Avoiding overflow in cross correlation in NetEq.

webrtc/modules/audio_coding/neteq/expand.cc

 /*
  *  Copyright (c) 2012 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/expand.h"
 
 #include <assert.h>
 #include <string.h>  // memset
 
 #include <algorithm>  // min, max
 #include <limits>  // numeric_limits<T>
 
 #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/dsp_helper.h"
 #include "webrtc/modules/audio_coding/neteq/random_vector.h"
 #include "webrtc/modules/audio_coding/neteq/statistics_calculator.h"
 #include "webrtc/modules/audio_coding/neteq/sync_buffer.h"
 
 namespace webrtc {
 
+namespace {
+
+// This function decides the overflow-protecting scaling and call
+// WebRtcSpl_CrossCorrelation.
+void CrossCorrelation(int32_t* cross_correlation,
+                      const int16_t* sequence_1,
+                      const int16_t* sequence_2,
+                      size_t sequence_1_length,
+                      size_t cross_correlation_length,
+                      int* right_shifts,
+                      int cross_correlation_step) {
+  // 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 * (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 + 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() /

[hlundin-webrtc, 2016/04/22 06:48:56]

Even though it is true that A*B/C will be evaluated as (A*B)/C, I think you can
help the reader by explicitly adding the parentheses around max_1*max_2.

[minyue-webrtc, 2016/04/22 13:58:30]

Done.

+      static_cast<int32_t>(sequence_1_length));
+  const int scaling = factor == 0 ? 0 : 31 - WebRtcSpl_NormW32(factor);
+
+  assert((double)max_1 * max_2 * sequence_1_length / (1 << scaling) <=
+         WEBRTC_SPL_WORD32_MAX);
+  assert(scaling == 0 ||
+      (double)max_1 * max_2 * sequence_1_length /(1 << scaling) * 2 >
+          WEBRTC_SPL_WORD32_MAX);
+
+  WebRtcSpl_CrossCorrelation(cross_correlation, sequence_1, sequence_2,
+                             sequence_1_length, cross_correlation_length,
+                             scaling, cross_correlation_step);
+  if (right_shifts)
+    *right_shifts = scaling;
+}
+
+}  // namespace
+
 Expand::Expand(BackgroundNoise* background_noise,
                SyncBuffer* sync_buffer,
                RandomVector* random_vector,
                StatisticsCalculator* statistics,
                int fs,
                size_t num_channels)
     : random_vector_(random_vector),
       sync_buffer_(sync_buffer),
       first_expand_(true),
       fs_hz_(fs),
@@ skipping 333 matching lines @@
   size_t fs_mult_lpc_analysis_len = fs_mult * kLpcAnalysisLength;
 
   const size_t signal_length = static_cast<size_t>(256 * fs_mult);
   const int16_t* audio_history =
       &(*sync_buffer_)[0][sync_buffer_->Size() - signal_length];
 
   // Initialize.
   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,
                            kNumCorrelationCandidates, fs_mult,
                            best_correlation_index, best_correlation);
 
   // Adjust peak locations; cross-correlation lags start at 2.5 ms
   // (20 * fs_mult samples).
   best_correlation_index[0] += fs_mult_20;
   best_correlation_index[1] += fs_mult_20;
@@ skipping 45 matching lines @@
       std::max(std::min(distortion_lag + 10, fs_mult_120),
                static_cast<size_t>(60 * fs_mult));
 
   size_t start_index = std::min(distortion_lag, correlation_lag);
   size_t correlation_lags = static_cast<size_t>(
       WEBRTC_SPL_ABS_W16((distortion_lag-correlation_lag)) + 1);
   assert(correlation_lags <= static_cast<size_t>(99 * fs_mult + 1));
 
   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);
+
+    int correlation_scale;
 
     // Calculate the correlation, store in |correlation_vector2|.
-    WebRtcSpl_CrossCorrelation(
+    CrossCorrelation(
         correlation_vector2,
         &(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, &correlation_scale, -1);
 
     // Find maximizing index.
     best_index = WebRtcSpl_MaxIndexW32(correlation_vector2, correlation_lags);
     int32_t max_correlation = correlation_vector2[best_index];
     // Compensate index with start offset.
     best_index = best_index + start_index;
 
     // Calculate energies.
     int32_t energy1 = WebRtcSpl_DotProductWithScale(
         &(audio_history[signal_length - correlation_length]),
@@ skipping 97 matching lines @@
       expand_lags_[1] = (distortion_lag + correlation_lag) / 2;
       // Third lag is the average again, but rounding towards |correlation_lag|.
       if (distortion_lag > correlation_lag) {
         expand_lags_[2] = (distortion_lag + correlation_lag - 1) / 2;
       } else {
         expand_lags_[2] = (distortion_lag + correlation_lag + 1) / 2;
       }
     }
 
     // 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 -
         kUnvoicedLpcOrder;
     // Copy signal to temporary vector to be able to pad with leading zeros.
     int16_t* temp_signal = new int16_t[fs_mult_lpc_analysis_len
                                        + kUnvoicedLpcOrder];
     memset(temp_signal, 0,
            sizeof(int16_t) * (fs_mult_lpc_analysis_len + kUnvoicedLpcOrder));
     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);
+    CrossCorrelation(auto_correlation,
+                     &temp_signal[kUnvoicedLpcOrder],
+                     &temp_signal[kUnvoicedLpcOrder],
+                     fs_mult_lpc_analysis_len, kUnvoicedLpcOrder + 1,
+                     &correlation_scale, -1);
     delete [] temp_signal;
 
     // Verify that variance is positive.
     if (auto_correlation[0] > 0) {
       // Estimate AR filter parameters using Levinson-Durbin algorithm;
       // kUnvoicedLpcOrder + 1 filter coefficients.
       int16_t stability = WebRtcSpl_LevinsonDurbin(auto_correlation,
                                                    parameters.ar_filter,
                                                    reflection_coeff,
                                                    kUnvoicedLpcOrder);
@@ skipping 140 matching lines @@
       voice_mix_factor(0),
       current_voice_mix_factor(0),
       onset(false),
       mute_slope(0) {
   memset(ar_filter, 0, sizeof(ar_filter));
   memset(ar_filter_state, 0, sizeof(ar_filter_state));
 }
 
 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;
   int16_t downsampling_factor;
   if (fs_hz_ == 8000) {
     num_coefficients = 3;
     downsampling_factor = 2;
     filter_coefficients = DspHelper::kDownsample8kHzTbl;
   } else if (fs_hz_ == 16000) {
     num_coefficients = 5;
@@ skipping 26 matching lines @@
       downsampling_factor, kFilterDelay);
 
   // Normalize |downsampled_input| to using all 16 bits.
   int16_t max_value = WebRtcSpl_MaxAbsValueW16(downsampled_input,
                                                kDownsampledLength);
   int16_t norm_shift = 16 - WebRtcSpl_NormW32(max_value);
   WebRtcSpl_VectorBitShiftW16(downsampled_input, kDownsampledLength,
                               downsampled_input, norm_shift);
 
   int32_t correlation[kNumCorrelationLags];
-  static const int kCorrelationShift = 6;
-  WebRtcSpl_CrossCorrelation(
+  CrossCorrelation(
       correlation,
       &downsampled_input[kDownsampledLength - kCorrelationLength],
       &downsampled_input[kDownsampledLength - kCorrelationLength
           - kCorrelationStartLag],
-      kCorrelationLength, kNumCorrelationLags, kCorrelationShift, -1);
+      kCorrelationLength, kNumCorrelationLags, nullptr, -1);
 
   // Normalize and move data from 32-bit to 16-bit vector.
   int32_t max_correlation = WebRtcSpl_MaxAbsValueW32(correlation,
                                                      kNumCorrelationLags);
   int16_t norm_shift2 = static_cast<int16_t>(
       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() {
   current_lag_index_ = current_lag_index_ + lag_index_direction_;
   // Change direction if needed.
   if (current_lag_index_ <= 0) {
     lag_index_direction_ = 1;
   }
   if (current_lag_index_ >= kNumLags - 1) {
     lag_index_direction_ = -1;
@@ skipping 109 matching lines @@
   const size_t kMaxRandSamples = RandomVector::kRandomTableSize;
   while (samples_generated < length) {
     size_t rand_length = std::min(length - samples_generated, kMaxRandSamples);
     random_vector_->IncreaseSeedIncrement(seed_increment);
     random_vector_->Generate(rand_length, &random_vector[samples_generated]);
     samples_generated += rand_length;
   }
 }
 
 }  // namespace webrtc