Avoiding overflow in cross correlation in NetEq.

BUG=

Review-Url: https://codereview.webrtc.org/1908623002
Cr-Commit-Position: refs/heads/master@{#12538}
This commit is contained in:
minyue
2016-04-27 15:06:10 -07:00
committed by Commit bot
parent 8034614b81
commit 3d09dfdbba
21 changed files with 186 additions and 112 deletions

View File

@ -1 +1 @@
2853ab577fe571adfc7b18f77bbe58f1253d2019
01be67dc4c3b8e74743a45cbd8684c0535dec9ad

View File

@ -1 +1 @@
dc2d9f584efb0111ebcd71a2c86f1fb09cd8c2bb
9f09f7a5be8e85ddee80ffbe20d75a6b86ecaaa2

View File

@ -1 +1 @@
c23004d91ffbe5e7a1f24620fc89b58c0426040f
9e490b6f347fd700395b70907d944684c4a8e595

View File

@ -1 +1 @@
c23004d91ffbe5e7a1f24620fc89b58c0426040f
9e490b6f347fd700395b70907d944684c4a8e595

View File

@ -1 +1 @@
c23004d91ffbe5e7a1f24620fc89b58c0426040f
9e490b6f347fd700395b70907d944684c4a8e595

View File

@ -1 +1 @@
f587883b7c371ee8d87dbf1b0f07525af7d959b8
037b597ada7d8b287ec3aea3e72671459455f445

View File

@ -1 +1 @@
a349bd71dba548029b05d1d2a6dc7caafab9a856
d2a6b6ff54b340cf9f961c7f07768d86b3761073

View File

@ -1 +1 @@
f587883b7c371ee8d87dbf1b0f07525af7d959b8
037b597ada7d8b287ec3aea3e72671459455f445

View File

@ -1 +1 @@
08266b198e7686b3cd9330813e0d2cd72fc8fdc2
bcd29d03325ad4db54399a89149db764b4c75287

View File

@ -10,6 +10,8 @@
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
#include <assert.h>
/* C version of WebRtcSpl_CrossCorrelation() for generic platforms. */
void WebRtcSpl_CrossCorrelationC(int32_t* cross_correlation,
const int16_t* seq1,
@ -23,7 +25,7 @@ void WebRtcSpl_CrossCorrelationC(int32_t* cross_correlation,
for (i = 0; i < dim_cross_correlation; i++) {
int32_t corr = 0;
for (j = 0; j < dim_seq; j++)
corr += (seq1[j] * seq2[j]) >> right_shifts;
corr += seq1[j] * seq2[j] >> right_shifts;
seq2 += step_seq2;
*cross_correlation++ = corr;
}

View File

@ -760,6 +760,8 @@ source_set("neteq") {
"neteq/buffer_level_filter.h",
"neteq/comfort_noise.cc",
"neteq/comfort_noise.h",
"neteq/cross_correlation.cc",
"neteq/cross_correlation.h",
"neteq/decision_logic.cc",
"neteq/decision_logic.h",
"neteq/decision_logic_fax.cc",

View File

@ -939,34 +939,34 @@ class AcmReceiverBitExactnessOldApi : public ::testing::Test {
#if (defined(WEBRTC_CODEC_ISAC) || defined(WEBRTC_CODEC_ISACFX)) && \
defined(WEBRTC_CODEC_ILBC) && defined(WEBRTC_CODEC_G722)
TEST_F(AcmReceiverBitExactnessOldApi, 8kHzOutput) {
Run(8000, PlatformChecksum("f34e5c0e4dd4cd6c82b23f6ed006dad0",
"67a1471049dc87e7498bc19bf130dd35",
"efb5a07480bad8afb184c4150f4b3f3a",
"51717ab374871cbfa2c6977ea2aa40f3"),
Run(8000, PlatformChecksum("39ef9b20cd8b58cad7e0a3d111635827",
"09ef9ae4e8890a9a2f7850f7326836a0",
"a607f7d0ba98683c9c236217f86aaa6b",
"4a54f6ec712bda58484a388e1a332b42"),
std::vector<ExternalDecoder>());
}
TEST_F(AcmReceiverBitExactnessOldApi, 16kHzOutput) {
Run(16000, PlatformChecksum("5066b412805f3050f65154d676006964",
"887905a40d37f213b76f64296871473e",
"f580bfd4e5e29f0399b61b7512d4e3b4",
"5b2ae32c590b41d0c601179e14eaae96"),
Run(16000, PlatformChecksum("53763516a209220a209e2287f54ecae6",
"37a61e657affd80c2b8f07cea9879411",
"cdc3d88f6d8e497d4e00c62c0e6dbb3c",
"83edb67c157d0e3a0fb9f7d7b1ce5dc7"),
std::vector<ExternalDecoder>());
}
TEST_F(AcmReceiverBitExactnessOldApi, 32kHzOutput) {
Run(32000, PlatformChecksum("2cb4784af507c45b9121e2315def36f2",
"d2392b3247095d894a49b74a1106f281",
"fdf5166b98c43235978685e40e28fea6",
"7f620312f2fa74a10048bbb7739d4bf3"),
Run(32000, PlatformChecksum("ef1a373c9d03e1832bf4eaf389b57fb7",
"c7f7372441ca69858cd56b9ef4d5dcf2",
"c4a0e0b2e031d62c693af2a9ff4337ac",
"4cbfc6ab4d704f5d9b4f10406437fda9"),
std::vector<ExternalDecoder>());
}
TEST_F(AcmReceiverBitExactnessOldApi, 48kHzOutput) {
Run(48000, PlatformChecksum("ce63f874a198621fa35398e412640fcf",
"2cf0b8fe9784e8c96db307e125beb723",
"71f89e87ee1bad594f529d6c036289ad",
"b64c891e99eccc9ff45541ef67c9e9bf"),
Run(48000, PlatformChecksum("fa4bd7c1d040e710d8b93c1a1fdbd2bb",
"41856a46d0604ad307df749598b6cdce",
"22242dd832824046d48db9ea8a01f84c",
"c7f46bf165400b266d9b57aee02d2747"),
std::vector<ExternalDecoder>());
}
@ -1021,10 +1021,10 @@ TEST_F(AcmReceiverBitExactnessOldApi, 48kHzOutputExternalDecoder) {
std::vector<ExternalDecoder> external_decoders;
external_decoders.push_back(ed);
Run(48000, PlatformChecksum("ce63f874a198621fa35398e412640fcf",
"2cf0b8fe9784e8c96db307e125beb723",
"71f89e87ee1bad594f529d6c036289ad",
"b64c891e99eccc9ff45541ef67c9e9bf"),
Run(48000, PlatformChecksum("fa4bd7c1d040e710d8b93c1a1fdbd2bb",
"41856a46d0604ad307df749598b6cdce",
"22242dd832824046d48db9ea8a01f84c",
"c7f46bf165400b266d9b57aee02d2747"),
external_decoders);
EXPECT_CALL(mock_decoder, Die());

View File

@ -17,6 +17,7 @@
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
#include "webrtc/modules/audio_coding/neteq/audio_multi_vector.h"
#include "webrtc/modules/audio_coding/neteq/cross_correlation.h"
#include "webrtc/modules/audio_coding/neteq/post_decode_vad.h"
namespace webrtc {
@ -169,15 +170,10 @@ int16_t BackgroundNoise::ScaleShift(size_t channel) const {
int32_t BackgroundNoise::CalculateAutoCorrelation(
const int16_t* signal, size_t length, int32_t* auto_correlation) const {
int16_t signal_max = WebRtcSpl_MaxAbsValueW16(signal, length);
int correlation_scale = kLogVecLen -
WebRtcSpl_NormW32(signal_max * signal_max);
correlation_scale = std::max(0, correlation_scale);
static const int kCorrelationStep = -1;
WebRtcSpl_CrossCorrelation(auto_correlation, signal, signal, length,
kMaxLpcOrder + 1, correlation_scale,
kCorrelationStep);
const int correlation_scale =
CrossCorrelationWithAutoShift(signal, signal, length, kMaxLpcOrder + 1,
kCorrelationStep, auto_correlation);
// Number of shifts to normalize energy to energy/sample.
int energy_sample_shift = kLogVecLen - correlation_scale;

View File

@ -0,0 +1,62 @@
/*
* Copyright (c) 2016 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 "webrtc/modules/audio_coding/neteq/cross_correlation.h"
#include <cstdlib>
#include <limits>
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
namespace webrtc {
// This function decides the overflow-protecting scaling and calls
// WebRtcSpl_CrossCorrelation.
int CrossCorrelationWithAutoShift(const int16_t* sequence_1,
const int16_t* sequence_2,
size_t sequence_1_length,
size_t cross_correlation_length,
int cross_correlation_step,
int32_t* cross_correlation) {
// Find the maximum absolute value of sequence_1 and 2.
const int16_t max_1 = WebRtcSpl_MaxAbsValueW16(sequence_1, sequence_1_length);
const int sequence_2_shift =
cross_correlation_step * (static_cast<int>(cross_correlation_length) - 1);
const int16_t* sequence_2_start =
sequence_2_shift >= 0 ? sequence_2 : sequence_2 + sequence_2_shift;
const size_t sequence_2_length =
sequence_1_length + std::abs(sequence_2_shift);
const int16_t max_2 =
WebRtcSpl_MaxAbsValueW16(sequence_2_start, sequence_2_length);
// In order to avoid overflow when computing the sum we should scale the
// samples so that (in_vector_length * max_1 * max_2) will not overflow.
// Expected scaling fulfills
// 1) sufficient:
// sequence_1_length * (max_1 * max_2 >> scaling) <= 0x7fffffff;
// 2) necessary:
// if (scaling > 0)
// sequence_1_length * (max_1 * max_2 >> (scaling - 1)) > 0x7fffffff;
// The following calculation fulfills 1) and almost fulfills 2).
// There are some corner cases that 2) is not satisfied, e.g.,
// max_1 = 17, max_2 = 30848, sequence_1_length = 4095, in such case,
// optimal scaling is 0, while the following calculation results in 1.
const int32_t factor = (max_1 * max_2) / (std::numeric_limits<int32_t>::max()
/ static_cast<int32_t>(sequence_1_length));
const int scaling = factor == 0 ? 0 : 31 - WebRtcSpl_NormW32(factor);
WebRtcSpl_CrossCorrelation(cross_correlation, sequence_1, sequence_2,
sequence_1_length, cross_correlation_length,
scaling, cross_correlation_step);
return scaling;
}
} // namespace webrtc

View File

@ -0,0 +1,50 @@
/*
* Copyright (c) 2016 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 WEBRTC_MODULES_AUDIO_CODING_NETEQ_CROSS_CORRELATION_H_
#define WEBRTC_MODULES_AUDIO_CODING_NETEQ_CROSS_CORRELATION_H_
#include "webrtc/common_types.h"
namespace webrtc {
// The function calculates the cross-correlation between two sequences
// |sequence_1| and |sequence_2|. |sequence_1| is taken as reference, with
// |sequence_1_length| as its length. |sequence_2| slides for the calculation of
// cross-correlation. The result will be saved in |cross_correlation|.
// |cross_correlation_length| correlation points are calculated.
// The corresponding lag starts from 0, and increases with a step of
// |cross_correlation_step|. The result is without normalization. To avoid
// overflow, the result will be right shifted. The amount of shifts will be
// returned.
//
// Input:
// - sequence_1 : First sequence (reference).
// - sequence_2 : Second sequence (sliding during calculation).
// - sequence_1_length : Length of |sequence_1|.
// - cross_correlation_length : Number of cross-correlations to calculate.
// - cross_correlation_step : Step in the lag for the cross-correlation.
//
// Output:
// - cross_correlation : The cross-correlation in Q(-right_shifts)
//
// Return:
// Number of right shifts in cross_correlation.
int CrossCorrelationWithAutoShift(const int16_t* sequence_1,
const int16_t* sequence_2,
size_t sequence_1_length,
size_t cross_correlation_length,
int cross_correlation_step,
int32_t* cross_correlation);
} // namespace webrtc
#endif // WEBRTC_MODULES_AUDIO_CODING_NETEQ_CROSS_CORRELATION_H_

View File

@ -19,6 +19,7 @@
#include "webrtc/base/safe_conversions.h"
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
#include "webrtc/modules/audio_coding/neteq/background_noise.h"
#include "webrtc/modules/audio_coding/neteq/cross_correlation.h"
#include "webrtc/modules/audio_coding/neteq/dsp_helper.h"
#include "webrtc/modules/audio_coding/neteq/random_vector.h"
#include "webrtc/modules/audio_coding/neteq/statistics_calculator.h"
@ -379,12 +380,10 @@ void Expand::AnalyzeSignal(int16_t* random_vector) {
InitializeForAnExpandPeriod();
// Calculate correlation in downsampled domain (4 kHz sample rate).
int correlation_scale;
size_t correlation_length = 51; // TODO(hlundin): Legacy bit-exactness.
// If it is decided to break bit-exactness |correlation_length| should be
// initialized to the return value of Correlation().
Correlation(audio_history, signal_length, correlation_vector,
&correlation_scale);
Correlation(audio_history, signal_length, correlation_vector);
// Find peaks in correlation vector.
DspHelper::PeakDetection(correlation_vector, correlation_length,
@ -450,21 +449,12 @@ void Expand::AnalyzeSignal(int16_t* random_vector) {
for (size_t channel_ix = 0; channel_ix < num_channels_; ++channel_ix) {
ChannelParameters& parameters = channel_parameters_[channel_ix];
// Calculate suitable scaling.
int16_t signal_max = WebRtcSpl_MaxAbsValueW16(
&audio_history[signal_length - correlation_length - start_index
- correlation_lags],
correlation_length + start_index + correlation_lags - 1);
correlation_scale = (31 - WebRtcSpl_NormW32(signal_max * signal_max)) +
(31 - WebRtcSpl_NormW32(static_cast<int32_t>(correlation_length))) - 31;
correlation_scale = std::max(0, correlation_scale);
// Calculate the correlation, store in |correlation_vector2|.
WebRtcSpl_CrossCorrelation(
correlation_vector2,
int correlation_scale = CrossCorrelationWithAutoShift(
&(audio_history[signal_length - correlation_length]),
&(audio_history[signal_length - correlation_length - start_index]),
correlation_length, correlation_lags, correlation_scale, -1);
correlation_length, correlation_lags, -1, correlation_vector2);
// Find maximizing index.
best_index = WebRtcSpl_MaxIndexW32(correlation_vector2, correlation_lags);
@ -582,13 +572,6 @@ void Expand::AnalyzeSignal(int16_t* random_vector) {
}
// Calculate the LPC and the gain of the filters.
// Calculate scale value needed for auto-correlation.
correlation_scale = WebRtcSpl_MaxAbsValueW16(
&(audio_history[signal_length - fs_mult_lpc_analysis_len]),
fs_mult_lpc_analysis_len);
correlation_scale = std::min(16 - WebRtcSpl_NormW32(correlation_scale), 0);
correlation_scale = std::max(correlation_scale * 2 + 7, 0);
// Calculate kUnvoicedLpcOrder + 1 lags of the auto-correlation function.
size_t temp_index = signal_length - fs_mult_lpc_analysis_len -
@ -601,11 +584,9 @@ void Expand::AnalyzeSignal(int16_t* random_vector) {
memcpy(&temp_signal[kUnvoicedLpcOrder],
&audio_history[temp_index + kUnvoicedLpcOrder],
sizeof(int16_t) * fs_mult_lpc_analysis_len);
WebRtcSpl_CrossCorrelation(auto_correlation,
&temp_signal[kUnvoicedLpcOrder],
&temp_signal[kUnvoicedLpcOrder],
fs_mult_lpc_analysis_len, kUnvoicedLpcOrder + 1,
correlation_scale, -1);
correlation_scale = CrossCorrelationWithAutoShift(
&temp_signal[kUnvoicedLpcOrder], &temp_signal[kUnvoicedLpcOrder],
fs_mult_lpc_analysis_len, kUnvoicedLpcOrder + 1, -1, auto_correlation);
delete [] temp_signal;
// Verify that variance is positive.
@ -766,8 +747,7 @@ Expand::ChannelParameters::ChannelParameters()
void Expand::Correlation(const int16_t* input,
size_t input_length,
int16_t* output,
int* output_scale) const {
int16_t* output) const {
// Set parameters depending on sample rate.
const int16_t* filter_coefficients;
size_t num_coefficients;
@ -814,13 +794,11 @@ void Expand::Correlation(const int16_t* input,
downsampled_input, norm_shift);
int32_t correlation[kNumCorrelationLags];
static const int kCorrelationShift = 6;
WebRtcSpl_CrossCorrelation(
correlation,
CrossCorrelationWithAutoShift(
&downsampled_input[kDownsampledLength - kCorrelationLength],
&downsampled_input[kDownsampledLength - kCorrelationLength
- kCorrelationStartLag],
kCorrelationLength, kNumCorrelationLags, kCorrelationShift, -1);
kCorrelationLength, kNumCorrelationLags, -1, correlation);
// Normalize and move data from 32-bit to 16-bit vector.
int32_t max_correlation = WebRtcSpl_MaxAbsValueW32(correlation,
@ -829,8 +807,6 @@ void Expand::Correlation(const int16_t* input,
std::max(18 - WebRtcSpl_NormW32(max_correlation), 0));
WebRtcSpl_VectorBitShiftW32ToW16(output, kNumCorrelationLags, correlation,
norm_shift2);
// Total scale factor (right shifts) of correlation value.
*output_scale = 2 * norm_shift + kCorrelationShift + norm_shift2;
}
void Expand::UpdateLagIndex() {

View File

@ -120,12 +120,10 @@ class Expand {
// Calculate the auto-correlation of |input|, with length |input_length|
// samples. The correlation is calculated from a downsampled version of
// |input|, and is written to |output|. The scale factor is written to
// |output_scale|.
// |input|, and is written to |output|.
void Correlation(const int16_t* input,
size_t input_length,
int16_t* output,
int* output_scale) const;
int16_t* output) const;
void UpdateLagIndex();

View File

@ -18,6 +18,7 @@
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
#include "webrtc/modules/audio_coding/neteq/audio_multi_vector.h"
#include "webrtc/modules/audio_coding/neteq/cross_correlation.h"
#include "webrtc/modules/audio_coding/neteq/dsp_helper.h"
#include "webrtc/modules/audio_coding/neteq/expand.h"
#include "webrtc/modules/audio_coding/neteq/sync_buffer.h"
@ -63,10 +64,8 @@ size_t Merge::Process(int16_t* input, size_t input_length,
for (size_t channel = 0; channel < num_channels_; ++channel) {
int16_t* input_channel = &input_vector[channel][0];
int16_t* expanded_channel = &expanded_[channel][0];
int16_t expanded_max, input_max;
int16_t new_mute_factor = SignalScaling(
input_channel, input_length_per_channel, expanded_channel,
&expanded_max, &input_max);
input_channel, input_length_per_channel, expanded_channel);
// Adjust muting factor (product of "main" muting factor and expand muting
// factor).
@ -89,8 +88,7 @@ size_t Merge::Process(int16_t* input, size_t input_length,
// Calculate the lag of the strongest correlation period.
best_correlation_index = CorrelateAndPeakSearch(
expanded_max, input_max, old_length,
input_length_per_channel, expand_period);
old_length, input_length_per_channel, expand_period);
}
static const int kTempDataSize = 3600;
@ -204,19 +202,19 @@ size_t Merge::GetExpandedSignal(size_t* old_length, size_t* expand_period) {
}
int16_t Merge::SignalScaling(const int16_t* input, size_t input_length,
const int16_t* expanded_signal,
int16_t* expanded_max, int16_t* input_max) const {
const int16_t* expanded_signal) const {
// Adjust muting factor if new vector is more or less of the BGN energy.
const size_t mod_input_length =
std::min(static_cast<size_t>(64 * fs_mult_), input_length);
*expanded_max = WebRtcSpl_MaxAbsValueW16(expanded_signal, mod_input_length);
*input_max = WebRtcSpl_MaxAbsValueW16(input, mod_input_length);
const int16_t expanded_max =
WebRtcSpl_MaxAbsValueW16(expanded_signal, mod_input_length);
const int16_t input_max = WebRtcSpl_MaxAbsValueW16(input, mod_input_length);
// Calculate energy of expanded signal.
// |log_fs_mult| is log2(fs_mult_), but is not exact for 48000 Hz.
int log_fs_mult = 30 - WebRtcSpl_NormW32(fs_mult_);
int expanded_shift = 6 + log_fs_mult
- WebRtcSpl_NormW32(*expanded_max * *expanded_max);
- WebRtcSpl_NormW32(expanded_max * expanded_max);
expanded_shift = std::max(expanded_shift, 0);
int32_t energy_expanded = WebRtcSpl_DotProductWithScale(expanded_signal,
expanded_signal,
@ -224,8 +222,7 @@ int16_t Merge::SignalScaling(const int16_t* input, size_t input_length,
expanded_shift);
// Calculate energy of input signal.
int input_shift = 6 + log_fs_mult -
WebRtcSpl_NormW32(*input_max * *input_max);
int input_shift = 6 + log_fs_mult - WebRtcSpl_NormW32(input_max * input_max);
input_shift = std::max(input_shift, 0);
int32_t energy_input = WebRtcSpl_DotProductWithScale(input, input,
mod_input_length,
@ -307,22 +304,17 @@ void Merge::Downsample(const int16_t* input, size_t input_length,
}
}
size_t Merge::CorrelateAndPeakSearch(int16_t expanded_max, int16_t input_max,
size_t start_position, size_t input_length,
size_t Merge::CorrelateAndPeakSearch(size_t start_position, size_t input_length,
size_t expand_period) const {
// Calculate correlation without any normalization.
const size_t max_corr_length = kMaxCorrelationLength;
size_t stop_position_downsamp =
std::min(max_corr_length, expand_->max_lag() / (fs_mult_ * 2) + 1);
int correlation_shift = 0;
if (expanded_max * input_max > 26843546) {
correlation_shift = 3;
}
int32_t correlation[kMaxCorrelationLength];
WebRtcSpl_CrossCorrelation(correlation, input_downsampled_,
expanded_downsampled_, kInputDownsampLength,
stop_position_downsamp, correlation_shift, 1);
CrossCorrelationWithAutoShift(input_downsampled_, expanded_downsampled_,
kInputDownsampLength, stop_position_downsamp, 1,
correlation);
// Normalize correlation to 14 bits and copy to a 16-bit array.
const size_t pad_length = expand_->overlap_length() - 1;

View File

@ -69,11 +69,10 @@ class Merge {
// of samples that were taken from the |sync_buffer_|.
size_t GetExpandedSignal(size_t* old_length, size_t* expand_period);
// Analyzes |input| and |expanded_signal| to find maximum values. Returns
// a muting factor (Q14) to be used on the new data.
// Analyzes |input| and |expanded_signal| and returns muting factor (Q14) to
// be used on the new data.
int16_t SignalScaling(const int16_t* input, size_t input_length,
const int16_t* expanded_signal,
int16_t* expanded_max, int16_t* input_max) const;
const int16_t* expanded_signal) const;
// Downsamples |input| (|input_length| samples) and |expanded_signal| to
// 4 kHz sample rate. The downsampled signals are written to
@ -84,8 +83,7 @@ class Merge {
// Calculates cross-correlation between |input_downsampled_| and
// |expanded_downsampled_|, and finds the correlation maximum. The maximizing
// lag is returned.
size_t CorrelateAndPeakSearch(int16_t expanded_max, int16_t input_max,
size_t start_position, size_t input_length,
size_t CorrelateAndPeakSearch(size_t start_position, size_t input_length,
size_t expand_period) const;
const int fs_mult_; // fs_hz_ / 8000.

View File

@ -73,6 +73,8 @@
'buffer_level_filter.h',
'comfort_noise.cc',
'comfort_noise.h',
'cross_correlation.cc',
'cross_correlation.h',
'decision_logic.cc',
'decision_logic.h',
'decision_logic_fax.cc',

View File

@ -16,6 +16,7 @@
#include "webrtc/base/safe_conversions.h"
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
#include "webrtc/modules/audio_coding/neteq/background_noise.h"
#include "webrtc/modules/audio_coding/neteq/cross_correlation.h"
#include "webrtc/modules/audio_coding/neteq/dsp_helper.h"
namespace webrtc {
@ -158,20 +159,15 @@ TimeStretch::ReturnCodes TimeStretch::Process(const int16_t* input,
}
void TimeStretch::AutoCorrelation() {
// Set scaling factor for cross correlation to protect against overflow.
int scaling = kLogCorrelationLen - WebRtcSpl_NormW32(
max_input_value_ * max_input_value_);
scaling = std::max(0, scaling);
// Calculate correlation from lag kMinLag to lag kMaxLag in 4 kHz domain.
int32_t auto_corr[kCorrelationLen];
WebRtcSpl_CrossCorrelation(auto_corr, &downsampled_input_[kMaxLag],
&downsampled_input_[kMaxLag - kMinLag],
kCorrelationLen, kMaxLag - kMinLag, scaling, -1);
CrossCorrelationWithAutoShift(
&downsampled_input_[kMaxLag], &downsampled_input_[kMaxLag - kMinLag],
kCorrelationLen, kMaxLag - kMinLag, -1, auto_corr);
// Normalize correlation to 14 bits and write to |auto_correlation_|.
int32_t max_corr = WebRtcSpl_MaxAbsValueW32(auto_corr, kCorrelationLen);
scaling = std::max(0, 17 - WebRtcSpl_NormW32(max_corr));
int scaling = std::max(0, 17 - WebRtcSpl_NormW32(max_corr));
WebRtcSpl_VectorBitShiftW32ToW16(auto_correlation_, kCorrelationLen,
auto_corr, scaling);
}