[Opus] Update to v1.2.1

third_party/opus/src/silk/float/encode_frame_FLP.c

 /***********************************************************************
 Copyright (c) 2006-2011, Skype Limited. All rights reserved.
 Redistribution and use in source and binary forms, with or without
 modification, are permitted provided that the following conditions
 are met:
 - Redistributions of source code must retain the above copyright notice,
 this list of conditions and the following disclaimer.
 - Redistributions in binary form must reproduce the above copyright
 notice, this list of conditions and the following disclaimer in the
 documentation and/or other materials provided with the distribution.
@@ skipping 11 matching lines @@
 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 POSSIBILITY OF SUCH DAMAGE.
 ***********************************************************************/
 
 #ifdef HAVE_CONFIG_H
 #include "config.h"
 #endif
 
+#include <stdlib.h>
 #include "main_FLP.h"
 #include "tuning_parameters.h"
 
 /* Low Bitrate Redundancy (LBRR) encoding. Reuse all parameters but encode with lower bitrate */
 static OPUS_INLINE void silk_LBRR_encode_FLP(
     silk_encoder_state_FLP          *psEnc,                             /* I/O  Encoder state FLP                           */
     silk_encoder_control_FLP        *psEncCtrl,                         /* I/O  Encoder control FLP                         */
     const silk_float                xfw[],                              /* I    Input signal                                */
     opus_int                        condCoding                          /* I    The type of conditional coding used so far for this frame */
 );
@@ skipping 36 matching lines @@
     opus_int32                      *pnBytesOut,                        /* O    Number of payload bytes;                    */
     ec_enc                          *psRangeEnc,                        /* I/O  compressor data structure                   */
     opus_int                        condCoding,                         /* I    The type of conditional coding to use       */
     opus_int                        maxBits,                            /* I    If > 0: maximum number of output bits       */
     opus_int                        useCBR                              /* I    Flag to force constant-bitrate operation    */
 )
 {
     silk_encoder_control_FLP sEncCtrl;
     opus_int     i, iter, maxIter, found_upper, found_lower, ret = 0;
     silk_float   *x_frame, *res_pitch_frame;
-    silk_float   xfw[ MAX_FRAME_LENGTH ];
     silk_float   res_pitch[ 2 * MAX_FRAME_LENGTH + LA_PITCH_MAX ];
     ec_enc       sRangeEnc_copy, sRangeEnc_copy2;
     silk_nsq_state sNSQ_copy, sNSQ_copy2;
     opus_int32   seed_copy, nBits, nBits_lower, nBits_upper, gainMult_lower, gainMult_upper;
     opus_int32   gainsID, gainsID_lower, gainsID_upper;
     opus_int16   gainMult_Q8;
     opus_int16   ec_prevLagIndex_copy;
     opus_int     ec_prevSignalType_copy;
     opus_int8    LastGainIndex_copy2;
     opus_int32   pGains_Q16[ MAX_NB_SUBFR ];
     opus_uint8   ec_buf_copy[ 1275 ];
+    opus_int     gain_lock[ MAX_NB_SUBFR ] = {0};
+    opus_int16   best_gain_mult[ MAX_NB_SUBFR ];
+    opus_int     best_sum[ MAX_NB_SUBFR ];
 
     /* This is totally unnecessary but many compilers (including gcc) are too dumb to realise it */
     LastGainIndex_copy2 = nBits_lower = nBits_upper = gainMult_lower = gainMult_upper = 0;
 
     psEnc->sCmn.indices.Seed = psEnc->sCmn.frameCounter++ & 3;
 
     /**************************************************************/
     /* Set up Input Pointers, and insert frame in input buffer    */
     /**************************************************************/
     /* pointers aligned with start of frame to encode */
@@ skipping 22 matching lines @@
         silk_find_pitch_lags_FLP( psEnc, &sEncCtrl, res_pitch, x_frame, psEnc->sCmn.arch );
 
         /************************/
         /* Noise shape analysis */
         /************************/
         silk_noise_shape_analysis_FLP( psEnc, &sEncCtrl, res_pitch_frame, x_frame );
 
         /***************************************************/
         /* Find linear prediction coefficients (LPC + LTP) */
         /***************************************************/
-        silk_find_pred_coefs_FLP( psEnc, &sEncCtrl, res_pitch, x_frame, condCoding );
+        silk_find_pred_coefs_FLP( psEnc, &sEncCtrl, res_pitch_frame, x_frame, condCoding );
 
         /****************************************/
         /* Process gains                        */
         /****************************************/
         silk_process_gains_FLP( psEnc, &sEncCtrl, condCoding );
 
-        /*****************************************/
-        /* Prefiltering for noise shaper         */
-        /*****************************************/
-        silk_prefilter_FLP( psEnc, &sEncCtrl, xfw, x_frame );
-
         /****************************************/
         /* Low Bitrate Redundant Encoding       */
         /****************************************/
-        silk_LBRR_encode_FLP( psEnc, &sEncCtrl, xfw, condCoding );
+        silk_LBRR_encode_FLP( psEnc, &sEncCtrl, x_frame, condCoding );
 
         /* Loop over quantizer and entroy coding to control bitrate */
         maxIter = 6;
         gainMult_Q8 = SILK_FIX_CONST( 1, 8 );
         found_lower = 0;
         found_upper = 0;
         gainsID = silk_gains_ID( psEnc->sCmn.indices.GainsIndices, psEnc->sCmn.nb_subfr );
         gainsID_lower = -1;
         gainsID_upper = -1;
         /* Copy part of the input state */
@@ skipping 13 matching lines @@
                     silk_memcpy( psRangeEnc, &sRangeEnc_copy, sizeof( ec_enc ) );
                     silk_memcpy( &psEnc->sCmn.sNSQ, &sNSQ_copy, sizeof( silk_nsq_state ) );
                     psEnc->sCmn.indices.Seed = seed_copy;
                     psEnc->sCmn.ec_prevLagIndex = ec_prevLagIndex_copy;
                     psEnc->sCmn.ec_prevSignalType = ec_prevSignalType_copy;
                 }
 
                 /*****************************************/
                 /* Noise shaping quantization            */
                 /*****************************************/
-                silk_NSQ_wrapper_FLP( psEnc, &sEncCtrl, &psEnc->sCmn.indices, &psEnc->sCmn.sNSQ, psEnc->sCmn.pulses, xfw );
+                silk_NSQ_wrapper_FLP( psEnc, &sEncCtrl, &psEnc->sCmn.indices, &psEnc->sCmn.sNSQ, psEnc->sCmn.pulses, x_frame );
+
+                if ( iter == maxIter && !found_lower ) {
+                    silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) );
+                }
 
                 /****************************************/
                 /* Encode Parameters                    */
                 /****************************************/
                 silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding );
 
                 /****************************************/
                 /* Encode Excitation Signal             */
                 /****************************************/
                 silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType,
                       psEnc->sCmn.pulses, psEnc->sCmn.frame_length );
 
                 nBits = ec_tell( psRangeEnc );
 
+                /* If we still bust after the last iteration, do some damage control. */
+                if ( iter == maxIter && !found_lower && nBits > maxBits ) {
+                    silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) );
+
+                    /* Keep gains the same as the last frame. */
+                    psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev;
+                    for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
+                        psEnc->sCmn.indices.GainsIndices[ i ] = 4;
+                    }
+                    if (condCoding != CODE_CONDITIONALLY) {
+                       psEnc->sCmn.indices.GainsIndices[ 0 ] = sEncCtrl.lastGainIndexPrev;
+                    }
+                    psEnc->sCmn.ec_prevLagIndex = ec_prevLagIndex_copy;
+                    psEnc->sCmn.ec_prevSignalType = ec_prevSignalType_copy;
+                    /* Clear all pulses. */
+                    for ( i = 0; i < psEnc->sCmn.frame_length; i++ ) {
+                        psEnc->sCmn.pulses[ i ] = 0;
+                    }
+
+                    silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding );
+
+                    silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType,
+                        psEnc->sCmn.pulses, psEnc->sCmn.frame_length );
+
+                    nBits = ec_tell( psRangeEnc );
+                }
+
                 if( useCBR == 0 && iter == 0 && nBits <= maxBits ) {
                     break;
                 }
             }
 
             if( iter == maxIter ) {
                 if( found_lower && ( gainsID == gainsID_lower || nBits > maxBits ) ) {
                     /* Restore output state from earlier iteration that did meet the bitrate budget */
                     silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) );
                     silk_assert( sRangeEnc_copy2.offs <= 1275 );
                     silk_memcpy( psRangeEnc->buf, ec_buf_copy, sRangeEnc_copy2.offs );
                     silk_memcpy( &psEnc->sCmn.sNSQ, &sNSQ_copy2, sizeof( silk_nsq_state ) );
                     psEnc->sShape.LastGainIndex = LastGainIndex_copy2;
                 }
                 break;
             }
 
             if( nBits > maxBits ) {
                 if( found_lower == 0 && iter >= 2 ) {
                     /* Adjust the quantizer's rate/distortion tradeoff and discard previous "upper" results */
-                    sEncCtrl.Lambda *= 1.5f;
+                    sEncCtrl.Lambda = silk_max_float(sEncCtrl.Lambda*1.5f, 1.5f);
+                    /* Reducing dithering can help us hit the target. */
+                    psEnc->sCmn.indices.quantOffsetType = 0;
                     found_upper = 0;
                     gainsID_upper = -1;
                 } else {
                     found_upper = 1;
                     nBits_upper = nBits;
                     gainMult_upper = gainMult_Q8;
                     gainsID_upper = gainsID;
                 }
             } else if( nBits < maxBits - 5 ) {
                 found_lower = 1;
                 nBits_lower = nBits;
                 gainMult_lower = gainMult_Q8;
                 if( gainsID != gainsID_lower ) {
                     gainsID_lower = gainsID;
                     /* Copy part of the output state */
                     silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) );
                     silk_assert( psRangeEnc->offs <= 1275 );
                     silk_memcpy( ec_buf_copy, psRangeEnc->buf, psRangeEnc->offs );
                     silk_memcpy( &sNSQ_copy2, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) );
                     LastGainIndex_copy2 = psEnc->sShape.LastGainIndex;
                 }
             } else {
                 /* Within 5 bits of budget: close enough */
                 break;
             }
 
+            if ( !found_lower && nBits > maxBits ) {
+                int j;
+                for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
+                    int sum=0;
+                    for ( j = i*psEnc->sCmn.subfr_length; j < (i+1)*psEnc->sCmn.subfr_length; j++ ) {
+                        sum += abs( psEnc->sCmn.pulses[j] );
+                    }
+                    if ( iter == 0 || (sum < best_sum[i] && !gain_lock[i]) ) {
+                        best_sum[i] = sum;
+                        best_gain_mult[i] = gainMult_Q8;
+                    } else {
+                        gain_lock[i] = 1;
+                    }
+                }
+            }
             if( ( found_lower & found_upper ) == 0 ) {
                 /* Adjust gain according to high-rate rate/distortion curve */
-                opus_int32 gain_factor_Q16;
-                gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) );
-                gain_factor_Q16 = silk_min_32( gain_factor_Q16, SILK_FIX_CONST( 2, 16 ) );
                 if( nBits > maxBits ) {
-                    gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) );
+                    if (gainMult_Q8 < 16384) {
+                        gainMult_Q8 *= 2;
+                    } else {
+                        gainMult_Q8 = 32767;
+                    }
+                } else {
+                    opus_int32 gain_factor_Q16;
+                    gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) );
+                    gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 );
                 }
-                gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 );
             } else {
                 /* Adjust gain by interpolating */
                 gainMult_Q8 = gainMult_lower + ( ( gainMult_upper - gainMult_lower ) * ( maxBits - nBits_lower ) ) / ( nBits_upper - nBits_lower );
                 /* New gain multplier must be between 25% and 75% of old range (note that gainMult_upper < gainMult_lower) */
                 if( gainMult_Q8 > silk_ADD_RSHIFT32( gainMult_lower, gainMult_upper - gainMult_lower, 2 ) ) {
                     gainMult_Q8 = silk_ADD_RSHIFT32( gainMult_lower, gainMult_upper - gainMult_lower, 2 );
                 } else
                 if( gainMult_Q8 < silk_SUB_RSHIFT32( gainMult_upper, gainMult_upper - gainMult_lower, 2 ) ) {
                     gainMult_Q8 = silk_SUB_RSHIFT32( gainMult_upper, gainMult_upper - gainMult_lower, 2 );
                 }
             }
 
             for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
-                pGains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q8 ), 8 );
+                opus_int16 tmp;
+                if ( gain_lock[i] ) {
+                    tmp = best_gain_mult[i];
+                } else {
+                    tmp = gainMult_Q8;
+                }
+                pGains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], tmp ), 8 );
             }
 
             /* Quantize gains */
             psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev;
             silk_gains_quant( psEnc->sCmn.indices.GainsIndices, pGains_Q16,
                   &psEnc->sShape.LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr );
 
             /* Unique identifier of gains vector */
             gainsID = silk_gains_ID( psEnc->sCmn.indices.GainsIndices, psEnc->sCmn.nb_subfr );
 
@@ skipping 77 matching lines @@
         /*****************************************/
         /* Noise shaping quantization            */
         /*****************************************/
         silk_NSQ_wrapper_FLP( psEnc, psEncCtrl, psIndices_LBRR, &sNSQ_LBRR,
             psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], xfw );
 
         /* Restore original gains */
         silk_memcpy( psEncCtrl->Gains, TempGains, psEnc->sCmn.nb_subfr * sizeof( silk_float ) );
     }
 }