admin/webrtc_m130
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add parameterization for three multi channel AEC3 unit tests

Browse Source 
Bug: webrtc:11295
Change-Id: I478aa02908c494cf9609db00021438a59a132b66
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/167202
Commit-Queue: Sam Zackrisson <saza@webrtc.org>
Reviewed-by: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#30370}
This commit is contained in:
Sam Zackrisson 2020-01-24 12:55:17 +01:00 committed by Commit Bot
parent 159c414ff8
commit b18c4eb0a9
10 changed files with 849 additions and 795 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

612
modules/audio_processing/aec3/adaptive_fir_filter_unittest.cc
View File

@ -51,15 +51,137 @@ std::string ProduceDebugText(size_t num_render_channels, size_t delay) {
} // namespace
class AdaptiveFirFilterOneTwoFourEightRenderChannels
: public ::testing::Test,
public ::testing::WithParamInterface<size_t> {};
INSTANTIATE_TEST_SUITE_P(MultiChannel,
AdaptiveFirFilterOneTwoFourEightRenderChannels,
::testing::Values(1, 2, 4, 8));
#if defined(WEBRTC_HAS_NEON)
// Verifies that the optimized methods for filter adaptation are similar to
// their reference counterparts.
TEST(AdaptiveFirFilter, FilterAdaptationNeonOptimizations) {
TEST_P(AdaptiveFirFilterOneTwoFourEightRenderChannels,
FilterAdaptationNeonOptimizations) {
const size_t num_render_channels = GetParam();
for (size_t num_partitions : {2, 5, 12, 30, 50}) {
for (size_t num_render_channels : {1, 2, 4, 8}) {
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(EchoCanceller3Config(), kSampleRateHz,
num_render_channels));
Random random_generator(42U);
std::vector<std::vector<std::vector<float>>> x(
kNumBands,
std::vector<std::vector<float>>(num_render_channels,
std::vector<float>(kBlockSize, 0.f)));
FftData S_C;
FftData S_Neon;
FftData G;
Aec3Fft fft;
std::vector<std::vector<FftData>> H_C(
num_partitions, std::vector<FftData>(num_render_channels));
std::vector<std::vector<FftData>> H_Neon(
num_partitions, std::vector<FftData>(num_render_channels));
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
H_C[p][ch].Clear();
H_Neon[p][ch].Clear();
}
}
for (size_t k = 0; k < 30; ++k) {
for (size_t band = 0; band < x.size(); ++band) {
for (size_t ch = 0; ch < x[band].size(); ++ch) {
RandomizeSampleVector(&random_generator, x[band][ch]);
}
}
render_delay_buffer->Insert(x);
if (k == 0) {
render_delay_buffer->Reset();
}
render_delay_buffer->PrepareCaptureProcessing();
}
auto* const render_buffer = render_delay_buffer->GetRenderBuffer();
for (size_t j = 0; j < G.re.size(); ++j) {
G.re[j] = j / 10001.f;
}
for (size_t j = 1; j < G.im.size() - 1; ++j) {
G.im[j] = j / 20001.f;
}
G.im[0] = 0.f;
G.im[G.im.size() - 1] = 0.f;
AdaptPartitions_Neon(*render_buffer, G, num_partitions, &H_Neon);
AdaptPartitions(*render_buffer, G, num_partitions, &H_C);
AdaptPartitions_Neon(*render_buffer, G, num_partitions, &H_Neon);
AdaptPartitions(*render_buffer, G, num_partitions, &H_C);
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t j = 0; j < H_C[p][ch].re.size(); ++j) {
EXPECT_FLOAT_EQ(H_C[p][ch].re[j], H_Neon[p][ch].re[j]);
EXPECT_FLOAT_EQ(H_C[p][ch].im[j], H_Neon[p][ch].im[j]);
}
}
}
ApplyFilter_Neon(*render_buffer, num_partitions, H_Neon, &S_Neon);
ApplyFilter(*render_buffer, num_partitions, H_C, &S_C);
for (size_t j = 0; j < S_C.re.size(); ++j) {
EXPECT_NEAR(S_C.re[j], S_Neon.re[j], fabs(S_C.re[j] * 0.00001f));
EXPECT_NEAR(S_C.im[j], S_Neon.im[j], fabs(S_C.re[j] * 0.00001f));
}
}
}
// Verifies that the optimized method for frequency response computation is
// bitexact to the reference counterpart.
TEST_P(AdaptiveFirFilterOneTwoFourEightRenderChannels,
ComputeFrequencyResponseNeonOptimization) {
const size_t num_render_channels = GetParam();
for (size_t num_partitions : {2, 5, 12, 30, 50}) {
std::vector<std::vector<FftData>> H(
num_partitions, std::vector<FftData>(num_render_channels));
std::vector<std::array<float, kFftLengthBy2Plus1>> H2(num_partitions);
std::vector<std::array<float, kFftLengthBy2Plus1>> H2_Neon(num_partitions);
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t k = 0; k < H[p][ch].re.size(); ++k) {
H[p][ch].re[k] = k + p / 3.f + ch;
H[p][ch].im[k] = p + k / 7.f - ch;
}
}
}
ComputeFrequencyResponse(num_partitions, H, &H2);
ComputeFrequencyResponse_Neon(num_partitions, H, &H2_Neon);
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t k = 0; k < H2[p].size(); ++k) {
EXPECT_FLOAT_EQ(H2[p][k], H2_Neon[p][k]);
}
}
}
}
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
// Verifies that the optimized methods for filter adaptation are bitexact to
// their reference counterparts.
TEST_P(AdaptiveFirFilterOneTwoFourEightRenderChannels,
FilterAdaptationSse2Optimizations) {
const size_t num_render_channels = GetParam();
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
bool use_sse2 = (WebRtc_GetCPUInfo(kSSE2) != 0);
if (use_sse2) {
for (size_t num_partitions : {2, 5, 12, 30, 50}) {
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(EchoCanceller3Config(), kSampleRateHz,
num_render_channels));
@ -69,21 +191,21 @@ TEST(AdaptiveFirFilter, FilterAdaptationNeonOptimizations) {
std::vector<std::vector<float>>(num_render_channels,
std::vector<float>(kBlockSize, 0.f)));
FftData S_C;
FftData S_Neon;
FftData S_Sse2;
FftData G;
Aec3Fft fft;
std::vector<std::vector<FftData>> H_C(
num_partitions, std::vector<FftData>(num_render_channels));
std::vector<std::vector<FftData>> H_Neon(
std::vector<std::vector<FftData>> H_Sse2(
num_partitions, std::vector<FftData>(num_render_channels));
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
H_C[p][ch].Clear();
H_Neon[p][ch].Clear();
H_Sse2[p][ch].Clear();
}
}
for (size_t k = 0; k < 30; ++k) {
for (size_t k = 0; k < 500; ++k) {
for (size_t band = 0; band < x.size(); ++band) {
for (size_t ch = 0; ch < x[band].size(); ++ch) {
RandomizeSampleVector(&random_generator, x[band][ch]);
@ -94,51 +216,48 @@ TEST(AdaptiveFirFilter, FilterAdaptationNeonOptimizations) {
render_delay_buffer->Reset();
}
render_delay_buffer->PrepareCaptureProcessing();
}
auto* const render_buffer = render_delay_buffer->GetRenderBuffer();
auto* const render_buffer = render_delay_buffer->GetRenderBuffer();
for (size_t j = 0; j < G.re.size(); ++j) {
G.re[j] = j / 10001.f;
}
for (size_t j = 1; j < G.im.size() - 1; ++j) {
G.im[j] = j / 20001.f;
}
G.im[0] = 0.f;
G.im[G.im.size() - 1] = 0.f;
ApplyFilter_Sse2(*render_buffer, num_partitions, H_Sse2, &S_Sse2);
ApplyFilter(*render_buffer, num_partitions, H_C, &S_C);
for (size_t j = 0; j < S_C.re.size(); ++j) {
EXPECT_FLOAT_EQ(S_C.re[j], S_Sse2.re[j]);
EXPECT_FLOAT_EQ(S_C.im[j], S_Sse2.im[j]);
}
AdaptPartitions_Neon(*render_buffer, G, num_partitions, &H_Neon);
AdaptPartitions(*render_buffer, G, num_partitions, &H_C);
AdaptPartitions_Neon(*render_buffer, G, num_partitions, &H_Neon);
AdaptPartitions(*render_buffer, G, num_partitions, &H_C);
std::for_each(G.re.begin(), G.re.end(),
[&](float& a) { a = random_generator.Rand<float>(); });
std::for_each(G.im.begin(), G.im.end(),
[&](float& a) { a = random_generator.Rand<float>(); });
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t j = 0; j < H_C[p][ch].re.size(); ++j) {
EXPECT_FLOAT_EQ(H_C[p][ch].re[j], H_Neon[p][ch].re[j]);
EXPECT_FLOAT_EQ(H_C[p][ch].im[j], H_Neon[p][ch].im[j]);
AdaptPartitions_Sse2(*render_buffer, G, num_partitions, &H_Sse2);
AdaptPartitions(*render_buffer, G, num_partitions, &H_C);
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t j = 0; j < H_C[p][ch].re.size(); ++j) {
EXPECT_FLOAT_EQ(H_C[p][ch].re[j], H_Sse2[p][ch].re[j]);
EXPECT_FLOAT_EQ(H_C[p][ch].im[j], H_Sse2[p][ch].im[j]);
}
}
}
}
ApplyFilter_Neon(*render_buffer, num_partitions, H_Neon, &S_Neon);
ApplyFilter(*render_buffer, num_partitions, H_C, &S_C);
for (size_t j = 0; j < S_C.re.size(); ++j) {
EXPECT_NEAR(S_C.re[j], S_Neon.re[j], fabs(S_C.re[j] * 0.00001f));
EXPECT_NEAR(S_C.im[j], S_Neon.im[j], fabs(S_C.re[j] * 0.00001f));
}
}
}
}
// Verifies that the optimized method for frequency response computation is
// bitexact to the reference counterpart.
TEST(AdaptiveFirFilter, ComputeFrequencyResponseNeonOptimization) {
for (size_t num_partitions : {2, 5, 12, 30, 50}) {
for (size_t num_render_channels : {1, 2, 4, 8}) {
TEST_P(AdaptiveFirFilterOneTwoFourEightRenderChannels,
ComputeFrequencyResponseSse2Optimization) {
const size_t num_render_channels = GetParam();
bool use_sse2 = (WebRtc_GetCPUInfo(kSSE2) != 0);
if (use_sse2) {
for (size_t num_partitions : {2, 5, 12, 30, 50}) {
std::vector<std::vector<FftData>> H(
num_partitions, std::vector<FftData>(num_render_channels));
std::vector<std::array<float, kFftLengthBy2Plus1>> H2(num_partitions);
std::vector<std::array<float, kFftLengthBy2Plus1>> H2_Neon(
std::vector<std::array<float, kFftLengthBy2Plus1>> H2_Sse2(
num_partitions);
for (size_t p = 0; p < num_partitions; ++p) {
@ -151,123 +270,11 @@ TEST(AdaptiveFirFilter, ComputeFrequencyResponseNeonOptimization) {
}
ComputeFrequencyResponse(num_partitions, H, &H2);
ComputeFrequencyResponse_Neon(num_partitions, H, &H2_Neon);
ComputeFrequencyResponse_Sse2(num_partitions, H, &H2_Sse2);
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t k = 0; k < H2[p].size(); ++k) {
EXPECT_FLOAT_EQ(H2[p][k], H2_Neon[p][k]);
}
}
}
}
}
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
// Verifies that the optimized methods for filter adaptation are bitexact to
// their reference counterparts.
TEST(AdaptiveFirFilter, FilterAdaptationSse2Optimizations) {
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
bool use_sse2 = (WebRtc_GetCPUInfo(kSSE2) != 0);
if (use_sse2) {
for (size_t num_partitions : {2, 5, 12, 30, 50}) {
for (size_t num_render_channels : {1, 2, 4, 8}) {
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(EchoCanceller3Config(), kSampleRateHz,
num_render_channels));
Random random_generator(42U);
std::vector<std::vector<std::vector<float>>> x(
kNumBands,
std::vector<std::vector<float>>(
num_render_channels, std::vector<float>(kBlockSize, 0.f)));
FftData S_C;
FftData S_Sse2;
FftData G;
Aec3Fft fft;
std::vector<std::vector<FftData>> H_C(
num_partitions, std::vector<FftData>(num_render_channels));
std::vector<std::vector<FftData>> H_Sse2(
num_partitions, std::vector<FftData>(num_render_channels));
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
H_C[p][ch].Clear();
H_Sse2[p][ch].Clear();
}
}
for (size_t k = 0; k < 500; ++k) {
for (size_t band = 0; band < x.size(); ++band) {
for (size_t ch = 0; ch < x[band].size(); ++ch) {
RandomizeSampleVector(&random_generator, x[band][ch]);
}
}
render_delay_buffer->Insert(x);
if (k == 0) {
render_delay_buffer->Reset();
}
render_delay_buffer->PrepareCaptureProcessing();
auto* const render_buffer = render_delay_buffer->GetRenderBuffer();
ApplyFilter_Sse2(*render_buffer, num_partitions, H_Sse2, &S_Sse2);
ApplyFilter(*render_buffer, num_partitions, H_C, &S_C);
for (size_t j = 0; j < S_C.re.size(); ++j) {
EXPECT_FLOAT_EQ(S_C.re[j], S_Sse2.re[j]);
EXPECT_FLOAT_EQ(S_C.im[j], S_Sse2.im[j]);
}
std::for_each(G.re.begin(), G.re.end(),
[&](float& a) { a = random_generator.Rand<float>(); });
std::for_each(G.im.begin(), G.im.end(),
[&](float& a) { a = random_generator.Rand<float>(); });
AdaptPartitions_Sse2(*render_buffer, G, num_partitions, &H_Sse2);
AdaptPartitions(*render_buffer, G, num_partitions, &H_C);
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t j = 0; j < H_C[p][ch].re.size(); ++j) {
EXPECT_FLOAT_EQ(H_C[p][ch].re[j], H_Sse2[p][ch].re[j]);
EXPECT_FLOAT_EQ(H_C[p][ch].im[j], H_Sse2[p][ch].im[j]);
}
}
}
}
}
}
}
}
// Verifies that the optimized method for frequency response computation is
// bitexact to the reference counterpart.
TEST(AdaptiveFirFilter, ComputeFrequencyResponseSse2Optimization) {
bool use_sse2 = (WebRtc_GetCPUInfo(kSSE2) != 0);
if (use_sse2) {
for (size_t num_partitions : {2, 5, 12, 30, 50}) {
for (size_t num_render_channels : {1, 2, 4, 8}) {
std::vector<std::vector<FftData>> H(
num_partitions, std::vector<FftData>(num_render_channels));
std::vector<std::array<float, kFftLengthBy2Plus1>> H2(num_partitions);
std::vector<std::array<float, kFftLengthBy2Plus1>> H2_Sse2(
num_partitions);
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t ch = 0; ch < num_render_channels; ++ch) {
for (size_t k = 0; k < H[p][ch].re.size(); ++k) {
H[p][ch].re[k] = k + p / 3.f + ch;
H[p][ch].im[k] = p + k / 7.f - ch;
}
}
}
ComputeFrequencyResponse(num_partitions, H, &H2);
ComputeFrequencyResponse_Sse2(num_partitions, H, &H2_Sse2);
for (size_t p = 0; p < num_partitions; ++p) {
for (size_t k = 0; k < H2[p].size(); ++k) {
EXPECT_FLOAT_EQ(H2[p][k], H2_Sse2[p][k]);
}
EXPECT_FLOAT_EQ(H2[p][k], H2_Sse2[p][k]);
}
}
}
@ -278,13 +285,13 @@ TEST(AdaptiveFirFilter, ComputeFrequencyResponseSse2Optimization) {
#if RTC_DCHECK_IS_ON && GTEST_HAS_DEATH_TEST && !defined(WEBRTC_ANDROID)
// Verifies that the check for non-null data dumper works.
TEST(AdaptiveFirFilter, NullDataDumper) {
TEST(AdaptiveFirFilterTest, NullDataDumper) {
EXPECT_DEATH(AdaptiveFirFilter(9, 9, 250, 1, DetectOptimization(), nullptr),
"");
}
// Verifies that the check for non-null filter output works.
TEST(AdaptiveFirFilter, NullFilterOutput) {
TEST(AdaptiveFirFilterTest, NullFilterOutput) {
ApmDataDumper data_dumper(42);
AdaptiveFirFilter filter(9, 9, 250, 1, DetectOptimization(), &data_dumper);
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
@ -297,7 +304,7 @@ TEST(AdaptiveFirFilter, NullFilterOutput) {
// Verifies that the filter statistics can be accessed when filter statistics
// are turned on.
TEST(AdaptiveFirFilter, FilterStatisticsAccess) {
TEST(AdaptiveFirFilterTest, FilterStatisticsAccess) {
ApmDataDumper data_dumper(42);
Aec3Optimization optimization = DetectOptimization();
AdaptiveFirFilter filter(9, 9, 250, 1, optimization, &data_dumper);
@ -314,7 +321,7 @@ TEST(AdaptiveFirFilter, FilterStatisticsAccess) {
}
// Verifies that the filter size if correctly repported.
TEST(AdaptiveFirFilter, FilterSize) {
TEST(AdaptiveFirFilterTest, FilterSize) {
ApmDataDumper data_dumper(42);
for (size_t filter_size = 1; filter_size < 5; ++filter_size) {
AdaptiveFirFilter filter(filter_size, filter_size, 250, 1,
@ -323,163 +330,166 @@ TEST(AdaptiveFirFilter, FilterSize) {
}
}
class AdaptiveFirFilterMultiChannel
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
INSTANTIATE_TEST_SUITE_P(MultiChannel,
AdaptiveFirFilterMultiChannel,
::testing::Combine(::testing::Values(1, 4),
::testing::Values(1, 8)));
// Verifies that the filter is being able to properly filter a signal and to
// adapt its coefficients.
TEST(AdaptiveFirFilter, FilterAndAdapt) {
TEST_P(AdaptiveFirFilterMultiChannel, FilterAndAdapt) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
constexpr size_t kNumBlocksToProcessPerRenderChannel = 1000;
for (size_t num_capture_channels : {1, 4}) {
for (size_t num_render_channels : {1, 8}) {
ApmDataDumper data_dumper(42);
EchoCanceller3Config config;
ApmDataDumper data_dumper(42);
EchoCanceller3Config config;
if (num_render_channels == 33) {
config.filter.main = {13, 0.00005f, 0.0005f, 0.0001f, 2.f, 20075344.f};
config.filter.shadow = {13, 0.1f, 20075344.f};
config.filter.main_initial = {12, 0.005f, 0.5f,
0.001f, 2.f, 20075344.f};
config.filter.shadow_initial = {12, 0.7f, 20075344.f};
}
if (num_render_channels == 33) {
config.filter.main = {13, 0.00005f, 0.0005f, 0.0001f, 2.f, 20075344.f};
config.filter.shadow = {13, 0.1f, 20075344.f};
config.filter.main_initial = {12, 0.005f, 0.5f, 0.001f, 2.f, 20075344.f};
config.filter.shadow_initial = {12, 0.7f, 20075344.f};
}
AdaptiveFirFilter filter(
config.filter.main.length_blocks, config.filter.main.length_blocks,
config.filter.config_change_duration_blocks, num_render_channels,
DetectOptimization(), &data_dumper);
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>> H2(
num_capture_channels,
std::vector<std::array<float, kFftLengthBy2Plus1>>(
filter.max_filter_size_partitions(),
std::array<float, kFftLengthBy2Plus1>()));
std::vector<std::vector<float>> h(
num_capture_channels,
std::vector<float>(
GetTimeDomainLength(filter.max_filter_size_partitions()), 0.f));
Aec3Fft fft;
config.delay.default_delay = 1;
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz,
num_render_channels));
ShadowFilterUpdateGain gain(config.filter.shadow,
config.filter.config_change_duration_blocks);
Random random_generator(42U);
std::vector<std::vector<std::vector<float>>> x(
kNumBands,
std::vector<std::vector<float>>(num_render_channels,
std::vector<float>(kBlockSize, 0.f)));
std::vector<float> n(kBlockSize, 0.f);
std::vector<float> y(kBlockSize, 0.f);
AecState aec_state(EchoCanceller3Config{}, num_capture_channels);
RenderSignalAnalyzer render_signal_analyzer(config);
absl::optional<DelayEstimate> delay_estimate;
std::vector<float> e(kBlockSize, 0.f);
std::array<float, kFftLength> s_scratch;
std::vector<SubtractorOutput> output(num_capture_channels);
FftData S;
FftData G;
FftData E;
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(
num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> E2_main(
num_capture_channels);
std::array<float, kFftLengthBy2Plus1> E2_shadow;
// [B,A] = butter(2,100/8000,'high')
constexpr CascadedBiQuadFilter::BiQuadCoefficients
kHighPassFilterCoefficients = {{0.97261f, -1.94523f, 0.97261f},
{-1.94448f, 0.94598f}};
for (auto& Y2_ch : Y2) {
Y2_ch.fill(0.f);
}
for (auto& E2_main_ch : E2_main) {
E2_main_ch.fill(0.f);
}
E2_shadow.fill(0.f);
for (auto& subtractor_output : output) {
subtractor_output.Reset();
}
AdaptiveFirFilter filter(
config.filter.main.length_blocks, config.filter.main.length_blocks,
config.filter.config_change_duration_blocks, num_render_channels,
DetectOptimization(), &data_dumper);
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>> H2(
num_capture_channels, std::vector<std::array<float, kFftLengthBy2Plus1>>(
filter.max_filter_size_partitions(),
std::array<float, kFftLengthBy2Plus1>()));
std::vector<std::vector<float>> h(
num_capture_channels,
std::vector<float>(
GetTimeDomainLength(filter.max_filter_size_partitions()), 0.f));
Aec3Fft fft;
config.delay.default_delay = 1;
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz, num_render_channels));
ShadowFilterUpdateGain gain(config.filter.shadow,
config.filter.config_change_duration_blocks);
Random random_generator(42U);
std::vector<std::vector<std::vector<float>>> x(
kNumBands, std::vector<std::vector<float>>(
num_render_channels, std::vector<float>(kBlockSize, 0.f)));
std::vector<float> n(kBlockSize, 0.f);
std::vector<float> y(kBlockSize, 0.f);
AecState aec_state(EchoCanceller3Config{}, num_capture_channels);
RenderSignalAnalyzer render_signal_analyzer(config);
absl::optional<DelayEstimate> delay_estimate;
std::vector<float> e(kBlockSize, 0.f);
std::array<float, kFftLength> s_scratch;
std::vector<SubtractorOutput> output(num_capture_channels);
FftData S;
FftData G;
FftData E;
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> E2_main(
num_capture_channels);
std::array<float, kFftLengthBy2Plus1> E2_shadow;
// [B,A] = butter(2,100/8000,'high')
constexpr CascadedBiQuadFilter::BiQuadCoefficients
kHighPassFilterCoefficients = {{0.97261f, -1.94523f, 0.97261f},
{-1.94448f, 0.94598f}};
for (auto& Y2_ch : Y2) {
Y2_ch.fill(0.f);
}
for (auto& E2_main_ch : E2_main) {
E2_main_ch.fill(0.f);
}
E2_shadow.fill(0.f);
for (auto& subtractor_output : output) {
subtractor_output.Reset();
}
constexpr float kScale = 1.0f / kFftLengthBy2;
constexpr float kScale = 1.0f / kFftLengthBy2;
for (size_t delay_samples : {0, 64, 150, 200, 301}) {
std::vector<DelayBuffer<float>> delay_buffer(
num_render_channels, DelayBuffer<float>(delay_samples));
std::vector<std::unique_ptr<CascadedBiQuadFilter>> x_hp_filter(
num_render_channels);
for (size_t ch = 0; ch < num_render_channels; ++ch) {
x_hp_filter[ch] = std::make_unique<CascadedBiQuadFilter>(
kHighPassFilterCoefficients, 1);
}
CascadedBiQuadFilter y_hp_filter(kHighPassFilterCoefficients, 1);
SCOPED_TRACE(ProduceDebugText(num_render_channels, delay_samples));
const size_t num_blocks_to_process =
kNumBlocksToProcessPerRenderChannel * num_render_channels;
for (size_t j = 0; j < num_blocks_to_process; ++j) {
std::fill(y.begin(), y.end(), 0.f);
for (size_t ch = 0; ch < num_render_channels; ++ch) {
RandomizeSampleVector(&random_generator, x[0][ch]);
std::array<float, kBlockSize> y_channel;
delay_buffer[ch].Delay(x[0][ch], y_channel);
for (size_t k = 0; k < y.size(); ++k) {
y[k] += y_channel[k] / num_render_channels;
}
}
RandomizeSampleVector(&random_generator, n);
const float noise_scaling = 1.f / 100.f / num_render_channels;
for (size_t k = 0; k < y.size(); ++k) {
y[k] += n[k] * noise_scaling;
}
for (size_t ch = 0; ch < num_render_channels; ++ch) {
x_hp_filter[ch]->Process(x[0][ch]);
}
y_hp_filter.Process(y);
render_delay_buffer->Insert(x);
if (j == 0) {
render_delay_buffer->Reset();
}
render_delay_buffer->PrepareCaptureProcessing();
auto* const render_buffer = render_delay_buffer->GetRenderBuffer();
render_signal_analyzer.Update(*render_buffer,
aec_state.MinDirectPathFilterDelay());
filter.Filter(*render_buffer, &S);
fft.Ifft(S, &s_scratch);
std::transform(y.begin(), y.end(), s_scratch.begin() + kFftLengthBy2,
e.begin(),
[&](float a, float b) { return a - b * kScale; });
std::for_each(e.begin(), e.end(), [](float& a) {
a = rtc::SafeClamp(a, -32768.f, 32767.f);
});
fft.ZeroPaddedFft(e, Aec3Fft::Window::kRectangular, &E);
for (auto& o : output) {
for (size_t k = 0; k < kBlockSize; ++k) {
o.s_main[k] = kScale * s_scratch[k + kFftLengthBy2];
}
}
std::array<float, kFftLengthBy2Plus1> render_power;
render_buffer->SpectralSum(filter.SizePartitions(), &render_power);
gain.Compute(render_power, render_signal_analyzer, E,
filter.SizePartitions(), false, &G);
filter.Adapt(*render_buffer, G, &h[0]);
aec_state.HandleEchoPathChange(EchoPathVariability(
false, EchoPathVariability::DelayAdjustment::kNone, false));
filter.ComputeFrequencyResponse(&H2[0]);
aec_state.Update(delay_estimate, H2, h, *render_buffer, E2_main, Y2,
output);
}
// Verify that the filter is able to perform well.
EXPECT_LT(1000 * std::inner_product(e.begin(), e.end(), e.begin(), 0.f),
std::inner_product(y.begin(), y.end(), y.begin(), 0.f));
}
for (size_t delay_samples : {0, 64, 150, 200, 301}) {
std::vector<DelayBuffer<float>> delay_buffer(
num_render_channels, DelayBuffer<float>(delay_samples));
std::vector<std::unique_ptr<CascadedBiQuadFilter>> x_hp_filter(
num_render_channels);
for (size_t ch = 0; ch < num_render_channels; ++ch) {
x_hp_filter[ch] = std::make_unique<CascadedBiQuadFilter>(
kHighPassFilterCoefficients, 1);
}
CascadedBiQuadFilter y_hp_filter(kHighPassFilterCoefficients, 1);
SCOPED_TRACE(ProduceDebugText(num_render_channels, delay_samples));
const size_t num_blocks_to_process =
kNumBlocksToProcessPerRenderChannel * num_render_channels;
for (size_t j = 0; j < num_blocks_to_process; ++j) {
std::fill(y.begin(), y.end(), 0.f);
for (size_t ch = 0; ch < num_render_channels; ++ch) {
RandomizeSampleVector(&random_generator, x[0][ch]);
std::array<float, kBlockSize> y_channel;
delay_buffer[ch].Delay(x[0][ch], y_channel);
for (size_t k = 0; k < y.size(); ++k) {
y[k] += y_channel[k] / num_render_channels;
}
}
RandomizeSampleVector(&random_generator, n);
const float noise_scaling = 1.f / 100.f / num_render_channels;
for (size_t k = 0; k < y.size(); ++k) {
y[k] += n[k] * noise_scaling;
}
for (size_t ch = 0; ch < num_render_channels; ++ch) {
x_hp_filter[ch]->Process(x[0][ch]);
}
y_hp_filter.Process(y);
render_delay_buffer->Insert(x);
if (j == 0) {
render_delay_buffer->Reset();
}
render_delay_buffer->PrepareCaptureProcessing();
auto* const render_buffer = render_delay_buffer->GetRenderBuffer();
render_signal_analyzer.Update(*render_buffer,
aec_state.MinDirectPathFilterDelay());
filter.Filter(*render_buffer, &S);
fft.Ifft(S, &s_scratch);
std::transform(y.begin(), y.end(), s_scratch.begin() + kFftLengthBy2,
e.begin(),
[&](float a, float b) { return a - b * kScale; });
std::for_each(e.begin(), e.end(),
[](float& a) { a = rtc::SafeClamp(a, -32768.f, 32767.f); });
fft.ZeroPaddedFft(e, Aec3Fft::Window::kRectangular, &E);
for (auto& o : output) {
for (size_t k = 0; k < kBlockSize; ++k) {
o.s_main[k] = kScale * s_scratch[k + kFftLengthBy2];
}
}
std::array<float, kFftLengthBy2Plus1> render_power;
render_buffer->SpectralSum(filter.SizePartitions(), &render_power);
gain.Compute(render_power, render_signal_analyzer, E,
filter.SizePartitions(), false, &G);
filter.Adapt(*render_buffer, G, &h[0]);
aec_state.HandleEchoPathChange(EchoPathVariability(
false, EchoPathVariability::DelayAdjustment::kNone, false));
filter.ComputeFrequencyResponse(&H2[0]);
aec_state.Update(delay_estimate, H2, h, *render_buffer, E2_main, Y2,
output);
}
// Verify that the filter is able to perform well.
EXPECT_LT(1000 * std::inner_product(e.begin(), e.end(), e.begin(), 0.f),
std::inner_product(y.begin(), y.end(), y.begin(), 0.f));
}
}
} // namespace aec3
} // namespace webrtc

27
modules/audio_processing/aec3/aec_state_unittest.cc
View File

@ -18,13 +18,6 @@
namespace webrtc {
namespace {
std::string ProduceDebugText(size_t num_render_channels,
size_t num_capture_channels) {
rtc::StringBuilder ss;
ss << "Render channels: " << num_render_channels;
ss << ", Capture channels: " << num_capture_channels;
return ss.Release();
}
void RunNormalUsageTest(size_t num_render_channels,
size_t num_capture_channels) {
@ -232,14 +225,20 @@ void RunNormalUsageTest(size_t num_render_channels,
} // namespace
class AecStateMultiChannel
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
INSTANTIATE_TEST_SUITE_P(MultiChannel,
AecStateMultiChannel,
::testing::Combine(::testing::Values(1, 2, 8),
::testing::Values(1, 2, 8)));
// Verify the general functionality of AecState
TEST(AecState, NormalUsage) {
for (size_t num_render_channels : {1, 2, 8}) {
for (size_t num_capture_channels : {1, 2, 8}) {
SCOPED_TRACE(ProduceDebugText(num_render_channels, num_capture_channels));
RunNormalUsageTest(num_render_channels, num_capture_channels);
}
}
TEST_P(AecStateMultiChannel, NormalUsage) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
RunNormalUsageTest(num_render_channels, num_capture_channels);
}
// Verifies the delay for a converged filter is correctly identified.

47
modules/audio_processing/aec3/echo_path_delay_estimator_unittest.cc
View File

@ -34,30 +34,35 @@ std::string ProduceDebugText(size_t delay, size_t down_sampling_factor) {
} // namespace
class EchoPathDelayEstimatorMultiChannel
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
INSTANTIATE_TEST_SUITE_P(MultiChannel,
EchoPathDelayEstimatorMultiChannel,
::testing::Combine(::testing::Values(1, 2, 3, 6, 8),
::testing::Values(1, 2, 4)));
// Verifies that the basic API calls work.
TEST(EchoPathDelayEstimator, BasicApiCalls) {
TEST_P(EchoPathDelayEstimatorMultiChannel, BasicApiCalls) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
for (size_t num_capture_channels : {1, 2, 4}) {
for (size_t num_render_channels : {1, 2, 3, 6, 8}) {
ApmDataDumper data_dumper(0);
EchoCanceller3Config config;
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz,
num_render_channels));
EchoPathDelayEstimator estimator(&data_dumper, config,
num_capture_channels);
std::vector<std::vector<std::vector<float>>> render(
kNumBands, std::vector<std::vector<float>>(
num_render_channels, std::vector<float>(kBlockSize)));
std::vector<std::vector<float>> capture(num_capture_channels,
std::vector<float>(kBlockSize));
for (size_t k = 0; k < 100; ++k) {
render_delay_buffer->Insert(render);
estimator.EstimateDelay(
render_delay_buffer->GetDownsampledRenderBuffer(), capture);
}
}
ApmDataDumper data_dumper(0);
EchoCanceller3Config config;
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz, num_render_channels));
EchoPathDelayEstimator estimator(&data_dumper, config, num_capture_channels);
std::vector<std::vector<std::vector<float>>> render(
kNumBands, std::vector<std::vector<float>>(
num_render_channels, std::vector<float>(kBlockSize)));
std::vector<std::vector<float>> capture(num_capture_channels,
std::vector<float>(kBlockSize));
for (size_t k = 0; k < 100; ++k) {
render_delay_buffer->Insert(render);
estimator.EstimateDelay(render_delay_buffer->GetDownsampledRenderBuffer(),
capture);
}
}

70
modules/audio_processing/aec3/echo_remover_unittest.cc
View File

@ -26,7 +26,6 @@
namespace webrtc {
namespace {
std::string ProduceDebugText(int sample_rate_hz) {
rtc::StringBuilder ss;
ss << "Sample rate: " << sample_rate_hz;
@ -41,43 +40,48 @@ std::string ProduceDebugText(int sample_rate_hz, int delay) {
} // namespace
class EchoRemoverMultiChannel
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
INSTANTIATE_TEST_SUITE_P(MultiChannel,
EchoRemoverMultiChannel,
::testing::Combine(::testing::Values(1, 2, 8),
::testing::Values(1, 2, 8)));
// Verifies the basic API call sequence
TEST(EchoRemover, BasicApiCalls) {
TEST_P(EchoRemoverMultiChannel, BasicApiCalls) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
absl::optional<DelayEstimate> delay_estimate;
for (auto rate : {16000, 32000, 48000}) {
for (size_t num_render_channels : {1, 2, 8}) {
for (size_t num_capture_channels : {1, 2, 8}) {
SCOPED_TRACE(ProduceDebugText(rate));
std::unique_ptr<EchoRemover> remover(
EchoRemover::Create(EchoCanceller3Config(), rate,
num_render_channels, num_capture_channels));
std::unique_ptr<RenderDelayBuffer> render_buffer(
RenderDelayBuffer::Create(EchoCanceller3Config(), rate,
num_render_channels));
SCOPED_TRACE(ProduceDebugText(rate));
std::unique_ptr<EchoRemover> remover(
EchoRemover::Create(EchoCanceller3Config(), rate, num_render_channels,
num_capture_channels));
std::unique_ptr<RenderDelayBuffer> render_buffer(RenderDelayBuffer::Create(
EchoCanceller3Config(), rate, num_render_channels));
std::vector<std::vector<std::vector<float>>> render(
NumBandsForRate(rate),
std::vector<std::vector<float>>(
num_render_channels, std::vector<float>(kBlockSize, 0.f)));
std::vector<std::vector<std::vector<float>>> capture(
NumBandsForRate(rate),
std::vector<std::vector<float>>(
num_capture_channels, std::vector<float>(kBlockSize, 0.f)));
for (size_t k = 0; k < 100; ++k) {
EchoPathVariability echo_path_variability(
k % 3 == 0 ? true : false,
k % 5 == 0
? EchoPathVariability::DelayAdjustment::kNewDetectedDelay
: EchoPathVariability::DelayAdjustment::kNone,
false);
render_buffer->Insert(render);
render_buffer->PrepareCaptureProcessing();
std::vector<std::vector<std::vector<float>>> render(
NumBandsForRate(rate),
std::vector<std::vector<float>>(num_render_channels,
std::vector<float>(kBlockSize, 0.f)));
std::vector<std::vector<std::vector<float>>> capture(
NumBandsForRate(rate),
std::vector<std::vector<float>>(num_capture_channels,
std::vector<float>(kBlockSize, 0.f)));
for (size_t k = 0; k < 100; ++k) {
EchoPathVariability echo_path_variability(
k % 3 == 0 ? true : false,
k % 5 == 0 ? EchoPathVariability::DelayAdjustment::kNewDetectedDelay
: EchoPathVariability::DelayAdjustment::kNone,
false);
render_buffer->Insert(render);
render_buffer->PrepareCaptureProcessing();
remover->ProcessCapture(
echo_path_variability, k % 2 == 0 ? true : false, delay_estimate,
render_buffer->GetRenderBuffer(), nullptr, &capture);
}
}
remover->ProcessCapture(echo_path_variability, k % 2 == 0 ? true : false,
delay_estimate, render_buffer->GetRenderBuffer(),
nullptr, &capture);
}
}
}

124
modules/audio_processing/aec3/erl_estimator_unittest.cc
View File

@ -34,67 +34,71 @@ void VerifyErl(const std::array<float, kFftLengthBy2Plus1>& erl,
} // namespace
class ErlEstimatorMultiChannel
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
INSTANTIATE_TEST_SUITE_P(MultiChannel,
ErlEstimatorMultiChannel,
::testing::Combine(::testing::Values(1, 2, 8),
::testing::Values(1, 2, 8)));
// Verifies that the correct ERL estimates are achieved.
TEST(ErlEstimator, Estimates) {
for (size_t num_render_channels : {1, 2, 8}) {
for (size_t num_capture_channels : {1, 2, 8}) {
SCOPED_TRACE(ProduceDebugText(num_render_channels, num_capture_channels));
std::vector<std::array<float, kFftLengthBy2Plus1>> X2(
num_render_channels);
for (auto& X2_ch : X2) {
X2_ch.fill(0.f);
}
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(
num_capture_channels);
for (auto& Y2_ch : Y2) {
Y2_ch.fill(0.f);
}
std::vector<bool> converged_filters(num_capture_channels, false);
const size_t converged_idx = num_capture_channels - 1;
converged_filters[converged_idx] = true;
ErlEstimator estimator(0);
// Verifies that the ERL estimate is properly reduced to lower values.
for (auto& X2_ch : X2) {
X2_ch.fill(500 * 1000.f * 1000.f);
}
Y2[converged_idx].fill(10 * X2[0][0]);
for (size_t k = 0; k < 200; ++k) {
estimator.Update(converged_filters, X2, Y2);
}
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 10.f);
// Verifies that the ERL is not immediately increased when the ERL in the
// data increases.
Y2[converged_idx].fill(10000 * X2[0][0]);
for (size_t k = 0; k < 998; ++k) {
estimator.Update(converged_filters, X2, Y2);
}
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 10.f);
// Verifies that the rate of increase is 3 dB.
estimator.Update(converged_filters, X2, Y2);
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 20.f);
// Verifies that the maximum ERL is achieved when there are no low RLE
// estimates.
for (size_t k = 0; k < 1000; ++k) {
estimator.Update(converged_filters, X2, Y2);
}
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 1000.f);
// Verifies that the ERL estimate is is not updated for low-level signals
for (auto& X2_ch : X2) {
X2_ch.fill(1000.f * 1000.f);
}
Y2[converged_idx].fill(10 * X2[0][0]);
for (size_t k = 0; k < 200; ++k) {
estimator.Update(converged_filters, X2, Y2);
}
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 1000.f);
}
TEST_P(ErlEstimatorMultiChannel, Estimates) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
SCOPED_TRACE(ProduceDebugText(num_render_channels, num_capture_channels));
std::vector<std::array<float, kFftLengthBy2Plus1>> X2(num_render_channels);
for (auto& X2_ch : X2) {
X2_ch.fill(0.f);
}
}
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(num_capture_channels);
for (auto& Y2_ch : Y2) {
Y2_ch.fill(0.f);
}
std::vector<bool> converged_filters(num_capture_channels, false);
const size_t converged_idx = num_capture_channels - 1;
converged_filters[converged_idx] = true;
ErlEstimator estimator(0);
// Verifies that the ERL estimate is properly reduced to lower values.
for (auto& X2_ch : X2) {
X2_ch.fill(500 * 1000.f * 1000.f);
}
Y2[converged_idx].fill(10 * X2[0][0]);
for (size_t k = 0; k < 200; ++k) {
estimator.Update(converged_filters, X2, Y2);
}
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 10.f);
// Verifies that the ERL is not immediately increased when the ERL in the
// data increases.
Y2[converged_idx].fill(10000 * X2[0][0]);
for (size_t k = 0; k < 998; ++k) {
estimator.Update(converged_filters, X2, Y2);
}
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 10.f);
// Verifies that the rate of increase is 3 dB.
estimator.Update(converged_filters, X2, Y2);
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 20.f);
// Verifies that the maximum ERL is achieved when there are no low RLE
// estimates.
for (size_t k = 0; k < 1000; ++k) {
estimator.Update(converged_filters, X2, Y2);
}
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 1000.f);
// Verifies that the ERL estimate is is not updated for low-level signals
for (auto& X2_ch : X2) {
X2_ch.fill(1000.f * 1000.f);
}
Y2[converged_idx].fill(10 * X2[0][0]);
for (size_t k = 0; k < 200; ++k) {
estimator.Update(converged_filters, X2, Y2);
}
VerifyErl(estimator.Erl(), estimator.ErlTimeDomain(), 1000.f);
}
} // namespace webrtc

238
modules/audio_processing/aec3/erle_estimator_unittest.cc
View File

@ -16,12 +16,12 @@
#include "modules/audio_processing/aec3/render_delay_buffer.h"
#include "modules/audio_processing/aec3/spectrum_buffer.h"
#include "rtc_base/random.h"
#include "rtc_base/strings/string_builder.h"
#include "test/gtest.h"
namespace webrtc {
namespace {
constexpr int kLowFrequencyLimit = kFftLengthBy2 / 2;
constexpr float kTrueErle = 10.f;
constexpr float kTrueErleOnsets = 1.0f;
@ -129,150 +129,140 @@ void GetFilterFreq(
} // namespace
TEST(ErleEstimator, VerifyErleIncreaseAndHold) {
class ErleEstimatorMultiChannel
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
INSTANTIATE_TEST_SUITE_P(MultiChannel,
ErleEstimatorMultiChannel,
::testing::Combine(::testing::Values(1, 2, 4, 8),
::testing::Values(1, 2, 8)));
TEST_P(ErleEstimatorMultiChannel, VerifyErleIncreaseAndHold) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
for (size_t num_render_channels : {1, 2, 4, 8}) {
for (size_t num_capture_channels : {1, 2, 4}) {
std::array<float, kFftLengthBy2Plus1> X2;
std::vector<std::array<float, kFftLengthBy2Plus1>> E2(
num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(
num_capture_channels);
std::vector<bool> converged_filters(num_capture_channels, true);
std::array<float, kFftLengthBy2Plus1> X2;
std::vector<std::array<float, kFftLengthBy2Plus1>> E2(num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(num_capture_channels);
std::vector<bool> converged_filters(num_capture_channels, true);
EchoCanceller3Config config;
config.erle.onset_detection = true;
EchoCanceller3Config config;
config.erle.onset_detection = true;
std::vector<std::vector<std::vector<float>>> x(
kNumBands,
std::vector<std::vector<float>>(num_render_channels,
std::vector<float>(kBlockSize, 0.f)));
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>>
filter_frequency_response(
config.filter.main.length_blocks,
std::vector<std::array<float, kFftLengthBy2Plus1>>(
num_capture_channels));
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz,
num_render_channels));
std::vector<std::vector<std::vector<float>>> x(
kNumBands, std::vector<std::vector<float>>(
num_render_channels, std::vector<float>(kBlockSize, 0.f)));
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>>
filter_frequency_response(
config.filter.main.length_blocks,
std::vector<std::array<float, kFftLengthBy2Plus1>>(num_capture_channels));
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz, num_render_channels));
GetFilterFreq(config.delay.delay_headroom_samples,
filter_frequency_response);
GetFilterFreq(config.delay.delay_headroom_samples, filter_frequency_response);
ErleEstimator estimator(0, config, num_capture_channels);
ErleEstimator estimator(0, config, num_capture_channels);
FormFarendTimeFrame(&x);
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
// Verifies that the ERLE estimate is properly increased to higher values.
FormFarendFrame(*render_delay_buffer->GetRenderBuffer(), kTrueErle, &X2,
E2, Y2);
for (size_t k = 0; k < 200; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2,
converged_filters);
}
VerifyErle(estimator.Erle(), std::pow(2.f, estimator.FullbandErleLog2()),
config.erle.max_l, config.erle.max_h);
FormNearendFrame(&x, &X2, E2, Y2);
// Verifies that the ERLE is not immediately decreased during nearend
// activity.
for (size_t k = 0; k < 50; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2,
converged_filters);
}
VerifyErle(estimator.Erle(), std::pow(2.f, estimator.FullbandErleLog2()),
config.erle.max_l, config.erle.max_h);
}
FormFarendTimeFrame(&x);
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
// Verifies that the ERLE estimate is properly increased to higher values.
FormFarendFrame(*render_delay_buffer->GetRenderBuffer(), kTrueErle, &X2, E2,
Y2);
for (size_t k = 0; k < 200; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2, converged_filters);
}
VerifyErle(estimator.Erle(), std::pow(2.f, estimator.FullbandErleLog2()),
config.erle.max_l, config.erle.max_h);
FormNearendFrame(&x, &X2, E2, Y2);
// Verifies that the ERLE is not immediately decreased during nearend
// activity.
for (size_t k = 0; k < 50; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2, converged_filters);
}
VerifyErle(estimator.Erle(), std::pow(2.f, estimator.FullbandErleLog2()),
config.erle.max_l, config.erle.max_h);
}
TEST(ErleEstimator, VerifyErleTrackingOnOnsets) {
TEST_P(ErleEstimatorMultiChannel, VerifyErleTrackingOnOnsets) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
for (size_t num_render_channels : {1, 2, 4, 8}) {
for (size_t num_capture_channels : {1, 2, 4}) {
std::array<float, kFftLengthBy2Plus1> X2;
std::vector<std::array<float, kFftLengthBy2Plus1>> E2(
num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(
num_capture_channels);
std::vector<bool> converged_filters(num_capture_channels, true);
EchoCanceller3Config config;
config.erle.onset_detection = true;
std::vector<std::vector<std::vector<float>>> x(
kNumBands,
std::vector<std::vector<float>>(num_render_channels,
std::vector<float>(kBlockSize, 0.f)));
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>>
filter_frequency_response(
config.filter.main.length_blocks,
std::vector<std::array<float, kFftLengthBy2Plus1>>(
num_capture_channels));
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz,
num_render_channels));
std::array<float, kFftLengthBy2Plus1> X2;
std::vector<std::array<float, kFftLengthBy2Plus1>> E2(num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(num_capture_channels);
std::vector<bool> converged_filters(num_capture_channels, true);
EchoCanceller3Config config;
config.erle.onset_detection = true;
std::vector<std::vector<std::vector<float>>> x(
kNumBands, std::vector<std::vector<float>>(
num_render_channels, std::vector<float>(kBlockSize, 0.f)));
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>>
filter_frequency_response(
config.filter.main.length_blocks,
std::vector<std::array<float, kFftLengthBy2Plus1>>(num_capture_channels));
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz, num_render_channels));
GetFilterFreq(config.delay.delay_headroom_samples,
filter_frequency_response);
GetFilterFreq(config.delay.delay_headroom_samples, filter_frequency_response);
ErleEstimator estimator(/*startup_phase_length_blocks=*/0, config,
num_capture_channels);
ErleEstimator estimator(/*startup_phase_length_blocks=*/0, config,
num_capture_channels);
FormFarendTimeFrame(&x);
FormFarendTimeFrame(&x);
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
for (size_t burst = 0; burst < 20; ++burst) {
FormFarendFrame(*render_delay_buffer->GetRenderBuffer(), kTrueErleOnsets,
&X2, E2, Y2);
for (size_t k = 0; k < 10; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
for (size_t burst = 0; burst < 20; ++burst) {
FormFarendFrame(*render_delay_buffer->GetRenderBuffer(),
kTrueErleOnsets, &X2, E2, Y2);
for (size_t k = 0; k < 10; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2,
converged_filters);
}
FormFarendFrame(*render_delay_buffer->GetRenderBuffer(), kTrueErle, &X2,
E2, Y2);
for (size_t k = 0; k < 200; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2,
converged_filters);
}
FormNearendFrame(&x, &X2, E2, Y2);
for (size_t k = 0; k < 300; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2,
converged_filters);
}
}
VerifyErleBands(estimator.ErleOnsets(), config.erle.min, config.erle.min);
FormNearendFrame(&x, &X2, E2, Y2);
for (size_t k = 0; k < 1000; k++) {
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2,
converged_filters);
}
// Verifies that during ne activity, Erle converges to the Erle for
// onsets.
VerifyErle(estimator.Erle(), std::pow(2.f, estimator.FullbandErleLog2()),
config.erle.min, config.erle.min);
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2,
converged_filters);
}
FormFarendFrame(*render_delay_buffer->GetRenderBuffer(), kTrueErle, &X2, E2,
Y2);
for (size_t k = 0; k < 200; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2,
converged_filters);
}
FormNearendFrame(&x, &X2, E2, Y2);
for (size_t k = 0; k < 300; ++k) {
render_delay_buffer->Insert(x);
render_delay_buffer->PrepareCaptureProcessing();
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2,
converged_filters);
}
}
VerifyErleBands(estimator.ErleOnsets(), config.erle.min, config.erle.min);
FormNearendFrame(&x, &X2, E2, Y2);
for (size_t k = 0; k < 1000; k++) {
estimator.Update(*render_delay_buffer->GetRenderBuffer(),
filter_frequency_response, X2, Y2, E2, converged_filters);
}
// Verifies that during ne activity, Erle converges to the Erle for
// onsets.
VerifyErle(estimator.Erle(), std::pow(2.f, estimator.FullbandErleLog2()),
config.erle.min, config.erle.min);
}
} // namespace webrtc

146
modules/audio_processing/aec3/residual_echo_estimator_unittest.cc
View File

@ -16,87 +16,91 @@
#include "modules/audio_processing/aec3/render_delay_buffer.h"
#include "modules/audio_processing/test/echo_canceller_test_tools.h"
#include "rtc_base/random.h"
#include "rtc_base/strings/string_builder.h"
#include "test/gtest.h"
namespace webrtc {
TEST(ResidualEchoEstimator, BasicTest) {
for (size_t num_render_channels : {1, 2, 4}) {
for (size_t num_capture_channels : {1, 2, 4}) {
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
class ResidualEchoEstimatorMultiChannel
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
EchoCanceller3Config config;
ResidualEchoEstimator estimator(config, num_render_channels);
AecState aec_state(config, num_capture_channels);
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz,
num_render_channels));
INSTANTIATE_TEST_SUITE_P(MultiChannel,
ResidualEchoEstimatorMultiChannel,
::testing::Combine(::testing::Values(1, 2, 4),
::testing::Values(1, 2, 4)));
std::vector<std::array<float, kFftLengthBy2Plus1>> E2_main(
num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> S2_linear(
num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(
num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> R2(
num_capture_channels);
std::vector<std::vector<std::vector<float>>> x(
kNumBands,
std::vector<std::vector<float>>(num_render_channels,
std::vector<float>(kBlockSize, 0.f)));
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>> H2(
num_capture_channels,
std::vector<std::array<float, kFftLengthBy2Plus1>>(10));
Random random_generator(42U);
std::vector<SubtractorOutput> output(num_capture_channels);
std::array<float, kBlockSize> y;
absl::optional<DelayEstimate> delay_estimate;
TEST_P(ResidualEchoEstimatorMultiChannel, BasicTest) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
for (auto& H2_ch : H2) {
for (auto& H2_k : H2_ch) {
H2_k.fill(0.01f);
}
H2_ch[2].fill(10.f);
H2_ch[2][0] = 0.1f;
}
EchoCanceller3Config config;
ResidualEchoEstimator estimator(config, num_render_channels);
AecState aec_state(config, num_capture_channels);
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz, num_render_channels));
std::vector<std::vector<float>> h(
num_capture_channels,
std::vector<float>(
GetTimeDomainLength(config.filter.main.length_blocks), 0.f));
std::vector<std::array<float, kFftLengthBy2Plus1>> E2_main(
num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> S2_linear(
num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> R2(num_capture_channels);
std::vector<std::vector<std::vector<float>>> x(
kNumBands, std::vector<std::vector<float>>(
num_render_channels, std::vector<float>(kBlockSize, 0.f)));
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>> H2(
num_capture_channels,
std::vector<std::array<float, kFftLengthBy2Plus1>>(10));
Random random_generator(42U);
std::vector<SubtractorOutput> output(num_capture_channels);
std::array<float, kBlockSize> y;
absl::optional<DelayEstimate> delay_estimate;
for (auto& subtractor_output : output) {
subtractor_output.Reset();
subtractor_output.s_main.fill(100.f);
}
y.fill(0.f);
constexpr float kLevel = 10.f;
for (auto& E2_main_ch : E2_main) {
E2_main_ch.fill(kLevel);
}
S2_linear[0].fill(kLevel);
for (auto& Y2_ch : Y2) {
Y2_ch.fill(kLevel);
}
for (int k = 0; k < 1993; ++k) {
RandomizeSampleVector(&random_generator, x[0][0]);
render_delay_buffer->Insert(x);
if (k == 0) {
render_delay_buffer->Reset();
}
render_delay_buffer->PrepareCaptureProcessing();
aec_state.Update(delay_estimate, H2, h,
*render_delay_buffer->GetRenderBuffer(), E2_main, Y2,
output);
estimator.Estimate(aec_state, *render_delay_buffer->GetRenderBuffer(),
S2_linear, Y2, R2);
}
for (auto& H2_ch : H2) {
for (auto& H2_k : H2_ch) {
H2_k.fill(0.01f);
}
H2_ch[2].fill(10.f);
H2_ch[2][0] = 0.1f;
}
std::vector<std::vector<float>> h(
num_capture_channels,
std::vector<float>(GetTimeDomainLength(config.filter.main.length_blocks),
0.f));
for (auto& subtractor_output : output) {
subtractor_output.Reset();
subtractor_output.s_main.fill(100.f);
}
y.fill(0.f);
constexpr float kLevel = 10.f;
for (auto& E2_main_ch : E2_main) {
E2_main_ch.fill(kLevel);
}
S2_linear[0].fill(kLevel);
for (auto& Y2_ch : Y2) {
Y2_ch.fill(kLevel);
}
for (int k = 0; k < 1993; ++k) {
RandomizeSampleVector(&random_generator, x[0][0]);
render_delay_buffer->Insert(x);
if (k == 0) {
render_delay_buffer->Reset();
}
render_delay_buffer->PrepareCaptureProcessing();
aec_state.Update(delay_estimate, H2, h,
*render_delay_buffer->GetRenderBuffer(), E2_main, Y2,
output);
estimator.Estimate(aec_state, *render_delay_buffer->GetRenderBuffer(),
S2_linear, Y2, R2);
}
}

166
modules/audio_processing/aec3/shadow_filter_update_gain_unittest.cc
View File

@ -27,7 +27,6 @@
namespace webrtc {
namespace {
// Method for performing the simulations needed to test the main filter update
// gain functionality.
void RunFilterUpdateTest(int num_blocks_to_process,
@ -153,102 +152,119 @@ TEST(ShadowFilterUpdateGain, NullDataOutputGain) {
#endif
class ShadowFilterUpdateGainOneTwoEightRenderChannels
: public ::testing::Test,
public ::testing::WithParamInterface<size_t> {};
INSTANTIATE_TEST_SUITE_P(MultiChannel,
ShadowFilterUpdateGainOneTwoEightRenderChannels,
::testing::Values(1, 2, 8));
// Verifies that the gain formed causes the filter using it to converge.
TEST(ShadowFilterUpdateGain, GainCausesFilterToConverge) {
TEST_P(ShadowFilterUpdateGainOneTwoEightRenderChannels,
GainCausesFilterToConverge) {
const size_t num_render_channels = GetParam();
std::vector<int> blocks_with_echo_path_changes;
std::vector<int> blocks_with_saturation;
for (size_t num_render_channels : {1, 2, 8}) {
for (size_t filter_length_blocks : {12, 20, 30}) {
for (size_t delay_samples : {0, 64, 150, 200, 301}) {
SCOPED_TRACE(ProduceDebugText(delay_samples, filter_length_blocks));
for (size_t filter_length_blocks : {12, 20, 30}) {
for (size_t delay_samples : {0, 64, 150, 200, 301}) {
SCOPED_TRACE(ProduceDebugText(delay_samples, filter_length_blocks));
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G;
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G;
RunFilterUpdateTest(5000, delay_samples, num_render_channels,
filter_length_blocks, blocks_with_saturation, &e,
&y, &G);
RunFilterUpdateTest(5000, delay_samples, num_render_channels,
filter_length_blocks, blocks_with_saturation, &e, &y,
&G);
// Verify that the main filter is able to perform well.
// Use different criteria to take overmodelling into account.
if (filter_length_blocks == 12) {
EXPECT_LT(
1000 * std::inner_product(e.begin(), e.end(), e.begin(), 0.f),
std::inner_product(y.begin(), y.end(), y.begin(), 0.f));
} else {
EXPECT_LT(std::inner_product(e.begin(), e.end(), e.begin(), 0.f),
std::inner_product(y.begin(), y.end(), y.begin(), 0.f));
}
// Verify that the main filter is able to perform well.
// Use different criteria to take overmodelling into account.
if (filter_length_blocks == 12) {
EXPECT_LT(1000 * std::inner_product(e.begin(), e.end(), e.begin(), 0.f),
std::inner_product(y.begin(), y.end(), y.begin(), 0.f));
} else {
EXPECT_LT(std::inner_product(e.begin(), e.end(), e.begin(), 0.f),
std::inner_product(y.begin(), y.end(), y.begin(), 0.f));
}
}
}
}
// Verifies that the magnitude of the gain on average decreases for a
// persistently exciting signal.
TEST(ShadowFilterUpdateGain, DecreasingGain) {
for (size_t num_render_channels : {1, 2, 4}) {
for (size_t filter_length_blocks : {12, 20, 30}) {
SCOPED_TRACE(ProduceDebugText(filter_length_blocks));
std::vector<int> blocks_with_echo_path_changes;
std::vector<int> blocks_with_saturation;
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G_a;
FftData G_b;
FftData G_c;
std::array<float, kFftLengthBy2Plus1> G_a_power;
std::array<float, kFftLengthBy2Plus1> G_b_power;
std::array<float, kFftLengthBy2Plus1> G_c_power;
RunFilterUpdateTest(100, 65, num_render_channels, filter_length_blocks,
blocks_with_saturation, &e, &y, &G_a);
RunFilterUpdateTest(200, 65, num_render_channels, filter_length_blocks,
blocks_with_saturation, &e, &y, &G_b);
RunFilterUpdateTest(300, 65, num_render_channels, filter_length_blocks,
blocks_with_saturation, &e, &y, &G_c);
G_a.Spectrum(Aec3Optimization::kNone, G_a_power);
G_b.Spectrum(Aec3Optimization::kNone, G_b_power);
G_c.Spectrum(Aec3Optimization::kNone, G_c_power);
EXPECT_GT(std::accumulate(G_a_power.begin(), G_a_power.end(), 0.),
std::accumulate(G_b_power.begin(), G_b_power.end(), 0.));
EXPECT_GT(std::accumulate(G_b_power.begin(), G_b_power.end(), 0.),
std::accumulate(G_c_power.begin(), G_c_power.end(), 0.));
}
}
}
// Verifies that the gain is zero when there is saturation.
TEST(ShadowFilterUpdateGain, SaturationBehavior) {
TEST_P(ShadowFilterUpdateGainOneTwoEightRenderChannels, SaturationBehavior) {
const size_t num_render_channels = GetParam();
std::vector<int> blocks_with_echo_path_changes;
std::vector<int> blocks_with_saturation;
for (int k = 99; k < 200; ++k) {
blocks_with_saturation.push_back(k);
}
for (size_t num_render_channels : {1, 2, 8}) {
for (size_t filter_length_blocks : {12, 20, 30}) {
SCOPED_TRACE(ProduceDebugText(filter_length_blocks));
for (size_t filter_length_blocks : {12, 20, 30}) {
SCOPED_TRACE(ProduceDebugText(filter_length_blocks));
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G_a;
FftData G_a_ref;
G_a_ref.re.fill(0.f);
G_a_ref.im.fill(0.f);
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G_a;
FftData G_a_ref;
G_a_ref.re.fill(0.f);
G_a_ref.im.fill(0.f);
RunFilterUpdateTest(100, 65, num_render_channels, filter_length_blocks,
blocks_with_saturation, &e, &y, &G_a);
RunFilterUpdateTest(100, 65, num_render_channels, filter_length_blocks,
blocks_with_saturation, &e, &y, &G_a);
EXPECT_EQ(G_a_ref.re, G_a.re);
EXPECT_EQ(G_a_ref.im, G_a.im);
}
EXPECT_EQ(G_a_ref.re, G_a.re);
EXPECT_EQ(G_a_ref.im, G_a.im);
}
}
class ShadowFilterUpdateGainOneTwoFourRenderChannels
: public ::testing::Test,
public ::testing::WithParamInterface<size_t> {};
INSTANTIATE_TEST_SUITE_P(
MultiChannel,
ShadowFilterUpdateGainOneTwoFourRenderChannels,
::testing::Values(1, 2, 4),
[](const ::testing::TestParamInfo<
ShadowFilterUpdateGainOneTwoFourRenderChannels::ParamType>& info) {
return (rtc::StringBuilder() << "Render" << info.param).str();
});
// Verifies that the magnitude of the gain on average decreases for a
// persistently exciting signal.
TEST_P(ShadowFilterUpdateGainOneTwoFourRenderChannels, DecreasingGain) {
const size_t num_render_channels = GetParam();
for (size_t filter_length_blocks : {12, 20, 30}) {
SCOPED_TRACE(ProduceDebugText(filter_length_blocks));
std::vector<int> blocks_with_echo_path_changes;
std::vector<int> blocks_with_saturation;
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G_a;
FftData G_b;
FftData G_c;
std::array<float, kFftLengthBy2Plus1> G_a_power;
std::array<float, kFftLengthBy2Plus1> G_b_power;
std::array<float, kFftLengthBy2Plus1> G_c_power;
RunFilterUpdateTest(100, 65, num_render_channels, filter_length_blocks,
blocks_with_saturation, &e, &y, &G_a);
RunFilterUpdateTest(200, 65, num_render_channels, filter_length_blocks,
blocks_with_saturation, &e, &y, &G_b);
RunFilterUpdateTest(300, 65, num_render_channels, filter_length_blocks,
blocks_with_saturation, &e, &y, &G_c);
G_a.Spectrum(Aec3Optimization::kNone, G_a_power);
G_b.Spectrum(Aec3Optimization::kNone, G_b_power);
G_c.Spectrum(Aec3Optimization::kNone, G_c_power);
EXPECT_GT(std::accumulate(G_a_power.begin(), G_a_power.end(), 0.),
std::accumulate(G_b_power.begin(), G_b_power.end(), 0.));
EXPECT_GT(std::accumulate(G_b_power.begin(), G_b_power.end(), 0.),
std::accumulate(G_c_power.begin(), G_c_power.end(), 0.));
}
}
} // namespace webrtc

112
modules/audio_processing/aec3/signal_dependent_erle_estimator_unittest.cc
View File

@ -138,36 +138,42 @@ void TestInputs::UpdateCurrentPowerSpectra() {
} // namespace
TEST(SignalDependentErleEstimator, SweepSettings) {
for (size_t num_render_channels : {1, 2, 4}) {
for (size_t num_capture_channels : {1, 2, 4}) {
EchoCanceller3Config cfg;
size_t max_length_blocks = 50;
for (size_t blocks = 1; blocks < max_length_blocks;
blocks = blocks + 10) {
for (size_t delay_headroom = 0; delay_headroom < 5; ++delay_headroom) {
for (size_t num_sections = 2; num_sections < max_length_blocks;
++num_sections) {
cfg.filter.main.length_blocks = blocks;
cfg.filter.main_initial.length_blocks =
std::min(cfg.filter.main_initial.length_blocks, blocks);
cfg.delay.delay_headroom_samples = delay_headroom * kBlockSize;
cfg.erle.num_sections = num_sections;
if (EchoCanceller3Config::Validate(&cfg)) {
SignalDependentErleEstimator s(cfg, num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> average_erle(
num_capture_channels);
for (auto& e : average_erle) {
e.fill(cfg.erle.max_l);
}
TestInputs inputs(cfg, num_render_channels, num_capture_channels);
for (size_t n = 0; n < 10; ++n) {
inputs.Update();
s.Update(inputs.GetRenderBuffer(), inputs.GetH2(),
inputs.GetX2(), inputs.GetY2(), inputs.GetE2(),
average_erle, inputs.GetConvergedFilters());
}
}
class SignalDependentErleEstimatorMultiChannel
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
INSTANTIATE_TEST_SUITE_P(MultiChannel,
SignalDependentErleEstimatorMultiChannel,
::testing::Combine(::testing::Values(1, 2, 4),
::testing::Values(1, 2, 4)));
TEST_P(SignalDependentErleEstimatorMultiChannel, SweepSettings) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
EchoCanceller3Config cfg;
size_t max_length_blocks = 50;
for (size_t blocks = 1; blocks < max_length_blocks; blocks = blocks + 10) {
for (size_t delay_headroom = 0; delay_headroom < 5; ++delay_headroom) {
for (size_t num_sections = 2; num_sections < max_length_blocks;
++num_sections) {
cfg.filter.main.length_blocks = blocks;
cfg.filter.main_initial.length_blocks =
std::min(cfg.filter.main_initial.length_blocks, blocks);
cfg.delay.delay_headroom_samples = delay_headroom * kBlockSize;
cfg.erle.num_sections = num_sections;
if (EchoCanceller3Config::Validate(&cfg)) {
SignalDependentErleEstimator s(cfg, num_capture_channels);
std::vector<std::array<float, kFftLengthBy2Plus1>> average_erle(
num_capture_channels);
for (auto& e : average_erle) {
e.fill(cfg.erle.max_l);
}
TestInputs inputs(cfg, num_render_channels, num_capture_channels);
for (size_t n = 0; n < 10; ++n) {
inputs.Update();
s.Update(inputs.GetRenderBuffer(), inputs.GetH2(), inputs.GetX2(),
inputs.GetY2(), inputs.GetE2(), average_erle,
inputs.GetConvergedFilters());
}
}
}
@ -175,30 +181,28 @@ TEST(SignalDependentErleEstimator, SweepSettings) {
}
}
TEST(SignalDependentErleEstimator, LongerRun) {
for (size_t num_render_channels : {1, 2, 4}) {
for (size_t num_capture_channels : {1, 2, 4}) {
EchoCanceller3Config cfg;
cfg.filter.main.length_blocks = 2;
cfg.filter.main_initial.length_blocks = 1;
cfg.delay.delay_headroom_samples = 0;
cfg.delay.hysteresis_limit_blocks = 0;
cfg.erle.num_sections = 2;
EXPECT_EQ(EchoCanceller3Config::Validate(&cfg), true);
std::vector<std::array<float, kFftLengthBy2Plus1>> average_erle(
num_capture_channels);
for (auto& e : average_erle) {
e.fill(cfg.erle.max_l);
}
SignalDependentErleEstimator s(cfg, num_capture_channels);
TestInputs inputs(cfg, num_render_channels, num_capture_channels);
for (size_t n = 0; n < 200; ++n) {
inputs.Update();
s.Update(inputs.GetRenderBuffer(), inputs.GetH2(), inputs.GetX2(),
inputs.GetY2(), inputs.GetE2(), average_erle,
inputs.GetConvergedFilters());
}
}
TEST_P(SignalDependentErleEstimatorMultiChannel, LongerRun) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
EchoCanceller3Config cfg;
cfg.filter.main.length_blocks = 2;
cfg.filter.main_initial.length_blocks = 1;
cfg.delay.delay_headroom_samples = 0;
cfg.delay.hysteresis_limit_blocks = 0;
cfg.erle.num_sections = 2;
EXPECT_EQ(EchoCanceller3Config::Validate(&cfg), true);
std::vector<std::array<float, kFftLengthBy2Plus1>> average_erle(
num_capture_channels);
for (auto& e : average_erle) {
e.fill(cfg.erle.max_l);
}
SignalDependentErleEstimator s(cfg, num_capture_channels);
TestInputs inputs(cfg, num_render_channels, num_capture_channels);
for (size_t n = 0; n < 200; ++n) {
inputs.Update();
s.Update(inputs.GetRenderBuffer(), inputs.GetH2(), inputs.GetX2(),
inputs.GetY2(), inputs.GetE2(), average_erle,
inputs.GetConvergedFilters());
}
}

102
modules/audio_processing/aec3/subtractor_unittest.cc
View File

@ -231,33 +231,6 @@ TEST(Subtractor, Convergence) {
}
}
// Verifies that the subtractor is able to converge on correlated data.
TEST(Subtractor, ConvergenceMultiChannel) {
#if defined(NDEBUG)
const size_t kNumRenderChannelsToTest[] = {1, 2, 8};
const size_t kNumCaptureChannelsToTest[] = {1, 2, 4};
#else
const size_t kNumRenderChannelsToTest[] = {1, 2};
const size_t kNumCaptureChannelsToTest[] = {1, 2};
#endif
std::vector<int> blocks_with_echo_path_changes;
for (size_t num_render_channels : kNumRenderChannelsToTest) {
for (size_t num_capture_channels : kNumCaptureChannelsToTest) {
SCOPED_TRACE(
ProduceDebugText(num_render_channels, num_render_channels, 64, 20));
size_t num_blocks_to_process = 2500 * num_render_channels;
std::vector<float> echo_to_nearend_powers = RunSubtractorTest(
num_render_channels, num_capture_channels, num_blocks_to_process, 64,
20, 20, false, blocks_with_echo_path_changes);
for (float echo_to_nearend_power : echo_to_nearend_powers) {
EXPECT_GT(0.1f, echo_to_nearend_power);
}
}
}
}
// Verifies that the subtractor is able to handle the case when the main filter
// is longer than the shadow filter.
TEST(Subtractor, MainFilterLongerThanShadowFilter) {
@ -297,23 +270,68 @@ TEST(Subtractor, NonConvergenceOnUncorrelatedSignals) {
}
}
// Verifies that the subtractor does not converge on uncorrelated signals.
TEST(Subtractor, NonConvergenceOnUncorrelatedSignalsMultiChannel) {
class SubtractorMultiChannelUpToEightRender
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
#if defined(NDEBUG)
INSTANTIATE_TEST_SUITE_P(NonDebugMultiChannel,
SubtractorMultiChannelUpToEightRender,
::testing::Combine(::testing::Values(1, 2, 8),
::testing::Values(1, 2, 4)));
#else
INSTANTIATE_TEST_SUITE_P(DebugMultiChannel,
SubtractorMultiChannelUpToEightRender,
::testing::Combine(::testing::Values(1, 2),
::testing::Values(1, 2)));
#endif
// Verifies that the subtractor is able to converge on correlated data.
TEST_P(SubtractorMultiChannelUpToEightRender, Convergence) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
std::vector<int> blocks_with_echo_path_changes;
for (size_t num_render_channels : {1, 2, 4}) {
for (size_t num_capture_channels : {1, 2, 4}) {
SCOPED_TRACE(
ProduceDebugText(num_render_channels, num_render_channels, 64, 20));
size_t num_blocks_to_process = 5000 * num_render_channels;
std::vector<float> echo_to_nearend_powers = RunSubtractorTest(
num_render_channels, num_capture_channels, num_blocks_to_process, 64,
20, 20, true, blocks_with_echo_path_changes);
for (float echo_to_nearend_power : echo_to_nearend_powers) {
EXPECT_LT(.8f, echo_to_nearend_power);
EXPECT_NEAR(1.f, echo_to_nearend_power, 0.25f);
}
}
size_t num_blocks_to_process = 2500 * num_render_channels;
std::vector<float> echo_to_nearend_powers = RunSubtractorTest(
num_render_channels, num_capture_channels, num_blocks_to_process, 64, 20,
20, false, blocks_with_echo_path_changes);
for (float echo_to_nearend_power : echo_to_nearend_powers) {
EXPECT_GT(0.1f, echo_to_nearend_power);
}
}
class SubtractorMultiChannelUpToFourRender
: public ::testing::Test,
public ::testing::WithParamInterface<std::tuple<size_t, size_t>> {};
#if defined(NDEBUG)
INSTANTIATE_TEST_SUITE_P(NonDebugMultiChannel,
SubtractorMultiChannelUpToFourRender,
::testing::Combine(::testing::Values(1, 2, 4),
::testing::Values(1, 2, 4)));
#else
INSTANTIATE_TEST_SUITE_P(DebugMultiChannel,
SubtractorMultiChannelUpToFourRender,
::testing::Combine(::testing::Values(1, 2),
::testing::Values(1, 2)));
#endif
// Verifies that the subtractor does not converge on uncorrelated signals.
TEST_P(SubtractorMultiChannelUpToFourRender,
NonConvergenceOnUncorrelatedSignals) {
const size_t num_render_channels = std::get<0>(GetParam());
const size_t num_capture_channels = std::get<1>(GetParam());
std::vector<int> blocks_with_echo_path_changes;
size_t num_blocks_to_process = 5000 * num_render_channels;
std::vector<float> echo_to_nearend_powers = RunSubtractorTest(
num_render_channels, num_capture_channels, num_blocks_to_process, 64, 20,
20, true, blocks_with_echo_path_changes);
for (float echo_to_nearend_power : echo_to_nearend_powers) {
EXPECT_LT(.8f, echo_to_nearend_power);
EXPECT_NEAR(1.f, echo_to_nearend_power, 0.25f);
}
}
} // namespace webrtc