opus_int silk_InitEncoder( /* O Returns error code */ void *encState, /* I/O State */ int arch, /* I Run-time architecture */ silk_EncControlStruct *encStatus /* O Encoder Status */ ) { silk_encoder *psEnc; opus_int n, ret = SILK_NO_ERROR; psEnc = (silk_encoder *)encState; /* Reset encoder */ silk_memset( psEnc, 0, sizeof( silk_encoder ) ); for( n = 0; n < ENCODER_NUM_CHANNELS; n++ ) { if( ret += silk_init_encoder( &psEnc->state_Fxx[ n ], arch ) ) { silk_assert( 0 ); } } psEnc->nChannelsAPI = 1; psEnc->nChannelsInternal = 1; /* Read control structure */ if( ret += silk_QueryEncoder( encState, encStatus ) ) { silk_assert( 0 ); } return ret; }
opus_int silk_VAD_Init( /* O Return value, 0 if success */ silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */ ) { opus_int b, ret = 0; /* reset state memory */ silk_memset( psSilk_VAD, 0, sizeof( silk_VAD_state ) ); /* init noise levels */ /* Initialize array with approx pink noise levels (psd proportional to inverse of frequency) */ for( b = 0; b < VAD_N_BANDS; b++ ) { psSilk_VAD->NoiseLevelBias[ b ] = silk_max_32( silk_DIV32_16( VAD_NOISE_LEVELS_BIAS, b + 1 ), 1 ); } /* Initialize state */ for( b = 0; b < VAD_N_BANDS; b++ ) { psSilk_VAD->NL[ b ] = silk_MUL( 100, psSilk_VAD->NoiseLevelBias[ b ] ); psSilk_VAD->inv_NL[ b ] = silk_DIV32( silk_int32_MAX, psSilk_VAD->NL[ b ] ); } psSilk_VAD->counter = 15; /* init smoothed energy-to-noise ratio*/ for( b = 0; b < VAD_N_BANDS; b++ ) { psSilk_VAD->NrgRatioSmth_Q8[ b ] = 100 * 256; /* 100 * 256 --> 20 dB SNR */ } return( ret ); }
/* Uses SMULL(), available on armv4 */ opus_int32 silk_schur64( /* O returns residual energy */ opus_int32 rc_Q16[], /* O Reflection coefficients [order] Q16 */ const opus_int32 c[], /* I Correlations [order+1] */ opus_int32 order /* I Prediction order */ ) { opus_int k, n; opus_int32 C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; opus_int32 Ctmp1_Q30, Ctmp2_Q30, rc_tmp_Q31; silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 ); /* Check for invalid input */ if( c[ 0 ] <= 0 ) { silk_memset( rc_Q16, 0, order * sizeof( opus_int32 ) ); return 0; } for( k = 0; k < order + 1; k++ ) { C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ]; } for( k = 0; k < order; k++ ) { /* Check that we won't be getting an unstable rc, otherwise stop here. */ if (silk_abs_int32(C[ k + 1 ][ 0 ]) >= C[ 0 ][ 1 ]) { if ( C[ k + 1 ][ 0 ] > 0 ) { rc_Q16[ k ] = -SILK_FIX_CONST( .99f, 16 ); } else { rc_Q16[ k ] = SILK_FIX_CONST( .99f, 16 ); } k++; break; } /* Get reflection coefficient: divide two Q30 values and get result in Q31 */ rc_tmp_Q31 = silk_DIV32_varQ( -C[ k + 1 ][ 0 ], C[ 0 ][ 1 ], 31 ); /* Save the output */ rc_Q16[ k ] = silk_RSHIFT_ROUND( rc_tmp_Q31, 15 ); /* Update correlations */ for( n = 0; n < order - k; n++ ) { Ctmp1_Q30 = C[ n + k + 1 ][ 0 ]; Ctmp2_Q30 = C[ n ][ 1 ]; /* Multiply and add the highest int32 */ C[ n + k + 1 ][ 0 ] = Ctmp1_Q30 + silk_SMMUL( silk_LSHIFT( Ctmp2_Q30, 1 ), rc_tmp_Q31 ); C[ n ][ 1 ] = Ctmp2_Q30 + silk_SMMUL( silk_LSHIFT( Ctmp1_Q30, 1 ), rc_tmp_Q31 ); } } for(; k < order; k++ ) { rc_Q16[ k ] = 0; } return silk_max_32( 1, C[ 0 ][ 1 ] ); }
void silk_LPC_analysis_filter( opus_int16 *out, /* O Output signal */ const opus_int16 *in, /* I Input signal */ const opus_int16 *B, /* I MA prediction coefficients, Q12 [order] */ const opus_int32 len, /* I Signal length */ const opus_int32 d /* I Filter order */ ) { opus_int ix, j; opus_int32 out32_Q12, out32; const opus_int16 *in_ptr; silk_assert( d >= 6 ); silk_assert( (d & 1) == 0 ); silk_assert( d <= len ); for( ix = d; ix < len; ix++ ) { in_ptr = &in[ ix - 1 ]; out32_Q12 = silk_SMULBB( in_ptr[ 0 ], B[ 0 ] ); /* Allowing wrap around so that two wraps can cancel each other. The rare cases where the result wraps around can only be triggered by invalid streams*/ out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -1 ], B[ 1 ] ); out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -2 ], B[ 2 ] ); out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -3 ], B[ 3 ] ); out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -4 ], B[ 4 ] ); out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -5 ], B[ 5 ] ); for( j = 6; j < d; j += 2 ) { out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -j ], B[ j ] ); out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -j - 1 ], B[ j + 1 ] ); } /* Subtract prediction */ out32_Q12 = silk_SUB32_ovflw( silk_LSHIFT( (opus_int32)in_ptr[ 1 ], 12 ), out32_Q12 ); /* Scale to Q0 */ out32 = silk_RSHIFT_ROUND( out32_Q12, 12 ); /* Saturate output */ out[ ix ] = (opus_int16)silk_SAT16( out32 ); } /* Set first d output samples to zero */ silk_memset( out, 0, d * sizeof( opus_int16 ) ); }
opus_int silk_init_decoder( silk_decoder_state *psDec /* I/O Decoder state pointer */ ) { /* Clear the entire encoder state, except anything copied */ silk_memset( psDec, 0, sizeof( silk_decoder_state ) ); /* Used to deactivate LSF interpolation */ psDec->first_frame_after_reset = 1; psDec->prev_gain_Q16 = 65536; /* Reset CNG state */ silk_CNG_Reset( psDec ); /* Reset PLC state */ silk_PLC_Reset( psDec ); return(0); }
opus_int silk_init_encoder( silk_encoder_state_Fxx *psEnc /* I/O Pointer to Silk FIX encoder state */ ) { opus_int ret = 0; /* Clear the entire encoder state */ silk_memset( psEnc, 0, sizeof( silk_encoder_state_Fxx ) ); psEnc->sCmn.variable_HP_smth1_Q15 = silk_LSHIFT( silk_lin2log( SILK_FIX_CONST( VARIABLE_HP_MIN_CUTOFF_HZ, 16 ) ) - ( 16 << 7 ), 8 ); psEnc->sCmn.variable_HP_smth2_Q15 = psEnc->sCmn.variable_HP_smth1_Q15; /* Used to deactivate LSF interpolation, pitch prediction */ psEnc->sCmn.first_frame_after_reset = 1; /* Initialize Silk VAD */ ret += silk_VAD_Init( &psEnc->sCmn.sVAD ); return ret; }
void silk_LPC_analysis_filter_FLP( silk_float r_LPC[], /* O LPC residual signal */ const silk_float PredCoef[], /* I LPC coefficients */ const silk_float s[], /* I Input signal */ const opus_int length, /* I Length of input signal */ const opus_int Order /* I LPC order */ ) { silk_assert( Order <= length ); switch( Order ) { case 6: silk_LPC_analysis_filter6_FLP( r_LPC, PredCoef, s, length ); break; case 8: silk_LPC_analysis_filter8_FLP( r_LPC, PredCoef, s, length ); break; case 10: silk_LPC_analysis_filter10_FLP( r_LPC, PredCoef, s, length ); break; case 12: silk_LPC_analysis_filter12_FLP( r_LPC, PredCoef, s, length ); break; case 16: silk_LPC_analysis_filter16_FLP( r_LPC, PredCoef, s, length ); break; default: silk_assert( 0 ); break; } /* Set first Order output samples to zero */ silk_memset( r_LPC, 0, Order * sizeof( silk_float ) ); }
static inline void silk_PLC_conceal( silk_decoder_state *psDec, /* I/O Decoder state */ silk_decoder_control *psDecCtrl, /* I/O Decoder control */ opus_int16 frame[] /* O LPC residual signal */ ) { opus_int i, j, k; opus_int lag, idx, sLTP_buf_idx, shift1, shift2; opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q16, inv_gain_Q30; opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr; opus_int32 LPC_exc_Q14, LPC_pred_Q10, LTP_pred_Q12; opus_int16 rand_scale_Q14; opus_int16 *B_Q14, *exc_buf_ptr; opus_int32 *sLPC_Q14_ptr; opus_int16 exc_buf[ 2 * MAX_SUB_FRAME_LENGTH ]; opus_int16 A_Q12[ MAX_LPC_ORDER ]; opus_int16 sLTP[ MAX_FRAME_LENGTH ]; opus_int32 sLTP_Q14[ 2 * MAX_FRAME_LENGTH ]; silk_PLC_struct *psPLC = &psDec->sPLC; if (psDec->first_frame_after_reset) silk_memset(psPLC->prevLPC_Q12, 0, MAX_LPC_ORDER*sizeof(psPLC->prevLPC_Q12[ 0 ])); /* Find random noise component */ /* Scale previous excitation signal */ exc_buf_ptr = exc_buf; for( k = 0; k < 2; k++ ) { for( i = 0; i < psPLC->subfr_length; i++ ) { exc_buf_ptr[ i ] = ( opus_int16 )silk_RSHIFT( silk_SMULWW( psDec->exc_Q10[ i + ( k + psPLC->nb_subfr - 2 ) * psPLC->subfr_length ], psPLC->prevGain_Q16[ k ] ), 10 ); } exc_buf_ptr += psPLC->subfr_length; } /* Find the subframe with lowest energy of the last two and use that as random noise generator */ silk_sum_sqr_shift( &energy1, &shift1, exc_buf, psPLC->subfr_length ); silk_sum_sqr_shift( &energy2, &shift2, &exc_buf[ psPLC->subfr_length ], psPLC->subfr_length ); if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) { /* First sub-frame has lowest energy */ rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, ( psPLC->nb_subfr - 1 ) * psPLC->subfr_length - RAND_BUF_SIZE ) ]; } else { /* Second sub-frame has lowest energy */ rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, psPLC->nb_subfr * psPLC->subfr_length - RAND_BUF_SIZE ) ]; } /* Setup Gain to random noise component */ B_Q14 = psPLC->LTPCoef_Q14; rand_scale_Q14 = psPLC->randScale_Q14; /* Setup attenuation gains */ harm_Gain_Q15 = HARM_ATT_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ]; if( psDec->prevSignalType == TYPE_VOICED ) { rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ]; } else { rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ]; } /* LPC concealment. Apply BWE to previous LPC */ silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) ); /* Preload LPC coeficients to array on stack. Gives small performance gain */ silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) ); /* First Lost frame */ if( psDec->lossCnt == 0 ) { rand_scale_Q14 = 1 << 14; /* Reduce random noise Gain for voiced frames */ if( psDec->prevSignalType == TYPE_VOICED ) { for( i = 0; i < LTP_ORDER; i++ ) { rand_scale_Q14 -= B_Q14[ i ]; } rand_scale_Q14 = silk_max_16( 3277, rand_scale_Q14 ); /* 0.2 */ rand_scale_Q14 = ( opus_int16 )silk_RSHIFT( silk_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 ); } else { /* Reduce random noise for unvoiced frames with high LPC gain */ opus_int32 invGain_Q30, down_scale_Q30; silk_LPC_inverse_pred_gain( &invGain_Q30, psPLC->prevLPC_Q12, psDec->LPC_order ); down_scale_Q30 = silk_min_32( silk_RSHIFT( 1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 ); down_scale_Q30 = silk_max_32( silk_RSHIFT( 1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 ); down_scale_Q30 = silk_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES ); rand_Gain_Q15 = silk_RSHIFT( silk_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 ); } } rand_seed = psPLC->rand_seed; lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 ); sLTP_buf_idx = psDec->ltp_mem_length; /* Rewhiten LTP state */ idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2; silk_assert( idx > 0 ); silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order ); /* Scale LTP state */ inv_gain_Q16 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 32 ); inv_gain_Q16 = silk_min( inv_gain_Q16, silk_int16_MAX ); inv_gain_Q30 = silk_LSHIFT( inv_gain_Q16, 14 ); for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) { sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] ); } /***************************/ /* LTP synthesis filtering */ /***************************/ for( k = 0; k < psDec->nb_subfr; k++ ) { /* Setup pointer */ pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ]; for( i = 0; i < psDec->subfr_length; i++ ) { /* Unrolled loop */ LTP_pred_Q12 = silk_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] ); LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] ); LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] ); LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] ); pred_lag_ptr++; /* Generate LPC excitation */ rand_seed = silk_RAND( rand_seed ); idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK; LPC_exc_Q14 = silk_LSHIFT32( silk_SMULWB( rand_ptr[ idx ], rand_scale_Q14 ), 6 ); /* Random noise part */ LPC_exc_Q14 = silk_ADD32( LPC_exc_Q14, silk_LSHIFT32( LTP_pred_Q12, 2 ) ); /* Harmonic part */ sLTP_Q14[ sLTP_buf_idx ] = LPC_exc_Q14; sLTP_buf_idx++; } /* Gradually reduce LTP gain */ for( j = 0; j < LTP_ORDER; j++ ) { B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 ); } /* Gradually reduce excitation gain */ rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 ); /* Slowly increase pitch lag */ psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 ); psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) ); lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 ); } /***************************/ /* LPC synthesis filtering */ /***************************/ sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ]; /* Copy LPC state */ silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) ); silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */ for( i = 0; i < psDec->frame_length; i++ ) { /* partly unrolled */ LPC_pred_Q10 = silk_SMULWB( sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 5 ], A_Q12[ 4 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 6 ], A_Q12[ 5 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 7 ], A_Q12[ 6 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 8 ], A_Q12[ 7 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 9 ], A_Q12[ 8 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] ); for( j = 10; j < psDec->LPC_order; j++ ) { LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] ); } /* Add prediction to LPC excitation */ sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_LSHIFT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], LPC_pred_Q10, 4 ); /* Scale with Gain */ frame[ i ] = ( opus_int16 )silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], psPLC->prevGain_Q16[ 1 ] ), 14 ) ); } /* Save LPC state */ silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); /**************************************/ /* Update states */ /**************************************/ psPLC->rand_seed = rand_seed; psPLC->randScale_Q14 = rand_scale_Q14; for( i = 0; i < MAX_NB_SUBFR; i++ ) { psDecCtrl->pitchL[ i ] = lag; } }
/* Encode quantization indices of excitation */ void silk_encode_pulses( ec_enc *psRangeEnc, /* I/O compressor data structure */ const opus_int signalType, /* I Signal type */ const opus_int quantOffsetType, /* I quantOffsetType */ opus_int8 pulses[], /* I quantization indices */ const opus_int frame_length /* I Frame length */ ) { opus_int i, k, j, iter, bit, nLS, scale_down, RateLevelIndex = 0; opus_int32 abs_q, minSumBits_Q5, sumBits_Q5; opus_int abs_pulses[ MAX_FRAME_LENGTH ]; opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ]; opus_int nRshifts[ MAX_NB_SHELL_BLOCKS ]; opus_int pulses_comb[ 8 ]; opus_int *abs_pulses_ptr; const opus_int8 *pulses_ptr; const opus_uint8 *cdf_ptr; const opus_uint8 *nBits_ptr; silk_memset( pulses_comb, 0, 8 * sizeof( opus_int ) ); /* Fixing Valgrind reported problem*/ /****************************/ /* Prepare for shell coding */ /****************************/ /* Calculate number of shell blocks */ silk_assert( 1 << LOG2_SHELL_CODEC_FRAME_LENGTH == SHELL_CODEC_FRAME_LENGTH ); iter = silk_RSHIFT( frame_length, LOG2_SHELL_CODEC_FRAME_LENGTH ); if( iter * SHELL_CODEC_FRAME_LENGTH < frame_length ) { silk_assert( frame_length == 12 * 10 ); /* Make sure only happens for 10 ms @ 12 kHz */ iter++; silk_memset( &pulses[ frame_length ], 0, SHELL_CODEC_FRAME_LENGTH * sizeof(opus_int8)); } /* Take the absolute value of the pulses */ for( i = 0; i < iter * SHELL_CODEC_FRAME_LENGTH; i+=4 ) { abs_pulses[i+0] = ( opus_int )silk_abs( pulses[ i + 0 ] ); abs_pulses[i+1] = ( opus_int )silk_abs( pulses[ i + 1 ] ); abs_pulses[i+2] = ( opus_int )silk_abs( pulses[ i + 2 ] ); abs_pulses[i+3] = ( opus_int )silk_abs( pulses[ i + 3 ] ); } /* Calc sum pulses per shell code frame */ abs_pulses_ptr = abs_pulses; for( i = 0; i < iter; i++ ) { nRshifts[ i ] = 0; while( 1 ) { /* 1+1 -> 2 */ scale_down = combine_and_check( pulses_comb, abs_pulses_ptr, silk_max_pulses_table[ 0 ], 8 ); /* 2+2 -> 4 */ scale_down += combine_and_check( pulses_comb, pulses_comb, silk_max_pulses_table[ 1 ], 4 ); /* 4+4 -> 8 */ scale_down += combine_and_check( pulses_comb, pulses_comb, silk_max_pulses_table[ 2 ], 2 ); /* 8+8 -> 16 */ scale_down += combine_and_check( &sum_pulses[ i ], pulses_comb, silk_max_pulses_table[ 3 ], 1 ); if( scale_down ) { /* We need to downscale the quantization signal */ nRshifts[ i ]++; for( k = 0; k < SHELL_CODEC_FRAME_LENGTH; k++ ) { abs_pulses_ptr[ k ] = silk_RSHIFT( abs_pulses_ptr[ k ], 1 ); } } else { /* Jump out of while(1) loop and go to next shell coding frame */ break; } } abs_pulses_ptr += SHELL_CODEC_FRAME_LENGTH; } /**************/ /* Rate level */ /**************/ /* find rate level that leads to fewest bits for coding of pulses per block info */ minSumBits_Q5 = silk_int32_MAX; for( k = 0; k < N_RATE_LEVELS - 1; k++ ) { nBits_ptr = silk_pulses_per_block_BITS_Q5[ k ]; sumBits_Q5 = silk_rate_levels_BITS_Q5[ signalType >> 1 ][ k ]; for( i = 0; i < iter; i++ ) { if( nRshifts[ i ] > 0 ) { sumBits_Q5 += nBits_ptr[ MAX_PULSES + 1 ]; } else { sumBits_Q5 += nBits_ptr[ sum_pulses[ i ] ]; } } if( sumBits_Q5 < minSumBits_Q5 ) { minSumBits_Q5 = sumBits_Q5; RateLevelIndex = k; } } ec_enc_icdf( psRangeEnc, RateLevelIndex, silk_rate_levels_iCDF[ signalType >> 1 ], 8 ); /***************************************************/ /* Sum-Weighted-Pulses Encoding */ /***************************************************/ cdf_ptr = silk_pulses_per_block_iCDF[ RateLevelIndex ]; for( i = 0; i < iter; i++ ) { if( nRshifts[ i ] == 0 ) { ec_enc_icdf( psRangeEnc, sum_pulses[ i ], cdf_ptr, 8 ); } else { ec_enc_icdf( psRangeEnc, MAX_PULSES + 1, cdf_ptr, 8 ); for( k = 0; k < nRshifts[ i ] - 1; k++ ) { ec_enc_icdf( psRangeEnc, MAX_PULSES + 1, silk_pulses_per_block_iCDF[ N_RATE_LEVELS - 1 ], 8 ); } ec_enc_icdf( psRangeEnc, sum_pulses[ i ], silk_pulses_per_block_iCDF[ N_RATE_LEVELS - 1 ], 8 ); } } /******************/ /* Shell Encoding */ /******************/ for( i = 0; i < iter; i++ ) { if( sum_pulses[ i ] > 0 ) { silk_shell_encoder( psRangeEnc, &abs_pulses[ i * SHELL_CODEC_FRAME_LENGTH ] ); } } /****************/ /* LSB Encoding */ /****************/ for( i = 0; i < iter; i++ ) { if( nRshifts[ i ] > 0 ) { pulses_ptr = &pulses[ i * SHELL_CODEC_FRAME_LENGTH ]; nLS = nRshifts[ i ] - 1; for( k = 0; k < SHELL_CODEC_FRAME_LENGTH; k++ ) { abs_q = (opus_int8)silk_abs( pulses_ptr[ k ] ); for( j = nLS; j > 0; j-- ) { bit = silk_RSHIFT( abs_q, j ) & 1; ec_enc_icdf( psRangeEnc, bit, silk_lsb_iCDF, 8 ); } bit = abs_q & 1; ec_enc_icdf( psRangeEnc, bit, silk_lsb_iCDF, 8 ); } } } /****************/ /* Encode signs */ /****************/ silk_encode_signs( psRangeEnc, pulses, frame_length, signalType, quantOffsetType, sum_pulses ); }
/* Find pitch lags */ void silk_find_pitch_lags_FIX( silk_encoder_state_FIX *psEnc, /* I/O encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ opus_int16 res[], /* O residual */ const opus_int16 x[], /* I Speech signal */ int arch /* I Run-time architecture */ ) { opus_int buf_len, i, scale; opus_int32 thrhld_Q13, res_nrg; const opus_int16 *x_buf, *x_buf_ptr; VARDECL( opus_int16, Wsig ); opus_int16 *Wsig_ptr; opus_int32 auto_corr[ MAX_FIND_PITCH_LPC_ORDER + 1 ]; opus_int16 rc_Q15[ MAX_FIND_PITCH_LPC_ORDER ]; opus_int32 A_Q24[ MAX_FIND_PITCH_LPC_ORDER ]; opus_int16 A_Q12[ MAX_FIND_PITCH_LPC_ORDER ]; SAVE_STACK; /******************************************/ /* Set up buffer lengths etc based on Fs */ /******************************************/ buf_len = psEnc->sCmn.la_pitch + psEnc->sCmn.frame_length + psEnc->sCmn.ltp_mem_length; /* Safety check */ silk_assert( buf_len >= psEnc->sCmn.pitch_LPC_win_length ); x_buf = x - psEnc->sCmn.ltp_mem_length; /*************************************/ /* Estimate LPC AR coefficients */ /*************************************/ /* Calculate windowed signal */ ALLOC( Wsig, psEnc->sCmn.pitch_LPC_win_length, opus_int16 ); /* First LA_LTP samples */ x_buf_ptr = x_buf + buf_len - psEnc->sCmn.pitch_LPC_win_length; Wsig_ptr = Wsig; silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 1, psEnc->sCmn.la_pitch ); /* Middle un - windowed samples */ Wsig_ptr += psEnc->sCmn.la_pitch; x_buf_ptr += psEnc->sCmn.la_pitch; silk_memcpy( Wsig_ptr, x_buf_ptr, ( psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ) ) * sizeof( opus_int16 ) ); /* Last LA_LTP samples */ Wsig_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ); x_buf_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ); silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 2, psEnc->sCmn.la_pitch ); /* Calculate autocorrelation sequence */ silk_autocorr( auto_corr, &scale, Wsig, psEnc->sCmn.pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1, arch ); /* Add white noise, as fraction of energy */ auto_corr[ 0 ] = silk_SMLAWB( auto_corr[ 0 ], auto_corr[ 0 ], SILK_FIX_CONST( FIND_PITCH_WHITE_NOISE_FRACTION, 16 ) ) + 1; /* Calculate the reflection coefficients using schur */ res_nrg = silk_schur( rc_Q15, auto_corr, psEnc->sCmn.pitchEstimationLPCOrder ); /* Prediction gain */ psEncCtrl->predGain_Q16 = silk_DIV32_varQ( auto_corr[ 0 ], silk_max_int( res_nrg, 1 ), 16 ); /* Convert reflection coefficients to prediction coefficients */ silk_k2a( A_Q24, rc_Q15, psEnc->sCmn.pitchEstimationLPCOrder ); /* Convert From 32 bit Q24 to 16 bit Q12 coefs */ for( i = 0; i < psEnc->sCmn.pitchEstimationLPCOrder; i++ ) { A_Q12[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT( A_Q24[ i ], 12 ) ); } /* Do BWE */ silk_bwexpander( A_Q12, psEnc->sCmn.pitchEstimationLPCOrder, SILK_FIX_CONST( FIND_PITCH_BANDWIDTH_EXPANSION, 16 ) ); /*****************************************/ /* LPC analysis filtering */ /*****************************************/ silk_LPC_analysis_filter( res, x_buf, A_Q12, buf_len, psEnc->sCmn.pitchEstimationLPCOrder ); if( psEnc->sCmn.indices.signalType != TYPE_NO_VOICE_ACTIVITY && psEnc->sCmn.first_frame_after_reset == 0 ) { /* Threshold for pitch estimator */ thrhld_Q13 = SILK_FIX_CONST( 0.6, 13 ); thrhld_Q13 = silk_SMLABB( thrhld_Q13, SILK_FIX_CONST( -0.004, 13 ), psEnc->sCmn.pitchEstimationLPCOrder ); thrhld_Q13 = silk_SMLAWB( thrhld_Q13, SILK_FIX_CONST( -0.1, 21 ), psEnc->sCmn.speech_activity_Q8 ); thrhld_Q13 = silk_SMLABB( thrhld_Q13, SILK_FIX_CONST( -0.15, 13 ), silk_RSHIFT( psEnc->sCmn.prevSignalType, 1 ) ); thrhld_Q13 = silk_SMLAWB( thrhld_Q13, SILK_FIX_CONST( -0.1, 14 ), psEnc->sCmn.input_tilt_Q15 ); thrhld_Q13 = silk_SAT16( thrhld_Q13 ); /*****************************************/ /* Call pitch estimator */ /*****************************************/ if( silk_pitch_analysis_core( res, psEncCtrl->pitchL, &psEnc->sCmn.indices.lagIndex, &psEnc->sCmn.indices.contourIndex, &psEnc->LTPCorr_Q15, psEnc->sCmn.prevLag, psEnc->sCmn.pitchEstimationThreshold_Q16, (opus_int)thrhld_Q13, psEnc->sCmn.fs_kHz, psEnc->sCmn.pitchEstimationComplexity, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch) == 0 ) { psEnc->sCmn.indices.signalType = TYPE_VOICED; } else { psEnc->sCmn.indices.signalType = TYPE_UNVOICED; } } else { silk_memset( psEncCtrl->pitchL, 0, sizeof( psEncCtrl->pitchL ) ); psEnc->sCmn.indices.lagIndex = 0; psEnc->sCmn.indices.contourIndex = 0; psEnc->LTPCorr_Q15 = 0; } RESTORE_STACK; }
/* Compute reflection coefficients from input signal */ void silk_burg_modified( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */ const opus_int subfr_length, /* I Input signal subframe length (incl. D preceeding samples) */ const opus_int nb_subfr, /* I Number of subframes stacked in x */ const opus_int32 WhiteNoiseFrac_Q32, /* I Fraction added to zero-lag autocorrelation */ const opus_int D /* I Order */ ) { opus_int k, n, s, lz, rshifts, rshifts_extra; opus_int32 C0, num, nrg, rc_Q31, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; const opus_int16 *x_ptr; opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); silk_assert( nb_subfr <= MAX_NB_SUBFR ); /* Compute autocorrelations, added over subframes */ silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length ); if( rshifts > MAX_RSHIFTS ) { C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS ); silk_assert( C0 > 0 ); rshifts = MAX_RSHIFTS; } else { lz = silk_CLZ32( C0 ) - 1; rshifts_extra = N_BITS_HEAD_ROOM - lz; if( rshifts_extra > 0 ) { rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts ); C0 = silk_RSHIFT32( C0, rshifts_extra ); } else { rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts ); C0 = silk_LSHIFT32( C0, -rshifts_extra ); } rshifts += rshifts_extra; } silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); if( rshifts > 0 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts ); } } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts ); } } } silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( WhiteNoiseFrac_Q32, C0 ) + 1; /* Q(-rshifts)*/ for( n = 0; n < D; n++ ) { /* Update first row of correlation matrix (without first element) */ /* Update last row of correlation matrix (without last element, stored in reversed order) */ /* Update C * Af */ /* Update C * flipud(Af) (stored in reversed order) */ if( rshifts > -2 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts)*/ x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts)*/ tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16)*/ tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16)*/ for( k = 0; k < n; k++ ) { C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts )*/ C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts )*/ Atmp_QA = Af_QA[ k ]; tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16)*/ tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16)*/ } tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts)*/ tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts)*/ for( k = 0; k <= n; k++ ) { CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift )*/ CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift )*/ } } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts )*/ x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts )*/ tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17*/ tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17*/ for( k = 0; k < n; k++ ) { C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts )*/ C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts )*/ Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17*/ tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17*/ tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17*/ } tmp1 = -tmp1; /* Q17*/ tmp2 = -tmp2; /* Q17*/ for( k = 0; k <= n; k++ ) { CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1, silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift )*/ CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2, silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift )*/ } } } /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ tmp1 = C_first_row[ n ]; /* Q( -rshifts )*/ tmp2 = C_last_row[ n ]; /* Q( -rshifts )*/ num = 0; /* Q( -rshifts )*/ nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts )*/ for( k = 0; k < n; k++ ) { Atmp_QA = Af_QA[ k ]; lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1; lz = silk_min( 32 - QA, lz ); Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz )*/ tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts )*/ tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts )*/ num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts )*/ nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ), Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts )*/ } CAf[ n + 1 ] = tmp1; /* Q( -rshifts )*/ CAb[ n + 1 ] = tmp2; /* Q( -rshifts )*/ num = silk_ADD32( num, tmp2 ); /* Q( -rshifts )*/ num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts )*/ /* Calculate the next order reflection (parcor) coefficient */ if( silk_abs( num ) < nrg ) { rc_Q31 = silk_DIV32_varQ( num, nrg, 31 ); } else { /* Negative energy or ratio too high; set remaining coefficients to zero and exit loop */ silk_memset( &Af_QA[ n ], 0, ( D - n ) * sizeof( opus_int32 ) ); silk_assert( 0 ); break; } /* Update the AR coefficients */ for( k = 0; k < (n + 1) >> 1; k++ ) { tmp1 = Af_QA[ k ]; /* QA*/ tmp2 = Af_QA[ n - k - 1 ]; /* QA*/ Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA*/ Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA*/ } Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA*/ /* Update C * Af and C * Ab */ for( k = 0; k <= n + 1; k++ ) { tmp1 = CAf[ k ]; /* Q( -rshifts )*/ tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts )*/ CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts )*/ CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts )*/ } } /* Return residual energy */ nrg = CAf[ 0 ]; /* Q( -rshifts )*/ tmp1 = 1 << 16; /* Q16*/ for( k = 0; k < D; k++ ) { Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16*/ nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts )*/ tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16*/ A_Q16[ k ] = -Atmp1; } *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( WhiteNoiseFrac_Q32, C0 ), -tmp1 ); /* Q( -rshifts )*/ *res_nrg_Q = -rshifts; }
/* encControl->payloadSize_ms is set to */ opus_int silk_Encode( /* O Returns error code */ void *encState, /* I/O State */ silk_EncControlStruct *encControl, /* I Control status */ const opus_int16 *samplesIn, /* I Speech sample input vector */ opus_int nSamplesIn, /* I Number of samples in input vector */ ec_enc *psRangeEnc, /* I/O Compressor data structure */ opus_int32 *nBytesOut, /* I/O Number of bytes in payload (input: Max bytes) */ const opus_int prefillFlag /* I Flag to indicate prefilling buffers no coding */ ) { opus_int n, i, nBits, flags, tmp_payloadSize_ms = 0, tmp_complexity = 0, ret = 0; opus_int nSamplesToBuffer, nSamplesToBufferMax, nBlocksOf10ms; opus_int nSamplesFromInput = 0, nSamplesFromInputMax; opus_int speech_act_thr_for_switch_Q8; opus_int32 TargetRate_bps, MStargetRates_bps[ 2 ], channelRate_bps, LBRR_symbol, sum; silk_encoder *psEnc = ( silk_encoder * )encState; VARDECL( opus_int16, buf ); opus_int transition, curr_block, tot_blocks; SAVE_STACK; if (encControl->reducedDependency) { psEnc->state_Fxx[0].sCmn.first_frame_after_reset = 1; psEnc->state_Fxx[1].sCmn.first_frame_after_reset = 1; } psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded = psEnc->state_Fxx[ 1 ].sCmn.nFramesEncoded = 0; /* Check values in encoder control structure */ if( ( ret = check_control_input( encControl ) ) != 0 ) { silk_assert( 0 ); RESTORE_STACK; return ret; } encControl->switchReady = 0; if( encControl->nChannelsInternal > psEnc->nChannelsInternal ) { /* Mono -> Stereo transition: init state of second channel and stereo state */ ret += silk_init_encoder( &psEnc->state_Fxx[ 1 ], psEnc->state_Fxx[ 0 ].sCmn.arch ); silk_memset( psEnc->sStereo.pred_prev_Q13, 0, sizeof( psEnc->sStereo.pred_prev_Q13 ) ); silk_memset( psEnc->sStereo.sSide, 0, sizeof( psEnc->sStereo.sSide ) ); psEnc->sStereo.mid_side_amp_Q0[ 0 ] = 0; psEnc->sStereo.mid_side_amp_Q0[ 1 ] = 1; psEnc->sStereo.mid_side_amp_Q0[ 2 ] = 0; psEnc->sStereo.mid_side_amp_Q0[ 3 ] = 1; psEnc->sStereo.width_prev_Q14 = 0; psEnc->sStereo.smth_width_Q14 = SILK_FIX_CONST( 1, 14 ); if( psEnc->nChannelsAPI == 2 ) { silk_memcpy( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, sizeof( silk_resampler_state_struct ) ); silk_memcpy( &psEnc->state_Fxx[ 1 ].sCmn.In_HP_State, &psEnc->state_Fxx[ 0 ].sCmn.In_HP_State, sizeof( psEnc->state_Fxx[ 1 ].sCmn.In_HP_State ) ); } } transition = (encControl->payloadSize_ms != psEnc->state_Fxx[ 0 ].sCmn.PacketSize_ms) || (psEnc->nChannelsInternal != encControl->nChannelsInternal); psEnc->nChannelsAPI = encControl->nChannelsAPI; 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 ) { silk_assert( 0 ); RESTORE_STACK; return SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES; } /* Reset Encoder */ for( n = 0; n < encControl->nChannelsInternal; n++ ) { ret = silk_init_encoder( &psEnc->state_Fxx[ n ], psEnc->state_Fxx[ n ].sCmn.arch ); silk_assert( !ret ); } tmp_payloadSize_ms = encControl->payloadSize_ms; encControl->payloadSize_ms = 10; tmp_complexity = encControl->complexity; encControl->complexity = 0; for( n = 0; n < encControl->nChannelsInternal; n++ ) { psEnc->state_Fxx[ n ].sCmn.controlled_since_last_payload = 0; psEnc->state_Fxx[ n ].sCmn.prefillFlag = 1; } } else { /* Only accept input lengths that are a multiple of 10 ms */ if( nBlocksOf10ms * encControl->API_sampleRate != 100 * nSamplesIn || nSamplesIn < 0 ) { silk_assert( 0 ); RESTORE_STACK; return SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES; } /* Make sure no more than one packet can be produced */ if( 1000 * (opus_int32)nSamplesIn > encControl->payloadSize_ms * encControl->API_sampleRate ) { silk_assert( 0 ); RESTORE_STACK; return SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES; } } TargetRate_bps = silk_RSHIFT32( encControl->bitRate, encControl->nChannelsInternal - 1 ); for( n = 0; n < encControl->nChannelsInternal; n++ ) { /* Force the side channel to the same rate as the mid */ 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 ); RESTORE_STACK; return ret; } 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; } } psEnc->state_Fxx[ n ].sCmn.inDTX = psEnc->state_Fxx[ n ].sCmn.useDTX; } silk_assert( encControl->nChannelsInternal == 1 || psEnc->state_Fxx[ 0 ].sCmn.fs_kHz == psEnc->state_Fxx[ 1 ].sCmn.fs_kHz ); /* Input buffering/resampling and encoding */ nSamplesToBufferMax = 10 * nBlocksOf10ms * psEnc->state_Fxx[ 0 ].sCmn.fs_kHz; nSamplesFromInputMax = silk_DIV32_16( nSamplesToBufferMax * psEnc->state_Fxx[ 0 ].sCmn.API_fs_Hz, psEnc->state_Fxx[ 0 ].sCmn.fs_kHz * 1000 ); ALLOC( buf, nSamplesFromInputMax, opus_int16 ); while( 1 ) { nSamplesToBuffer = psEnc->state_Fxx[ 0 ].sCmn.frame_length - psEnc->state_Fxx[ 0 ].sCmn.inputBufIx; nSamplesToBuffer = silk_min( nSamplesToBuffer, nSamplesToBufferMax ); 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 ) { opus_int id = psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded; for( n = 0; n < nSamplesFromInput; n++ ) { buf[ n ] = samplesIn[ 2 * n ]; } /* Making sure to start both resamplers from the same state when switching from mono to stereo */ if( psEnc->nPrevChannelsInternal == 1 && id==0 ) { silk_memcpy( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, sizeof(psEnc->state_Fxx[ 1 ].sCmn.resampler_state)); } ret += silk_resampler( &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); psEnc->state_Fxx[ 0 ].sCmn.inputBufIx += nSamplesToBuffer; nSamplesToBuffer = psEnc->state_Fxx[ 1 ].sCmn.frame_length - psEnc->state_Fxx[ 1 ].sCmn.inputBufIx; nSamplesToBuffer = silk_min( nSamplesToBuffer, 10 * nBlocksOf10ms * psEnc->state_Fxx[ 1 ].sCmn.fs_kHz ); for( n = 0; n < nSamplesFromInput; n++ ) { buf[ n ] = samplesIn[ 2 * n + 1 ]; } ret += silk_resampler( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ psEnc->state_Fxx[ 1 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); psEnc->state_Fxx[ 1 ].sCmn.inputBufIx += nSamplesToBuffer; } else if( encControl->nChannelsAPI == 2 && encControl->nChannelsInternal == 1 ) { /* Combine left and right channels before resampling */ for( n = 0; n < nSamplesFromInput; n++ ) { sum = samplesIn[ 2 * n ] + samplesIn[ 2 * n + 1 ]; buf[ n ] = (opus_int16)silk_RSHIFT_ROUND( sum, 1 ); } ret += silk_resampler( &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); /* On the first mono frame, average the results for the two resampler states */ if( psEnc->nPrevChannelsInternal == 2 && psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded == 0 ) { ret += silk_resampler( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ psEnc->state_Fxx[ 1 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); for( n = 0; n < psEnc->state_Fxx[ 0 ].sCmn.frame_length; n++ ) { psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx+n+2 ] = silk_RSHIFT(psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx+n+2 ] + psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ psEnc->state_Fxx[ 1 ].sCmn.inputBufIx+n+2 ], 1); } } psEnc->state_Fxx[ 0 ].sCmn.inputBufIx += nSamplesToBuffer; } else { silk_assert( encControl->nChannelsAPI == 1 && encControl->nChannelsInternal == 1 ); silk_memcpy(buf, samplesIn, nSamplesFromInput*sizeof(opus_int16)); ret += silk_resampler( &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); psEnc->state_Fxx[ 0 ].sCmn.inputBufIx += nSamplesToBuffer; } samplesIn += nSamplesFromInput * encControl->nChannelsAPI; nSamplesIn -= nSamplesFromInput; /* Default */ psEnc->allowBandwidthSwitch = 0; /* Silk encoder */ if( psEnc->state_Fxx[ 0 ].sCmn.inputBufIx >= psEnc->state_Fxx[ 0 ].sCmn.frame_length ) { /* Enough data in input buffer, so encode */ silk_assert( psEnc->state_Fxx[ 0 ].sCmn.inputBufIx == psEnc->state_Fxx[ 0 ].sCmn.frame_length ); silk_assert( encControl->nChannelsInternal == 1 || psEnc->state_Fxx[ 1 ].sCmn.inputBufIx == psEnc->state_Fxx[ 1 ].sCmn.frame_length ); /* Deal with LBRR data */ if( psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded == 0 && !prefillFlag ) { /* Create space at start of payload for VAD and FEC flags */ opus_uint8 iCDF[ 2 ] = { 0, 0 }; iCDF[ 0 ] = 256 - silk_RSHIFT( 256, ( psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket + 1 ) * encControl->nChannelsInternal ); ec_enc_icdf( psRangeEnc, 0, iCDF, 8 ); /* Encode any LBRR data from previous packet */ /* Encode LBRR flags */ for( n = 0; n < encControl->nChannelsInternal; n++ ) { LBRR_symbol = 0; for( i = 0; i < psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket; i++ ) { LBRR_symbol |= silk_LSHIFT( psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i ], i ); } psEnc->state_Fxx[ n ].sCmn.LBRR_flag = LBRR_symbol > 0 ? 1 : 0; if( LBRR_symbol && psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket > 1 ) { ec_enc_icdf( psRangeEnc, LBRR_symbol - 1, silk_LBRR_flags_iCDF_ptr[ psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket - 2 ], 8 ); } } /* Code LBRR indices and excitation signals */ for( i = 0; i < psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket; i++ ) { for( n = 0; n < encControl->nChannelsInternal; n++ ) { if( psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i ] ) { opus_int condCoding; if( encControl->nChannelsInternal == 2 && n == 0 ) { silk_stereo_encode_pred( psRangeEnc, psEnc->sStereo.predIx[ i ] ); /* For LBRR data there's no need to code the mid-only flag if the side-channel LBRR flag is set */ if( psEnc->state_Fxx[ 1 ].sCmn.LBRR_flags[ i ] == 0 ) { silk_stereo_encode_mid_only( psRangeEnc, psEnc->sStereo.mid_only_flags[ i ] ); } } /* Use conditional coding if previous frame available */ if( i > 0 && psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i - 1 ] ) { condCoding = CODE_CONDITIONALLY; } else { condCoding = CODE_INDEPENDENTLY; } silk_encode_indices( &psEnc->state_Fxx[ n ].sCmn, psRangeEnc, i, 1, condCoding ); silk_encode_pulses( psRangeEnc, psEnc->state_Fxx[ n ].sCmn.indices_LBRR[i].signalType, psEnc->state_Fxx[ n ].sCmn.indices_LBRR[i].quantOffsetType, psEnc->state_Fxx[ n ].sCmn.pulses_LBRR[ i ], psEnc->state_Fxx[ n ].sCmn.frame_length ); } } } /* Reset LBRR flags */ for( n = 0; n < encControl->nChannelsInternal; n++ ) { silk_memset( psEnc->state_Fxx[ n ].sCmn.LBRR_flags, 0, sizeof( psEnc->state_Fxx[ n ].sCmn.LBRR_flags ) ); } psEnc->nBitsUsedLBRR = ec_tell( psRangeEnc ); } silk_HP_variable_cutoff( psEnc->state_Fxx ); /* Total target bits for packet */ nBits = silk_DIV32_16( silk_MUL( encControl->bitRate, encControl->payloadSize_ms ), 1000 ); /* Subtract bits used for LBRR */ if( !prefillFlag ) { nBits -= psEnc->nBitsUsedLBRR; } /* Divide by number of uncoded frames left in packet */ nBits = silk_DIV32_16( nBits, psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket ); /* Convert to bits/second */ if( encControl->payloadSize_ms == 10 ) { TargetRate_bps = silk_SMULBB( nBits, 100 ); } else { TargetRate_bps = silk_SMULBB( nBits, 50 ); } /* Subtract fraction of bits in excess of target in previous frames and packets */ TargetRate_bps -= silk_DIV32_16( silk_MUL( psEnc->nBitsExceeded, 1000 ), BITRESERVOIR_DECAY_TIME_MS ); if( !prefillFlag && psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded > 0 ) { /* Compare actual vs target bits so far in this packet */ opus_int32 bitsBalance = ec_tell( psRangeEnc ) - psEnc->nBitsUsedLBRR - nBits * psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded; TargetRate_bps -= silk_DIV32_16( silk_MUL( bitsBalance, 1000 ), BITRESERVOIR_DECAY_TIME_MS ); } /* Never exceed input bitrate */ TargetRate_bps = silk_LIMIT( TargetRate_bps, encControl->bitRate, 5000 ); /* Convert Left/Right to Mid/Side */ if( encControl->nChannelsInternal == 2 ) { silk_stereo_LR_to_MS( &psEnc->sStereo, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ 2 ], &psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ 2 ], psEnc->sStereo.predIx[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ], &psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ], MStargetRates_bps, TargetRate_bps, psEnc->state_Fxx[ 0 ].sCmn.speech_activity_Q8, encControl->toMono, psEnc->state_Fxx[ 0 ].sCmn.fs_kHz, psEnc->state_Fxx[ 0 ].sCmn.frame_length ); if( psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] == 0 ) { /* Reset side channel encoder memory for first frame with side coding */ if( psEnc->prev_decode_only_middle == 1 ) { silk_memset( &psEnc->state_Fxx[ 1 ].sShape, 0, sizeof( psEnc->state_Fxx[ 1 ].sShape ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sPrefilt, 0, sizeof( psEnc->state_Fxx[ 1 ].sPrefilt ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sNSQ, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sNSQ ) ); silk_memset( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15 ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State ) ); psEnc->state_Fxx[ 1 ].sCmn.prevLag = 100; psEnc->state_Fxx[ 1 ].sCmn.sNSQ.lagPrev = 100; psEnc->state_Fxx[ 1 ].sShape.LastGainIndex = 10; psEnc->state_Fxx[ 1 ].sCmn.prevSignalType = TYPE_NO_VOICE_ACTIVITY; psEnc->state_Fxx[ 1 ].sCmn.sNSQ.prev_gain_Q16 = 65536; psEnc->state_Fxx[ 1 ].sCmn.first_frame_after_reset = 1; } silk_encode_do_VAD_Fxx( &psEnc->state_Fxx[ 1 ] ); } else { psEnc->state_Fxx[ 1 ].sCmn.VAD_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] = 0; } if( !prefillFlag ) { silk_stereo_encode_pred( psRangeEnc, psEnc->sStereo.predIx[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] ); if( psEnc->state_Fxx[ 1 ].sCmn.VAD_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] == 0 ) { silk_stereo_encode_mid_only( psRangeEnc, psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] ); } } } else { /* Buffering */ silk_memcpy( psEnc->state_Fxx[ 0 ].sCmn.inputBuf, psEnc->sStereo.sMid, 2 * sizeof( opus_int16 ) ); silk_memcpy( psEnc->sStereo.sMid, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.frame_length ], 2 * sizeof( opus_int16 ) ); } silk_encode_do_VAD_Fxx( &psEnc->state_Fxx[ 0 ] ); /* Encode */ for( n = 0; n < encControl->nChannelsInternal; n++ ) { opus_int maxBits, useCBR; /* Handling rate constraints */ maxBits = encControl->maxBits; 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 ) { opus_int condCoding; silk_control_SNR( &psEnc->state_Fxx[ n ].sCmn, channelRate_bps ); /* Use independent coding if no previous frame available */ if( psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded - n <= 0 ) { condCoding = CODE_INDEPENDENTLY; } else if( n > 0 && psEnc->prev_decode_only_middle ) { /* If we skipped a side frame in this packet, we don't need LTP scaling; the LTP state is well-defined. */ condCoding = CODE_INDEPENDENTLY_NO_LTP_SCALING; } else { condCoding = CODE_CONDITIONALLY; } if( ( ret = silk_encode_frame_Fxx( &psEnc->state_Fxx[ n ], nBytesOut, psRangeEnc, condCoding, maxBits, useCBR ) ) != 0 ) { silk_assert( 0 ); } } psEnc->state_Fxx[ n ].sCmn.controlled_since_last_payload = 0; psEnc->state_Fxx[ n ].sCmn.inputBufIx = 0; psEnc->state_Fxx[ n ].sCmn.nFramesEncoded++; } psEnc->prev_decode_only_middle = psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded - 1 ]; /* Insert VAD and FEC flags at beginning of bitstream */ if( *nBytesOut > 0 && psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded == psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket) { flags = 0; for( n = 0; n < encControl->nChannelsInternal; n++ ) { for( i = 0; i < psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket; i++ ) { flags = silk_LSHIFT( flags, 1 ); flags |= psEnc->state_Fxx[ n ].sCmn.VAD_flags[ i ]; } flags = silk_LSHIFT( flags, 1 ); flags |= psEnc->state_Fxx[ n ].sCmn.LBRR_flag; } if( !prefillFlag ) { ec_enc_patch_initial_bits( psRangeEnc, flags, ( psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket + 1 ) * encControl->nChannelsInternal ); } /* Return zero bytes if all channels DTXed */ if( psEnc->state_Fxx[ 0 ].sCmn.inDTX && ( encControl->nChannelsInternal == 1 || psEnc->state_Fxx[ 1 ].sCmn.inDTX ) ) { *nBytesOut = 0; } psEnc->nBitsExceeded += *nBytesOut * 8; psEnc->nBitsExceeded -= silk_DIV32_16( silk_MUL( encControl->bitRate, encControl->payloadSize_ms ), 1000 ); psEnc->nBitsExceeded = silk_LIMIT( psEnc->nBitsExceeded, 0, 10000 ); /* Update flag indicating if bandwidth switching is allowed */ speech_act_thr_for_switch_Q8 = (opus_int) silk_SMLAWB( SILK_FIX_CONST( SPEECH_ACTIVITY_DTX_THRES, 8 ), SILK_FIX_CONST( ( 1 - SPEECH_ACTIVITY_DTX_THRES ) / MAX_BANDWIDTH_SWITCH_DELAY_MS, 16 + 8 ), psEnc->timeSinceSwitchAllowed_ms ); if( psEnc->state_Fxx[ 0 ].sCmn.speech_activity_Q8 < speech_act_thr_for_switch_Q8 ) { psEnc->allowBandwidthSwitch = 1; psEnc->timeSinceSwitchAllowed_ms = 0; } else { psEnc->allowBandwidthSwitch = 0; psEnc->timeSinceSwitchAllowed_ms += encControl->payloadSize_ms; } } if( nSamplesIn == 0 ) { break; } } else { break; } curr_block++; } psEnc->nPrevChannelsInternal = encControl->nChannelsInternal; encControl->allowBandwidthSwitch = psEnc->allowBandwidthSwitch; encControl->inWBmodeWithoutVariableLP = psEnc->state_Fxx[ 0 ].sCmn.fs_kHz == 16 && psEnc->state_Fxx[ 0 ].sCmn.sLP.mode == 0; encControl->internalSampleRate = silk_SMULBB( psEnc->state_Fxx[ 0 ].sCmn.fs_kHz, 1000 ); encControl->stereoWidth_Q14 = encControl->toMono ? 0 : psEnc->sStereo.smth_width_Q14; if( prefillFlag ) { encControl->payloadSize_ms = tmp_payloadSize_ms; encControl->complexity = tmp_complexity; for( n = 0; n < encControl->nChannelsInternal; n++ ) { psEnc->state_Fxx[ n ].sCmn.controlled_since_last_payload = 0; psEnc->state_Fxx[ n ].sCmn.prefillFlag = 0; } } RESTORE_STACK; return ret; }
/* Compute reflection coefficients from input signal */ silk_float silk_burg_modified_FLP( /* O returns residual energy */ silk_float A[], /* O prediction coefficients (length order) */ const silk_float x[], /* I input signal, length: nb_subfr*(D+L_sub) */ const silk_float minInvGain, /* I minimum inverse prediction gain */ const opus_int subfr_length, /* I input signal subframe length (incl. D preceding samples) */ const opus_int nb_subfr, /* I number of subframes stacked in x */ const opus_int D /* I order */ ) { opus_int k, n, s, reached_max_gain; double C0, invGain, num, nrg_f, nrg_b, rc, Atmp, tmp1, tmp2; const silk_float *x_ptr; double C_first_row[ SILK_MAX_ORDER_LPC ], C_last_row[ SILK_MAX_ORDER_LPC ]; double CAf[ SILK_MAX_ORDER_LPC + 1 ], CAb[ SILK_MAX_ORDER_LPC + 1 ]; double Af[ SILK_MAX_ORDER_LPC ]; silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); /* Compute autocorrelations, added over subframes */ C0 = silk_energy_FLP( x, nb_subfr * subfr_length ); silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( double ) ); for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += silk_inner_product_FLP( x_ptr, x_ptr + n, subfr_length - n ); } } silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( double ) ); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + FIND_LPC_COND_FAC * C0 + 1e-9f; invGain = 1.0f; reached_max_gain = 0; for( n = 0; n < D; n++ ) { /* Update first row of correlation matrix (without first element) */ /* Update last row of correlation matrix (without last element, stored in reversed order) */ /* Update C * Af */ /* Update C * flipud(Af) (stored in reversed order) */ for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; tmp1 = x_ptr[ n ]; tmp2 = x_ptr[ subfr_length - n - 1 ]; for( k = 0; k < n; k++ ) { C_first_row[ k ] -= x_ptr[ n ] * x_ptr[ n - k - 1 ]; C_last_row[ k ] -= x_ptr[ subfr_length - n - 1 ] * x_ptr[ subfr_length - n + k ]; Atmp = Af[ k ]; tmp1 += x_ptr[ n - k - 1 ] * Atmp; tmp2 += x_ptr[ subfr_length - n + k ] * Atmp; } for( k = 0; k <= n; k++ ) { CAf[ k ] -= tmp1 * x_ptr[ n - k ]; CAb[ k ] -= tmp2 * x_ptr[ subfr_length - n + k - 1 ]; } } tmp1 = C_first_row[ n ]; tmp2 = C_last_row[ n ]; for( k = 0; k < n; k++ ) { Atmp = Af[ k ]; tmp1 += C_last_row[ n - k - 1 ] * Atmp; tmp2 += C_first_row[ n - k - 1 ] * Atmp; } CAf[ n + 1 ] = tmp1; CAb[ n + 1 ] = tmp2; /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ num = CAb[ n + 1 ]; nrg_b = CAb[ 0 ]; nrg_f = CAf[ 0 ]; for( k = 0; k < n; k++ ) { Atmp = Af[ k ]; num += CAb[ n - k ] * Atmp; nrg_b += CAb[ k + 1 ] * Atmp; nrg_f += CAf[ k + 1 ] * Atmp; } silk_assert( nrg_f > 0.0 ); silk_assert( nrg_b > 0.0 ); /* Calculate the next order reflection (parcor) coefficient */ rc = -2.0 * num / ( nrg_f + nrg_b ); silk_assert( rc > -1.0 && rc < 1.0 ); /* Update inverse prediction gain */ tmp1 = invGain * ( 1.0 - rc * rc ); if( tmp1 <= minInvGain ) { /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ rc = sqrt( 1.0 - minInvGain / invGain ); if( num > 0 ) { /* Ensure adjusted reflection coefficients has the original sign */ rc = -rc; } invGain = minInvGain; reached_max_gain = 1; } else { invGain = tmp1; } /* Update the AR coefficients */ for( k = 0; k < (n + 1) >> 1; k++ ) { tmp1 = Af[ k ]; tmp2 = Af[ n - k - 1 ]; Af[ k ] = tmp1 + rc * tmp2; Af[ n - k - 1 ] = tmp2 + rc * tmp1; } Af[ n ] = rc; if( reached_max_gain ) { /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ for( k = n + 1; k < D; k++ ) { Af[ k ] = 0.0; } break; } /* Update C * Af and C * Ab */ for( k = 0; k <= n + 1; k++ ) { tmp1 = CAf[ k ]; CAf[ k ] += rc * CAb[ n - k + 1 ]; CAb[ n - k + 1 ] += rc * tmp1; } } if( reached_max_gain ) { /* Convert to silk_float */ for( k = 0; k < D; k++ ) { A[ k ] = (silk_float)( -Af[ k ] ); } /* Subtract energy of preceding samples from C0 */ for( s = 0; s < nb_subfr; s++ ) { C0 -= silk_energy_FLP( x + s * subfr_length, D ); } /* Approximate residual energy */ nrg_f = C0 * invGain; } else { /* Compute residual energy and store coefficients as silk_float */ nrg_f = CAf[ 0 ]; tmp1 = 1.0; for( k = 0; k < D; k++ ) { Atmp = Af[ k ]; nrg_f += CAf[ k + 1 ] * Atmp; tmp1 += Atmp * Atmp; A[ k ] = (silk_float)(-Atmp); } nrg_f -= FIND_LPC_COND_FAC * C0 * tmp1; } /* Return residual energy */ return (silk_float)nrg_f; }
void silk_find_pred_coefs_FIX( silk_encoder_state_FIX *psEnc, /* I/O encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ const opus_int16 res_pitch[], /* I Residual from pitch analysis */ const opus_int16 x[], /* I Speech signal */ opus_int condCoding /* I The type of conditional coding to use */ ) { opus_int i; opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ], Wght_Q15[ MAX_NB_SUBFR ]; opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const opus_int16 *x_ptr; opus_int16 *x_pre_ptr; VARDECL( opus_int16, LPC_in_pre ); opus_int32 tmp, min_gain_Q16, minInvGain_Q30; opus_int LTP_corrs_rshift[ MAX_NB_SUBFR ]; SAVE_STACK; /* weighting for weighted least squares */ min_gain_Q16 = silk_int32_MAX >> 6; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { min_gain_Q16 = silk_min( min_gain_Q16, psEncCtrl->Gains_Q16[ i ] ); } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { /* Divide to Q16 */ silk_assert( psEncCtrl->Gains_Q16[ i ] > 0 ); /* Invert and normalize gains, and ensure that maximum invGains_Q16 is within range of a 16 bit int */ invGains_Q16[ i ] = silk_DIV32_varQ( min_gain_Q16, psEncCtrl->Gains_Q16[ i ], 16 - 2 ); /* Ensure Wght_Q15 a minimum value 1 */ invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 363 ); /* Square the inverted gains */ silk_assert( invGains_Q16[ i ] == silk_SAT16( invGains_Q16[ i ] ) ); tmp = silk_SMULWB( invGains_Q16[ i ], invGains_Q16[ i ] ); Wght_Q15[ i ] = silk_RSHIFT( tmp, 1 ); /* Invert the inverted and normalized gains */ local_gains[ i ] = silk_DIV32( ( (opus_int32)1 << 16 ), invGains_Q16[ i ] ); } ALLOC( LPC_in_pre, psEnc->sCmn.nb_subfr * psEnc->sCmn.predictLPCOrder + psEnc->sCmn.frame_length, opus_int16 ); if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { VARDECL( opus_int32, WLTP ); /**********/ /* VOICED */ /**********/ silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); ALLOC( WLTP, psEnc->sCmn.nb_subfr * LTP_ORDER * LTP_ORDER, opus_int32 ); /* LTP analysis */ silk_find_LTP_FIX( psEncCtrl->LTPCoef_Q14, WLTP, &psEncCtrl->LTPredCodGain_Q7, res_pitch, psEncCtrl->pitchL, Wght_Q15, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length, LTP_corrs_rshift ); /* Quantize LTP gain parameters */ silk_quant_LTP_gains( psEncCtrl->LTPCoef_Q14, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, &psEnc->sCmn.sum_log_gain_Q7, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr); /* Control LTP scaling */ silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding ); /* Create LTP residual */ silk_LTP_analysis_filter_FIX( LPC_in_pre, x - psEnc->sCmn.predictLPCOrder, psEncCtrl->LTPCoef_Q14, psEncCtrl->pitchL, invGains_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); } else { /************/ /* UNVOICED */ /************/ /* Create signal with prepended subframes, scaled by inverse gains */ x_ptr = x - psEnc->sCmn.predictLPCOrder; x_pre_ptr = LPC_in_pre; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { silk_scale_copy_vector16( x_pre_ptr, x_ptr, invGains_Q16[ i ], psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder ); x_pre_ptr += psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder; x_ptr += psEnc->sCmn.subfr_length; } silk_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( opus_int16 ) ); psEncCtrl->LTPredCodGain_Q7 = 0; psEnc->sCmn.sum_log_gain_Q7 = 0; } /* Limit on total predictive coding gain */ if( psEnc->sCmn.first_frame_after_reset ) { minInvGain_Q30 = SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN_AFTER_RESET, 30 ); } else { minInvGain_Q30 = silk_log2lin( silk_SMLAWB( 16 << 7, (opus_int32)psEncCtrl->LTPredCodGain_Q7, SILK_FIX_CONST( 1.0 / 3, 16 ) ) ); /* Q16 */ minInvGain_Q30 = silk_DIV32_varQ( minInvGain_Q30, silk_SMULWW( SILK_FIX_CONST( MAX_PREDICTION_POWER_GAIN, 0 ), silk_SMLAWB( SILK_FIX_CONST( 0.25, 18 ), SILK_FIX_CONST( 0.75, 18 ), psEncCtrl->coding_quality_Q14 ) ), 14 ); } /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ silk_find_LPC_FIX( &psEnc->sCmn, NLSF_Q15, LPC_in_pre, minInvGain_Q30 ); /* Quantize LSFs */ silk_process_NLSFs( &psEnc->sCmn, psEncCtrl->PredCoef_Q12, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 ); /* Calculate residual energy using quantized LPC coefficients */ 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 interpolation */ silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); RESTORE_STACK; }
/* Decode parameters from payload */ void silk_decode_parameters( silk_decoder_state *psDec, /* I/O State */ silk_decoder_control *psDecCtrl, /* I/O Decoder control */ opus_int condCoding /* I The type of conditional coding to use */ ) { opus_int i, k, Ix; opus_int16 pNLSF_Q15[ MAX_LPC_ORDER ], pNLSF0_Q15[ MAX_LPC_ORDER ]; const opus_int8 *cbk_ptr_Q7; /* Dequant Gains */ silk_gains_dequant( psDecCtrl->Gains_Q16, psDec->indices.GainsIndices, &psDec->LastGainIndex, condCoding == CODE_CONDITIONALLY, psDec->nb_subfr ); /****************/ /* Decode NLSFs */ /****************/ silk_NLSF_decode( pNLSF_Q15, psDec->indices.NLSFIndices, psDec->psNLSF_CB ); /* Convert NLSF parameters to AR prediction filter coefficients */ silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 1 ], pNLSF_Q15, psDec->LPC_order ); /* If just reset, e.g., because internal Fs changed, do not allow interpolation */ /* improves the case of packet loss in the first frame after a switch */ if( psDec->first_frame_after_reset == 1 ) { psDec->indices.NLSFInterpCoef_Q2 = 4; } if( psDec->indices.NLSFInterpCoef_Q2 < 4 ) { /* Calculation of the interpolated NLSF0 vector from the interpolation factor, */ /* the previous NLSF1, and the current NLSF1 */ for( i = 0; i < psDec->LPC_order; i++ ) { pNLSF0_Q15[ i ] = psDec->prevNLSF_Q15[ i ] + silk_RSHIFT( silk_MUL( psDec->indices.NLSFInterpCoef_Q2, pNLSF_Q15[ i ] - psDec->prevNLSF_Q15[ i ] ), 2 ); } /* Convert NLSF parameters to AR prediction filter coefficients */ silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 0 ], pNLSF0_Q15, psDec->LPC_order ); } else { /* Copy LPC coefficients for first half from second half */ silk_memcpy( psDecCtrl->PredCoef_Q12[ 0 ], psDecCtrl->PredCoef_Q12[ 1 ], psDec->LPC_order * sizeof( opus_int16 ) ); } silk_memcpy( psDec->prevNLSF_Q15, pNLSF_Q15, psDec->LPC_order * sizeof( opus_int16 ) ); /* After a packet loss do BWE of LPC coefs */ if( psDec->lossCnt ) { silk_bwexpander( psDecCtrl->PredCoef_Q12[ 0 ], psDec->LPC_order, BWE_AFTER_LOSS_Q16 ); silk_bwexpander( psDecCtrl->PredCoef_Q12[ 1 ], psDec->LPC_order, BWE_AFTER_LOSS_Q16 ); } if( psDec->indices.signalType == TYPE_VOICED ) { /*********************/ /* Decode pitch lags */ /*********************/ /* Decode pitch values */ silk_decode_pitch( psDec->indices.lagIndex, psDec->indices.contourIndex, psDecCtrl->pitchL, psDec->fs_kHz, psDec->nb_subfr ); /* Decode Codebook Index */ cbk_ptr_Q7 = silk_LTP_vq_ptrs_Q7[ psDec->indices.PERIndex ]; /* set pointer to start of codebook */ for( k = 0; k < psDec->nb_subfr; k++ ) { Ix = psDec->indices.LTPIndex[ k ]; for( i = 0; i < LTP_ORDER; i++ ) { psDecCtrl->LTPCoef_Q14[ k * LTP_ORDER + i ] = silk_LSHIFT( cbk_ptr_Q7[ Ix * LTP_ORDER + i ], 7 ); } } /**********************/ /* Decode LTP scaling */ /**********************/ Ix = psDec->indices.LTP_scaleIndex; psDecCtrl->LTP_scale_Q14 = silk_LTPScales_table_Q14[ Ix ]; } else { silk_memset( psDecCtrl->pitchL, 0, psDec->nb_subfr * sizeof( opus_int ) ); silk_memset( psDecCtrl->LTPCoef_Q14, 0, LTP_ORDER * psDec->nb_subfr * sizeof( opus_int16 ) ); psDec->indices.PERIndex = 0; psDecCtrl->LTP_scale_Q14 = 0; } }
/* Compute reflection coefficients from input signal */ void silk_burg_modified_sse4_1( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ const opus_int16 x[], /* I Input signal, length: nb_subfr * (D + subfr_length) */ const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ const opus_int nb_subfr, /* I Number of subframes stacked in x */ const opus_int D, /* I Order */ int arch /* I Run-time architecture */ ) { opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain; opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; const opus_int16 *x_ptr; opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 xcorr[ SILK_MAX_ORDER_LPC ]; __m128i FIRST_3210, LAST_3210, ATMP_3210, TMP1_3210, TMP2_3210, T1_3210, T2_3210, PTR_3210, SUBFR_3210, X1_3210, X2_3210; __m128i CONST1 = _mm_set1_epi32(1); silk_assert(subfr_length * nb_subfr <= MAX_FRAME_SIZE); /* Compute autocorrelations, added over subframes */ silk_sum_sqr_shift(&C0, &rshifts, x, nb_subfr * subfr_length); if(rshifts > MAX_RSHIFTS) { C0 = silk_LSHIFT32(C0, rshifts - MAX_RSHIFTS); silk_assert(C0 > 0); rshifts = MAX_RSHIFTS; } else { lz = silk_CLZ32(C0) - 1; rshifts_extra = N_BITS_HEAD_ROOM - lz; if(rshifts_extra > 0) { rshifts_extra = silk_min(rshifts_extra, MAX_RSHIFTS - rshifts); C0 = silk_RSHIFT32(C0, rshifts_extra); } else { rshifts_extra = silk_max(rshifts_extra, MIN_RSHIFTS - rshifts); C0 = silk_LSHIFT32(C0, -rshifts_extra); } rshifts += rshifts_extra; } CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */ silk_memset(C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof(opus_int32)); if(rshifts > 0) { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; for(n = 1; n < D + 1; n++) { C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64(x_ptr, x_ptr + n, subfr_length - n, arch), rshifts); } } } else { for(s = 0; s < nb_subfr; s++) { int i; opus_int32 d; x_ptr = x + s * subfr_length; celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch); for(n = 1; n < D + 1; n++) { for (i = n + subfr_length - D, d = 0; i < subfr_length; i++) d = MAC16_16(d, x_ptr[ i ], x_ptr[ i - n ]); xcorr[ n - 1 ] += d; } for(n = 1; n < D + 1; n++) { C_first_row[ n - 1 ] += silk_LSHIFT32(xcorr[ n - 1 ], -rshifts); } } } silk_memcpy(C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof(opus_int32)); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */ invGain_Q30 = (opus_int32)1 << 30; reached_max_gain = 0; for(n = 0; n < D; n++) { /* Update first row of correlation matrix (without first element) */ /* Update last row of correlation matrix (without last element, stored in reversed order) */ /* Update C * Af */ /* Update C * flipud(Af) (stored in reversed order) */ if(rshifts > -2) { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], 16 - rshifts); /* Q(16-rshifts) */ x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts); /* Q(16-rshifts) */ tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], QA - 16); /* Q(QA-16) */ tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16); /* Q(QA-16) */ for(k = 0; k < n; k++) { C_first_row[ k ] = silk_SMLAWB(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */ C_last_row[ k ] = silk_SMLAWB(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */ Atmp_QA = Af_QA[ k ]; tmp1 = silk_SMLAWB(tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ tmp2 = silk_SMLAWB(tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ]); /* Q(QA-16) */ } tmp1 = silk_LSHIFT32(-tmp1, 32 - QA - rshifts); /* Q(16-rshifts) */ tmp2 = silk_LSHIFT32(-tmp2, 32 - QA - rshifts); /* Q(16-rshifts) */ for(k = 0; k <= n; k++) { CAf[ k ] = silk_SMLAWB(CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q(-rshift) */ CAb[ k ] = silk_SMLAWB(CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ]); /* Q(-rshift) */ } } } else { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], -rshifts); /* Q(-rshifts) */ x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts); /* Q(-rshifts) */ tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], 17); /* Q17 */ tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 17); /* Q17 */ X1_3210 = _mm_set1_epi32(x1); X2_3210 = _mm_set1_epi32(x2); TMP1_3210 = _mm_setzero_si128(); TMP2_3210 = _mm_setzero_si128(); for(k = 0; k < n - 3; k += 4) { PTR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 1 - 3 ]); SUBFR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k ]); FIRST_3210 = _mm_loadu_si128((__m128i *)&C_first_row[ k ]); PTR_3210 = _mm_shuffle_epi32(PTR_3210, _MM_SHUFFLE(0, 1, 2, 3)); LAST_3210 = _mm_loadu_si128((__m128i *)&C_last_row[ k ]); ATMP_3210 = _mm_loadu_si128((__m128i *)&Af_QA[ k ]); T1_3210 = _mm_mullo_epi32(PTR_3210, X1_3210); T2_3210 = _mm_mullo_epi32(SUBFR_3210, X2_3210); ATMP_3210 = _mm_srai_epi32(ATMP_3210, 7); ATMP_3210 = _mm_add_epi32(ATMP_3210, CONST1); ATMP_3210 = _mm_srai_epi32(ATMP_3210, 1); FIRST_3210 = _mm_add_epi32(FIRST_3210, T1_3210); LAST_3210 = _mm_add_epi32(LAST_3210, T2_3210); PTR_3210 = _mm_mullo_epi32(ATMP_3210, PTR_3210); SUBFR_3210 = _mm_mullo_epi32(ATMP_3210, SUBFR_3210); _mm_storeu_si128((__m128i *)&C_first_row[ k ], FIRST_3210); _mm_storeu_si128((__m128i *)&C_last_row[ k ], LAST_3210); TMP1_3210 = _mm_add_epi32(TMP1_3210, PTR_3210); TMP2_3210 = _mm_add_epi32(TMP2_3210, SUBFR_3210); } TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_unpackhi_epi64(TMP1_3210, TMP1_3210)); TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_unpackhi_epi64(TMP2_3210, TMP2_3210)); TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_shufflelo_epi16(TMP1_3210, 0x0E)); TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_shufflelo_epi16(TMP2_3210, 0x0E)); tmp1 += _mm_cvtsi128_si32(TMP1_3210); tmp2 += _mm_cvtsi128_si32(TMP2_3210); for(; k < n; k++) { C_first_row[ k ] = silk_MLA(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */ C_last_row[ k ] = silk_MLA(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */ Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 17); /* Q17 */ tmp1 = silk_MLA(tmp1, x_ptr[ n - k - 1 ], Atmp1); /* Q17 */ tmp2 = silk_MLA(tmp2, x_ptr[ subfr_length - n + k ], Atmp1); /* Q17 */ } tmp1 = -tmp1; /* Q17 */ tmp2 = -tmp2; /* Q17 */ { __m128i xmm_tmp1, xmm_tmp2; __m128i xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1; __m128i xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1; xmm_tmp1 = _mm_set1_epi32(tmp1); xmm_tmp2 = _mm_set1_epi32(tmp2); for(k = 0; k <= n - 3; k += 4) { xmm_x_ptr_n_k_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 3 ]); xmm_x_ptr_sub_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k - 1 ]); xmm_x_ptr_n_k_x2x0 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 1, 2, 3)); xmm_x_ptr_n_k_x2x0 = _mm_slli_epi32(xmm_x_ptr_n_k_x2x0, -rshifts - 1); xmm_x_ptr_sub_x2x0 = _mm_slli_epi32(xmm_x_ptr_sub_x2x0, -rshifts - 1); /* equal shift right 4 bytes, xmm_x_ptr_n_k_x3x1 = _mm_srli_si128(xmm_x_ptr_n_k_x2x0, 4)*/ xmm_x_ptr_n_k_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 3, 2, 1)); xmm_x_ptr_sub_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_sub_x2x0, _MM_SHUFFLE(0, 3, 2, 1)); xmm_x_ptr_n_k_x2x0 = _mm_mul_epi32(xmm_x_ptr_n_k_x2x0, xmm_tmp1); xmm_x_ptr_n_k_x3x1 = _mm_mul_epi32(xmm_x_ptr_n_k_x3x1, xmm_tmp1); xmm_x_ptr_sub_x2x0 = _mm_mul_epi32(xmm_x_ptr_sub_x2x0, xmm_tmp2); xmm_x_ptr_sub_x3x1 = _mm_mul_epi32(xmm_x_ptr_sub_x3x1, xmm_tmp2); xmm_x_ptr_n_k_x2x0 = _mm_srli_epi64(xmm_x_ptr_n_k_x2x0, 16); xmm_x_ptr_n_k_x3x1 = _mm_slli_epi64(xmm_x_ptr_n_k_x3x1, 16); xmm_x_ptr_sub_x2x0 = _mm_srli_epi64(xmm_x_ptr_sub_x2x0, 16); xmm_x_ptr_sub_x3x1 = _mm_slli_epi64(xmm_x_ptr_sub_x3x1, 16); xmm_x_ptr_n_k_x2x0 = _mm_blend_epi16(xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1, 0xCC); xmm_x_ptr_sub_x2x0 = _mm_blend_epi16(xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1, 0xCC); X1_3210 = _mm_loadu_si128((__m128i *)&CAf[ k ]); PTR_3210 = _mm_loadu_si128((__m128i *)&CAb[ k ]); X1_3210 = _mm_add_epi32(X1_3210, xmm_x_ptr_n_k_x2x0); PTR_3210 = _mm_add_epi32(PTR_3210, xmm_x_ptr_sub_x2x0); _mm_storeu_si128((__m128i *)&CAf[ k ], X1_3210); _mm_storeu_si128((__m128i *)&CAb[ k ], PTR_3210); } for(; k <= n; k++) { CAf[ k ] = silk_SMLAWW(CAf[ k ], tmp1, silk_LSHIFT32((opus_int32)x_ptr[ n - k ], -rshifts - 1)); /* Q(-rshift) */ CAb[ k ] = silk_SMLAWW(CAb[ k ], tmp2, silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1)); /* Q(-rshift) */ } } } } /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ tmp1 = C_first_row[ n ]; /* Q(-rshifts) */ tmp2 = C_last_row[ n ]; /* Q(-rshifts) */ num = 0; /* Q(-rshifts) */ nrg = silk_ADD32(CAb[ 0 ], CAf[ 0 ]); /* Q(1-rshifts) */ for(k = 0; k < n; k++) { Atmp_QA = Af_QA[ k ]; lz = silk_CLZ32(silk_abs(Atmp_QA)) - 1; lz = silk_min(32 - QA, lz); Atmp1 = silk_LSHIFT32(Atmp_QA, lz); /* Q(QA + lz) */ tmp1 = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(C_last_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */ tmp2 = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(C_first_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */ num = silk_ADD_LSHIFT32(num, silk_SMMUL(CAb[ n - k ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */ nrg = silk_ADD_LSHIFT32(nrg, silk_SMMUL(silk_ADD32(CAb[ k + 1 ], CAf[ k + 1 ]), Atmp1), 32 - QA - lz); /* Q(1-rshifts) */ } CAf[ n + 1 ] = tmp1; /* Q(-rshifts) */ CAb[ n + 1 ] = tmp2; /* Q(-rshifts) */ num = silk_ADD32(num, tmp2); /* Q(-rshifts) */ num = silk_LSHIFT32(-num, 1); /* Q(1-rshifts) */ /* Calculate the next order reflection (parcor) coefficient */ if(silk_abs(num) < nrg) { rc_Q31 = silk_DIV32_varQ(num, nrg, 31); } else { rc_Q31 = (num > 0) ? silk_int32_MAX : silk_int32_MIN; } /* Update inverse prediction gain */ tmp1 = ((opus_int32)1 << 30) - silk_SMMUL(rc_Q31, rc_Q31); tmp1 = silk_LSHIFT(silk_SMMUL(invGain_Q30, tmp1), 2); if(tmp1 <= minInvGain_Q30) { /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ tmp2 = ((opus_int32)1 << 30) - silk_DIV32_varQ(minInvGain_Q30, invGain_Q30, 30); /* Q30 */ rc_Q31 = silk_SQRT_APPROX(tmp2); /* Q15 */ /* Newton-Raphson iteration */ rc_Q31 = silk_RSHIFT32(rc_Q31 + silk_DIV32(tmp2, rc_Q31), 1); /* Q15 */ rc_Q31 = silk_LSHIFT32(rc_Q31, 16); /* Q31 */ if(num < 0) { /* Ensure adjusted reflection coefficients has the original sign */ rc_Q31 = -rc_Q31; } invGain_Q30 = minInvGain_Q30; reached_max_gain = 1; } else { invGain_Q30 = tmp1; } /* Update the AR coefficients */ for(k = 0; k < (n + 1) >> 1; k++) { tmp1 = Af_QA[ k ]; /* QA */ tmp2 = Af_QA[ n - k - 1 ]; /* QA */ Af_QA[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* QA */ Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* QA */ } Af_QA[ n ] = silk_RSHIFT32(rc_Q31, 31 - QA); /* QA */ if(reached_max_gain) { /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ for(k = n + 1; k < D; k++) { Af_QA[ k ] = 0; } break; } /* Update C * Af and C * Ab */ for(k = 0; k <= n + 1; k++) { tmp1 = CAf[ k ]; /* Q(-rshifts) */ tmp2 = CAb[ n - k + 1 ]; /* Q(-rshifts) */ CAf[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* Q(-rshifts) */ CAb[ n - k + 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* Q(-rshifts) */ } } if(reached_max_gain) { for(k = 0; k < D; k++) { /* Scale coefficients */ A_Q16[ k ] = -silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); } /* Subtract energy of preceding samples from C0 */ if(rshifts > 0) { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; C0 -= (opus_int32)silk_RSHIFT64(silk_inner_prod16_aligned_64(x_ptr, x_ptr, D, arch), rshifts); } } else { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; C0 -= silk_LSHIFT32(silk_inner_prod_aligned(x_ptr, x_ptr, D, arch), -rshifts); } } /* Approximate residual energy */ *res_nrg = silk_LSHIFT(silk_SMMUL(invGain_Q30, C0), 2); *res_nrg_Q = -rshifts; } else { /* Return residual energy */ nrg = CAf[ 0 ]; /* Q(-rshifts) */ tmp1 = (opus_int32)1 << 16; /* Q16 */ for(k = 0; k < D; k++) { Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); /* Q16 */ nrg = silk_SMLAWW(nrg, CAf[ k + 1 ], Atmp1); /* Q(-rshifts) */ tmp1 = silk_SMLAWW(tmp1, Atmp1, Atmp1); /* Q16 */ A_Q16[ k ] = -Atmp1; } *res_nrg = silk_SMLAWW(nrg, silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0), -tmp1);/* Q(-rshifts) */ *res_nrg_Q = -rshifts; } }
/* Compute reflection coefficients from input signal */ void silk_burg_modified( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */ const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ const opus_int nb_subfr, /* I Number of subframes stacked in x */ const opus_int D /* I Order */ ) { opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain; opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; const opus_int16 *x_ptr; opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); /* Compute autocorrelations, added over subframes */ silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length ); if( rshifts > MAX_RSHIFTS ) { C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS ); silk_assert( C0 > 0 ); rshifts = MAX_RSHIFTS; } else { lz = silk_CLZ32( C0 ) - 1; rshifts_extra = N_BITS_HEAD_ROOM - lz; if( rshifts_extra > 0 ) { rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts ); C0 = silk_RSHIFT32( C0, rshifts_extra ); } else { rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts ); C0 = silk_LSHIFT32( C0, -rshifts_extra ); } rshifts += rshifts_extra; } CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); if( rshifts > 0 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts ); } } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts ); } } } silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ invGain_Q30 = (opus_int32)1 << 30; reached_max_gain = 0; for( n = 0; n < D; n++ ) { /* Update first row of correlation matrix (without first element) */ /* Update last row of correlation matrix (without last element, stored in reversed order) */ /* Update C * Af */ /* Update C * flipud(Af) (stored in reversed order) */ if( rshifts > -2 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts) */ x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts) */ tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16) */ tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16) */ for( k = 0; k < n; k++ ) { C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ Atmp_QA = Af_QA[ k ]; tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16) */ } tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts) */ tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts) */ for( k = 0; k <= n; k++ ) { CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift ) */ CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift ) */ } } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts ) */ x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts ) */ tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17 */ tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17 */ for( k = 0; k < n; k++ ) { C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */ tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */ tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */ } tmp1 = -tmp1; /* Q17 */ tmp2 = -tmp2; /* Q17 */ for( k = 0; k <= n; k++ ) { CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1, silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift ) */ CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2, silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */ } } } /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ tmp1 = C_first_row[ n ]; /* Q( -rshifts ) */ tmp2 = C_last_row[ n ]; /* Q( -rshifts ) */ num = 0; /* Q( -rshifts ) */ nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts ) */ for( k = 0; k < n; k++ ) { Atmp_QA = Af_QA[ k ]; lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1; lz = silk_min( 32 - QA, lz ); Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz ) */ tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ), Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts ) */ } CAf[ n + 1 ] = tmp1; /* Q( -rshifts ) */ CAb[ n + 1 ] = tmp2; /* Q( -rshifts ) */ num = silk_ADD32( num, tmp2 ); /* Q( -rshifts ) */ num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts ) */ /* Calculate the next order reflection (parcor) coefficient */ if( silk_abs( num ) < nrg ) { rc_Q31 = silk_DIV32_varQ( num, nrg, 31 ); } else { rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN; } /* Update inverse prediction gain */ tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 ); if( tmp1 <= minInvGain_Q30 ) { /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */ rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */ /* Newton-Raphson iteration */ rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */ rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */ if( num < 0 ) { /* Ensure adjusted reflection coefficients has the original sign */ rc_Q31 = -rc_Q31; } invGain_Q30 = minInvGain_Q30; reached_max_gain = 1; } else { invGain_Q30 = tmp1; } /* Update the AR coefficients */ for( k = 0; k < (n + 1) >> 1; k++ ) { tmp1 = Af_QA[ k ]; /* QA */ tmp2 = Af_QA[ n - k - 1 ]; /* QA */ Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA */ Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA */ } Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA */ if( reached_max_gain ) { /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ for( k = n + 1; k < D; k++ ) { Af_QA[ k ] = 0; } break; } /* Update C * Af and C * Ab */ for( k = 0; k <= n + 1; k++ ) { tmp1 = CAf[ k ]; /* Q( -rshifts ) */ tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts ) */ CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts ) */ CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts ) */ } } if( reached_max_gain ) { for( k = 0; k < D; k++ ) { /* Scale coefficients */ A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); } /* Subtract energy of preceding samples from C0 */ if( rshifts > 0 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D ), rshifts ); } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D ), -rshifts ); } } /* Approximate residual energy */ *res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 ); *res_nrg_Q = -rshifts; } else { /* Return residual energy */ nrg = CAf[ 0 ]; /* Q( -rshifts ) */ tmp1 = (opus_int32)1 << 16; /* Q16 */ for( k = 0; k < D; k++ ) { Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16 */ nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts ) */ tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16 */ A_Q16[ k ] = -Atmp1; } *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( FIND_LPC_COND_FAC, C0 ), -tmp1 ); /* Q( -rshifts ) */ *res_nrg_Q = -rshifts; } }
/* Set decoder sampling rate */ opus_int silk_decoder_set_fs( silk_decoder_state *psDec, /* I/O Decoder state pointer */ opus_int fs_kHz, /* I Sampling frequency (kHz) */ opus_int fs_API_Hz /* I API Sampling frequency (Hz) */ ) { opus_int frame_length, ret = 0; silk_assert( fs_kHz == 8 || fs_kHz == 12 || fs_kHz == 16 ); silk_assert( psDec->nb_subfr == MAX_NB_SUBFR || psDec->nb_subfr == MAX_NB_SUBFR/2 ); /* New (sub)frame length */ psDec->subfr_length = silk_SMULBB( SUB_FRAME_LENGTH_MS, fs_kHz ); frame_length = silk_SMULBB( psDec->nb_subfr, psDec->subfr_length ); /* Initialize resampler when switching internal or external sampling frequency */ if( psDec->fs_kHz != fs_kHz || psDec->fs_API_hz != fs_API_Hz ) { /* Initialize the resampler for dec_API.c preparing resampling from fs_kHz to API_fs_Hz */ ret += silk_resampler_init( &psDec->resampler_state, silk_SMULBB( fs_kHz, 1000 ), fs_API_Hz, 0 ); psDec->fs_API_hz = fs_API_Hz; } if( psDec->fs_kHz != fs_kHz || frame_length != psDec->frame_length ) { if( fs_kHz == 8 ) { if( psDec->nb_subfr == MAX_NB_SUBFR ) { psDec->pitch_contour_iCDF = silk_pitch_contour_NB_iCDF; } else { psDec->pitch_contour_iCDF = silk_pitch_contour_10_ms_NB_iCDF; } } else { if( psDec->nb_subfr == MAX_NB_SUBFR ) { psDec->pitch_contour_iCDF = silk_pitch_contour_iCDF; } else { psDec->pitch_contour_iCDF = silk_pitch_contour_10_ms_iCDF; } } if( psDec->fs_kHz != fs_kHz ) { psDec->ltp_mem_length = silk_SMULBB( LTP_MEM_LENGTH_MS, fs_kHz ); if( fs_kHz == 8 || fs_kHz == 12 ) { psDec->LPC_order = MIN_LPC_ORDER; psDec->psNLSF_CB = &silk_NLSF_CB_NB_MB; } else { psDec->LPC_order = MAX_LPC_ORDER; psDec->psNLSF_CB = &silk_NLSF_CB_WB; } if( fs_kHz == 16 ) { psDec->pitch_lag_low_bits_iCDF = silk_uniform8_iCDF; } else if( fs_kHz == 12 ) { psDec->pitch_lag_low_bits_iCDF = silk_uniform6_iCDF; } else if( fs_kHz == 8 ) { psDec->pitch_lag_low_bits_iCDF = silk_uniform4_iCDF; } else { /* unsupported sampling rate */ silk_assert( 0 ); } psDec->first_frame_after_reset = 1; psDec->lagPrev = 100; psDec->LastGainIndex = 10; psDec->prevSignalType = TYPE_NO_VOICE_ACTIVITY; silk_memset( psDec->outBuf, 0, sizeof(psDec->outBuf)); silk_memset( psDec->sLPC_Q14_buf, 0, sizeof(psDec->sLPC_Q14_buf) ); } psDec->fs_kHz = fs_kHz; psDec->frame_length = frame_length; } /* Check that settings are valid */ silk_assert( psDec->frame_length > 0 && psDec->frame_length <= MAX_FRAME_LENGTH ); return ret; }
/* Initialize/reset the resampler state for a given pair of input/output sampling rates */ opus_int silk_resampler_init( silk_resampler_state_struct *S, /* I/O Resampler state */ opus_int32 Fs_Hz_in, /* I Input sampling rate (Hz) */ opus_int32 Fs_Hz_out /* I Output sampling rate (Hz) */ ) { opus_int32 up2 = 0, down2 = 0; /* Clear state */ silk_memset( S, 0, sizeof( silk_resampler_state_struct ) ); /* Input checking */ if( ( Fs_Hz_in != 8000 && Fs_Hz_in != 12000 && Fs_Hz_in != 16000 && Fs_Hz_in != 24000 && Fs_Hz_in != 48000 ) || ( Fs_Hz_out != 8000 && Fs_Hz_out != 12000 && Fs_Hz_out != 16000 && Fs_Hz_out != 24000 && Fs_Hz_out != 48000 ) ) { silk_assert( 0 ); return -1; } /* Number of samples processed per batch */ S->batchSize = silk_DIV32_16( Fs_Hz_in, 100 ); /* Find resampler with the right sampling ratio */ if( Fs_Hz_out > Fs_Hz_in ) { /* Upsample */ if( Fs_Hz_out == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 1 */ /* Special case: directly use 2x upsampler */ S->resampler_function = USE_silk_resampler_private_up2_HQ_wrapper; } else { /* Default resampler */ S->resampler_function = USE_silk_resampler_private_IIR_FIR; up2 = 1; } } else if ( Fs_Hz_out < Fs_Hz_in ) { /* Downsample */ if( silk_MUL( Fs_Hz_out, 4 ) == silk_MUL( Fs_Hz_in, 3 ) ) { /* Fs_out : Fs_in = 3 : 4 */ S->FIR_Fracs = 3; S->Coefs = silk_Resampler_3_4_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 3 ) == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 3 */ S->FIR_Fracs = 2; S->Coefs = silk_Resampler_2_3_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 2 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 2 */ S->FIR_Fracs = 1; S->Coefs = silk_Resampler_1_2_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 3 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 3 */ S->FIR_Fracs = 1; S->Coefs = silk_Resampler_1_3_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 4 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 4 */ S->FIR_Fracs = 1; down2 = 1; S->Coefs = silk_Resampler_1_2_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 6 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 6 */ S->FIR_Fracs = 1; down2 = 1; S->Coefs = silk_Resampler_1_3_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else { /* None available */ silk_assert( 0 ); return -1; } } else { /* Input and output sampling rates are equal: copy */ S->resampler_function = USE_silk_resampler_copy; } S->input2x = up2 | down2; /* Ratio of input/output samples */ S->invRatio_Q16 = silk_LSHIFT32( silk_DIV32( silk_LSHIFT32( Fs_Hz_in, 14 + up2 - down2 ), Fs_Hz_out ), 2 ); /* Make sure the ratio is rounded up */ while( silk_SMULWW( S->invRatio_Q16, silk_LSHIFT32( Fs_Hz_out, down2 ) ) < silk_LSHIFT32( Fs_Hz_in, up2 ) ) { S->invRatio_Q16++; } return 0; }
static OPUS_INLINE void silk_PLC_conceal( silk_decoder_state *psDec, /* I/O Decoder state */ silk_decoder_control *psDecCtrl, /* I/O Decoder control */ opus_int16 frame[], /* O LPC residual signal */ int arch /* I Run-time architecture */ ) { opus_int i, j, k; opus_int lag, idx, sLTP_buf_idx, shift1, shift2; opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q30; opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr; opus_int32 LPC_pred_Q10, LTP_pred_Q12; opus_int16 rand_scale_Q14; opus_int16 *B_Q14; opus_int32 *sLPC_Q14_ptr; opus_int16 A_Q12[ MAX_LPC_ORDER ]; #ifdef SMALL_FOOTPRINT opus_int16 *sLTP; #else VARDECL( opus_int16, sLTP ); #endif VARDECL( opus_int32, sLTP_Q14 ); silk_PLC_struct *psPLC = &psDec->sPLC; opus_int32 prevGain_Q10[2]; SAVE_STACK; ALLOC( sLTP_Q14, psDec->ltp_mem_length + psDec->frame_length, opus_int32 ); #ifdef SMALL_FOOTPRINT /* Ugly hack that breaks aliasing rules to save stack: put sLTP at the very end of sLTP_Q14. */ sLTP = ((opus_int16*)&sLTP_Q14[psDec->ltp_mem_length + psDec->frame_length])-psDec->ltp_mem_length; #else ALLOC( sLTP, psDec->ltp_mem_length, opus_int16 ); #endif prevGain_Q10[0] = silk_RSHIFT( psPLC->prevGain_Q16[ 0 ], 6); prevGain_Q10[1] = silk_RSHIFT( psPLC->prevGain_Q16[ 1 ], 6); if( psDec->first_frame_after_reset ) { silk_memset( psPLC->prevLPC_Q12, 0, sizeof( psPLC->prevLPC_Q12 ) ); } silk_PLC_energy(&energy1, &shift1, &energy2, &shift2, psDec->exc_Q14, prevGain_Q10, psDec->subfr_length, psDec->nb_subfr); if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) { /* First sub-frame has lowest energy */ rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, ( psPLC->nb_subfr - 1 ) * psPLC->subfr_length - RAND_BUF_SIZE ) ]; } else { /* Second sub-frame has lowest energy */ rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, psPLC->nb_subfr * psPLC->subfr_length - RAND_BUF_SIZE ) ]; } /* Set up Gain to random noise component */ B_Q14 = psPLC->LTPCoef_Q14; rand_scale_Q14 = psPLC->randScale_Q14; /* Set up attenuation gains */ harm_Gain_Q15 = HARM_ATT_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ]; if( psDec->prevSignalType == TYPE_VOICED ) { rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ]; } else { rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ]; } /* LPC concealment. Apply BWE to previous LPC */ silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) ); /* Preload LPC coeficients to array on stack. Gives small performance gain */ silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) ); /* First Lost frame */ if( psDec->lossCnt == 0 ) { rand_scale_Q14 = 1 << 14; /* Reduce random noise Gain for voiced frames */ if( psDec->prevSignalType == TYPE_VOICED ) { for( i = 0; i < LTP_ORDER; i++ ) { rand_scale_Q14 -= B_Q14[ i ]; } rand_scale_Q14 = silk_max_16( 3277, rand_scale_Q14 ); /* 0.2 */ rand_scale_Q14 = (opus_int16)silk_RSHIFT( silk_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 ); } else { /* Reduce random noise for unvoiced frames with high LPC gain */ opus_int32 invGain_Q30, down_scale_Q30; invGain_Q30 = silk_LPC_inverse_pred_gain( psPLC->prevLPC_Q12, psDec->LPC_order, arch ); down_scale_Q30 = silk_min_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 ); down_scale_Q30 = silk_max_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 ); down_scale_Q30 = silk_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES ); rand_Gain_Q15 = silk_RSHIFT( silk_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 ); } } rand_seed = psPLC->rand_seed; lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 ); sLTP_buf_idx = psDec->ltp_mem_length; /* Rewhiten LTP state */ idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2; silk_assert( idx > 0 ); silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order, arch ); /* Scale LTP state */ inv_gain_Q30 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 46 ); inv_gain_Q30 = silk_min( inv_gain_Q30, silk_int32_MAX >> 1 ); for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) { sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] ); } /***************************/ /* LTP synthesis filtering */ /***************************/ for( k = 0; k < psDec->nb_subfr; k++ ) { /* Set up pointer */ pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ]; for( i = 0; i < psDec->subfr_length; i++ ) { /* Unrolled loop */ /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ LTP_pred_Q12 = 2; LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] ); LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] ); LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] ); LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] ); pred_lag_ptr++; /* Generate LPC excitation */ rand_seed = silk_RAND( rand_seed ); idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK; sLTP_Q14[ sLTP_buf_idx ] = silk_LSHIFT32( silk_SMLAWB( LTP_pred_Q12, rand_ptr[ idx ], rand_scale_Q14 ), 2 ); sLTP_buf_idx++; } /* Gradually reduce LTP gain */ for( j = 0; j < LTP_ORDER; j++ ) { B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 ); } if ( psDec->indices.signalType != TYPE_NO_VOICE_ACTIVITY ) { /* Gradually reduce excitation gain */ rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 ); } /* Slowly increase pitch lag */ psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 ); psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) ); lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 ); } /***************************/ /* LPC synthesis filtering */ /***************************/ sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ]; /* Copy LPC state */ silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) ); silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */ for( i = 0; i < psDec->frame_length; i++ ) { /* partly unrolled */ /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ LPC_pred_Q10 = silk_RSHIFT( psDec->LPC_order, 1 ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 5 ], A_Q12[ 4 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 6 ], A_Q12[ 5 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 7 ], A_Q12[ 6 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 8 ], A_Q12[ 7 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 9 ], A_Q12[ 8 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] ); for( j = 10; j < psDec->LPC_order; j++ ) { LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] ); } /* Add prediction to LPC excitation */ sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_SAT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], silk_LSHIFT_SAT32( LPC_pred_Q10, 4 )); /* Scale with Gain */ frame[ i ] = (opus_int16)silk_SAT16( silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], prevGain_Q10[ 1 ] ), 8 ) ) ); } /* Save LPC state */ silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); /**************************************/ /* Update states */ /**************************************/ psPLC->rand_seed = rand_seed; psPLC->randScale_Q14 = rand_scale_Q14; for( i = 0; i < MAX_NB_SUBFR; i++ ) { psDecCtrl->pitchL[ i ] = lag; } RESTORE_STACK; }
/* Find LPC and LTP coefficients */ void silk_find_pred_coefs_FLP( silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ silk_encoder_control_FLP *psEncCtrl, /* I/O Encoder control FLP */ const silk_float res_pitch[], /* I Residual from pitch analysis */ const silk_float x[], /* I Speech signal */ opus_int condCoding /* I The type of conditional coding to use */ ) { opus_int i; silk_float WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; silk_float invGains[ MAX_NB_SUBFR ], Wght[ MAX_NB_SUBFR ]; opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const silk_float *x_ptr; silk_float *x_pre_ptr, LPC_in_pre[ MAX_NB_SUBFR * MAX_LPC_ORDER + MAX_FRAME_LENGTH ]; silk_float minInvGain; /* Weighting for weighted least squares */ for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { silk_assert( psEncCtrl->Gains[ i ] > 0.0f ); invGains[ i ] = 1.0f / psEncCtrl->Gains[ i ]; Wght[ i ] = invGains[ i ] * invGains[ i ]; } if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /**********/ /* VOICED */ /**********/ silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); /* LTP analysis */ silk_find_LTP_FLP( psEncCtrl->LTPCoef, WLTP, &psEncCtrl->LTPredCodGain, res_pitch, psEncCtrl->pitchL, Wght, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length ); /* Quantize LTP gain parameters */ silk_quant_LTP_gains_FLP( psEncCtrl->LTPCoef, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr ); /* Control LTP scaling */ silk_LTP_scale_ctrl_FLP( psEnc, psEncCtrl, condCoding ); /* Create LTP residual */ silk_LTP_analysis_filter_FLP( LPC_in_pre, x - psEnc->sCmn.predictLPCOrder, psEncCtrl->LTPCoef, psEncCtrl->pitchL, invGains, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); } else { /************/ /* UNVOICED */ /************/ /* Create signal with prepended subframes, scaled by inverse gains */ x_ptr = x - psEnc->sCmn.predictLPCOrder; x_pre_ptr = LPC_in_pre; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { silk_scale_copy_vector_FLP( x_pre_ptr, x_ptr, invGains[ i ], psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder ); x_pre_ptr += psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder; x_ptr += psEnc->sCmn.subfr_length; } silk_memset( psEncCtrl->LTPCoef, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( silk_float ) ); psEncCtrl->LTPredCodGain = 0.0f; } /* Limit on total predictive coding gain */ if( psEnc->sCmn.first_frame_after_reset ) { minInvGain = 1.0f / MAX_PREDICTION_POWER_GAIN_AFTER_RESET; } else { minInvGain = (silk_float)pow( 2, psEncCtrl->LTPredCodGain / 3 ) / MAX_PREDICTION_POWER_GAIN; minInvGain /= 0.25f + 0.75f * psEncCtrl->coding_quality; } /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ silk_find_LPC_FLP( &psEnc->sCmn, NLSF_Q15, LPC_in_pre, minInvGain ); /* Quantize LSFs */ silk_process_NLSFs_FLP( &psEnc->sCmn, psEncCtrl->PredCoef, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 ); /* Calculate residual energy using quantized LPC coefficients */ 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 interpolation */ silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); }
void silk_find_pred_coefs_FIX( silk_encoder_state_FIX *psEnc, /* I/O encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ const opus_int16 res_pitch[], /* I Residual from pitch analysis */ const opus_int16 x[], /* I Speech signal */ opus_int condCoding /* I The type of conditional coding to use */ ) { opus_int i; opus_int32 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ], Wght_Q15[ MAX_NB_SUBFR ]; opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const opus_int16 *x_ptr; opus_int16 *x_pre_ptr, LPC_in_pre[ MAX_NB_SUBFR * MAX_LPC_ORDER + MAX_FRAME_LENGTH ]; opus_int32 tmp, min_gain_Q16; opus_int LTP_corrs_rshift[ MAX_NB_SUBFR ]; /* weighting for weighted least squares */ min_gain_Q16 = silk_int32_MAX >> 6; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { min_gain_Q16 = silk_min( min_gain_Q16, psEncCtrl->Gains_Q16[ i ] ); } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { /* Divide to Q16 */ silk_assert( psEncCtrl->Gains_Q16[ i ] > 0 ); /* Invert and normalize gains, and ensure that maximum invGains_Q16 is within range of a 16 bit int */ invGains_Q16[ i ] = silk_DIV32_varQ( min_gain_Q16, psEncCtrl->Gains_Q16[ i ], 16 - 2 ); /* Ensure Wght_Q15 a minimum value 1 */ invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 363 ); /* Square the inverted gains */ silk_assert( invGains_Q16[ i ] == silk_SAT16( invGains_Q16[ i ] ) ); tmp = silk_SMULWB( invGains_Q16[ i ], invGains_Q16[ i ] ); Wght_Q15[ i ] = silk_RSHIFT( tmp, 1 ); /* Invert the inverted and normalized gains */ local_gains[ i ] = silk_DIV32( ( 1 << 16 ), invGains_Q16[ i ] ); } if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /**********/ /* VOICED */ /**********/ silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); /* LTP analysis */ silk_find_LTP_FIX( psEncCtrl->LTPCoef_Q14, WLTP, &psEncCtrl->LTPredCodGain_Q7, res_pitch, psEncCtrl->pitchL, Wght_Q15, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length, LTP_corrs_rshift ); /* Quantize LTP gain parameters */ silk_quant_LTP_gains( psEncCtrl->LTPCoef_Q14, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr); /* Control LTP scaling */ silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding ); /* Create LTP residual */ silk_LTP_analysis_filter_FIX( LPC_in_pre, x - psEnc->sCmn.predictLPCOrder, psEncCtrl->LTPCoef_Q14, psEncCtrl->pitchL, invGains_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); } else { /************/ /* UNVOICED */ /************/ /* Create signal with prepended subframes, scaled by inverse gains */ x_ptr = x - psEnc->sCmn.predictLPCOrder; x_pre_ptr = LPC_in_pre; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { silk_scale_copy_vector16( x_pre_ptr, x_ptr, invGains_Q16[ i ], psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder ); x_pre_ptr += psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder; x_ptr += psEnc->sCmn.subfr_length; } silk_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( opus_int16 ) ); psEncCtrl->LTPredCodGain_Q7 = 0; } /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ silk_find_LPC_FIX( NLSF_Q15, &psEnc->sCmn.indices.NLSFInterpCoef_Q2, psEnc->sCmn.prev_NLSFq_Q15, psEnc->sCmn.useInterpolatedNLSFs, psEnc->sCmn.first_frame_after_reset, psEnc->sCmn.predictLPCOrder, LPC_in_pre, psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder, psEnc->sCmn.nb_subfr ); /* Quantize LSFs */ silk_process_NLSFs( &psEnc->sCmn, psEncCtrl->PredCoef_Q12, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 ); /* Calculate residual energy using quantized LPC coefficients */ 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 interpolation */ silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); }
/* Control internal sampling rate */ opus_int silk_control_audio_bandwidth( silk_encoder_state *psEncC, /* I/O Pointer to Silk encoder state */ silk_EncControlStruct *encControl /* I Control structure */ ) { opus_int fs_kHz; opus_int32 fs_Hz; fs_kHz = psEncC->fs_kHz; fs_Hz = silk_SMULBB( fs_kHz, 1000 ); if( fs_Hz == 0 ) { /* Encoder has just been initialized */ fs_Hz = silk_min( psEncC->desiredInternal_fs_Hz, psEncC->API_fs_Hz ); fs_kHz = silk_DIV32_16( fs_Hz, 1000 ); } else if( fs_Hz > psEncC->API_fs_Hz || fs_Hz > psEncC->maxInternal_fs_Hz || fs_Hz < psEncC->minInternal_fs_Hz ) { /* Make sure internal rate is not higher than external rate or maximum allowed, or lower than minimum allowed */ fs_Hz = psEncC->API_fs_Hz; fs_Hz = silk_min( fs_Hz, psEncC->maxInternal_fs_Hz ); fs_Hz = silk_max( fs_Hz, psEncC->minInternal_fs_Hz ); fs_kHz = silk_DIV32_16( fs_Hz, 1000 ); } else { /* State machine for the internal sampling rate switching */ if( psEncC->sLP.transition_frame_no >= TRANSITION_FRAMES ) { /* Stop transition phase */ psEncC->sLP.mode = 0; } if( psEncC->allow_bandwidth_switch || encControl->opusCanSwitch ) { /* Check if we should switch down */ if( silk_SMULBB( psEncC->fs_kHz, 1000 ) > psEncC->desiredInternal_fs_Hz ) { /* Switch down */ if( psEncC->sLP.mode == 0 ) { /* New transition */ psEncC->sLP.transition_frame_no = TRANSITION_FRAMES; /* Reset transition filter state */ silk_memset( psEncC->sLP.In_LP_State, 0, sizeof( psEncC->sLP.In_LP_State ) ); } if( encControl->opusCanSwitch ) { /* Stop transition phase */ psEncC->sLP.mode = 0; /* Switch to a lower sample frequency */ fs_kHz = psEncC->fs_kHz == 16 ? 12 : 8; } else { if( psEncC->sLP.transition_frame_no <= 0 ) { encControl->switchReady = 1; } else { /* Direction: down (at double speed) */ psEncC->sLP.mode = -2; } } } else /* Check if we should switch up */ if( silk_SMULBB( psEncC->fs_kHz, 1000 ) < psEncC->desiredInternal_fs_Hz ) { /* Switch up */ if( encControl->opusCanSwitch ) { /* Switch to a higher sample frequency */ fs_kHz = psEncC->fs_kHz == 8 ? 12 : 16; /* New transition */ psEncC->sLP.transition_frame_no = 0; /* Reset transition filter state */ silk_memset( psEncC->sLP.In_LP_State, 0, sizeof( psEncC->sLP.In_LP_State ) ); /* Direction: up */ psEncC->sLP.mode = 1; } else { if( psEncC->sLP.mode == 0 ) { encControl->switchReady = 1; } else { /* Direction: up */ psEncC->sLP.mode = 1; } } } } } return fs_kHz; }
void silk_decode_pulses( ec_dec *psRangeDec, /* I/O Compressor data structure */ opus_int pulses[], /* O Excitation signal */ const opus_int signalType, /* I Sigtype */ const opus_int quantOffsetType, /* I quantOffsetType */ const opus_int frame_length /* I Frame length */ ) { opus_int i, j, k, iter, abs_q, nLS, RateLevelIndex; opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ], nLshifts[ MAX_NB_SHELL_BLOCKS ]; opus_int *pulses_ptr; const opus_uint8 *cdf_ptr; /*********************/ /* Decode rate level */ /*********************/ RateLevelIndex = ec_dec_icdf( psRangeDec, silk_rate_levels_iCDF[ signalType >> 1 ], 8 ); /* Calculate number of shell blocks */ silk_assert( 1 << LOG2_SHELL_CODEC_FRAME_LENGTH == SHELL_CODEC_FRAME_LENGTH ); iter = silk_RSHIFT( frame_length, LOG2_SHELL_CODEC_FRAME_LENGTH ); if( iter * SHELL_CODEC_FRAME_LENGTH < frame_length ) { silk_assert( frame_length == 12 * 10 ); /* Make sure only happens for 10 ms @ 12 kHz */ iter++; } /***************************************************/ /* Sum-Weighted-Pulses Decoding */ /***************************************************/ cdf_ptr = silk_pulses_per_block_iCDF[ RateLevelIndex ]; for( i = 0; i < iter; i++ ) { nLshifts[ i ] = 0; sum_pulses[ i ] = ec_dec_icdf( psRangeDec, cdf_ptr, 8 ); /* LSB indication */ while( sum_pulses[ i ] == MAX_PULSES + 1 ) { nLshifts[ i ]++; /* When we've already got 10 LSBs, we shift the table to not allow (MAX_PULSES + 1) */ sum_pulses[ i ] = ec_dec_icdf( psRangeDec, silk_pulses_per_block_iCDF[ N_RATE_LEVELS - 1] + ( nLshifts[ i ] == 10 ), 8 ); } } /***************************************************/ /* Shell decoding */ /***************************************************/ for( i = 0; i < iter; i++ ) { if( sum_pulses[ i ] > 0 ) { silk_shell_decoder( &pulses[ silk_SMULBB( i, SHELL_CODEC_FRAME_LENGTH ) ], psRangeDec, sum_pulses[ i ] ); } else { silk_memset( &pulses[ silk_SMULBB( i, SHELL_CODEC_FRAME_LENGTH ) ], 0, SHELL_CODEC_FRAME_LENGTH * sizeof( opus_int ) ); } } /***************************************************/ /* LSB Decoding */ /***************************************************/ for( i = 0; i < iter; i++ ) { if( nLshifts[ i ] > 0 ) { nLS = nLshifts[ i ]; pulses_ptr = &pulses[ silk_SMULBB( i, SHELL_CODEC_FRAME_LENGTH ) ]; for( k = 0; k < SHELL_CODEC_FRAME_LENGTH; k++ ) { abs_q = pulses_ptr[ k ]; for( j = 0; j < nLS; j++ ) { abs_q = silk_LSHIFT( abs_q, 1 ); abs_q += ec_dec_icdf( psRangeDec, silk_lsb_iCDF, 8 ); } pulses_ptr[ k ] = abs_q; } /* Mark the number of pulses non-zero for sign decoding. */ sum_pulses[ i ] |= nLS << 5; } } /****************************************/ /* Decode and add signs to pulse signal */ /****************************************/ silk_decode_signs( psRangeDec, pulses, frame_length, signalType, quantOffsetType, sum_pulses ); }
opus_int silk_setup_fs( silk_encoder_state_Fxx *psEnc, /* I/O */ opus_int fs_kHz, /* I */ opus_int PacketSize_ms /* I */ ) { opus_int ret = SILK_NO_ERROR; /* Set packet size */ if( PacketSize_ms != psEnc->sCmn.PacketSize_ms ) { if( ( PacketSize_ms != 10 ) && ( PacketSize_ms != 20 ) && ( PacketSize_ms != 40 ) && ( PacketSize_ms != 60 ) ) { ret = SILK_ENC_PACKET_SIZE_NOT_SUPPORTED; } if( PacketSize_ms <= 10 ) { psEnc->sCmn.nFramesPerPacket = 1; psEnc->sCmn.nb_subfr = PacketSize_ms == 10 ? 2 : 1; psEnc->sCmn.frame_length = silk_SMULBB( PacketSize_ms, fs_kHz ); psEnc->sCmn.pitch_LPC_win_length = silk_SMULBB( FIND_PITCH_LPC_WIN_MS_2_SF, fs_kHz ); if( psEnc->sCmn.fs_kHz == 8 ) { psEnc->sCmn.pitch_contour_iCDF = silk_pitch_contour_10_ms_NB_iCDF; } else { psEnc->sCmn.pitch_contour_iCDF = silk_pitch_contour_10_ms_iCDF; } } else { psEnc->sCmn.nFramesPerPacket = silk_DIV32_16( PacketSize_ms, MAX_FRAME_LENGTH_MS ); psEnc->sCmn.nb_subfr = MAX_NB_SUBFR; psEnc->sCmn.frame_length = silk_SMULBB( 20, fs_kHz ); psEnc->sCmn.pitch_LPC_win_length = silk_SMULBB( FIND_PITCH_LPC_WIN_MS, fs_kHz ); if( psEnc->sCmn.fs_kHz == 8 ) { psEnc->sCmn.pitch_contour_iCDF = silk_pitch_contour_NB_iCDF; } else { psEnc->sCmn.pitch_contour_iCDF = silk_pitch_contour_iCDF; } } psEnc->sCmn.PacketSize_ms = PacketSize_ms; psEnc->sCmn.TargetRate_bps = 0; /* trigger new SNR computation */ } /* Set internal sampling frequency */ silk_assert( fs_kHz == 8 || fs_kHz == 12 || fs_kHz == 16 ); silk_assert( psEnc->sCmn.nb_subfr == 2 || psEnc->sCmn.nb_subfr == 4 ); if( psEnc->sCmn.fs_kHz != fs_kHz ) { /* reset part of the state */ silk_memset( &psEnc->sShape, 0, sizeof( psEnc->sShape ) ); silk_memset( &psEnc->sPrefilt, 0, sizeof( psEnc->sPrefilt ) ); silk_memset( &psEnc->sCmn.sNSQ, 0, sizeof( psEnc->sCmn.sNSQ ) ); silk_memset( psEnc->sCmn.prev_NLSFq_Q15, 0, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); silk_memset( &psEnc->sCmn.sLP.In_LP_State, 0, sizeof( psEnc->sCmn.sLP.In_LP_State ) ); psEnc->sCmn.inputBufIx = 0; psEnc->sCmn.nFramesEncoded = 0; psEnc->sCmn.TargetRate_bps = 0; /* trigger new SNR computation */ /* Initialize non-zero parameters */ psEnc->sCmn.prevLag = 100; psEnc->sCmn.first_frame_after_reset = 1; psEnc->sPrefilt.lagPrev = 100; psEnc->sShape.LastGainIndex = 10; psEnc->sCmn.sNSQ.lagPrev = 100; psEnc->sCmn.sNSQ.prev_gain_Q16 = 65536; psEnc->sCmn.prevSignalType = TYPE_NO_VOICE_ACTIVITY; psEnc->sCmn.fs_kHz = fs_kHz; if( psEnc->sCmn.fs_kHz == 8 ) { if( psEnc->sCmn.nb_subfr == MAX_NB_SUBFR ) { psEnc->sCmn.pitch_contour_iCDF = silk_pitch_contour_NB_iCDF; } else { psEnc->sCmn.pitch_contour_iCDF = silk_pitch_contour_10_ms_NB_iCDF; } } else { if( psEnc->sCmn.nb_subfr == MAX_NB_SUBFR ) { psEnc->sCmn.pitch_contour_iCDF = silk_pitch_contour_iCDF; } else { psEnc->sCmn.pitch_contour_iCDF = silk_pitch_contour_10_ms_iCDF; } } if( psEnc->sCmn.fs_kHz == 8 || psEnc->sCmn.fs_kHz == 12 ) { psEnc->sCmn.predictLPCOrder = MIN_LPC_ORDER; psEnc->sCmn.psNLSF_CB = &silk_NLSF_CB_NB_MB; } else { psEnc->sCmn.predictLPCOrder = MAX_LPC_ORDER; psEnc->sCmn.psNLSF_CB = &silk_NLSF_CB_WB; } psEnc->sCmn.subfr_length = SUB_FRAME_LENGTH_MS * fs_kHz; psEnc->sCmn.frame_length = silk_SMULBB( psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr ); psEnc->sCmn.ltp_mem_length = silk_SMULBB( LTP_MEM_LENGTH_MS, fs_kHz ); psEnc->sCmn.la_pitch = silk_SMULBB( LA_PITCH_MS, fs_kHz ); psEnc->sCmn.max_pitch_lag = silk_SMULBB( 18, fs_kHz ); if( psEnc->sCmn.nb_subfr == MAX_NB_SUBFR ) { psEnc->sCmn.pitch_LPC_win_length = silk_SMULBB( FIND_PITCH_LPC_WIN_MS, fs_kHz ); } else { psEnc->sCmn.pitch_LPC_win_length = silk_SMULBB( FIND_PITCH_LPC_WIN_MS_2_SF, fs_kHz ); } if( psEnc->sCmn.fs_kHz == 16 ) { psEnc->sCmn.mu_LTP_Q9 = SILK_FIX_CONST( MU_LTP_QUANT_WB, 9 ); psEnc->sCmn.pitch_lag_low_bits_iCDF = silk_uniform8_iCDF; } else if( psEnc->sCmn.fs_kHz == 12 ) { psEnc->sCmn.mu_LTP_Q9 = SILK_FIX_CONST( MU_LTP_QUANT_MB, 9 ); psEnc->sCmn.pitch_lag_low_bits_iCDF = silk_uniform6_iCDF; } else { psEnc->sCmn.mu_LTP_Q9 = SILK_FIX_CONST( MU_LTP_QUANT_NB, 9 ); psEnc->sCmn.pitch_lag_low_bits_iCDF = silk_uniform4_iCDF; } } /* Check that settings are valid */ silk_assert( ( psEnc->sCmn.subfr_length * psEnc->sCmn.nb_subfr ) == psEnc->sCmn.frame_length ); return ret; }
/* Initialize/reset the resampler state for a given pair of input/output sampling rates */ opus_int silk_resampler_init( silk_resampler_state_struct *S, /* I/O Resampler state */ opus_int32 Fs_Hz_in, /* I Input sampling rate (Hz) */ opus_int32 Fs_Hz_out, /* I Output sampling rate (Hz) */ opus_int forEnc /* I If 1: encoder; if 0: decoder */ ) { opus_int up2x; /* Clear state */ silk_memset( S, 0, sizeof( silk_resampler_state_struct ) ); /* Input checking */ if( forEnc ) { if( ( Fs_Hz_in != 8000 && Fs_Hz_in != 12000 && Fs_Hz_in != 16000 && Fs_Hz_in != 24000 && Fs_Hz_in != 48000 ) || ( Fs_Hz_out != 8000 && Fs_Hz_out != 12000 && Fs_Hz_out != 16000 ) ) { silk_assert( 0 ); return -1; } S->inputDelay = delay_matrix_enc[ rateID( Fs_Hz_in ) ][ rateID( Fs_Hz_out ) ]; } else { if( ( Fs_Hz_in != 8000 && Fs_Hz_in != 12000 && Fs_Hz_in != 16000 ) || ( Fs_Hz_out != 8000 && Fs_Hz_out != 12000 && Fs_Hz_out != 16000 && Fs_Hz_out != 24000 && Fs_Hz_out != 48000 ) ) { silk_assert( 0 ); return -1; } S->inputDelay = delay_matrix_dec[ rateID( Fs_Hz_in ) ][ rateID( Fs_Hz_out ) ]; } S->Fs_in_kHz = silk_DIV32_16( Fs_Hz_in, 1000 ); S->Fs_out_kHz = silk_DIV32_16( Fs_Hz_out, 1000 ); /* Number of samples processed per batch */ S->batchSize = S->Fs_in_kHz * RESAMPLER_MAX_BATCH_SIZE_MS; /* Find resampler with the right sampling ratio */ up2x = 0; if( Fs_Hz_out > Fs_Hz_in ) { /* Upsample */ if( Fs_Hz_out == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 1 */ /* Special case: directly use 2x upsampler */ S->resampler_function = USE_silk_resampler_private_up2_HQ_wrapper; } else { /* Default resampler */ S->resampler_function = USE_silk_resampler_private_IIR_FIR; up2x = 1; } } else if ( Fs_Hz_out < Fs_Hz_in ) { /* Downsample */ S->resampler_function = USE_silk_resampler_private_down_FIR; if( silk_MUL( Fs_Hz_out, 4 ) == silk_MUL( Fs_Hz_in, 3 ) ) { /* Fs_out : Fs_in = 3 : 4 */ S->FIR_Fracs = 3; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR0; S->Coefs = silk_Resampler_3_4_COEFS; } else if( silk_MUL( Fs_Hz_out, 3 ) == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 3 */ S->FIR_Fracs = 2; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR0; S->Coefs = silk_Resampler_2_3_COEFS; } else if( silk_MUL( Fs_Hz_out, 2 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 2 */ S->FIR_Fracs = 1; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR1; S->Coefs = silk_Resampler_1_2_COEFS; } else if( silk_MUL( Fs_Hz_out, 3 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 3 */ S->FIR_Fracs = 1; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR2; S->Coefs = silk_Resampler_1_3_COEFS; } else if( silk_MUL( Fs_Hz_out, 4 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 4 */ S->FIR_Fracs = 1; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR2; S->Coefs = silk_Resampler_1_4_COEFS; } else if( silk_MUL( Fs_Hz_out, 6 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 6 */ S->FIR_Fracs = 1; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR2; S->Coefs = silk_Resampler_1_6_COEFS; } else { /* None available */ silk_assert( 0 ); return -1; } } else { /* Input and output sampling rates are equal: copy */ S->resampler_function = USE_silk_resampler_copy; } /* Ratio of input/output samples */ S->invRatio_Q16 = silk_LSHIFT32( silk_DIV32( silk_LSHIFT32( Fs_Hz_in, 14 + up2x ), Fs_Hz_out ), 2 ); /* Make sure the ratio is rounded up */ while( silk_SMULWW( S->invRatio_Q16, Fs_Hz_out ) < silk_LSHIFT32( Fs_Hz_in, up2x ) ) { S->invRatio_Q16++; } return 0; }
static OPUS_INLINE void silk_PLC_update( silk_decoder_state *psDec, /* I/O Decoder state */ silk_decoder_control *psDecCtrl /* I/O Decoder control */ ) { opus_int32 LTP_Gain_Q14, temp_LTP_Gain_Q14; opus_int i, j; silk_PLC_struct *psPLC; psPLC = &psDec->sPLC; /* Update parameters used in case of packet loss */ psDec->prevSignalType = psDec->indices.signalType; LTP_Gain_Q14 = 0; if( psDec->indices.signalType == TYPE_VOICED ) { /* Find the parameters for the last subframe which contains a pitch pulse */ for( j = 0; j * psDec->subfr_length < psDecCtrl->pitchL[ psDec->nb_subfr - 1 ]; j++ ) { if( j == psDec->nb_subfr ) { break; } temp_LTP_Gain_Q14 = 0; for( i = 0; i < LTP_ORDER; i++ ) { temp_LTP_Gain_Q14 += psDecCtrl->LTPCoef_Q14[ ( psDec->nb_subfr - 1 - j ) * LTP_ORDER + i ]; } if( temp_LTP_Gain_Q14 > LTP_Gain_Q14 ) { LTP_Gain_Q14 = temp_LTP_Gain_Q14; silk_memcpy( psPLC->LTPCoef_Q14, &psDecCtrl->LTPCoef_Q14[ silk_SMULBB( psDec->nb_subfr - 1 - j, LTP_ORDER ) ], LTP_ORDER * sizeof( opus_int16 ) ); psPLC->pitchL_Q8 = silk_LSHIFT( psDecCtrl->pitchL[ psDec->nb_subfr - 1 - j ], 8 ); } } silk_memset( psPLC->LTPCoef_Q14, 0, LTP_ORDER * sizeof( opus_int16 ) ); psPLC->LTPCoef_Q14[ LTP_ORDER / 2 ] = LTP_Gain_Q14; /* Limit LT coefs */ if( LTP_Gain_Q14 < V_PITCH_GAIN_START_MIN_Q14 ) { opus_int scale_Q10; opus_int32 tmp; tmp = silk_LSHIFT( V_PITCH_GAIN_START_MIN_Q14, 10 ); scale_Q10 = silk_DIV32( tmp, silk_max( LTP_Gain_Q14, 1 ) ); for( i = 0; i < LTP_ORDER; i++ ) { psPLC->LTPCoef_Q14[ i ] = silk_RSHIFT( silk_SMULBB( psPLC->LTPCoef_Q14[ i ], scale_Q10 ), 10 ); } } else if( LTP_Gain_Q14 > V_PITCH_GAIN_START_MAX_Q14 ) { opus_int scale_Q14; opus_int32 tmp; tmp = silk_LSHIFT( V_PITCH_GAIN_START_MAX_Q14, 14 ); scale_Q14 = silk_DIV32( tmp, silk_max( LTP_Gain_Q14, 1 ) ); for( i = 0; i < LTP_ORDER; i++ ) { psPLC->LTPCoef_Q14[ i ] = silk_RSHIFT( silk_SMULBB( psPLC->LTPCoef_Q14[ i ], scale_Q14 ), 14 ); } } } else { psPLC->pitchL_Q8 = silk_LSHIFT( silk_SMULBB( psDec->fs_kHz, 18 ), 8 ); silk_memset( psPLC->LTPCoef_Q14, 0, LTP_ORDER * sizeof( opus_int16 )); } /* Save LPC coeficients */ silk_memcpy( psPLC->prevLPC_Q12, psDecCtrl->PredCoef_Q12[ 1 ], psDec->LPC_order * sizeof( opus_int16 ) ); psPLC->prevLTP_scale_Q14 = psDecCtrl->LTP_scale_Q14; /* Save last two gains */ silk_memcpy( psPLC->prevGain_Q16, &psDecCtrl->Gains_Q16[ psDec->nb_subfr - 2 ], 2 * sizeof( opus_int32 ) ); psPLC->subfr_length = psDec->subfr_length; psPLC->nb_subfr = psDec->nb_subfr; }
/* Compute reflection coefficients from input signal */ silk_float silk_burg_modified_FLP( /* O returns residual energy */ silk_float A[], /* O prediction coefficients (length order) */ const silk_float x[], /* I input signal, length: nb_subfr*(D+L_sub) */ const opus_int subfr_length, /* I input signal subframe length (incl. D preceeding samples) */ const opus_int nb_subfr, /* I number of subframes stacked in x */ const silk_float WhiteNoiseFrac, /* I fraction added to zero-lag autocorrelation */ const opus_int D /* I order */ ) { opus_int k, n, s; double C0, num, nrg_f, nrg_b, rc, Atmp, tmp1, tmp2; const silk_float *x_ptr; double C_first_row[ SILK_MAX_ORDER_LPC ], C_last_row[ SILK_MAX_ORDER_LPC ]; double CAf[ SILK_MAX_ORDER_LPC + 1 ], CAb[ SILK_MAX_ORDER_LPC + 1 ]; double Af[ SILK_MAX_ORDER_LPC ]; silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); silk_assert( nb_subfr <= MAX_NB_SUBFR ); /* Compute autocorrelations, added over subframes */ C0 = silk_energy_FLP( x, nb_subfr * subfr_length ); silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( double ) ); for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += silk_inner_product_FLP( x_ptr, x_ptr + n, subfr_length - n ); } } silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( double ) ); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + WhiteNoiseFrac * C0 + 1e-9f; for( n = 0; n < D; n++ ) { /* Update first row of correlation matrix (without first element) */ /* Update last row of correlation matrix (without last element, stored in reversed order) */ /* Update C * Af */ /* Update C * flipud(Af) (stored in reversed order) */ for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; tmp1 = x_ptr[ n ]; tmp2 = x_ptr[ subfr_length - n - 1 ]; for( k = 0; k < n; k++ ) { C_first_row[ k ] -= x_ptr[ n ] * x_ptr[ n - k - 1 ]; C_last_row[ k ] -= x_ptr[ subfr_length - n - 1 ] * x_ptr[ subfr_length - n + k ]; Atmp = Af[ k ]; tmp1 += x_ptr[ n - k - 1 ] * Atmp; tmp2 += x_ptr[ subfr_length - n + k ] * Atmp; } for( k = 0; k <= n; k++ ) { CAf[ k ] -= tmp1 * x_ptr[ n - k ]; CAb[ k ] -= tmp2 * x_ptr[ subfr_length - n + k - 1 ]; } } tmp1 = C_first_row[ n ]; tmp2 = C_last_row[ n ]; for( k = 0; k < n; k++ ) { Atmp = Af[ k ]; tmp1 += C_last_row[ n - k - 1 ] * Atmp; tmp2 += C_first_row[ n - k - 1 ] * Atmp; } CAf[ n + 1 ] = tmp1; CAb[ n + 1 ] = tmp2; /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ num = CAb[ n + 1 ]; nrg_b = CAb[ 0 ]; nrg_f = CAf[ 0 ]; for( k = 0; k < n; k++ ) { Atmp = Af[ k ]; num += CAb[ n - k ] * Atmp; nrg_b += CAb[ k + 1 ] * Atmp; nrg_f += CAf[ k + 1 ] * Atmp; } silk_assert( nrg_f > 0.0 ); silk_assert( nrg_b > 0.0 ); /* Calculate the next order reflection (parcor) coefficient */ rc = -2.0 * num / ( nrg_f + nrg_b ); silk_assert( rc > -1.0 && rc < 1.0 ); /* Update the AR coefficients */ for( k = 0; k < (n + 1) >> 1; k++ ) { tmp1 = Af[ k ]; tmp2 = Af[ n - k - 1 ]; Af[ k ] = tmp1 + rc * tmp2; Af[ n - k - 1 ] = tmp2 + rc * tmp1; } Af[ n ] = rc; /* Update C * Af and C * Ab */ for( k = 0; k <= n + 1; k++ ) { tmp1 = CAf[ k ]; CAf[ k ] += rc * CAb[ n - k + 1 ]; CAb[ n - k + 1 ] += rc * tmp1; } } /* Return residual energy */ nrg_f = CAf[ 0 ]; tmp1 = 1.0; for( k = 0; k < D; k++ ) { Atmp = Af[ k ]; nrg_f += CAf[ k + 1 ] * Atmp; tmp1 += Atmp * Atmp; A[ k ] = (silk_float)(-Atmp); } nrg_f -= WhiteNoiseFrac * C0 * tmp1; return (silk_float)nrg_f; }