cng exitation signal can saturate. Use same scaling as PLC and normal decoding

1 files changed, 32 insertions, 29 deletions

diff --git a/silk/CNG.c b/silk/CNG.c
index 61787c23..8443ad63 100644
--- a/silk/CNG.c
+++ b/silk/CNG.c
@@ -34,9 +34,8 @@ POSSIBILITY OF SUCH DAMAGE.
 
 /* Generates excitation for CNG LPC synthesis */
 static OPUS_INLINE void silk_CNG_exc(
-    opus_int32                       exc_Q10[],          /* O    CNG excitation signal Q10                   */
+    opus_int32                       exc_Q14[],          /* O    CNG excitation signal Q10                   */
     opus_int32                       exc_buf_Q14[],      /* I    Random samples buffer Q10                   */
-    opus_int32                       Gain_Q16,           /* I    Gain to apply                               */
     opus_int                         length,             /* I    Length                                      */
     opus_int32                       *rand_seed          /* I/O  Seed to random index generator              */
 )
@@ -55,7 +54,7 @@ static OPUS_INLINE void silk_CNG_exc(
         idx = (opus_int)( silk_RSHIFT( seed, 24 ) & exc_mask );
         silk_assert( idx >= 0 );
         silk_assert( idx <= CNG_BUF_MASK_MAX );
-        exc_Q10[ i ] = (opus_int16)silk_SAT16( silk_SMULWW( exc_buf_Q14[ idx ], Gain_Q16 >> 4 ) );
+        exc_Q14[ i ] = exc_buf_Q14[ idx ];
     }
     *rand_seed = seed;
 }
@@ -85,7 +84,7 @@ void silk_CNG(
 )
 {
     opus_int   i, subfr;
-    opus_int32 sum_Q6, max_Gain_Q16, gain_Q16;
+    opus_int32 LPC_pred_Q10, max_Gain_Q16, gain_Q16, gain_Q10;
     opus_int16 A_Q12[ MAX_LPC_ORDER ];
     silk_CNG_struct *psCNG = &psDec->sCNG;
     SAVE_STACK;
@@ -124,8 +123,8 @@ void silk_CNG(
 
     /* Add CNG when packet is lost or during DTX */
     if( psDec->lossCnt ) {
-        VARDECL( opus_int32, CNG_sig_Q10 );
-        ALLOC( CNG_sig_Q10, length + MAX_LPC_ORDER, opus_int32 );
+        VARDECL( opus_int32, CNG_sig_Q14 );
+        ALLOC( CNG_sig_Q14, length + MAX_LPC_ORDER, opus_int32 );
 
         /* Generate CNG excitation */
         gain_Q16 = silk_SMULWW( psDec->sPLC.randScale_Q14, psDec->sPLC.prevGain_Q16[1] );
@@ -138,42 +137,46 @@ void silk_CNG(
             gain_Q16 = silk_SUB_LSHIFT32(silk_SMULWW( psCNG->CNG_smth_Gain_Q16, psCNG->CNG_smth_Gain_Q16 ), gain_Q16, 5 );
             gain_Q16 = silk_LSHIFT32( silk_SQRT_APPROX( gain_Q16 ), 8 );
         }
-        silk_CNG_exc( CNG_sig_Q10 + MAX_LPC_ORDER, psCNG->CNG_exc_buf_Q14, gain_Q16, length, &psCNG->rand_seed );
+        gain_Q10 = silk_RSHIFT( gain_Q16, 6 );
+        
+        silk_CNG_exc( CNG_sig_Q14 + MAX_LPC_ORDER, psCNG->CNG_exc_buf_Q14, length, &psCNG->rand_seed );
 
         /* Convert CNG NLSF to filter representation */
         silk_NLSF2A( A_Q12, psCNG->CNG_smth_NLSF_Q15, psDec->LPC_order );
 
         /* Generate CNG signal, by synthesis filtering */
-        silk_memcpy( CNG_sig_Q10, psCNG->CNG_synth_state, MAX_LPC_ORDER * sizeof( opus_int32 ) );
+        silk_memcpy( CNG_sig_Q14, psCNG->CNG_synth_state, MAX_LPC_ORDER * sizeof( opus_int32 ) );
         for( i = 0; i < length; i++ ) {
             silk_assert( psDec->LPC_order == 10 || psDec->LPC_order == 16 );
             /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
-            sum_Q6 = silk_RSHIFT( psDec->LPC_order, 1 );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i -  1 ], A_Q12[ 0 ] );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i -  2 ], A_Q12[ 1 ] );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i -  3 ], A_Q12[ 2 ] );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i -  4 ], A_Q12[ 3 ] );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i -  5 ], A_Q12[ 4 ] );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i -  6 ], A_Q12[ 5 ] );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i -  7 ], A_Q12[ 6 ] );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i -  8 ], A_Q12[ 7 ] );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i -  9 ], A_Q12[ 8 ] );
-            sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
+            LPC_pred_Q10 = silk_RSHIFT( psDec->LPC_order, 1 );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  1 ], A_Q12[ 0 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  2 ], A_Q12[ 1 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  3 ], A_Q12[ 2 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  4 ], A_Q12[ 3 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  5 ], A_Q12[ 4 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  6 ], A_Q12[ 5 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  7 ], A_Q12[ 6 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  8 ], A_Q12[ 7 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  9 ], A_Q12[ 8 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
             if( psDec->LPC_order == 16 ) {
-                sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 11 ], A_Q12[ 10 ] );
-                sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 12 ], A_Q12[ 11 ] );
-                sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 13 ], A_Q12[ 12 ] );
-                sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 14 ], A_Q12[ 13 ] );
-                sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 15 ], A_Q12[ 14 ] );
-                sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 16 ], A_Q12[ 15 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 11 ], A_Q12[ 10 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 12 ], A_Q12[ 11 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 13 ], A_Q12[ 12 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 14 ], A_Q12[ 13 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 15 ], A_Q12[ 14 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 16 ], A_Q12[ 15 ] );
             }
 
             /* Update states */
-            CNG_sig_Q10[ MAX_LPC_ORDER + i ] = silk_ADD_LSHIFT( CNG_sig_Q10[ MAX_LPC_ORDER + i ], sum_Q6, 4 );
-
-            frame[ i ] = silk_ADD_SAT16( frame[ i ], silk_RSHIFT_ROUND( CNG_sig_Q10[ MAX_LPC_ORDER + i ], 10 ) );
+            CNG_sig_Q14[ MAX_LPC_ORDER + i ] = silk_ADD_LSHIFT( CNG_sig_Q14[ MAX_LPC_ORDER + i ], LPC_pred_Q10, 4 );
+            
+            /* Scale with Gain and add to input signal */
+            frame[ i ] = (opus_int16)silk_ADD_SAT16( frame[ i ], silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( CNG_sig_Q14[ MAX_LPC_ORDER + i ], gain_Q10 ), 8 ) ) );
+            
         }
-        silk_memcpy( psCNG->CNG_synth_state, &CNG_sig_Q10[ length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
+        silk_memcpy( psCNG->CNG_synth_state, &CNG_sig_Q14[ length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
     } else {
         silk_memset( psCNG->CNG_synth_state, 0, psDec->LPC_order *  sizeof( opus_int32 ) );
     }