Avoiding overflow in cross correlation in NetEq.

webrtc/modules/audio_coding/neteq/expand.cc

Index: webrtc/modules/audio_coding/neteq/expand.cc
diff --git a/webrtc/modules/audio_coding/neteq/expand.cc b/webrtc/modules/audio_coding/neteq/expand.cc
index ef7af46597e4344156dc48b553d467934622f4b2..965ae89428c12a27204b53d91f06537c3110d062 100644
--- a/webrtc/modules/audio_coding/neteq/expand.cc
+++ b/webrtc/modules/audio_coding/neteq/expand.cc
@@ -26,6 +26,49 @@
 
 namespace webrtc {
 
+namespace {
+
+// This function decides the overflow-protecting scaling and call

[hlundin-webrtc, 2016/04/21 08:52:01]

call -> calls

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

Still, call -> calls

[minyue-webrtc, 2016/04/22 07:39:49]

Yes, I realized it when I uploaded the patch.

+// WebRtcSpl_CrossCorrelation.s

[hlundin-webrtc, 2016/04/21 08:52:01]

Remove the 's' at the end of the line.

[minyue-webrtc, 2016/04/21 11:06:50]

oh, sorry, typo

+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) {

[hlundin-webrtc, 2016/04/21 08:52:01]

You are always calling this function with cross_correlation_step == -1. You can
simplify the code below if you remove the cross_correlation_step parameter and
simply assume -1.

[minyue-webrtc, 2016/04/21 11:06:49]

I am about to discuss with you whether we'd put this method to
signal_processing_library, since other calls to WebRtcSpl_CrossCorrelation are
prone to overflow as well.

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

OK. Keep it like this.

[minyue-webrtc, 2016/04/22 07:39:49]

Do you think it is good that we try to replace WebRtcSpl_CrossCorrelation with
CrossCorrelation now (at least for all uses in neteq), to avoid changing
bitexactness test twice?

[hlundin-webrtc, 2016/04/22 07:48:20]

Replace all in NetEq now. Leave iLBC as it is.

+  // 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.
+  int scaling;
+  const int32_t factor = WEBRTC_SPL_MUL(max_1, max_2) /

[hlundin-webrtc, 2016/04/21 08:52:01]

I think you can omit WEBRTC_SPL_MUL and simply do max_1 * max_2.

[minyue-webrtc, 2016/04/21 11:06:50]

Yes, I should.

+      (WEBRTC_SPL_WORD32_MAX / (int32_t)sequence_1_length);

[hlundin-webrtc, 2016/04/21 08:52:00]

Use c++-style cast.

[hlundin-webrtc, 2016/04/21 08:52:01]

Use std::numeric_limits<int32_t>::max() instead of WEBRTC_SPL_WORD32_MAX.

[minyue-webrtc, 2016/04/21 11:06:50]

Done.

[minyue-webrtc, 2016/04/21 11:06:50]

Done.

+  scaling = factor == 0 ? 0 : 31 - WebRtcSpl_NormW32(factor);
+
+  assert((double)max_1 * max_2 * sequence_1_length / (1 << scaling) <=

[minyue-webrtc, 2016/04/20 17:58:59]

test code, will remove

[hlundin-webrtc, 2016/04/21 08:52:00]

Acknowledged.

+         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)

[hlundin-webrtc, 2016/04/21 08:52:00]

You are not calling this function without a valid pointer to |right_shifts|. I
suggest you simply DCHECK that |right_shifts| is not nullptr, skip the internal
variable |scaling|, and use |*right_shifts| instead of |scaling|.

[minyue-webrtc, 2016/04/21 11:06:50]

ok, I was thinking of reducing some uses of the returned scale. This is true and
see my new patch. If you think the additional change is not needed, I can come
back.

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

I see. But then I suggest you let the scaling factor be a return value of the
method instead. Callers that do not care can simply not use the return value.
WDYT?

[minyue-webrtc, 2016/04/22 07:39:49]

Good idea.

+    *right_shifts = scaling;
+}
+
+}  // namespace
+
 Expand::Expand(BackgroundNoise* background_noise,
                SyncBuffer* sync_buffer,
                RandomVector* random_vector,
@@ -450,21 +493,13 @@ 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(
+    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);
@@ -582,13 +617,6 @@ 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 +629,11 @@ 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);
+    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.
@@ -814,13 +842,13 @@ void Expand::Correlation(const int16_t* input,
                               downsampled_input, norm_shift);
 
   int32_t correlation[kNumCorrelationLags];
-  static const int kCorrelationShift = 6;
-  WebRtcSpl_CrossCorrelation(
+  int correlation_shift;
+  CrossCorrelation(
       correlation,
       &downsampled_input[kDownsampledLength - kCorrelationLength],
       &downsampled_input[kDownsampledLength - kCorrelationLength
           - kCorrelationStartLag],
-      kCorrelationLength, kNumCorrelationLags, kCorrelationShift, -1);
+      kCorrelationLength, kNumCorrelationLags, &correlation_shift, -1);
 
   // Normalize and move data from 32-bit to 16-bit vector.
   int32_t max_correlation = WebRtcSpl_MaxAbsValueW32(correlation,
@@ -830,7 +858,7 @@ void Expand::Correlation(const int16_t* input,
   WebRtcSpl_VectorBitShiftW32ToW16(output, kNumCorrelationLags, correlation,
                                    norm_shift2);
   // Total scale factor (right shifts) of correlation value.
-  *output_scale = 2 * norm_shift + kCorrelationShift + norm_shift2;
+  *output_scale = 2 * norm_shift + correlation_shift + norm_shift2;
 }
 
 void Expand::UpdateLagIndex() {