO_SPARSE_FIR_FILTER_H_
-
-#include <cstring>
-#include <vector>
-
-#include "rtc_base/constructor_magic.h"
-
-namespace webrtc {
-
-// A Finite Impulse Response filter implementation which takes advantage of a
-// sparse structure with uniformly distributed non-zero coefficients.
-class SparseFIRFilter final {
- public:
- // |num_nonzero_coeffs| is the number of non-zero coefficients,
- // |nonzero_coeffs|. They are assumed to be uniformly distributed every
- // |sparsity| samples and with an initial |offset|. The rest of the filter
- // coefficients will be assumed zeros. For example, with sparsity = 3, and
- // offset = 1 the filter coefficients will be:
- // B = [0 coeffs[0] 0 0 coeffs[1] 0 0 coeffs[2] ... ]
- // All initial state values will be zeros.
- SparseFIRFilter(const float* nonzero_coeffs,
- size_t num_nonzero_coeffs,
- size_t sparsity,
- size_t offset);
- ~SparseFIRFilter();
-
- // Filters the |in| data supplied.
- // |out| must be previously allocated and it must be at least of |length|.
- void Filter(const float* in, size_t length, float* out);
-
- private:
- const size_t sparsity_;
- const size_t offset_;
- const std::vector<float> nonzero_coeffs_;
- std::vector<float> state_;
-
- RTC_DISALLOW_COPY_AND_ASSIGN(SparseFIRFilter);
-};
-
-} // namespace webrtc
-
-#endif // COMMON_AUDIO_SPARSE_FIR_FILTER_H_
diff --git a/common_audio/sparse_fir_filter_unittest.cc b/common_audio/sparse_fir_filter_unittest.cc
deleted file mode 100644
index 5dc7b6d..0000000
--- a/common_audio/sparse_fir_filter_unittest.cc
+++ /dev/null
@@ -1,219 +0,0 @@
-/*
- * Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "common_audio/sparse_fir_filter.h"
-
-#include <memory>
-
-#include "common_audio/fir_filter.h"
-#include "common_audio/fir_filter_factory.h"
-#include "rtc_base/arraysize.h"
-#include "test/gtest.h"
-
-namespace webrtc {
-namespace {
-
-static const float kCoeffs[] = {0.2f, 0.3f, 0.5f, 0.7f, 0.11f};
-static const float kInput[] = {1.f, 2.f, 3.f, 4.f, 5.f,
- 6.f, 7.f, 8.f, 9.f, 10.f};
-
-template <size_t N>
-void VerifyOutput(const float (&expected_output)[N], const float (&output)[N]) {
- EXPECT_EQ(0, memcmp(expected_output, output, sizeof(output)));
-}
-
-} // namespace
-
-TEST(SparseFIRFilterTest, FilterAsIdentity) {
- const float kCoeff = 1.f;
- const size_t kNumCoeff = 1;
- const size_t kSparsity = 3;
- const size_t kOffset = 0;
- float output[arraysize(kInput)];
- SparseFIRFilter filter(&kCoeff, kNumCoeff, kSparsity, kOffset);
- filter.Filter(kInput, arraysize(kInput), output);
- VerifyOutput(kInput, output);
-}
-
-TEST(SparseFIRFilterTest, SameOutputForScalarCoefficientAndDifferentSparsity) {
- const float kCoeff = 2.f;
- const size_t kNumCoeff = 1;
- const size_t kLowSparsity = 1;
- const size_t kHighSparsity = 7;
- const size_t kOffset = 0;
- float low_sparsity_output[arraysize(kInput)];
- float high_sparsity_output[arraysize(kInput)];
- SparseFIRFilter low_sparsity_filter(&kCoeff, kNumCoeff, kLowSparsity,
- kOffset);
- SparseFIRFilter high_sparsity_filter(&kCoeff, kNumCoeff, kHighSparsity,
- kOffset);
- low_sparsity_filter.Filter(kInput, arraysize(kInput), low_sparsity_output);
- high_sparsity_filter.Filter(kInput, arraysize(kInput), high_sparsity_output);
- VerifyOutput(low_sparsity_output, high_sparsity_output);
-}
-
-TEST(SparseFIRFilterTest, FilterUsedAsScalarMultiplication) {
- const float kCoeff = 5.f;
- const size_t kNumCoeff = 1;
- const size_t kSparsity = 5;
- const size_t kOffset = 0;
- float output[arraysize(kInput)];
- SparseFIRFilter filter(&kCoeff, kNumCoeff, kSparsity, kOffset);
- filter.Filter(kInput, arraysize(kInput), output);
- EXPECT_FLOAT_EQ(5.f, output[0]);
- EXPECT_FLOAT_EQ(20.f, output[3]);
- EXPECT_FLOAT_EQ(25.f, output[4]);
- EXPECT_FLOAT_EQ(50.f, output[arraysize(kInput) - 1]);
-}
-
-TEST(SparseFIRFilterTest, FilterUsedAsInputShifting) {
- const float kCoeff = 1.f;
- const size_t kNumCoeff = 1;
- const size_t kSparsity = 1;
- const size_t kOffset = 4;
- float output[arraysize(kInput)];
- SparseFIRFilter filter(&kCoeff, kNumCoeff, kSparsity, kOffset);
- filter.Filter(kInput, arraysize(kInput), output);
- EXPECT_FLOAT_EQ(0.f, output[0]);
- EXPECT_FLOAT_EQ(0.f, output[3]);
- EXPECT_FLOAT_EQ(1.f, output[4]);
- EXPECT_FLOAT_EQ(2.f, output[5]);
- EXPECT_FLOAT_EQ(6.f, output[arraysize(kInput) - 1]);
-}
-
-TEST(SparseFIRFilterTest, FilterUsedAsArbitraryWeighting) {
- const size_t kSparsity = 2;
- const size_t kOffset = 1;
- float output[arraysize(kInput)];
- SparseFIRFilter filter(kCoeffs, arraysize(kCoeffs), kSparsity, kOffset);
- filter.Filter(kInput, arraysize(kInput), output);
- EXPECT_FLOAT_EQ(0.f, output[0]);
- EXPECT_FLOAT_EQ(0.9f, output[3]);
- EXPECT_FLOAT_EQ(1.4f, output[4]);
- EXPECT_FLOAT_EQ(2.4f, output[5]);
- EXPECT_FLOAT_EQ(8.61f, output[arraysize(kInput) - 1]);
-}
-
-TEST(SparseFIRFilterTest, FilterInLengthLesserOrEqualToCoefficientsLength) {
- const size_t kSparsity = 1;
- const size_t kOffset = 0;
- float output[arraysize(kInput)];
- SparseFIRFilter filter(kCoeffs, arraysize(kCoeffs), kSparsity, kOffset);
- filter.Filter(kInput, 2, output);
- EXPECT_FLOAT_EQ(0.2f, output[0]);
- EXPECT_FLOAT_EQ(0.7f, output[1]);
-}
-
-TEST(SparseFIRFilterTest, MultipleFilterCalls) {
- const size_t kSparsity = 1;
- const size_t kOffset = 0;
- float output[arraysize(kInput)];
- SparseFIRFilter filter(kCoeffs, arraysize(kCoeffs), kSparsity, kOffset);
- filter.Filter(kInput, 2, output);
- EXPECT_FLOAT_EQ(0.2f, output[0]);
- EXPECT_FLOAT_EQ(0.7f, output[1]);
- filter.Filter(kInput, 2, output);
- EXPECT_FLOAT_EQ(1.3f, output[0]);
- EXPECT_FLOAT_EQ(2.4f, output[1]);
- filter.Filter(kInput, 2, output);
- EXPECT_FLOAT_EQ(2.81f, output[0]);
- EXPECT_FLOAT_EQ(2.62f, output[1]);
- filter.Filter(kInput, 2, output);
- EXPECT_FLOAT_EQ(2.81f, output[0]);
- EXPECT_FLOAT_EQ(2.62f, output[1]);
- filter.Filter(&kInput[3], 3, output);
- EXPECT_FLOAT_EQ(3.41f, output[0]);
- EXPECT_FLOAT_EQ(4.12f, output[1]);
- EXPECT_FLOAT_EQ(6.21f, output[2]);
- filter.Filter(&kInput[3], 3, output);
- EXPECT_FLOAT_EQ(8.12f, output[0]);
- EXPECT_FLOAT_EQ(9.14f, output[1]);
- EXPECT_FLOAT_EQ(9.45f, output[2]);
-}
-
-TEST(SparseFIRFilterTest, VerifySampleBasedVsBlockBasedFiltering) {
- const size_t kSparsity = 3;
- const size_t kOffset = 1;
- float output_block_based[arraysize(kInput)];
- SparseFIRFilter filter_block(kCoeffs, arraysize(kCoeffs), kSparsity, kOffset);
- filter_block.Filter(kInput, arraysize(kInput), output_block_based);
- float output_sample_based[arraysize(kInput)];
- SparseFIRFilter filter_sample(kCoeffs, arraysize(kCoeffs), kSparsity,
- kOffset);
- for (size_t i = 0; i < arraysize(kInput); ++i)
- filter_sample.Filter(&kInput[i], 1, &output_sample_based[i]);
- VerifyOutput(output_block_based, output_sample_based);
-}
-
-TEST(SparseFIRFilterTest, SimpleHighPassFilter) {
- const size_t kSparsity = 2;
- const size_t kOffset = 2;
- const float kHPCoeffs[] = {1.f, -1.f};
- const float kConstantInput[] = {1.f, 1.f, 1.f, 1.f, 1.f,
- 1.f, 1.f, 1.f, 1.f, 1.f};
- float output[arraysize(kConstantInput)];
- SparseFIRFilter filter(kHPCoeffs, arraysize(kHPCoeffs), kSparsity, kOffset);
- filter.Filter(kConstantInput, arraysize(kConstantInput), output);
- EXPECT_FLOAT_EQ(0.f, output[0]);
- EXPECT_FLOAT_EQ(0.f, output[1]);
- EXPECT_FLOAT_EQ(1.f, output[2]);
- EXPECT_FLOAT_EQ(1.f, output[3]);
- for (size_t i = kSparsity + kOffset; i < arraysize(kConstantInput); ++i)
- EXPECT_FLOAT_EQ(0.f, output[i]);
-}
-
-TEST(SparseFIRFilterTest, SimpleLowPassFilter) {
- const size_t kSparsity = 2;
- const size_t kOffset = 2;
- const float kLPCoeffs[] = {1.f, 1.f};
- const float kHighFrequencyInput[] = {1.f, 1.f, -1.f, -1.f, 1.f,
- 1.f, -1.f, -1.f, 1.f, 1.f};
- float output[arraysize(kHighFrequencyInput)];
- SparseFIRFilter filter(kLPCoeffs, arraysize(kLPCoeffs), kSparsity, kOffset);
- filter.Filter(kHighFrequencyInput, arraysize(kHighFrequencyInput), output);
- EXPECT_FLOAT_EQ(0.f, output[0]);
- EXPECT_FLOAT_EQ(0.f, output[1]);
- EXPECT_FLOAT_EQ(1.f, output[2]);
- EXPECT_FLOAT_EQ(1.f, output[3]);
- for (size_t i = kSparsity + kOffset; i < arraysize(kHighFrequencyInput); ++i)
- EXPECT_FLOAT_EQ(0.f, output[i]);
-}
-
-TEST(SparseFIRFilterTest, SameOutputWhenSwappedCoefficientsAndInput) {
- const size_t kSparsity = 1;
- const size_t kOffset = 0;
- float output[arraysize(kCoeffs)];
- float output_swapped[arraysize(kCoeffs)];
- SparseFIRFilter filter(kCoeffs, arraysize(kCoeffs), kSparsity, kOffset);
- // Use arraysize(kCoeffs) for in_length to get same-length outputs.
- filter.Filter(kInput, arraysize(kCoeffs), output);
- SparseFIRFilter filter_swapped(kInput, arraysize(kCoeffs), kSparsity,
- kOffset);
- filter_swapped.Filter(kCoeffs, arraysize(kCoeffs), output_swapped);
- VerifyOutput(output, output_swapped);
-}
-
-TEST(SparseFIRFilterTest, SameOutputAsFIRFilterWhenSparsityOneAndOffsetZero) {
- const size_t kSparsity = 1;
- const size_t kOffset = 0;
- float output[arraysize(kInput)];
- float sparse_output[arraysize(kInput)];
- std::unique_ptr<FIRFilter> filter(
- CreateFirFilter(kCoeffs, arraysize(kCoeffs), arraysize(kInput)));
- SparseFIRFilter sparse_filter(kCoeffs, arraysize(kCoeffs), kSparsity,
- kOffset);
- filter->Filter(kInput, arraysize(kInput), output);
- sparse_filter.Filter(kInput, arraysize(kInput), sparse_output);
- for (size_t i = 0; i < arraysize(kInput); ++i) {
- EXPECT_FLOAT_EQ(output[i], sparse_output[i]);
- }
-}
-
-} // namespace webrtc