From 294bfec27b82f879e3c3004d31bb91bcb34014f4 Mon Sep 17 00:00:00 2001 From: Jean-Marc Valin Date: Thu, 20 Oct 2011 00:39:41 -0400 Subject: Implements hard CBR for SILK This is achieved by running the encoding process in a loop and padding when we don't reach the exact rate. It also implements VBR-with-cap, which means we no longer need to artificially decrease the SILK bandwidth when it's close to the cap. --- silk/API.h | 4 +- silk/SigProc_FIX.h | 2 +- silk/biquad_alt.c | 6 +- silk/control.h | 3 + silk/control_codec.c | 2 +- silk/enc_API.c | 39 ++++++-- silk/fixed/LTP_scale_ctrl_FIX.c | 1 - silk/fixed/encode_frame_FIX.c | 166 ++++++++++++++++++++++++++++------- silk/fixed/find_pred_coefs_FIX.c | 2 +- silk/fixed/main_FIX.h | 4 +- silk/fixed/process_gains_FIX.c | 4 + silk/fixed/structs_FIX.h | 3 + silk/float/encode_frame_FLP.c | 161 ++++++++++++++++++++++++++++----- silk/float/find_pred_coefs_FLP.c | 2 +- silk/float/inner_product_FLP.c | 12 +-- silk/float/main_FLP.h | 4 +- silk/float/pitch_analysis_core_FLP.c | 6 +- silk/float/process_gains_FLP.c | 4 + silk/float/silk_float.vcxproj | 2 +- silk/float/structs_FLP.h | 4 + silk/process_NLSFs.c | 4 +- silk/structs.h | 2 +- src/opus_decoder.c | 2 +- src/opus_encoder.c | 93 ++++++++++++++++++-- src/test_opus.c | 4 +- 25 files changed, 442 insertions(+), 94 deletions(-) diff --git a/silk/API.h b/silk/API.h index ca24808f..a96bd543 100644 --- a/silk/API.h +++ b/silk/API.h @@ -56,7 +56,7 @@ typedef struct { /* Get size in bytes of the Silk encoder state */ /***********************************************/ opus_int silk_Get_Encoder_Size( /* O: Returns error code */ - int *encSizeBytes /* O: Number of bytes in SILK encoder state */ + opus_int *encSizeBytes /* O: Number of bytes in SILK encoder state */ ); /*************************/ @@ -98,7 +98,7 @@ opus_int silk_Encode( /* O: Returns error c /* Get size in bytes of the Silk decoder state */ /***********************************************/ opus_int silk_Get_Decoder_Size( /* O: Returns error code */ - int *decSizeBytes /* O: Number of bytes in SILK decoder state */ + opus_int *decSizeBytes /* O: Number of bytes in SILK decoder state */ ); /*************************/ diff --git a/silk/SigProc_FIX.h b/silk/SigProc_FIX.h index f9e2664a..24052ff9 100644 --- a/silk/SigProc_FIX.h +++ b/silk/SigProc_FIX.h @@ -99,7 +99,7 @@ void silk_biquad_alt( opus_int32 *S, /* I/O: State vector [2] */ opus_int16 *out, /* O: output signal */ const opus_int32 len, /* I: signal length (must be even) */ - int stride + opus_int stride /* I: Operate on interleaved signal if > 1 */ ); /* Variable order MA prediction error filter. */ diff --git a/silk/biquad_alt.c b/silk/biquad_alt.c index 617b33ab..74a66a5c 100644 --- a/silk/biquad_alt.c +++ b/silk/biquad_alt.c @@ -47,7 +47,7 @@ void silk_biquad_alt( opus_int32 *S, /* I/O: State vector [2] */ opus_int16 *out, /* O: Output signal */ const opus_int32 len, /* I: Signal length (must be even) */ - int stride + opus_int stride /* I: Operate on interleaved signal if > 1 */ ) { /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */ @@ -62,7 +62,7 @@ void silk_biquad_alt( for( k = 0; k < len; k++ ) { /* S[ 0 ], S[ 1 ]: Q12 */ - inval = in[ k*stride ]; + inval = in[ k * stride ]; out32_Q14 = silk_LSHIFT( silk_SMLAWB( S[ 0 ], B_Q28[ 0 ], inval ), 2 ); S[ 0 ] = S[1] + silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14, A0_L_Q28 ), 14 ); @@ -74,6 +74,6 @@ void silk_biquad_alt( S[ 1 ] = silk_SMLAWB( S[ 1 ], B_Q28[ 2 ], inval ); /* Scale back to Q0 and saturate */ - out[ k*stride ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14 + (1<<14) - 1, 14 ) ); + out[ k * stride ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14 + (1<<14) - 1, 14 ) ); } } diff --git a/silk/control.h b/silk/control.h index ac1b288e..c2f906b8 100644 --- a/silk/control.h +++ b/silk/control.h @@ -83,6 +83,9 @@ typedef struct { /* I: Flag to use constant bitrate */ opus_int useCBR; + /* I: Maximum number of bits allowed for the frame */ + opus_int maxBits; + /* I: Causes a smooth downmix to mono */ opus_int toMono; diff --git a/silk/control_codec.c b/silk/control_codec.c index 796c3ef6..a942336c 100644 --- a/silk/control_codec.c +++ b/silk/control_codec.c @@ -38,7 +38,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #include "tuning_parameters.h" -static const int enc_delay_matrix[3][5] = { +static const opus_int enc_delay_matrix[3][5] = { /*SILK API 8 12 16 24 48 */ /* 8 */ {5, 0, 3, 4, 8}, /*12 */ {0, 6, 0, 0, 0}, diff --git a/silk/enc_API.c b/silk/enc_API.c index fb4b717e..149befa6 100644 --- a/silk/enc_API.c +++ b/silk/enc_API.c @@ -44,7 +44,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. /* Encoder functions */ /****************************************/ -opus_int silk_Get_Encoder_Size( int *encSizeBytes ) +opus_int silk_Get_Encoder_Size( opus_int *encSizeBytes ) { opus_int ret = SILK_NO_ERROR; @@ -139,7 +139,7 @@ opus_int silk_Encode( opus_int32 TargetRate_bps, MStargetRates_bps[ 2 ], channelRate_bps, LBRR_symbol; silk_encoder *psEnc = ( silk_encoder * )encState; opus_int16 buf[ MAX_FRAME_LENGTH_MS * MAX_API_FS_KHZ + MAX_ENCODER_DELAY]; - opus_int transition, delay; + opus_int transition, delay, curr_block, tot_blocks; psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded = psEnc->state_Fxx[ 1 ].sCmn.nFramesEncoded = 0; @@ -172,6 +172,8 @@ opus_int silk_Encode( psEnc->nChannelsInternal = encControl->nChannelsInternal; nBlocksOf10ms = silk_DIV32( 100 * nSamplesIn, encControl->API_sampleRate ); + tot_blocks = ( nBlocksOf10ms > 1 ) ? nBlocksOf10ms >> 1 : 1; + curr_block = 0; if( prefillFlag ) { /* Only accept input length of 10 ms */ if( nBlocksOf10ms != 1 ) { @@ -211,13 +213,12 @@ opus_int silk_Encode( TargetRate_bps = silk_RSHIFT32( encControl->bitRate, encControl->nChannelsInternal - 1 ); for( n = 0; n < encControl->nChannelsInternal; n++ ) { /* JMV: Force the side channel to the same rate as the mid. Is this the right way? */ - int force_fs_kHz = (n==1) ? psEnc->state_Fxx[0].sCmn.fs_kHz : 0; + opus_int force_fs_kHz = (n==1) ? psEnc->state_Fxx[0].sCmn.fs_kHz : 0; if( ( ret = silk_control_encoder( &psEnc->state_Fxx[ n ], encControl, TargetRate_bps, psEnc->allowBandwidthSwitch, n, force_fs_kHz ) ) != 0 ) { silk_assert( 0 ); return ret; } - if (psEnc->state_Fxx[n].sCmn.first_frame_after_reset || transition) - { + if( psEnc->state_Fxx[n].sCmn.first_frame_after_reset || transition ) { for( i = 0; i < psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket; i++ ) { psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i ] = 0; } @@ -234,7 +235,7 @@ opus_int silk_Encode( nSamplesFromInput = silk_DIV32_16( nSamplesToBuffer * psEnc->state_Fxx[ 0 ].sCmn.API_fs_Hz, psEnc->state_Fxx[ 0 ].sCmn.fs_kHz * 1000 ); /* Resample and write to buffer */ if( encControl->nChannelsAPI == 2 && encControl->nChannelsInternal == 2 ) { - int id = psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded; + opus_int id = psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded; for( n = 0; n < nSamplesFromInput; n++ ) { buf[ n+delay ] = samplesIn[ 2 * n ]; } @@ -419,10 +420,33 @@ opus_int silk_Encode( /* Encode */ for( n = 0; n < encControl->nChannelsInternal; n++ ) { + opus_int maxBits, useCBR; + + /* Handling rate constraints */ + maxBits = encControl->maxBits; + /*if( encControl->useCBR ) { + maxBits = (encControl->bitRate * nBlocksOf10ms / 800) * 8; + }*/ + if( tot_blocks == 2 && curr_block == 0 ) { + maxBits = maxBits * 3 / 5; + } else if( tot_blocks == 3 ) { + if( curr_block == 0 ) { + maxBits = maxBits * 2 / 5; + } else if( curr_block == 1 ) { + maxBits = maxBits * 3 / 4; + } + } + useCBR = encControl->useCBR && curr_block == tot_blocks - 1; + if( encControl->nChannelsInternal == 1 ) { channelRate_bps = TargetRate_bps; } else { channelRate_bps = MStargetRates_bps[ n ]; + if( n == 0 && MStargetRates_bps[ 1 ] > 0 ) { + useCBR = 0; + /* Give mid up to 1/2 of the max bits for that frame */ + maxBits -= encControl->maxBits / ( tot_blocks * 2 ); + } } if( channelRate_bps > 0 ) { @@ -440,7 +464,7 @@ opus_int silk_Encode( } else { condCoding = CODE_CONDITIONALLY; } - if( ( ret = silk_encode_frame_Fxx( &psEnc->state_Fxx[ n ], nBytesOut, psRangeEnc, condCoding ) ) != 0 ) { + if( ( ret = silk_encode_frame_Fxx( &psEnc->state_Fxx[ n ], nBytesOut, psRangeEnc, condCoding, maxBits, useCBR ) ) != 0 ) { silk_assert( 0 ); } } @@ -492,6 +516,7 @@ opus_int silk_Encode( } else { break; } + curr_block++; } psEnc->nPrevChannelsInternal = encControl->nChannelsInternal; diff --git a/silk/fixed/LTP_scale_ctrl_FIX.c b/silk/fixed/LTP_scale_ctrl_FIX.c index 5e85cc74..dd4923f8 100644 --- a/silk/fixed/LTP_scale_ctrl_FIX.c +++ b/silk/fixed/LTP_scale_ctrl_FIX.c @@ -35,7 +35,6 @@ void silk_LTP_scale_ctrl_FIX( silk_encoder_state_FIX *psEnc, /* I/O encoder state FIX */ silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control FIX */ opus_int condCoding /* I The type of conditional coding to use */ - ) { opus_int round_loss; diff --git a/silk/fixed/encode_frame_FIX.c b/silk/fixed/encode_frame_FIX.c index e351def9..a8633f5a 100644 --- a/silk/fixed/encode_frame_FIX.c +++ b/silk/fixed/encode_frame_FIX.c @@ -79,14 +79,23 @@ opus_int silk_encode_frame_FIX( silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */ 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 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_FIX sEncCtrl; - opus_int ret = 0; + opus_int i, iter, maxIter, found_upper, found_lower, ret = 0; opus_int16 *x_frame, *res_pitch_frame; opus_int16 xfw[ MAX_FRAME_LENGTH ]; opus_int16 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_Q10, gainMult_lower, gainMult_upper; + opus_int16 ec_prevLagIndex_copy; + opus_int ec_prevSignalType_copy; + opus_int8 LastGainIndex_copy2; + opus_uint8 ec_buf_copy[ 512 ]; TIC(ENCODE_FRAME) @@ -151,21 +160,132 @@ TIC(LBRR) silk_LBRR_encode_FIX( psEnc, &sEncCtrl, xfw, condCoding ); TOC(LBRR) - /*****************************************/ - /* Noise shaping quantization */ - /*****************************************/ + /* Loop over quantizer and entroy coding to control bitrate */ + maxIter = 5; + gainMult_Q10 = SILK_FIX_CONST( 1, 10 ); + found_lower = 0; + found_upper = 0; + for( iter = 0; ; iter++ ) { + if( maxBits > 0 && !psEnc->sCmn.prefillFlag ) { + /* Copy part of the input state */ + silk_memcpy( &sRangeEnc_copy, psRangeEnc, sizeof( ec_enc ) ); + silk_memcpy( &sNSQ_copy, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) ); + seed_copy = psEnc->sCmn.indices.Seed; + ec_prevLagIndex_copy = psEnc->sCmn.ec_prevLagIndex; + ec_prevSignalType_copy = psEnc->sCmn.ec_prevSignalType; + } + + /*****************************************/ + /* Noise shaping quantization */ + /*****************************************/ TIC(NSQ) - if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { - silk_NSQ_del_dec( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, xfw, psEnc->sCmn.pulses, - sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR2_Q13, sEncCtrl.HarmShapeGain_Q14, - sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14 ); - } else { - silk_NSQ( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, xfw, psEnc->sCmn.pulses, - sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR2_Q13, sEncCtrl.HarmShapeGain_Q14, - sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14 ); - } + if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { + silk_NSQ_del_dec( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, xfw, psEnc->sCmn.pulses, + sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR2_Q13, sEncCtrl.HarmShapeGain_Q14, + sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14 ); + } else { + silk_NSQ( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, xfw, psEnc->sCmn.pulses, + sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR2_Q13, sEncCtrl.HarmShapeGain_Q14, + sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14 ); + } TOC(NSQ) + if( psEnc->sCmn.prefillFlag ) { + break; + } else { + /****************************************/ + /* Encode Parameters */ + /****************************************/ +TIC(ENCODE_PARAMS) + silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding ); +TOC(ENCODE_PARAMS) + + /****************************************/ + /* Encode Excitation Signal */ + /****************************************/ +TIC(ENCODE_PULSES) + silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType, + psEnc->sCmn.pulses, psEnc->sCmn.frame_length ); +TOC(ENCODE_PULSES) + + } + + nBits = ec_tell( psRangeEnc ); + + if( maxBits == 0 || ( useCBR == 0 && iter == 0 && nBits <= maxBits ) ) { + break; + } + + if( iter == maxIter ) { + if( nBits > maxBits && found_lower ) { + /* Restore output state from earlier iteration that did meet the bitrate budget */ + silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) ); + 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 ) { + found_upper = 1; + nBits_upper = nBits; + gainMult_upper = gainMult_Q10; + if( found_lower == 0 && iter >= 2 ) { + /* Adjust the quantizer's rate/distortion tradeoff */ + sEncCtrl.Lambda_Q10 = silk_ADD_RSHIFT32( sEncCtrl.Lambda_Q10, sEncCtrl.Lambda_Q10, 1 ); + } + } else if( nBits < maxBits - 5 ) { + found_lower = 1; + nBits_lower = nBits; + gainMult_lower = gainMult_Q10; + /* Copy part of the output state */ + silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) ); + 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 & 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 ) ); + if( nBits > maxBits ) { + gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) ); + } + gainMult_Q10 = silk_SMULWW( gainMult_Q10, gain_factor_Q16 ); + } else { + /* Adjust gain by interpolating */ + gainMult_Q10 = gainMult_lower + silk_DIV32_16( silk_MUL( 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_Q10 > gainMult_lower + silk_RSHIFT32( gainMult_upper - gainMult_lower, 2 ) ) { + gainMult_Q10 = gainMult_lower + silk_RSHIFT32( gainMult_upper - gainMult_lower, 2 ); + } else + if( gainMult_Q10 < gainMult_upper - silk_RSHIFT32( gainMult_upper - gainMult_lower, 2 ) ) { + gainMult_Q10 = gainMult_upper - silk_RSHIFT32( gainMult_upper - gainMult_lower, 2 ); + } + } + + for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + sEncCtrl.Gains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWW( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q10 ), 6 ); + } + psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev; + + /* Noise shaping quantization */ + silk_gains_quant( psEnc->sCmn.indices.GainsIndices, sEncCtrl.Gains_Q16, + &psEnc->sShape.LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr ); + + /* Restore part of the input state */ + 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; + } + /* Update input buffer */ silk_memmove( psEnc->x_buf, &psEnc->x_buf[ psEnc->sCmn.frame_length ], ( psEnc->sCmn.ltp_mem_length + LA_SHAPE_MS * psEnc->sCmn.fs_kHz ) * sizeof( opus_int16 ) ); @@ -181,28 +301,14 @@ TOC(NSQ) return ret; } - /****************************************/ - /* Encode Parameters */ - /****************************************/ -TIC(ENCODE_PARAMS) - silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding ); -TOC(ENCODE_PARAMS) - - /****************************************/ - /* Encode Excitation Signal */ - /****************************************/ -TIC(ENCODE_PULSES) - silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType, - psEnc->sCmn.pulses, psEnc->sCmn.frame_length ); -TOC(ENCODE_PULSES) - /****************************************/ /* Finalize payload */ /****************************************/ psEnc->sCmn.first_frame_after_reset = 0; /* Payload size */ *pnBytesOut = silk_RSHIFT( ec_tell( psRangeEnc ) + 7, 3 ); -TOC(ENCODE_FRAME) + + TOC(ENCODE_FRAME) #ifdef SAVE_ALL_INTERNAL_DATA { diff --git a/silk/fixed/find_pred_coefs_FIX.c b/silk/fixed/find_pred_coefs_FIX.c index e9891cd4..2cefd905 100644 --- a/silk/fixed/find_pred_coefs_FIX.c +++ b/silk/fixed/find_pred_coefs_FIX.c @@ -127,6 +127,6 @@ void silk_find_pred_coefs_FIX( silk_residual_energy_FIX( psEncCtrl->ResNrg, psEncCtrl->ResNrgQ, LPC_in_pre, psEncCtrl->PredCoef_Q12, local_gains, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); - /* Copy to prediction struct for use in next frame for fluctuation reduction */ + /* Copy to prediction struct for use in next frame for interpolation */ silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); } diff --git a/silk/fixed/main_FIX.h b/silk/fixed/main_FIX.h index 57b648eb..f4b0270c 100644 --- a/silk/fixed/main_FIX.h +++ b/silk/fixed/main_FIX.h @@ -66,7 +66,9 @@ opus_int silk_encode_frame_FIX( silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ opus_int32 *pnBytesOut, /* O Pointer to 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 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 */ ); /* Initializes the Silk encoder state */ diff --git a/silk/fixed/process_gains_FIX.c b/silk/fixed/process_gains_FIX.c index 25a13a57..c3d5664d 100644 --- a/silk/fixed/process_gains_FIX.c +++ b/silk/fixed/process_gains_FIX.c @@ -86,6 +86,10 @@ void silk_process_gains_FIX( } } + /* Save unquantized gains and gain Index */ + silk_memcpy( psEncCtrl->GainsUnq_Q16, psEncCtrl->Gains_Q16, psEnc->sCmn.nb_subfr * sizeof( opus_int32 ) ); + psEncCtrl->lastGainIndexPrev = psShapeSt->LastGainIndex; + /* Noise shaping quantization */ silk_gains_quant( psEnc->sCmn.indices.GainsIndices, psEncCtrl->Gains_Q16, &psShapeSt->LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr ); diff --git a/silk/fixed/structs_FIX.h b/silk/fixed/structs_FIX.h index 47e845d1..8803c7f2 100644 --- a/silk/fixed/structs_FIX.h +++ b/silk/fixed/structs_FIX.h @@ -109,6 +109,9 @@ typedef struct { opus_int32 ResNrg[ MAX_NB_SUBFR ]; /* Residual energy per subframe */ opus_int ResNrgQ[ MAX_NB_SUBFR ]; /* Q domain for the residual energy > 0 */ + /* Parameters for CBR mode */ + opus_int32 GainsUnq_Q16[ MAX_NB_SUBFR ]; + opus_int8 lastGainIndexPrev; } silk_encoder_control_FIX; /************************/ diff --git a/silk/float/encode_frame_FLP.c b/silk/float/encode_frame_FLP.c index 6c9e322e..a695d8a1 100644 --- a/silk/float/encode_frame_FLP.c +++ b/silk/float/encode_frame_FLP.c @@ -82,17 +82,31 @@ opus_int silk_encode_frame_FLP( silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ 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 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, ret = 0; + 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_Q10, gainMult_lower, gainMult_upper; + 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[ 512 ]; TIC(ENCODE_FRAME) + /* 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; /**************************************************************/ @@ -159,12 +173,128 @@ TIC(LBRR) silk_LBRR_encode_FLP( psEnc, &sEncCtrl, xfw, condCoding ); TOC(LBRR) - /*****************************************/ - /* Noise shaping quantization */ - /*****************************************/ + if ( psEnc->sCmn.prefillFlag ) + { TIC(NSQ) - 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, xfw ); TOC(NSQ) + } else { + /* Loop over quantizer and entroy coding to control bitrate */ + maxIter = 5; + gainMult_Q10 = SILK_FIX_CONST( 1, 10 ); + found_lower = 0; + found_upper = 0; + for( iter = 0; ; iter++ ) { + /* Copy part of the input state */ + silk_memcpy( &sRangeEnc_copy, psRangeEnc, sizeof( ec_enc ) ); + silk_memcpy( &sNSQ_copy, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) ); + seed_copy = psEnc->sCmn.indices.Seed; + ec_prevLagIndex_copy = psEnc->sCmn.ec_prevLagIndex; + ec_prevSignalType_copy = psEnc->sCmn.ec_prevSignalType; + + /*****************************************/ + /* Noise shaping quantization */ + /*****************************************/ +TIC(NSQ) + silk_NSQ_wrapper_FLP( psEnc, &sEncCtrl, &psEnc->sCmn.indices, &psEnc->sCmn.sNSQ, psEnc->sCmn.pulses, xfw ); +TOC(NSQ) + + /****************************************/ + /* Encode Parameters */ + /****************************************/ +TIC(ENCODE_PARAMS) + silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding ); +TOC(ENCODE_PARAMS) + + /****************************************/ + /* Encode Excitation Signal */ + /****************************************/ +TIC(ENCODE_PULSES) + silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType, + psEnc->sCmn.pulses, psEnc->sCmn.frame_length ); +TOC(ENCODE_PULSES) + + nBits = ec_tell( psRangeEnc ); + + if( useCBR == 0 && iter == 0 && nBits <= maxBits ) { + break; + } + + if( iter == maxIter ) { + if( nBits > maxBits && found_lower ) { + /* Restore output state from earlier iteration that did meet the bitrate budget */ + silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) ); + 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 ) { + found_upper = 1; + nBits_upper = nBits; + gainMult_upper = gainMult_Q10; + if( found_lower == 0 && iter >= 3 ) { + /* Adjust the quantizer's rate/distortion tradeoff */ + sEncCtrl.Lambda *= 1.5f; + } + } else if( nBits < maxBits - 5 ) { + found_lower = 1; + nBits_lower = nBits; + gainMult_lower = gainMult_Q10; + /* Copy part of the output state */ + silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) ); + 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 & 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 ) ); + if( nBits > maxBits ) { + gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) ); + } + gainMult_Q10 = silk_SMULWW( gainMult_Q10, gain_factor_Q16 ); + } else { + /* Adjust gain by interpolating */ + gainMult_Q10 = 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_Q10 > gainMult_lower + silk_RSHIFT32( gainMult_upper - gainMult_lower, 2 ) ) { + gainMult_Q10 = gainMult_lower + silk_RSHIFT32( gainMult_upper - gainMult_lower, 2 ); + } else + if( gainMult_Q10 < gainMult_upper - silk_RSHIFT32( gainMult_upper - gainMult_lower, 2 ) ) { + gainMult_Q10 = gainMult_upper - silk_RSHIFT32( gainMult_upper - gainMult_lower, 2 ); + } + } + + for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + pGains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWW( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q10 ), 6 ); + } + psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev; + + /* Noise shaping quantization */ + silk_gains_quant( psEnc->sCmn.indices.GainsIndices, pGains_Q16, + &psEnc->sShape.LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr ); + + /* Overwrite unquantized gains with quantized gains and convert back to Q0 from Q16 */ + for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + sEncCtrl.Gains[ i ] = pGains_Q16[ i ] / 65536.0f; + } + + /* Restore part of the input state */ + 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; + } + } /* Update input buffer */ silk_memmove( psEnc->x_buf, &psEnc->x_buf[ psEnc->sCmn.frame_length ], @@ -181,32 +311,16 @@ TOC(NSQ) return ret; } - /****************************************/ - /* Encode Parameters */ - /****************************************/ -TIC(ENCODE_PARAMS) - silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding ); -TOC(ENCODE_PARAMS) - - /****************************************/ - /* Encode Excitation Signal */ - /****************************************/ -TIC(ENCODE_PULSES) - silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType, - psEnc->sCmn.pulses, psEnc->sCmn.frame_length ); -TOC(ENCODE_PULSES) - /****************************************/ /* Finalize payload */ /****************************************/ psEnc->sCmn.first_frame_after_reset = 0; /* Payload size */ *pnBytesOut = silk_RSHIFT( ec_tell( psRangeEnc ) + 7, 3 ); + TOC(ENCODE_FRAME) #ifdef SAVE_ALL_INTERNAL_DATA - /*DEBUG_STORE_DATA( xf.dat, pIn_HP_LP, psEnc->sCmn.frame_length * sizeof( opus_int16 ) );*/ - /*DEBUG_STORE_DATA( xfw.dat, xfw, psEnc->sCmn.frame_length * sizeof( silk_float ) );*/ DEBUG_STORE_DATA( pitchL.dat, sEncCtrl.pitchL, MAX_NB_SUBFR * sizeof( opus_int ) ); DEBUG_STORE_DATA( pitchG_quantized.dat, sEncCtrl.LTPCoef, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( silk_float ) ); DEBUG_STORE_DATA( LTPcorr.dat, &psEnc->LTPCorr, sizeof( silk_float ) ); @@ -220,7 +334,6 @@ TOC(ENCODE_FRAME) DEBUG_STORE_DATA( per_index.dat, &psEnc->sCmn.indices.PERIndex, sizeof( opus_int8 ) ); DEBUG_STORE_DATA( PredCoef.dat, &sEncCtrl.PredCoef[ 1 ], psEnc->sCmn.predictLPCOrder * sizeof( silk_float ) ); DEBUG_STORE_DATA( ltp_scale_idx.dat, &psEnc->sCmn.indices.LTP_scaleIndex, sizeof( opus_int8 ) ); - /*DEBUG_STORE_DATA( xq.dat, psEnc->sCmn.sNSQ.xqBuf, psEnc->sCmn.frame_length * sizeof( silk_float ) );*/ #endif return ret; } diff --git a/silk/float/find_pred_coefs_FLP.c b/silk/float/find_pred_coefs_FLP.c index 04881855..90a4e467 100644 --- a/silk/float/find_pred_coefs_FLP.c +++ b/silk/float/find_pred_coefs_FLP.c @@ -112,7 +112,7 @@ TOC(LSF_quant); silk_residual_energy_FLP( psEncCtrl->ResNrg, LPC_in_pre, psEncCtrl->PredCoef, psEncCtrl->Gains, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); - /* Copy to prediction struct for use in next frame for fluctuation reduction */ + /* Copy to prediction struct for use in next frame for interpolation */ silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); } diff --git a/silk/float/inner_product_FLP.c b/silk/float/inner_product_FLP.c index 4ed56a35..a8f04d73 100644 --- a/silk/float/inner_product_FLP.c +++ b/silk/float/inner_product_FLP.c @@ -42,18 +42,18 @@ double silk_inner_product_FLP( /* O result */ double result; /* 4x unrolled loop */ - result = 0.0f; + result = 0.0; dataSize4 = dataSize & 0xFFFC; for( i = 0; i < dataSize4; i += 4 ) { - result += data1[ i + 0 ] * data2[ i + 0 ] + - data1[ i + 1 ] * data2[ i + 1 ] + - data1[ i + 2 ] * data2[ i + 2 ] + - data1[ i + 3 ] * data2[ i + 3 ]; + result += data1[ i + 0 ] * (double)data2[ i + 0 ] + + data1[ i + 1 ] * (double)data2[ i + 1 ] + + data1[ i + 2 ] * (double)data2[ i + 2 ] + + data1[ i + 3 ] * (double)data2[ i + 3 ]; } /* add any remaining products */ for( ; i < dataSize; i++ ) { - result += data1[ i ] * data2[ i ]; + result += data1[ i ] * (double)data2[ i ]; } return result; diff --git a/silk/float/main_FLP.h b/silk/float/main_FLP.h index 20a766b1..07e31ab6 100644 --- a/silk/float/main_FLP.h +++ b/silk/float/main_FLP.h @@ -64,7 +64,9 @@ opus_int silk_encode_frame_FLP( silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ 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 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 */ ); /* Initializes the Silk encoder state */ diff --git a/silk/float/pitch_analysis_core_FLP.c b/silk/float/pitch_analysis_core_FLP.c index 8e661590..4108d63d 100644 --- a/silk/float/pitch_analysis_core_FLP.c +++ b/silk/float/pitch_analysis_core_FLP.c @@ -202,8 +202,8 @@ opus_int silk_pitch_analysis_core_FLP( /* O voicing estimate: 0 voiced, 1 unvoic /* Add contribution of new sample and remove contribution from oldest sample */ normalizer += - basis_ptr[ 0 ] * basis_ptr[ 0 ] - - basis_ptr[ sf_length_8kHz ] * basis_ptr[ sf_length_8kHz ]; + basis_ptr[ 0 ] * (double)basis_ptr[ 0 ] - + basis_ptr[ sf_length_8kHz ] * (double)basis_ptr[ sf_length_8kHz ]; C[ 0 ][ d ] += (silk_float)(cross_corr / sqrt( normalizer )); } /* Update target pointer */ @@ -225,7 +225,7 @@ opus_int silk_pitch_analysis_core_FLP( /* O voicing estimate: 0 voiced, 1 unvoic target_ptr = &frame_4kHz[ silk_SMULBB( sf_length_4kHz, nb_subfr ) ]; energy = 1000.0f; for( i = 0; i < silk_LSHIFT( sf_length_4kHz, 2 ); i++ ) { - energy += target_ptr[i] * target_ptr[i]; + energy += target_ptr[i] * (double)target_ptr[i]; } threshold = Cmax * Cmax; if( energy / 16.0f > threshold ) { diff --git a/silk/float/process_gains_FLP.c b/silk/float/process_gains_FLP.c index 2c7dfdd0..0708a5c2 100644 --- a/silk/float/process_gains_FLP.c +++ b/silk/float/process_gains_FLP.c @@ -67,6 +67,10 @@ void silk_process_gains_FLP( pGains_Q16[ k ] = ( opus_int32 ) ( psEncCtrl->Gains[ k ] * 65536.0f ); } + /* Save unquantized gains and gain Index */ + silk_memcpy( psEncCtrl->GainsUnq_Q16, pGains_Q16, psEnc->sCmn.nb_subfr * sizeof( opus_int32 ) ); + psEncCtrl->lastGainIndexPrev = psShapeSt->LastGainIndex; + /* Noise shaping quantization */ silk_gains_quant( psEnc->sCmn.indices.GainsIndices, pGains_Q16, &psShapeSt->LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr ); diff --git a/silk/float/silk_float.vcxproj b/silk/float/silk_float.vcxproj index abcdaf05..44539478 100644 --- a/silk/float/silk_float.vcxproj +++ b/silk/float/silk_float.vcxproj @@ -127,4 +127,4 @@ - + \ No newline at end of file diff --git a/silk/float/structs_FLP.h b/silk/float/structs_FLP.h index 99264c97..dd776b5c 100644 --- a/silk/float/structs_FLP.h +++ b/silk/float/structs_FLP.h @@ -107,6 +107,10 @@ typedef struct { silk_float predGain; silk_float LTPredCodGain; silk_float ResNrg[ MAX_NB_SUBFR ]; /* Residual energy per subframe */ + + /* Parameters for CBR mode */ + opus_int32 GainsUnq_Q16[ MAX_NB_SUBFR ]; + opus_int8 lastGainIndexPrev; } silk_encoder_control_FLP; /************************/ diff --git a/silk/process_NLSFs.c b/silk/process_NLSFs.c index e077627c..6e1cc06f 100644 --- a/silk/process_NLSFs.c +++ b/silk/process_NLSFs.c @@ -54,14 +54,14 @@ void silk_process_NLSFs( /* Calculate mu values */ /***********************/ /* NLSF_mu = 0.003 - 0.0015 * psEnc->speech_activity; */ - NLSF_mu_Q20 = silk_SMLAWB( SILK_FIX_CONST( 0.0025, 20 ), SILK_FIX_CONST( -0.001, 28 ), psEncC->speech_activity_Q8 ); + NLSF_mu_Q20 = silk_SMLAWB( SILK_FIX_CONST( 0.003, 20 ), SILK_FIX_CONST( -0.001, 28 ), psEncC->speech_activity_Q8 ); if( psEncC->nb_subfr == 2 ) { /* Multiply by 1.5 for 10 ms packets */ NLSF_mu_Q20 = silk_ADD_RSHIFT( NLSF_mu_Q20, NLSF_mu_Q20, 1 ); } silk_assert( NLSF_mu_Q20 > 0 ); - silk_assert( NLSF_mu_Q20 <= SILK_FIX_CONST( 0.0045, 20 ) ); + silk_assert( NLSF_mu_Q20 <= SILK_FIX_CONST( 0.005, 20 ) ); /* Calculate NLSF weights */ silk_NLSF_VQ_weights_laroia( pNLSFW_QW, pNLSF_Q15, psEncC->predictLPCOrder ); diff --git a/silk/structs.h b/silk/structs.h index d11aa281..51c0f866 100644 --- a/silk/structs.h +++ b/silk/structs.h @@ -189,7 +189,7 @@ typedef struct { opus_int8 LBRR_flag; opus_int LBRR_flags[ MAX_FRAMES_PER_PACKET ]; - SideInfoIndices indices; + SideInfoIndices indices; opus_int8 pulses[ MAX_FRAME_LENGTH ]; /* Input/output buffering */ diff --git a/src/opus_decoder.c b/src/opus_decoder.c index 38672b14..4f4d6487 100644 --- a/src/opus_decoder.c +++ b/src/opus_decoder.c @@ -726,7 +726,7 @@ int opus_decode_float(OpusDecoder *st, const unsigned char *data, if (ret > 0) { for (i=0;ichannels;i++) - pcm[i] = (1./32768.)*(out[i]); + pcm[i] = (1.f/32768.f)*(out[i]); } RESTORE_STACK; return ret; diff --git a/src/opus_encoder.c b/src/opus_encoder.c index c1cb2cf6..9259fab0 100644 --- a/src/opus_encoder.c +++ b/src/opus_encoder.c @@ -219,6 +219,44 @@ int opus_encoder_init(OpusEncoder* st, opus_int32 Fs, int channels, int applicat return OPUS_OK; } +static int pad_frame(unsigned char *data, int len, int new_len) +{ + if (len == new_len) + return 0; + if (len > new_len) + return 1; + + if ((data[0]&0x3)==0) + { + int i; + int padding, nb_255s; + + padding = new_len - len; + if (padding >= 2) + { + nb_255s = (padding-2)/255; + + for (i=len-1;i>=1;i--) + data[i+nb_255s+2] = data[i]; + data[0] |= 0x3; + data[1] = 0x41; + for (i=0;i=1;i--) + data[i+1] = data[i]; + data[0] |= 0x3; + data[1] = 1; + } + return 0; + } else { + return 1; + } +} + static unsigned char gen_toc(int mode, int framerate, int bandwidth, int channels) { int period; @@ -449,6 +487,15 @@ int opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_size, st->bitrate_bps = user_bitrate_to_bitrate(st, frame_size, max_data_bytes); + if (!st->use_vbr) + { + int cbrBytes, frame_rate; + frame_rate = st->Fs/frame_size; + cbrBytes = IMIN( (st->bitrate_bps + 4*frame_rate)/(8*frame_rate) , max_data_bytes); + st->bitrate_bps = cbrBytes * (8*frame_rate); + max_data_bytes = cbrBytes; + } + /* Equivalent 20-ms rate for mode/channel/bandwidth decisions */ equiv_rate = st->bitrate_bps - 60*(st->Fs/frame_size - 50); @@ -614,6 +661,21 @@ int opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_size, if (st->user_bandwidth != OPUS_AUTO) st->bandwidth = st->user_bandwidth; + /* This enforces safe rates for CBR SILK */ + if (!st->use_vbr && st->mode != MODE_CELT_ONLY) + { + int frame_rate; + opus_int32 max_rate; + frame_rate = st->Fs/frame_size; + max_rate = frame_rate*max_data_bytes; + if (max_rate < 15000) + st->bandwidth = IMIN(st->bandwidth, OPUS_BANDWIDTH_WIDEBAND); + if (max_rate < 12000) + st->bandwidth = IMIN(st->bandwidth, OPUS_BANDWIDTH_MEDIUMBAND); + if (max_rate < 9000) + st->bandwidth = OPUS_BANDWIDTH_NARROWBAND; + } + /* Prevents Opus from wasting bits on frequencies that are above the Nyquist rate of the input signal */ if (st->Fs <= 24000 && st->bandwidth > OPUS_BANDWIDTH_SUPERWIDEBAND) @@ -625,6 +687,9 @@ int opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_size, if (st->Fs <= 8000 && st->bandwidth > OPUS_BANDWIDTH_NARROWBAND) st->bandwidth = OPUS_BANDWIDTH_NARROWBAND; + /* If max_data_bytes represents less than 8 kb/s, switch to CELT-only mode */ + if (max_data_bytes < 8000*frame_size / (st->Fs * 8)) + st->mode = MODE_CELT_ONLY; /* Can't support higher than wideband for >20 ms frames */ if (frame_size > st->Fs/50 && (st->mode == MODE_CELT_ONLY || st->bandwidth > OPUS_BANDWIDTH_WIDEBAND)) @@ -688,9 +753,6 @@ int opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_size, if (st->mode == MODE_HYBRID && st->bandwidth <= OPUS_BANDWIDTH_WIDEBAND) st->mode = MODE_SILK_ONLY; - /* If max_data_bytes represents less than 8 kb/s, switch to CELT-only mode */ - if (max_data_bytes < 8000*frame_size / (st->Fs * 8)) - st->mode = MODE_CELT_ONLY; /* printf("%d %d %d %d\n", st->bitrate_bps, st->stream_channels, st->mode, st->bandwidth); */ bytes_target = IMIN(max_data_bytes, st->bitrate_bps * frame_size / (st->Fs * 8)) - 1; @@ -756,9 +818,6 @@ int opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_size, st->silk_mode.bitRate = ( st->bitrate_bps - 8*st->Fs/frame_size ) * 4/5; } } - /* SILK is not allow to use more than 50% of max_data_bytes */ - if (max_data_bytes < st->silk_mode.bitRate*frame_size / (st->Fs * 4)) - st->silk_mode.bitRate = max_data_bytes*st->Fs*4/frame_size; st->silk_mode.payloadSize_ms = 1000 * frame_size / st->Fs; st->silk_mode.nChannelsAPI = st->channels; @@ -779,8 +838,24 @@ int opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_size, } st->silk_mode.maxInternalSampleRate = 16000; + st->silk_mode.useCBR = !st->use_vbr; + /* Call SILK encoder for the low band */ nBytes = IMIN(1275, max_data_bytes-1); + + st->silk_mode.maxBits = nBytes*8; + /* Only allow up to 90% of the bits for hybrid mode*/ + if (st->mode == MODE_HYBRID) + st->silk_mode.maxBits = (opus_int32)st->silk_mode.maxBits*9/10; + if (st->silk_mode.useCBR) + { + st->silk_mode.maxBits = (st->silk_mode.bitRate * frame_size / (st->Fs * 8))*8; + /* Reduce the initial target to make it easier to reach the CBR rate */ + st->silk_mode.bitRate = 1; + } + if (redundancy) + st->silk_mode.maxBits -= st->silk_mode.maxBits/(1 + frame_size/(st->Fs/200)); + if (prefill) { int zero=0; @@ -1029,6 +1104,12 @@ int opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_size, st->prev_framesize = frame_size; st->first = 0; + if (!redundancy && st->mode==MODE_SILK_ONLY && !st->use_vbr && ret >= 2) + { + nb_compr_bytes = st->bitrate_bps * frame_size / (st->Fs * 8); + pad_frame(data, ret+1, nb_compr_bytes); + return nb_compr_bytes; + } RESTORE_STACK; return ret+1+redundancy_bytes; } diff --git a/src/test_opus.c b/src/test_opus.c index ae5d4c8b..c1ef4f92 100644 --- a/src/test_opus.c +++ b/src/test_opus.c @@ -110,7 +110,7 @@ int main(int argc, char *argv[]) int stop=0; short *in, *out; int application=OPUS_APPLICATION_AUDIO; - double bits=0.0, bits_act=0.0, bits2=0.0, nrg; + double bits=0.0, bits_max=0.0, bits_act=0.0, bits2=0.0, nrg; int bandwidth=-1; const char *bandwidth_string; int lost = 0, lost_prev = 1; @@ -446,6 +446,7 @@ int main(int argc, char *argv[]) /* count bits */ bits += len[toggle]*8; + bits_max = ( len[toggle]*8 > bits_max ) ? len[toggle]*8 : bits_max; if( count >= use_inbandfec ) { nrg = 0.0; if (!decode_only) @@ -465,6 +466,7 @@ int main(int argc, char *argv[]) toggle = (toggle + use_inbandfec) & 1; } fprintf (stderr, "average bitrate: %7.3f kb/s\n", 1e-3*bits*sampling_rate/(frame_size*(double)count)); + fprintf (stderr, "maximum bitrate: %7.3f bkp/s\n", 1e-3*bits_max*sampling_rate/frame_size); if (!decode_only) fprintf (stderr, "active bitrate: %7.3f kb/s\n", 1e-3*bits_act*sampling_rate/(frame_size*(double)count_act)); fprintf (stderr, "bitrate standard deviation: %7.3f kb/s\n", 1e-3*sqrt(bits2/count - bits*bits/(count*(double)count))*sampling_rate/frame_size); -- cgit v1.2.3