NLSF decoding now uses tables for NLSF weights

8 files changed, 109 insertions, 21 deletions

diff --git a/silk/A2NLSF.c b/silk/A2NLSF.c
index b6e9e5ff..a1edde1e 100644
--- a/silk/A2NLSF.c
+++ b/silk/A2NLSF.c
@@ -130,7 +130,7 @@ void silk_A2NLSF(
     const opus_int              d                   /* I    Filter order (must be even)                                 */
 )
 {
-    opus_int      i, k, m, dd, root_ix, ffrac;
+    opus_int   i, k, m, dd, root_ix, ffrac;
     opus_int32 xlo, xhi, xmid;
     opus_int32 ylo, yhi, ymid, thr;
     opus_int32 nom, den;
diff --git a/silk/NLSF_decode.c b/silk/NLSF_decode.c
index 9f715060..93c86539 100644
--- a/silk/NLSF_decode.c
+++ b/silk/NLSF_decode.c
@@ -60,25 +60,20 @@ static OPUS_INLINE void silk_NLSF_residual_dequant(               /* O    Return
 /***********************/
 /* NLSF vector decoder */
 /***********************/
-void silk_NLSF_decode(
+opus_int silk_NLSF_decode(                                      /* O    Number of bits (Q5), if signalType >= 0     */
           opus_int16            *pNLSF_Q15,                     /* O    Quantized NLSF vector [ LPC_ORDER ]         */
           opus_int8             *NLSFIndices,                   /* I    Codebook path vector [ LPC_ORDER + 1 ]      */
-    const silk_NLSF_CB_struct   *psNLSF_CB                      /* I    Codebook object                             */
+    const silk_NLSF_CB_struct   *psNLSF_CB,                     /* I    Codebook object                             */
+    const opus_int              signalType                      /* I    SignalType, to determine number of bits     */
 )
 {
     opus_int         i;
     opus_uint8       pred_Q8[  MAX_LPC_ORDER ];
     opus_int16       ec_ix[    MAX_LPC_ORDER ];
     opus_int16       res_Q10[  MAX_LPC_ORDER ];
-    opus_int16       W_tmp_QW[ MAX_LPC_ORDER ];
-    opus_int32       W_tmp_Q9, NLSF_Q15_tmp;
+    opus_int32       NLSF_Q15_tmp;
     const opus_uint8 *pCB_element;
-
-    /* Decode first stage */
-    pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ NLSFIndices[ 0 ] * psNLSF_CB->order ];
-    for( i = 0; i < psNLSF_CB->order; i++ ) {
-        pNLSF_Q15[ i ] = silk_LSHIFT( (opus_int16)pCB_element[ i ], 7 );
-    }
+    const opus_int16 *pCB_Wght_Q9;
 
     /* Unpack entropy table indices and predictor for current CB1 index */
     silk_NLSF_unpack( ec_ix, pred_Q8, psNLSF_CB, NLSFIndices[ 0 ] );
@@ -86,16 +81,35 @@ void silk_NLSF_decode(
     /* Predictive residual dequantizer */
     silk_NLSF_residual_dequant( res_Q10, &NLSFIndices[ 1 ], pred_Q8, psNLSF_CB->quantStepSize_Q16, psNLSF_CB->order );
 
-    /* Weights from codebook vector */
-    silk_NLSF_VQ_weights_laroia( W_tmp_QW, pNLSF_Q15, psNLSF_CB->order );
-
-    /* Apply inverse square-rooted weights and add to output */
+    /* Apply inverse square-rooted weights to first stage and add to output */
+    pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ NLSFIndices[ 0 ] * psNLSF_CB->order ];
+    pCB_Wght_Q9 = &psNLSF_CB->CB1_Wght_Q9[ NLSFIndices[ 0 ] * psNLSF_CB->order ];
     for( i = 0; i < psNLSF_CB->order; i++ ) {
-        W_tmp_Q9 = silk_SQRT_APPROX( silk_LSHIFT( (opus_int32)W_tmp_QW[ i ], 18 - NLSF_W_Q ) );
-        NLSF_Q15_tmp = silk_ADD32( pNLSF_Q15[ i ], silk_DIV32_16( silk_LSHIFT( (opus_int32)res_Q10[ i ], 14 ), W_tmp_Q9 ) );
+        NLSF_Q15_tmp = silk_ADD_LSHIFT32( silk_DIV32_16( silk_LSHIFT( (opus_int32)res_Q10[ i ], 14 ), pCB_Wght_Q9[ i ] ), (opus_int16)pCB_element[ i ], 7 );
         pNLSF_Q15[ i ] = (opus_int16)silk_LIMIT( NLSF_Q15_tmp, 0, 32767 );
     }
 
     /* NLSF stabilization */
     silk_NLSF_stabilize( pNLSF_Q15, psNLSF_CB->deltaMin_Q15, psNLSF_CB->order );
+
+    if( signalType >= 0 ) {
+        opus_int         prob_Q8, bits_Q5;
+        const opus_uint8 *iCDF_ptr;
+        bits_Q5 = 0;
+        iCDF_ptr = &psNLSF_CB->CB1_iCDF[ ( signalType >> 1 ) * psNLSF_CB->nVectors ];
+        if( NLSFIndices[ 0 ] == 0 ) {
+            prob_Q8 = 256 - iCDF_ptr[ NLSFIndices[ 0 ] ];
+        } else {
+            prob_Q8 = iCDF_ptr[ NLSFIndices[ 0 ] - 1 ] - iCDF_ptr[ NLSFIndices[ 0 ] ];
+        }
+        bits_Q5 = ( 8 << 5 ) - ( silk_lin2log( prob_Q8 ) >> 2 );
+        for( i = 0; i < psNLSF_CB->order; i++ ) {
+            const opus_uint8 *rates_Q5;
+            rates_Q5 = &psNLSF_CB->ec_Rates_Q5[ ec_ix[ i ] ];
+            bits_Q5 += rates_Q5[ NLSFIndices[ i + 1 ] + NLSF_QUANT_MAX_AMPLITUDE ];
+        }
+        return bits_Q5;
+    }
+
+    return 0;
 }
diff --git a/silk/NLSF_encode.c b/silk/NLSF_encode.c
index f03c3f1c..9bd48cbe 100644
--- a/silk/NLSF_encode.c
+++ b/silk/NLSF_encode.c
@@ -129,7 +129,7 @@ opus_int32 silk_NLSF_encode(                                    /* O    Returns
     silk_memcpy( &NLSFIndices[ 1 ], &tempIndices2[ bestIndex * MAX_LPC_ORDER ], psNLSF_CB->order * sizeof( opus_int8 ) );
 
     /* Decode */
-    silk_NLSF_decode( pNLSF_Q15, NLSFIndices, psNLSF_CB );
+    silk_NLSF_decode( pNLSF_Q15, NLSFIndices, psNLSF_CB, -1 );
 
     ret = RD_Q25[ 0 ];
     RESTORE_STACK;
diff --git a/silk/decode_parameters.c b/silk/decode_parameters.c
index e345b1dc..07d6d6f4 100644
--- a/silk/decode_parameters.c
+++ b/silk/decode_parameters.c
@@ -49,7 +49,7 @@ void silk_decode_parameters(
     /****************/
     /* Decode NLSFs */
     /****************/
-    silk_NLSF_decode( pNLSF_Q15, psDec->indices.NLSFIndices, psDec->psNLSF_CB );
+    silk_NLSF_decode( pNLSF_Q15, psDec->indices.NLSFIndices, psDec->psNLSF_CB, -1 );
 
     /* Convert NLSF parameters to AR prediction filter coefficients */
     silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 1 ], pNLSF_Q15, psDec->LPC_order );
diff --git a/silk/main.h b/silk/main.h
index 2f90d68f..60153359 100644
--- a/silk/main.h
+++ b/silk/main.h
@@ -375,10 +375,11 @@ void silk_NLSF_unpack(
 /***********************/
 /* NLSF vector decoder */
 /***********************/
-void silk_NLSF_decode(
+opus_int silk_NLSF_decode(                                      /* O    Number of bits (Q5), if signalType >= 0     */
           opus_int16            *pNLSF_Q15,                     /* O    Quantized NLSF vector [ LPC_ORDER ]         */
           opus_int8             *NLSFIndices,                   /* I    Codebook path vector [ LPC_ORDER + 1 ]      */
-    const silk_NLSF_CB_struct   *psNLSF_CB                      /* I    Codebook object                             */
+    const silk_NLSF_CB_struct   *psNLSF_CB,                     /* I    Codebook object                             */
+    const opus_int              signalType                      /* I    SignalType, to determine number of bits     */
 );
 
 /****************************************************/
diff --git a/silk/structs.h b/silk/structs.h
index 827829dc..1a410d0d 100644
--- a/silk/structs.h
+++ b/silk/structs.h
@@ -86,6 +86,7 @@ typedef struct {
     const opus_int16             quantStepSize_Q16;
     const opus_int16             invQuantStepSize_Q6;
     const opus_uint8             *CB1_NLSF_Q8;
+    const opus_int16             *CB1_Wght_Q9;
     const opus_uint8             *CB1_iCDF;
     const opus_uint8             *pred_Q8;
     const opus_uint8             *ec_sel;
diff --git a/silk/tables_NLSF_CB_NB_MB.c b/silk/tables_NLSF_CB_NB_MB.c
index 8c59d207..195d5b95 100644
--- a/silk/tables_NLSF_CB_NB_MB.c
+++ b/silk/tables_NLSF_CB_NB_MB.c
@@ -74,6 +74,41 @@ static const opus_uint8 silk_NLSF_CB1_NB_MB_Q8[ 320 ] = {
         64,     84,    104,    118,    156,    177,    201,    230
 };
 
+static const opus_int16 silk_NLSF_CB1_Wght_Q9[ 320 ] = {
+     2897, 2314, 2314, 2314, 2287, 2287, 2314, 2300, 2327, 2287,
+     2888, 2580, 2394, 2367, 2314, 2274, 2274, 2274, 2274, 2194,
+     2487, 2340, 2340, 2314, 2314, 2314, 2340, 2340, 2367, 2354,
+     3216, 2766, 2340, 2340, 2314, 2274, 2221, 2207, 2261, 2194,
+     2460, 2474, 2367, 2394, 2394, 2394, 2394, 2367, 2407, 2314,
+     3479, 3056, 2127, 2207, 2274, 2274, 2274, 2287, 2314, 2261,
+     3282, 3141, 2580, 2394, 2247, 2221, 2207, 2194, 2194, 2114,
+     4096, 3845, 2221, 2620, 2620, 2407, 2314, 2394, 2367, 2074,
+     3178, 3244, 2367, 2221, 2553, 2434, 2340, 2314, 2167, 2221,
+     3338, 3488, 2726, 2194, 2261, 2460, 2354, 2367, 2207, 2101,
+     2354, 2420, 2327, 2367, 2394, 2420, 2420, 2420, 2460, 2367,
+     3779, 3629, 2434, 2527, 2367, 2274, 2274, 2300, 2207, 2048,
+     3254, 3225, 2713, 2846, 2447, 2327, 2300, 2300, 2274, 2127,
+     3263, 3300, 2753, 2806, 2447, 2261, 2261, 2247, 2127, 2101,
+     2873, 2981, 2633, 2367, 2407, 2354, 2194, 2247, 2247, 2114,
+     3225, 3197, 2633, 2580, 2274, 2181, 2247, 2221, 2221, 2141,
+     3178, 3310, 2740, 2407, 2274, 2274, 2274, 2287, 2194, 2114,
+     3141, 3272, 2460, 2061, 2287, 2500, 2367, 2487, 2434, 2181,
+     3507, 3282, 2314, 2700, 2647, 2474, 2367, 2394, 2340, 2127,
+     3423, 3535, 3038, 3056, 2300, 1950, 2221, 2274, 2274, 2274,
+     3404, 3366, 2087, 2687, 2873, 2354, 2420, 2274, 2474, 2540,
+     3760, 3488, 1950, 2660, 2897, 2527, 2394, 2367, 2460, 2261,
+     3028, 3272, 2740, 2888, 2740, 2154, 2127, 2287, 2234, 2247,
+     3695, 3657, 2025, 1969, 2660, 2700, 2580, 2500, 2327, 2367,
+     3207, 3413, 2354, 2074, 2888, 2888, 2340, 2487, 2247, 2167,
+     3338, 3366, 2846, 2780, 2327, 2154, 2274, 2287, 2114, 2061,
+     2327, 2300, 2181, 2167, 2181, 2367, 2633, 2700, 2700, 2553,
+     2407, 2434, 2221, 2261, 2221, 2221, 2340, 2420, 2607, 2700,
+     3038, 3244, 2806, 2888, 2474, 2074, 2300, 2314, 2354, 2380,
+     2221, 2154, 2127, 2287, 2500, 2793, 2793, 2620, 2580, 2367,
+     3676, 3713, 2234, 1838, 2181, 2753, 2726, 2673, 2513, 2207,
+     2793, 3160, 2726, 2553, 2846, 2513, 2181, 2394, 2221, 2181
+};
+
 static const opus_uint8 silk_NLSF_CB1_iCDF_NB_MB[ 64 ] = {
        212,    178,    148,    129,    108,     96,     85,     82,
         79,     77,     61,     59,     57,     56,     51,     49,
@@ -150,6 +185,7 @@ const silk_NLSF_CB_struct silk_NLSF_CB_NB_MB =
     SILK_FIX_CONST( 0.18, 16 ),
     SILK_FIX_CONST( 1.0 / 0.18, 6 ),
     silk_NLSF_CB1_NB_MB_Q8,
+    silk_NLSF_CB1_Wght_Q9,
     silk_NLSF_CB1_iCDF_NB_MB,
     silk_NLSF_PRED_NB_MB_Q8,
     silk_NLSF_CB2_SELECT_NB_MB,
diff --git a/silk/tables_NLSF_CB_WB.c b/silk/tables_NLSF_CB_WB.c
index 50af87eb..a15f7eb4 100644
--- a/silk/tables_NLSF_CB_WB.c
+++ b/silk/tables_NLSF_CB_WB.c
@@ -98,6 +98,41 @@ static const opus_uint8 silk_NLSF_CB1_WB_Q8[ 512 ] = {
        110,    119,    129,    141,    175,    198,    218,    237
 };
 
+static const opus_int16 silk_NLSF_CB1_Wght_Q9[ 512 ] = {
+     3657, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2963, 2963, 2925, 2846,
+     3216, 3085, 2972, 3056, 3056, 3010, 3010, 3010, 2963, 2963, 3010, 2972, 2888, 2846, 2846, 2726,
+     3920, 4014, 2981, 3207, 3207, 2934, 3056, 2846, 3122, 3244, 2925, 2846, 2620, 2553, 2780, 2925,
+     3516, 3197, 3010, 3103, 3019, 2888, 2925, 2925, 2925, 2925, 2888, 2888, 2888, 2888, 2888, 2753,
+     5054, 5054, 2934, 3573, 3385, 3056, 3085, 2793, 3160, 3160, 2972, 2846, 2513, 2540, 2753, 2888,
+     4428, 4149, 2700, 2753, 2972, 3010, 2925, 2846, 2981, 3019, 2925, 2925, 2925, 2925, 2888, 2726,
+     3620, 3019, 2972, 3056, 3056, 2873, 2806, 3056, 3216, 3047, 2981, 3291, 3291, 2981, 3310, 2991,
+     5227, 5014, 2540, 3338, 3526, 3385, 3197, 3094, 3376, 2981, 2700, 2647, 2687, 2793, 2846, 2673,
+     5081, 5174, 4615, 4428, 2460, 2897, 3047, 3207, 3169, 2687, 2740, 2888, 2846, 2793, 2846, 2700,
+     3122, 2888, 2963, 2925, 2925, 2925, 2925, 2963, 2963, 2963, 2963, 2925, 2925, 2963, 2963, 2963,
+     4202, 3207, 2981, 3103, 3010, 2888, 2888, 2925, 2972, 2873, 2916, 3019, 2972, 3010, 3197, 2873,
+     3760, 3760, 3244, 3103, 2981, 2888, 2925, 2888, 2972, 2934, 2793, 2793, 2846, 2888, 2888, 2660,
+     3854, 4014, 3207, 3122, 3244, 2934, 3047, 2963, 2963, 3085, 2846, 2793, 2793, 2793, 2793, 2580,
+     3845, 4080, 3357, 3516, 3094, 2740, 3010, 2934, 3122, 3085, 2846, 2846, 2647, 2647, 2846, 2806,
+     5147, 4894, 3225, 3845, 3441, 3169, 2897, 3413, 3451, 2700, 2580, 2673, 2740, 2846, 2806, 2753,
+     4109, 3789, 3291, 3160, 2925, 2888, 2888, 2925, 2793, 2740, 2793, 2740, 2793, 2846, 2888, 2806,
+     5081, 5054, 3047, 3545, 3244, 3056, 3085, 2944, 3103, 2897, 2740, 2740, 2740, 2846, 2793, 2620,
+     4309, 4309, 2860, 2527, 3207, 3376, 3376, 3075, 3075, 3376, 3056, 2846, 2647, 2580, 2726, 2753,
+     3056, 2916, 2806, 2888, 2740, 2687, 2897, 3103, 3150, 3150, 3216, 3169, 3056, 3010, 2963, 2846,
+     4375, 3882, 2925, 2888, 2846, 2888, 2846, 2846, 2888, 2888, 2888, 2846, 2888, 2925, 2888, 2846,
+     2981, 2916, 2916, 2981, 2981, 3056, 3122, 3216, 3150, 3056, 3010, 2972, 2972, 2972, 2925, 2740,
+     4229, 4149, 3310, 3347, 2925, 2963, 2888, 2981, 2981, 2846, 2793, 2740, 2846, 2846, 2846, 2793,
+     4080, 4014, 3103, 3010, 2925, 2925, 2925, 2888, 2925, 2925, 2846, 2846, 2846, 2793, 2888, 2780,
+     4615, 4575, 3169, 3441, 3207, 2981, 2897, 3038, 3122, 2740, 2687, 2687, 2687, 2740, 2793, 2700,
+     4149, 4269, 3789, 3657, 2726, 2780, 2888, 2888, 3010, 2972, 2925, 2846, 2687, 2687, 2793, 2888,
+     4215, 3554, 2753, 2846, 2846, 2888, 2888, 2888, 2925, 2925, 2888, 2925, 2925, 2925, 2963, 2888,
+     5174, 4921, 2261, 3432, 3789, 3479, 3347, 2846, 3310, 3479, 3150, 2897, 2460, 2487, 2753, 2925,
+     3451, 3685, 3122, 3197, 3357, 3047, 3207, 3207, 2981, 3216, 3085, 2925, 2925, 2687, 2540, 2434,
+     2981, 3010, 2793, 2793, 2740, 2793, 2846, 2972, 3056, 3103, 3150, 3150, 3150, 3103, 3010, 3010,
+     2944, 2873, 2687, 2726, 2780, 3010, 3432, 3545, 3357, 3244, 3056, 3010, 2963, 2925, 2888, 2846,
+     3019, 2944, 2897, 3010, 3010, 2972, 3019, 3103, 3056, 3056, 3010, 2888, 2846, 2925, 2925, 2888,
+     3920, 3967, 3010, 3197, 3357, 3216, 3291, 3291, 3479, 3704, 3441, 2726, 2181, 2460, 2580, 2607
+};
+
 static const opus_uint8 silk_NLSF_CB1_iCDF_WB[ 64 ] = {
        225,    204,    201,    184,    183,    175,    158,    154,
        153,    135,    119,    115,    113,    110,    109,     99,
@@ -188,6 +223,7 @@ const silk_NLSF_CB_struct silk_NLSF_CB_WB =
     SILK_FIX_CONST( 0.15, 16 ),
     SILK_FIX_CONST( 1.0 / 0.15, 6 ),
     silk_NLSF_CB1_WB_Q8,
+    silk_NLSF_CB1_Wght_Q9,
     silk_NLSF_CB1_iCDF_WB,
     silk_NLSF_PRED_WB_Q8,
     silk_NLSF_CB2_SELECT_WB,