/* every other sample of window is linearly interpolated, for speed */ void SKP_Silk_apply_sine_window( SKP_int16 px_win[], /* O Pointer to windowed signal */ const SKP_int16 px[], /* I Pointer to input signal */ const SKP_int win_type, /* I Selects a window type */ const SKP_int length /* I Window length, multiple of 4 */ ) { SKP_int k; SKP_int32 f_Q16, c_Q20, S0_Q16, S1_Q16; /* Length must be multiple of 4 */ SKP_assert( ( length & 3 ) == 0 ); /* Input pointer must be 4-byte aligned */ SKP_assert( ( (SKP_int64)px & 3 ) == 0 ); if( win_type == 0 ) { f_Q16 = SKP_DIV32_16( 411775, length + 1 ); // 411775 = 2 * 65536 * pi } else { f_Q16 = SKP_DIV32_16( 205887, length + 1 ); // 205887 = 65536 * pi } /* factor used for cosine approximation */ c_Q20 = -SKP_RSHIFT( SKP_MUL( f_Q16, f_Q16 ), 12 ); /* c_Q20 becomes too large if length is too small */ SKP_assert( c_Q20 >= -32768 ); /* initialize state */ if( win_type < 2 ) { /* start from 0 */ S0_Q16 = 0; /* approximation of sin(f) */ S1_Q16 = f_Q16; } else { /* start from 1 */ S0_Q16 = ( 1 << 16 ); /* approximation of cos(f) */ S1_Q16 = ( 1 << 16 ) + SKP_RSHIFT( c_Q20, 5 ); } /* Uses the recursive equation: sin(n*f) = 2 * cos(f) * sin((n-1)*f) - sin((n-2)*f) */ /* 4 samples at a time */ for( k = 0; k < length; k += 4 ) { px_win[ k ] = (SKP_int16)SKP_SMULWB( SKP_RSHIFT( S0_Q16 + S1_Q16, 1 ), px[ k ] ); px_win[ k + 1 ] = (SKP_int16)SKP_SMULWB( S1_Q16, px[ k + 1] ); S0_Q16 = SKP_RSHIFT( SKP_MUL( c_Q20, S1_Q16 ), 20 ) + SKP_LSHIFT( S1_Q16, 1 ) - S0_Q16 + 1; S0_Q16 = SKP_min( S0_Q16, ( 1 << 16 ) ); px_win[ k + 2 ] = (SKP_int16)SKP_SMULWB( SKP_RSHIFT( S0_Q16 + S1_Q16, 1 ), px[ k + 2] ); px_win[ k + 3 ] = (SKP_int16)SKP_SMULWB( S0_Q16, px[ k + 3 ] ); S1_Q16 = SKP_RSHIFT( SKP_MUL( c_Q20, S0_Q16 ), 20 ) + SKP_LSHIFT( S0_Q16, 1 ) - S1_Q16; S1_Q16 = SKP_min( S1_Q16, ( 1 << 16 ) ); } }
SKP_INLINE opus_int silk_setup_LBRR( silk_encoder_state *psEncC, /* I/O */ const opus_int32 TargetRate_bps /* I */ ) { opus_int ret = SILK_NO_ERROR; opus_int32 LBRR_rate_thres_bps; psEncC->LBRR_enabled = 0; if( psEncC->useInBandFEC && psEncC->PacketLoss_perc > 0 ) { if( psEncC->fs_kHz == 8 ) { LBRR_rate_thres_bps = LBRR_NB_MIN_RATE_BPS; } else if( psEncC->fs_kHz == 12 ) { LBRR_rate_thres_bps = LBRR_MB_MIN_RATE_BPS; } else { LBRR_rate_thres_bps = LBRR_WB_MIN_RATE_BPS; } LBRR_rate_thres_bps = SKP_SMULWB( SKP_MUL( LBRR_rate_thres_bps, 125 - SKP_min( psEncC->PacketLoss_perc, 25 ) ), SILK_FIX_CONST( 0.01, 16 ) ); if( TargetRate_bps > LBRR_rate_thres_bps ) { /* Set gain increase for coding LBRR excitation */ psEncC->LBRR_enabled = 1; psEncC->LBRR_GainIncreases = SKP_max_int( 7 - SKP_SMULWB( psEncC->PacketLoss_perc, SILK_FIX_CONST( 0.4, 16 ) ), 2 ); } } return ret; }
/* Encodes signs of excitation */ void silk_encode_signs( ec_enc *psRangeEnc, /* I/O Compressor data structure */ const SKP_int8 pulses[], /* I pulse signal */ SKP_int length, /* I length of input */ const SKP_int signalType, /* I Signal type */ const SKP_int quantOffsetType, /* I Quantization offset type */ const SKP_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ ) { SKP_int i, j, p; SKP_uint8 icdf[ 2 ]; const SKP_int8 *q_ptr; const SKP_uint8 *icdf_ptr; icdf[ 1 ] = 0; q_ptr = pulses; i = SKP_SMULBB( 6, SKP_ADD_LSHIFT( quantOffsetType, signalType, 1 ) ); icdf_ptr = &silk_sign_iCDF[ i ]; length = SKP_RSHIFT( length + SHELL_CODEC_FRAME_LENGTH/2, LOG2_SHELL_CODEC_FRAME_LENGTH ); for( i = 0; i < length; i++ ) { p = sum_pulses[ i ]; if( p > 0 ) { icdf[ 0 ] = icdf_ptr[ SKP_min( p - 1, 5 ) ]; for( j = 0; j < SHELL_CODEC_FRAME_LENGTH; j++ ) { if( q_ptr[ j ] != 0 ) { ec_enc_icdf( psRangeEnc, SKP_enc_map( q_ptr[ j ]), icdf, 8 ); } } } q_ptr += SHELL_CODEC_FRAME_LENGTH; } }
/* Resample with a 2x downsampler (optional), a 2nd order AR filter followed by FIR interpolation */ void silk_resampler_private_down_FIR( void *SS, /* I/O: Resampler state */ opus_int16 out[], /* O: Output signal */ const opus_int16 in[], /* I: Input signal */ opus_int32 inLen /* I: Number of input samples */ ) { silk_resampler_state_struct *S = (silk_resampler_state_struct *)SS; opus_int32 nSamplesIn; opus_int32 max_index_Q16, index_increment_Q16; opus_int16 buf1[ RESAMPLER_MAX_BATCH_SIZE_IN / 2 ]; opus_int32 buf2[ RESAMPLER_MAX_BATCH_SIZE_IN + RESAMPLER_DOWN_ORDER_FIR ]; const opus_int16 *FIR_Coefs; /* Copy buffered samples to start of buffer */ SKP_memcpy( buf2, S->sFIR, RESAMPLER_DOWN_ORDER_FIR * sizeof( opus_int32 ) ); FIR_Coefs = &S->Coefs[ 2 ]; /* Iterate over blocks of frameSizeIn input samples */ index_increment_Q16 = S->invRatio_Q16; while( 1 ) { nSamplesIn = SKP_min( inLen, S->batchSize ); if( S->input2x == 1 ) { /* Downsample 2x */ silk_resampler_down2( S->sDown2, buf1, in, nSamplesIn ); nSamplesIn = SKP_RSHIFT32( nSamplesIn, 1 ); /* Second-order AR filter (output in Q8) */ silk_resampler_private_AR2( S->sIIR, &buf2[ RESAMPLER_DOWN_ORDER_FIR ], buf1, S->Coefs, nSamplesIn ); } else { /* Second-order AR filter (output in Q8) */ silk_resampler_private_AR2( S->sIIR, &buf2[ RESAMPLER_DOWN_ORDER_FIR ], in, S->Coefs, nSamplesIn ); } max_index_Q16 = SKP_LSHIFT32( nSamplesIn, 16 ); /* Interpolate filtered signal */ if( S->FIR_Fracs == 1 ) { out = silk_resampler_private_down_FIR_INTERPOL0(out, buf2, FIR_Coefs, max_index_Q16, index_increment_Q16); } else { out = silk_resampler_private_down_FIR_INTERPOL1(out, buf2, FIR_Coefs, max_index_Q16, index_increment_Q16, S->FIR_Fracs); } in += nSamplesIn << S->input2x; inLen -= nSamplesIn << S->input2x; if( inLen > S->input2x ) { /* More iterations to do; copy last part of filtered signal to beginning of buffer */ SKP_memcpy( buf2, &buf2[ nSamplesIn ], RESAMPLER_DOWN_ORDER_FIR * sizeof( opus_int32 ) ); } else { break; } } /* Copy last part of filtered signal to the state for the next call */ SKP_memcpy( S->sFIR, &buf2[ nSamplesIn ], RESAMPLER_DOWN_ORDER_FIR * sizeof( opus_int32 ) ); }
/* Upsample using a combination of allpass-based 2x upsampling and FIR interpolation */ void SKP_Silk_resampler_private_IIR_FIR( void *SS, /* I/O: Resampler state */ SKP_int16 out[], /* O: Output signal */ const SKP_int16 in[], /* I: Input signal */ SKP_int32 inLen /* I: Number of input samples */ ) { SKP_Silk_resampler_state_struct *S = (SKP_Silk_resampler_state_struct *)SS; SKP_int32 nSamplesIn, table_index; SKP_int32 max_index_Q16, index_Q16, index_increment_Q16, res_Q15; SKP_int16 buf[ 2 * RESAMPLER_MAX_BATCH_SIZE_IN + RESAMPLER_ORDER_FIR_144 ]; SKP_int16 *buf_ptr; /* Copy buffered samples to start of buffer */ SKP_memcpy( buf, S->sFIR, RESAMPLER_ORDER_FIR_144 * sizeof( SKP_int32 ) ); /* Iterate over blocks of frameSizeIn input samples */ index_increment_Q16 = S->invRatio_Q16; while( 1 ) { nSamplesIn = SKP_min( inLen, S->batchSize ); if( S->input2x == 1 ) { /* Upsample 2x */ S->up2_function( S->sIIR, &buf[ RESAMPLER_ORDER_FIR_144 ], in, nSamplesIn ); } else { /* Fourth-order ARMA filter */ SKP_Silk_resampler_private_ARMA4( S->sIIR, &buf[ RESAMPLER_ORDER_FIR_144 ], in, S->Coefs, nSamplesIn ); } max_index_Q16 = SKP_LSHIFT32( nSamplesIn, 16 + S->input2x ); /* +1 if 2x upsampling */ /* Interpolate upsampled signal and store in output array */ for( index_Q16 = 0; index_Q16 < max_index_Q16; index_Q16 += index_increment_Q16 ) { table_index = SKP_SMULWB( index_Q16 & 0xFFFF, 144 ); buf_ptr = &buf[ index_Q16 >> 16 ]; res_Q15 = SKP_SMULBB( buf_ptr[ 0 ], SKP_Silk_resampler_frac_FIR_144[ table_index ][ 0 ] ); res_Q15 = SKP_SMLABB( res_Q15, buf_ptr[ 1 ], SKP_Silk_resampler_frac_FIR_144[ table_index ][ 1 ] ); res_Q15 = SKP_SMLABB( res_Q15, buf_ptr[ 2 ], SKP_Silk_resampler_frac_FIR_144[ table_index ][ 2 ] ); res_Q15 = SKP_SMLABB( res_Q15, buf_ptr[ 3 ], SKP_Silk_resampler_frac_FIR_144[ 143 - table_index ][ 2 ] ); res_Q15 = SKP_SMLABB( res_Q15, buf_ptr[ 4 ], SKP_Silk_resampler_frac_FIR_144[ 143 - table_index ][ 1 ] ); res_Q15 = SKP_SMLABB( res_Q15, buf_ptr[ 5 ], SKP_Silk_resampler_frac_FIR_144[ 143 - table_index ][ 0 ] ); *out++ = (SKP_int16)SKP_SAT16( SKP_RSHIFT_ROUND( res_Q15, 15 ) ); } in += nSamplesIn; inLen -= nSamplesIn; if( inLen > 0 ) { /* More iterations to do; copy last part of filtered signal to beginning of buffer */ SKP_memcpy( buf, &buf[ nSamplesIn << S->input2x ], RESAMPLER_ORDER_FIR_144 * sizeof( SKP_int32 ) ); } else { break; } } /* Copy last part of filtered signal to the state for the next call */ SKP_memcpy( S->sFIR, &buf[nSamplesIn << S->input2x ], RESAMPLER_ORDER_FIR_144 * sizeof( SKP_int32 ) ); }
/* Glues concealed frames with new good recieved frames */ void SKP_Silk_PLC_glue_frames( SKP_Silk_decoder_state *psDec, /* I/O decoder state */ SKP_Silk_decoder_control *psDecCtrl, /* I/O Decoder control */ SKP_int16 signal[], /* I/O signal */ SKP_int length /* I length of residual */ ) { SKP_int i, energy_shift; SKP_int32 energy; SKP_Silk_PLC_struct *psPLC; psPLC = &psDec->sPLC; if( psDec->lossCnt ) { /* Calculate energy in concealed residual */ SKP_Silk_sum_sqr_shift( &psPLC->conc_energy, &psPLC->conc_energy_shift, signal, length ); psPLC->last_frame_lost = 1; } else { if( psDec->sPLC.last_frame_lost ) { /* Calculate residual in decoded signal if last frame was lost */ SKP_Silk_sum_sqr_shift( &energy, &energy_shift, signal, length ); /* Normalize energies */ if( energy_shift > psPLC->conc_energy_shift ) { psPLC->conc_energy = SKP_RSHIFT( psPLC->conc_energy, energy_shift - psPLC->conc_energy_shift ); } else if( energy_shift < psPLC->conc_energy_shift ) { energy = SKP_RSHIFT( energy, psPLC->conc_energy_shift - energy_shift ); } /* Fade in the energy difference */ if( energy > psPLC->conc_energy ) { SKP_int32 frac_Q24, LZ; SKP_int32 gain_Q12, slope_Q12; LZ = SKP_Silk_CLZ32( psPLC->conc_energy ); LZ = LZ - 1; psPLC->conc_energy = SKP_LSHIFT( psPLC->conc_energy, LZ ); energy = SKP_RSHIFT( energy, SKP_max_32( 24 - LZ, 0 ) ); frac_Q24 = SKP_DIV32( psPLC->conc_energy, SKP_max( energy, 1 ) ); gain_Q12 = SKP_Silk_SQRT_APPROX( frac_Q24 ); slope_Q12 = SKP_DIV32_16( ( 1 << 12 ) - gain_Q12, length ); for( i = 0; i < length; i++ ) { signal[ i ] = SKP_RSHIFT( SKP_MUL( gain_Q12, signal[ i ] ), 12 ); gain_Q12 += slope_Q12; gain_Q12 = SKP_min( gain_Q12, ( 1 << 12 ) ); } } } psPLC->last_frame_lost = 0; } }
/* Downsample by a factor 3, low quality */ void SKP_Silk_resampler_down3( SKP_int32 *S, /* I/O: State vector [ 8 ] */ SKP_int16 *out, /* O: Output signal [ floor(inLen/3) ] */ const SKP_int16 *in, /* I: Input signal [ inLen ] */ SKP_int32 inLen /* I: Number of input samples */ ) { SKP_int32 nSamplesIn, counter, res_Q6; SKP_int32 buf[ RESAMPLER_MAX_BATCH_SIZE_IN + ORDER_FIR ]; SKP_int32 *buf_ptr; /* Copy buffered samples to start of buffer */ SKP_memcpy( buf, S, ORDER_FIR * sizeof( SKP_int32 ) ); /* Iterate over blocks of frameSizeIn input samples */ while( 1 ) { nSamplesIn = SKP_min( inLen, RESAMPLER_MAX_BATCH_SIZE_IN ); /* Second-order AR filter (output in Q8) */ SKP_Silk_resampler_private_AR2( &S[ ORDER_FIR ], &buf[ ORDER_FIR ], in, SKP_Silk_Resampler_1_3_COEFS_LQ, nSamplesIn ); /* Interpolate filtered signal */ buf_ptr = buf; counter = nSamplesIn; while( counter > 2 ) { /* Inner product */ res_Q6 = SKP_SMULWB( SKP_ADD32( buf_ptr[ 0 ], buf_ptr[ 5 ] ), SKP_Silk_Resampler_1_3_COEFS_LQ[ 2 ] ); res_Q6 = SKP_SMLAWB( res_Q6, SKP_ADD32( buf_ptr[ 1 ], buf_ptr[ 4 ] ), SKP_Silk_Resampler_1_3_COEFS_LQ[ 3 ] ); res_Q6 = SKP_SMLAWB( res_Q6, SKP_ADD32( buf_ptr[ 2 ], buf_ptr[ 3 ] ), SKP_Silk_Resampler_1_3_COEFS_LQ[ 4 ] ); /* Scale down, saturate and store in output array */ *out++ = (SKP_int16)SKP_SAT16( SKP_RSHIFT_ROUND( res_Q6, 6 ) ); buf_ptr += 3; counter -= 3; } in += nSamplesIn; inLen -= nSamplesIn; if( inLen > 0 ) { /* More iterations to do; copy last part of filtered signal to beginning of buffer */ SKP_memcpy( buf, &buf[ nSamplesIn ], ORDER_FIR * sizeof( SKP_int32 ) ); } else { break; } } /* Copy last part of filtered signal to the state for the next call */ SKP_memcpy( S, &buf[ nSamplesIn ], ORDER_FIR * sizeof( SKP_int32 ) ); }
/* Upsample using a combination of allpass-based 2x upsampling and FIR interpolation */ void SKP_Silk_resampler_private_IIR_FIR( void *SS, /* I/O: Resampler state */ SKP_int16 out[], /* O: Output signal */ const SKP_int16 in[], /* I: Input signal */ SKP_int32 inLen /* I: Number of input samples */ ) { SKP_Silk_resampler_state_struct *S = (SKP_Silk_resampler_state_struct *)SS; SKP_int32 nSamplesIn; SKP_int32 max_index_Q16, index_increment_Q16; SKP_int16 buf[ 2 * RESAMPLER_MAX_BATCH_SIZE_IN + RESAMPLER_ORDER_FIR_144 ]; /* Copy buffered samples to start of buffer */ SKP_memcpy( buf, S->sFIR, RESAMPLER_ORDER_FIR_144 * sizeof( SKP_int32 ) ); /* Iterate over blocks of frameSizeIn input samples */ index_increment_Q16 = S->invRatio_Q16; while( 1 ) { nSamplesIn = SKP_min( inLen, S->batchSize ); if( S->input2x == 1 ) { /* Upsample 2x */ S->up2_function( S->sIIR, &buf[ RESAMPLER_ORDER_FIR_144 ], in, nSamplesIn ); } else { /* Fourth-order ARMA filter */ SKP_Silk_resampler_private_ARMA4( S->sIIR, &buf[ RESAMPLER_ORDER_FIR_144 ], in, S->Coefs, nSamplesIn ); } max_index_Q16 = SKP_LSHIFT32( nSamplesIn, 16 + S->input2x ); /* +1 if 2x upsampling */ out = SKP_Silk_resampler_private_IIR_FIR_INTERPOL(out, buf, max_index_Q16, index_increment_Q16); in += nSamplesIn; inLen -= nSamplesIn; if( inLen > 0 ) { /* More iterations to do; copy last part of filtered signal to beginning of buffer */ SKP_memcpy( buf, &buf[ nSamplesIn << S->input2x ], RESAMPLER_ORDER_FIR_144 * sizeof( SKP_int32 ) ); } else { break; } } /* Copy last part of filtered signal to the state for the next call */ SKP_memcpy( S->sFIR, &buf[nSamplesIn << S->input2x ], RESAMPLER_ORDER_FIR_144 * sizeof( SKP_int32 ) ); }
/* Decodes signs of excitation */ void silk_decode_signs( ec_dec *psRangeDec, /* I/O Compressor data structure */ SKP_int pulses[], /* I/O pulse signal */ SKP_int length, /* I length of input */ const SKP_int signalType, /* I Signal type */ const SKP_int quantOffsetType, /* I Quantization offset type */ const SKP_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ ) { SKP_int i, j, p; SKP_uint8 icdf[ 2 ]; SKP_int *q_ptr; const SKP_uint8 *icdf_ptr; icdf[ 1 ] = 0; q_ptr = pulses; i = SKP_SMULBB( 6, SKP_ADD_LSHIFT( quantOffsetType, signalType, 1 ) ); icdf_ptr = &silk_sign_iCDF[ i ]; length = SKP_RSHIFT( length + SHELL_CODEC_FRAME_LENGTH/2, LOG2_SHELL_CODEC_FRAME_LENGTH ); for( i = 0; i < length; i++ ) { p = sum_pulses[ i ]; if( p > 0 ) { icdf[ 0 ] = icdf_ptr[ SKP_min( p - 1, 5 ) ]; for( j = 0; j < SHELL_CODEC_FRAME_LENGTH; j++ ) { if( q_ptr[ j ] > 0 ) { /* attach sign */ #if 0 /* conditional implementation */ if( ec_dec_icdf( psRangeDec, icdf, 8 ) == 0 ) { q_ptr[ j ] = -q_ptr[ j ]; } #else /* implementation with shift, subtraction, multiplication */ q_ptr[ j ] *= SKP_dec_map( ec_dec_icdf( psRangeDec, icdf, 8 ) ); #endif } } } q_ptr += SHELL_CODEC_FRAME_LENGTH; } }
void SKP_Silk_find_pred_coefs_FIX(SKP_Silk_encoder_state_FIX * psEnc, /* I/O encoder state */ SKP_Silk_encoder_control_FIX * psEncCtrl, /* I/O encoder control */ const int16_t res_pitch[] /* I Residual from pitch analysis */ ) { int i; int32_t WLTP[NB_SUBFR * LTP_ORDER * LTP_ORDER]; int32_t invGains_Q16[NB_SUBFR], local_gains_Qx[NB_SUBFR], Wght_Q15[NB_SUBFR]; int NLSF_Q15[MAX_LPC_ORDER]; const int16_t *x_ptr; int16_t *x_pre_ptr, LPC_in_pre[NB_SUBFR * MAX_LPC_ORDER + MAX_FRAME_LENGTH]; int32_t tmp, min_gain_Q16; #if !VARQ int LZ; #endif int LTP_corrs_rshift[NB_SUBFR]; /* weighting for weighted least squares */ min_gain_Q16 = int32_t_MAX >> 6; for (i = 0; i < NB_SUBFR; i++) { min_gain_Q16 = SKP_min(min_gain_Q16, psEncCtrl->Gains_Q16[i]); } #if !VARQ LZ = SKP_Silk_CLZ32(min_gain_Q16) - 1; LZ = SKP_LIMIT(LZ, 0, 16); min_gain_Q16 = SKP_RSHIFT(min_gain_Q16, 2); /* Ensure that maximum invGains_Q16 is within range of a 16 bit int */ #endif for (i = 0; i < NB_SUBFR; i++) { /* Divide to Q16 */ assert(psEncCtrl->Gains_Q16[i] > 0); #if VARQ /* Invert and normalize gains, and ensure that maximum invGains_Q16 is within range of a 16 bit int */ invGains_Q16[i] = SKP_DIV32_varQ(min_gain_Q16, psEncCtrl->Gains_Q16[i], 16 - 2); #else invGains_Q16[i] = SKP_DIV32(SKP_LSHIFT(min_gain_Q16, LZ), SKP_RSHIFT(psEncCtrl->Gains_Q16[i], 16 - LZ)); #endif /* Ensure Wght_Q15 a minimum value 1 */ invGains_Q16[i] = SKP_max(invGains_Q16[i], 363); /* Square the inverted gains */ assert(invGains_Q16[i] == SKP_SAT16(invGains_Q16[i])); tmp = SKP_SMULWB(invGains_Q16[i], invGains_Q16[i]); Wght_Q15[i] = SKP_RSHIFT(tmp, 1); /* Invert the inverted and normalized gains */ local_gains_Qx[i] = SKP_DIV32((1 << (16 + Qx)), invGains_Q16[i]); } if (psEncCtrl->sCmn.sigtype == SIG_TYPE_VOICED) { /**********/ /* VOICED */ /**********/ assert(psEnc->sCmn.frame_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->sCmn.pitchL[0] + LTP_ORDER / 2); /* LTP analysis */ SKP_Silk_find_LTP_FIX(psEncCtrl->LTPCoef_Q14, WLTP, &psEncCtrl->LTPredCodGain_Q7, res_pitch, res_pitch + SKP_RSHIFT(psEnc->sCmn.frame_length, 1), psEncCtrl->sCmn.pitchL, Wght_Q15, psEnc->sCmn.subfr_length, psEnc->sCmn.frame_length, LTP_corrs_rshift); /* Quantize LTP gain parameters */ SKP_Silk_quant_LTP_gains_FIX(psEncCtrl->LTPCoef_Q14, psEncCtrl->sCmn.LTPIndex, &psEncCtrl->sCmn.PERIndex, WLTP, psEnc->mu_LTP_Q8, psEnc->sCmn.LTPQuantLowComplexity); /* Control LTP scaling */ SKP_Silk_LTP_scale_ctrl_FIX(psEnc, psEncCtrl); /* Create LTP residual */ SKP_Silk_LTP_analysis_filter_FIX(LPC_in_pre, psEnc->x_buf + psEnc->sCmn.frame_length - psEnc->sCmn.predictLPCOrder, psEncCtrl->LTPCoef_Q14, psEncCtrl->sCmn.pitchL, invGains_Q16, 16, psEnc->sCmn.subfr_length, psEnc->sCmn.predictLPCOrder); } else { /************/ /* UNVOICED */ /************/ /* Create signal with prepended subframes, scaled by inverse gains */ x_ptr = psEnc->x_buf + psEnc->sCmn.frame_length - psEnc->sCmn.predictLPCOrder; x_pre_ptr = LPC_in_pre; for (i = 0; i < NB_SUBFR; i++) { SKP_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; } SKP_memset(psEncCtrl->LTPCoef_Q14, 0, NB_SUBFR * LTP_ORDER * sizeof(int16_t)); psEncCtrl->LTPredCodGain_Q7 = 0; } /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ TIC(FIND_LPC) SKP_Silk_find_LPC_FIX(NLSF_Q15, &psEncCtrl->sCmn.NLSFInterpCoef_Q2, psEnc->sPred.prev_NLSFq_Q15, psEnc->sCmn.useInterpolatedNLSFs * (1 - psEnc-> sCmn. first_frame_after_reset), psEnc->sCmn.predictLPCOrder, LPC_in_pre, psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder); TOC(FIND_LPC) /* Quantize LSFs */ TIC(PROCESS_LSFS) SKP_Silk_process_NLSFs_FIX(psEnc, psEncCtrl, NLSF_Q15); TOC(PROCESS_LSFS) /* Calculate residual energy using quantized LPC coefficients */ SKP_Silk_residual_energy_FIX(psEncCtrl->ResNrg, psEncCtrl->ResNrgQ, LPC_in_pre, (const int16_t(*)[])psEncCtrl->PredCoef_Q12, local_gains_Qx, Qx, psEnc->sCmn.subfr_length, psEnc->sCmn.predictLPCOrder); /* Copy to prediction struct for use in next frame for fluctuation reduction */ SKP_memcpy(psEnc->sPred.prev_NLSFq_Q15, NLSF_Q15, psEnc->sCmn.predictLPCOrder * sizeof(int)); }
void SKP_Silk_find_pred_coefs_FIX( SKP_Silk_encoder_state_FIX *psEnc, /* I/O encoder state */ SKP_Silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ const SKP_int16 res_pitch[], /* I Residual from pitch analysis */ const SKP_int16 x[] /* I Speech signal */ ) { SKP_int i; SKP_int32 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; SKP_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ], Wght_Q15[ MAX_NB_SUBFR ]; SKP_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const SKP_int16 *x_ptr; SKP_int16 *x_pre_ptr, LPC_in_pre[ MAX_NB_SUBFR * MAX_LPC_ORDER + MAX_FRAME_LENGTH ]; SKP_int32 tmp, min_gain_Q16; SKP_int LTP_corrs_rshift[ MAX_NB_SUBFR ]; /* weighting for weighted least squares */ min_gain_Q16 = SKP_int32_MAX >> 6; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { min_gain_Q16 = SKP_min( min_gain_Q16, psEncCtrl->Gains_Q16[ i ] ); } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { /* Divide to Q16 */ SKP_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 ] = SKP_DIV32_varQ( min_gain_Q16, psEncCtrl->Gains_Q16[ i ], 16 - 2 ); /* Ensure Wght_Q15 a minimum value 1 */ invGains_Q16[ i ] = SKP_max( invGains_Q16[ i ], 363 ); /* Square the inverted gains */ SKP_assert( invGains_Q16[ i ] == SKP_SAT16( invGains_Q16[ i ] ) ); tmp = SKP_SMULWB( invGains_Q16[ i ], invGains_Q16[ i ] ); Wght_Q15[ i ] = SKP_RSHIFT( tmp, 1 ); /* Invert the inverted and normalized gains */ local_gains[ i ] = SKP_DIV32( ( 1 << 16 ), invGains_Q16[ i ] ); } if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /**********/ /* VOICED */ /**********/ SKP_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); /* LTP analysis */ SKP_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 */ SKP_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 */ SKP_Silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl ); /* Create LTP residual */ SKP_Silk_LTP_analysis_filter_FIX( LPC_in_pre, psEnc->x_buf + psEnc->sCmn.ltp_mem_length - 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++ ) { SKP_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; } SKP_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( SKP_int16 ) ); psEncCtrl->LTPredCodGain_Q7 = 0; } /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ TIC(FIND_LPC) SKP_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 ); TOC(FIND_LPC) /* Quantize LSFs */ TIC(PROCESS_LSFS) SKP_Silk_process_NLSFs( &psEnc->sCmn, psEncCtrl->PredCoef_Q12, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 ); TOC(PROCESS_LSFS) /* Calculate residual energy using quantized LPC coefficients */ SKP_Silk_residual_energy_FIX( psEncCtrl->ResNrg, psEncCtrl->ResNrgQ, LPC_in_pre, psEncCtrl->PredCoef_Q12, local_gains, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); /* Copy to prediction struct for use in next frame for fluctuation reduction */ SKP_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); }
/* Control encoder SNR */ SKP_int SKP_Silk_control_encoder_FIX( SKP_Silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk encoder state */ const SKP_int API_fs_kHz, /* I External (API) sampling rate (kHz) */ const SKP_int PacketSize_ms, /* I Packet length (ms) */ SKP_int32 TargetRate_bps, /* I Target max bitrate (bps) (used if SNR_dB == 0) */ const SKP_int PacketLoss_perc, /* I Packet loss rate (in percent) */ const SKP_int INBandFec_enabled, /* I Enable (1) / disable (0) inband FEC */ const SKP_int DTX_enabled, /* I Enable / disable DTX */ const SKP_int InputFramesize_ms, /* I Inputframe in ms */ const SKP_int Complexity /* I Complexity (0->low; 1->medium; 2->high) */ ) { SKP_int32 LBRRRate_thres_bps; SKP_int k, fs_kHz, ret = 0; SKP_int32 frac_Q6; const SKP_int32 *rateTable; /* State machine for the SWB/WB switching */ fs_kHz = psEnc->sCmn.fs_kHz; /* Only switch during low speech activity, when no frames are sitting in the payload buffer */ if( API_fs_kHz == 8 || fs_kHz == 0 || API_fs_kHz < fs_kHz ) { // Switching is not possible, encoder just initialized, or internal mode higher than external fs_kHz = API_fs_kHz; } else { /* Resample all valid data in x_buf. Resampling the last part gets rid of a click, 5ms after switching */ /* this is because the same state is used when downsampling in API.c and is then up to date */ /* the click immidiatly after switching is most of the time still there */ if( psEnc->sCmn.fs_kHz == 24 ) { /* Accumulate the difference between the target rate and limit */ if( psEnc->sCmn.fs_kHz_changed == 0 ) { psEnc->sCmn.bitrateDiff += SKP_MUL( InputFramesize_ms, TargetRate_bps - SWB2WB_BITRATE_BPS_INITIAL ); } else { psEnc->sCmn.bitrateDiff += SKP_MUL( InputFramesize_ms, TargetRate_bps - SWB2WB_BITRATE_BPS ); } psEnc->sCmn.bitrateDiff = SKP_min( psEnc->sCmn.bitrateDiff, 0 ); /* Check if we should switch from 24 to 16 kHz */ #if SWITCH_TRANSITION_FILTERING if( ( psEnc->sCmn.sLP.transition_frame_no == 0 ) && /* Transition phase not active */ ( psEnc->sCmn.bitrateDiff <= -ACCUM_BITS_DIFF_THRESHOLD || psEnc->sCmn.sSWBdetect.WB_detected == 1 ) && ( psEnc->speech_activity_Q8 < 128 && psEnc->sCmn.nFramesInPayloadBuf == 0 ) ) { psEnc->sCmn.sLP.transition_frame_no = 1; /* Begin transition phase */ psEnc->sCmn.sLP.mode = 0; /* Switch down */ } if( ( psEnc->sCmn.sLP.transition_frame_no >= TRANSITION_FRAMES_DOWN ) && ( psEnc->sCmn.sLP.mode == 0 ) && /* Transition phase complete, ready to switch */ #else if( ( psEnc->sCmn.bitrateDiff <= -ACCUM_BITS_DIFF_THRESHOLD || psEnc->sCmn.sSWBdetect.WB_detected == 1 ) && #endif ( psEnc->speech_activity_Q8 < 128 && psEnc->sCmn.nFramesInPayloadBuf == 0 ) ) { SKP_int16 x_buf[ 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ]; SKP_int16 x_bufout[ 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ]; psEnc->sCmn.bitrateDiff = 0; fs_kHz = 16; SKP_memcpy( x_buf, psEnc->x_buf, ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) * sizeof( SKP_int16 ) ); SKP_memset( psEnc->sCmn.resample24To16state, 0, sizeof( psEnc->sCmn.resample24To16state ) ); #if LOW_COMPLEXITY_ONLY { SKP_int16 scratch[ ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) + SigProc_Resample_2_3_coarse_NUM_FIR_COEFS - 1 ]; SKP_Silk_resample_2_3_coarse( &x_bufout[ 0 ], psEnc->sCmn.resample24To16state, &x_buf[ 0 ], SKP_LSHIFT( psEnc->sCmn.frame_length, 1 ) + psEnc->sCmn.la_shape, (SKP_int16*)scratch ); } #else SKP_Silk_resample_2_3( &x_bufout[ 0 ], psEnc->sCmn.resample24To16state, &x_buf[ 0 ], SKP_LSHIFT( psEnc->sCmn.frame_length, 1 ) + psEnc->sCmn.la_shape ); #endif /* set the first frame to zero, no performance difference was noticed though */ SKP_memset( x_bufout, 0, 320 * sizeof( SKP_int16 ) ); SKP_memcpy( psEnc->x_buf, x_bufout, ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) * sizeof( SKP_int16 ) ); #if SWITCH_TRANSITION_FILTERING psEnc->sCmn.sLP.transition_frame_no = 0; /* Transition phase complete */ #endif } } else if( psEnc->sCmn.fs_kHz == 16 ) { /* Check if we should switch from 16 to 24 kHz */ #if SWITCH_TRANSITION_FILTERING if( ( psEnc->sCmn.sLP.transition_frame_no == 0 ) && /* No transition phase running, ready to switch */ #else if( #endif ( API_fs_kHz > psEnc->sCmn.fs_kHz && TargetRate_bps >= WB2SWB_BITRATE_BPS && psEnc->sCmn.sSWBdetect.WB_detected == 0 ) && ( psEnc->speech_activity_Q8 < 128 && psEnc->sCmn.nFramesInPayloadBuf == 0 ) ) { SKP_int16 x_buf[ 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ]; SKP_int16 x_bufout[ 3 * ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) / 2 ]; SKP_int32 resample16To24state[ 11 ]; psEnc->sCmn.bitrateDiff = 0; fs_kHz = 24; SKP_memcpy( x_buf, psEnc->x_buf, ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) * sizeof( SKP_int16 ) ); SKP_memset( resample16To24state, 0, sizeof(resample16To24state) ); SKP_Silk_resample_3_2( &x_bufout[ 0 ], resample16To24state, &x_buf[ 0 ], SKP_LSHIFT( psEnc->sCmn.frame_length, 1 ) + psEnc->sCmn.la_shape ); /* set the first frame to zero, no performance difference was noticed though */ SKP_memset( x_bufout, 0, 480 * sizeof( SKP_int16 ) ); SKP_memcpy( psEnc->x_buf, x_bufout, ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) * sizeof( SKP_int16 ) ); #if SWITCH_TRANSITION_FILTERING psEnc->sCmn.sLP.mode = 1; /* Switch up */ #endif } else { /* accumulate the difference between the target rate and limit */ psEnc->sCmn.bitrateDiff += SKP_MUL( InputFramesize_ms, TargetRate_bps - WB2MB_BITRATE_BPS ); psEnc->sCmn.bitrateDiff = SKP_min( psEnc->sCmn.bitrateDiff, 0 ); /* Check if we should switch from 16 to 12 kHz */ #if SWITCH_TRANSITION_FILTERING if( ( psEnc->sCmn.sLP.transition_frame_no == 0 ) && /* Transition phase not active */ ( psEnc->sCmn.bitrateDiff <= -ACCUM_BITS_DIFF_THRESHOLD ) && ( psEnc->speech_activity_Q8 < 128 && psEnc->sCmn.nFramesInPayloadBuf == 0 ) ) { psEnc->sCmn.sLP.transition_frame_no = 1; /* Begin transition phase */ psEnc->sCmn.sLP.mode = 0; /* Switch down */ } if( ( psEnc->sCmn.sLP.transition_frame_no >= TRANSITION_FRAMES_DOWN ) && ( psEnc->sCmn.sLP.mode == 0 ) && /* Transition phase complete, ready to switch */ #else if( ( psEnc->sCmn.bitrateDiff <= -ACCUM_BITS_DIFF_THRESHOLD ) && #endif ( psEnc->speech_activity_Q8 < 128 && psEnc->sCmn.nFramesInPayloadBuf == 0 ) ) { SKP_int16 x_buf[ 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ]; SKP_memcpy( x_buf, psEnc->x_buf, ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) * sizeof( SKP_int16 ) ); psEnc->sCmn.bitrateDiff = 0; fs_kHz = 12; if( API_fs_kHz == 24 ) { /* Intermediate upsampling of x_bufFIX from 16 to 24 kHz */ SKP_int16 x_buf24[ 3 * ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) / 2 ]; SKP_int32 scratch[ 3 * ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) ]; SKP_int32 resample16To24state[ 11 ]; SKP_memset( resample16To24state, 0, sizeof( resample16To24state ) ); SKP_Silk_resample_3_2( &x_buf24[ 0 ], resample16To24state, &x_buf[ 0 ], SKP_LSHIFT( psEnc->sCmn.frame_length, 1 ) + psEnc->sCmn.la_shape ); /* Update the state of the resampler used in API.c, from 24 to 12 kHz */ SKP_memset( psEnc->sCmn.resample24To12state, 0, sizeof( psEnc->sCmn.resample24To12state ) ); SKP_Silk_resample_1_2_coarse( &x_buf24[ 0 ], psEnc->sCmn.resample24To12state, &x_buf[ 0 ], scratch, SKP_RSHIFT( SKP_SMULBB( 3, SKP_LSHIFT( psEnc->sCmn.frame_length, 1 ) + psEnc->sCmn.la_shape ), 2 ) ); /* set the first frame to zero, no performance difference was noticed though */ SKP_memset( x_buf, 0, 240 * sizeof( SKP_int16 ) ); SKP_memcpy( psEnc->x_buf, x_buf, ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) * sizeof( SKP_int16 ) ); } else if( API_fs_kHz == 16 ) { SKP_int16 x_bufout[ 3 * ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) / 4 ]; SKP_memset( psEnc->sCmn.resample16To12state, 0, sizeof( psEnc->sCmn.resample16To12state ) ); SKP_Silk_resample_3_4( &x_bufout[ 0 ], psEnc->sCmn.resample16To12state, &x_buf[ 0 ], SKP_LSHIFT( psEnc->sCmn.frame_length, 1 ) + psEnc->sCmn.la_shape ); /* set the first frame to zero, no performance difference was noticed though */ SKP_memset( x_bufout, 0, 240 * sizeof( SKP_int16 ) ); SKP_memcpy( psEnc->x_buf, x_bufout, ( 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ) * sizeof( SKP_int16 ) ); } #if SWITCH_TRANSITION_FILTERING psEnc->sCmn.sLP.transition_frame_no = 0; /* Transition phase complete */ #endif } } } else if( psEnc->sCmn.fs_kHz == 12 ) {
SKP_int SKP_Silk_SDK_Encode( void *encState, /* I/O: State */ SKP_Silk_EncodeControlStruct *encControl, /* I: Control structure */ const SKP_int16 *samplesIn, /* I: Speech sample input vector */ SKP_int nSamplesIn, /* I: Number of samples in input vector */ ec_enc *psRangeEnc, /* I/O Compressor data structure */ SKP_int32 *nBytesOut, /* I/O: Number of bytes in payload (input: Max bytes) */ const SKP_int prefillFlag /* I: Flag to indicate prefilling buffers no coding */ ) { SKP_int tmp_payloadSize_ms, tmp_complexity, ret = 0; SKP_int nSamplesToBuffer, nBlocksOf10ms, nSamplesFromInput = 0; SKP_Silk_encoder_state_Fxx *psEnc = ( SKP_Silk_encoder_state_Fxx* )encState; ret = process_enc_control_struct( psEnc, encControl ); nBlocksOf10ms = SKP_DIV32( 100 * nSamplesIn, psEnc->sCmn.API_fs_Hz ); if( prefillFlag ) { /* Only accept input length of 10 ms */ if( nBlocksOf10ms != 1 ) { ret = SKP_SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES; SKP_assert( 0 ); return( ret ); } /* Reset Encoder */ if( ret = SKP_Silk_init_encoder_Fxx( psEnc ) ) { SKP_assert( 0 ); } tmp_payloadSize_ms = encControl->payloadSize_ms; encControl->payloadSize_ms = 10; tmp_complexity = encControl->complexity; encControl->complexity = 0; ret = process_enc_control_struct( psEnc, encControl ); psEnc->sCmn.prefillFlag = 1; } else { /* Only accept input lengths that are a multiple of 10 ms */ if( nBlocksOf10ms * psEnc->sCmn.API_fs_Hz != 100 * nSamplesIn || nSamplesIn < 0 ) { ret = SKP_SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES; SKP_assert( 0 ); return( ret ); } /* Make sure no more than one packet can be produced */ if( 1000 * (SKP_int32)nSamplesIn > psEnc->sCmn.PacketSize_ms * psEnc->sCmn.API_fs_Hz ) { ret = SKP_SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES; SKP_assert( 0 ); return( ret ); } } /* Input buffering/resampling and encoding */ while( 1 ) { nSamplesToBuffer = psEnc->sCmn.frame_length - psEnc->sCmn.inputBufIx; if( psEnc->sCmn.API_fs_Hz == SKP_SMULBB( 1000, psEnc->sCmn.fs_kHz ) ) { nSamplesToBuffer = SKP_min_int( nSamplesToBuffer, nSamplesIn ); nSamplesFromInput = nSamplesToBuffer; /* Copy to buffer */ SKP_memcpy( &psEnc->sCmn.inputBuf[ psEnc->sCmn.inputBufIx ], samplesIn, nSamplesFromInput * sizeof( SKP_int16 ) ); } else { nSamplesToBuffer = SKP_min( nSamplesToBuffer, 10 * nBlocksOf10ms * psEnc->sCmn.fs_kHz ); nSamplesFromInput = SKP_DIV32_16( nSamplesToBuffer * psEnc->sCmn.API_fs_Hz, psEnc->sCmn.fs_kHz * 1000 ); /* Resample and write to buffer */ ret += SKP_Silk_resampler( &psEnc->sCmn.resampler_state, &psEnc->sCmn.inputBuf[ psEnc->sCmn.inputBufIx ], samplesIn, nSamplesFromInput ); } samplesIn += nSamplesFromInput; nSamplesIn -= nSamplesFromInput; psEnc->sCmn.inputBufIx += nSamplesToBuffer; /* Silk encoder */ if( psEnc->sCmn.inputBufIx >= psEnc->sCmn.frame_length ) { SKP_assert( psEnc->sCmn.inputBufIx == psEnc->sCmn.frame_length ); /* Enough data in input buffer, so encode */ if( ( ret = SKP_Silk_encode_frame_Fxx( psEnc, nBytesOut, psRangeEnc ) ) != 0 ) { SKP_assert( 0 ); } psEnc->sCmn.inputBufIx = 0; psEnc->sCmn.controlled_since_last_payload = 0; if( nSamplesIn == 0 ) { break; } } else { break; } } if( prefillFlag ) { encControl->payloadSize_ms = tmp_payloadSize_ms; encControl->complexity = tmp_complexity; ret = process_enc_control_struct( psEnc, encControl ); psEnc->sCmn.prefillFlag = 0; } return ret; }
/* Control internal sampling rate */ SKP_int SKP_Silk_control_audio_bandwidth( SKP_Silk_encoder_state *psEncC, /* I/O Pointer to Silk encoder state */ const SKP_int32 TargetRate_bps /* I Target max bitrate (bps) */ ) { SKP_int fs_kHz; fs_kHz = psEncC->fs_kHz; if( fs_kHz == 0 ) { /* Encoder has just been initialized */ if( TargetRate_bps >= SWB2WB_BITRATE_BPS ) { fs_kHz = 24; } else if( TargetRate_bps >= WB2MB_BITRATE_BPS ) { fs_kHz = 16; } else if( TargetRate_bps >= MB2NB_BITRATE_BPS ) { fs_kHz = 12; } else { fs_kHz = 8; } /* Make sure internal rate is not higher than external rate or maximum allowed, or lower than minimum allowed */ fs_kHz = SKP_min( fs_kHz, SKP_DIV32_16( psEncC->API_fs_Hz, 1000 ) ); fs_kHz = SKP_min( fs_kHz, psEncC->maxInternal_fs_kHz ); } else if( SKP_SMULBB( fs_kHz, 1000 ) > psEncC->API_fs_Hz || fs_kHz > psEncC->maxInternal_fs_kHz ) { /* Make sure internal rate is not higher than external rate or maximum allowed */ fs_kHz = SKP_DIV32_16( psEncC->API_fs_Hz, 1000 ); fs_kHz = SKP_min( fs_kHz, psEncC->maxInternal_fs_kHz ); } else { /* State machine for the internal sampling rate switching */ if( psEncC->API_fs_Hz > 8000 ) { /* Accumulate the difference between the target rate and limit for switching down */ psEncC->bitrateDiff += SKP_MUL( psEncC->PacketSize_ms, psEncC->TargetRate_bps - psEncC->bitrate_threshold_down ); psEncC->bitrateDiff = SKP_min( psEncC->bitrateDiff, 0 ); if( psEncC->vadFlag == NO_VOICE_ACTIVITY ) { /* Low speech activity */ /* Check if we should switch down */ #if SWITCH_TRANSITION_FILTERING if( ( psEncC->sLP.transition_frame_no == 0 ) && /* Transition phase not active */ ( psEncC->bitrateDiff <= -ACCUM_BITS_DIFF_THRESHOLD || /* Bitrate threshold is met */ ( psEncC->sSWBdetect.WB_detected * psEncC->fs_kHz == 24 ) ) ) { /* Forced down-switching due to WB input */ psEncC->sLP.transition_frame_no = 1; /* Begin transition phase */ psEncC->sLP.mode = 0; /* Switch down */ } else if( ( psEncC->sLP.transition_frame_no >= TRANSITION_FRAMES_DOWN ) && /* Transition phase complete */ ( psEncC->sLP.mode == 0 ) ) { /* Ready to switch down */ psEncC->sLP.transition_frame_no = 0; /* Ready for new transition phase */ #else if( psEncC->bitrateDiff <= -ACCUM_BITS_DIFF_THRESHOLD ) { /* Bitrate threshold is met */ #endif psEncC->bitrateDiff = 0; /* Switch to a lower sample frequency */ if( psEncC->fs_kHz == 24 ) { fs_kHz = 16; } else if( psEncC->fs_kHz == 16 ) { fs_kHz = 12; } else { SKP_assert( psEncC->fs_kHz == 12 ); fs_kHz = 8; } } /* Check if we should switch up */ if( ( ( psEncC->fs_kHz * 1000 < psEncC->API_fs_Hz ) && ( psEncC->TargetRate_bps >= psEncC->bitrate_threshold_up ) && ( psEncC->sSWBdetect.WB_detected * psEncC->fs_kHz < 16 ) ) && ( ( ( psEncC->fs_kHz == 16 ) && ( psEncC->maxInternal_fs_kHz >= 24 ) ) || ( ( psEncC->fs_kHz == 12 ) && ( psEncC->maxInternal_fs_kHz >= 16 ) ) || ( ( psEncC->fs_kHz == 8 ) && ( psEncC->maxInternal_fs_kHz >= 12 ) ) ) #if SWITCH_TRANSITION_FILTERING && ( psEncC->sLP.transition_frame_no == 0 ) ) { /* No transition phase running, ready to switch */ psEncC->sLP.mode = 1; /* Switch up */ #else ) { #endif psEncC->bitrateDiff = 0; /* Switch to a higher sample frequency */ if( psEncC->fs_kHz == 8 ) { fs_kHz = 12; } else if( psEncC->fs_kHz == 12 ) { fs_kHz = 16; } else { SKP_assert( psEncC->fs_kHz == 16 ); fs_kHz = 24; } } } }
/* Control encoder SNR */ SKP_int SKP_Silk_control_encoder_FIX( SKP_Silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk encoder state */ const SKP_int32 API_fs_Hz, /* I External (API) sampling rate (Hz) */ const SKP_int max_internal_fs_kHz,/* I Maximum internal sampling rate (kHz) */ const SKP_int PacketSize_ms, /* I Packet length (ms) */ SKP_int32 TargetRate_bps, /* I Target max bitrate (bps) (used if SNR_dB == 0) */ const SKP_int PacketLoss_perc, /* I Packet loss rate (in percent) */ const SKP_int INBandFEC_enabled, /* I Enable (1) / disable (0) inband FEC */ const SKP_int DTX_enabled, /* I Enable / disable DTX */ const SKP_int InputFramesize_ms, /* I Inputframe in ms */ const SKP_int Complexity /* I Complexity (0->low; 1->medium; 2->high) */ ) { SKP_int32 LBRRRate_thres_bps; SKP_int k, fs_kHz, ret = 0; SKP_int32 frac_Q6; const SKP_int32 *rateTable; /* State machine for the SWB/WB switching */ fs_kHz = psEnc->sCmn.fs_kHz; /* Only switch during low speech activity, when no frames are sitting in the payload buffer */ if( API_fs_Hz == 8000 || fs_kHz == 0 || API_fs_Hz < SKP_SMULBB( fs_kHz, 1000 ) || fs_kHz > max_internal_fs_kHz ) { /* Switching is not possible, encoder just initialized, internal mode higher than external, */ /* or internal mode higher than maximum allowed internal mode */ fs_kHz = SKP_min( SKP_DIV32_16( API_fs_Hz, 1000 ), max_internal_fs_kHz ); } else { /* Accumulate the difference between the target rate and limit for switching down */ psEnc->sCmn.bitrateDiff += SKP_MUL( InputFramesize_ms, TargetRate_bps - psEnc->sCmn.bitrate_threshold_down ); psEnc->sCmn.bitrateDiff = SKP_min( psEnc->sCmn.bitrateDiff, 0 ); if( psEnc->speech_activity_Q8 < 128 && psEnc->sCmn.nFramesInPayloadBuf == 0 ) { /* Low speech activity and payload buffer empty */ /* Check if we should switch down */ #if SWITCH_TRANSITION_FILTERING if( ( psEnc->sCmn.sLP.transition_frame_no == 0 ) && /* Transition phase not active */ ( psEnc->sCmn.bitrateDiff <= -ACCUM_BITS_DIFF_THRESHOLD || /* Bitrate threshold is met */ ( psEnc->sCmn.sSWBdetect.WB_detected * psEnc->sCmn.fs_kHz == 24 ) ) ) { /* Forced down-switching due to WB input */ psEnc->sCmn.sLP.transition_frame_no = 1; /* Begin transition phase */ psEnc->sCmn.sLP.mode = 0; /* Switch down */ } else if( ( psEnc->sCmn.sLP.transition_frame_no >= TRANSITION_FRAMES_DOWN ) && /* Transition phase complete */ ( psEnc->sCmn.sLP.mode == 0 ) ) { /* Ready to switch down */ psEnc->sCmn.sLP.transition_frame_no = 0; /* Ready for new transition phase */ #else if( psEnc->sCmn.bitrateDiff <= -ACCUM_BITS_DIFF_THRESHOLD ) { /* Bitrate threshold is met */ #endif psEnc->sCmn.bitrateDiff = 0; /* Switch to a lower sample frequency */ if( psEnc->sCmn.fs_kHz == 24 ) { fs_kHz = 16; } else if( psEnc->sCmn.fs_kHz == 16 ) { fs_kHz = 12; } else { SKP_assert( psEnc->sCmn.fs_kHz == 12 ); fs_kHz = 8; } } /* Check if we should switch up */ if( ( ( SKP_SMULBB( psEnc->sCmn.fs_kHz, 1000 ) < API_fs_Hz ) && ( TargetRate_bps >= psEnc->sCmn.bitrate_threshold_up ) && ( psEnc->sCmn.sSWBdetect.WB_detected * psEnc->sCmn.fs_kHz != 16 ) ) && ( ( psEnc->sCmn.fs_kHz == 16 ) && ( max_internal_fs_kHz >= 24 ) || ( psEnc->sCmn.fs_kHz == 12 ) && ( max_internal_fs_kHz >= 16 ) || ( psEnc->sCmn.fs_kHz == 8 ) && ( max_internal_fs_kHz >= 12 ) ) #if SWITCH_TRANSITION_FILTERING && ( psEnc->sCmn.sLP.transition_frame_no == 0 ) ) { /* No transition phase running, ready to switch */ psEnc->sCmn.sLP.mode = 1; /* Switch up */ #else ) { #endif psEnc->sCmn.bitrateDiff = 0; /* Switch to a higher sample frequency */ if( psEnc->sCmn.fs_kHz == 8 ) { fs_kHz = 12; } else if( psEnc->sCmn.fs_kHz == 12 ) { fs_kHz = 16; } else { SKP_assert( psEnc->sCmn.fs_kHz == 16 ); fs_kHz = 24; } } } }
/* Compute reflection coefficients from input signal */ void SKP_Silk_burg_modified( SKP_int32 *res_nrg, /* O residual energy */ SKP_int *res_nrg_Q, /* O residual energy Q value */ SKP_int32 A_Q16[], /* O prediction coefficients (length order) */ const SKP_int16 x[], /* I input signal, length: nb_subfr * ( D + subfr_length ) */ const SKP_int subfr_length, /* I input signal subframe length (including D preceeding samples) */ const SKP_int nb_subfr, /* I number of subframes stacked in x */ const SKP_int32 WhiteNoiseFrac_Q32, /* I fraction added to zero-lag autocorrelation */ const SKP_int D /* I order */ ) { SKP_int k, n, s, lz, rshifts, rshifts_extra; SKP_int32 C0, num, nrg, rc_Q31, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; const SKP_int16 *x_ptr; SKP_int32 C_first_row[ SKP_Silk_MAX_ORDER_LPC ]; SKP_int32 C_last_row[ SKP_Silk_MAX_ORDER_LPC ]; SKP_int32 Af_QA[ SKP_Silk_MAX_ORDER_LPC ]; SKP_int32 CAf[ SKP_Silk_MAX_ORDER_LPC + 1 ]; SKP_int32 CAb[ SKP_Silk_MAX_ORDER_LPC + 1 ]; SKP_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); SKP_assert( nb_subfr <= MAX_NB_SUBFR ); /* Compute autocorrelations, added over subframes */ SKP_Silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length ); if( rshifts > MAX_RSHIFTS ) { C0 = SKP_LSHIFT32( C0, rshifts - MAX_RSHIFTS ); SKP_assert( C0 > 0 ); rshifts = MAX_RSHIFTS; } else { lz = SKP_Silk_CLZ32( C0 ) - 1; rshifts_extra = N_BITS_HEAD_ROOM - lz; if( rshifts_extra > 0 ) { rshifts_extra = SKP_min( rshifts_extra, MAX_RSHIFTS - rshifts ); C0 = SKP_RSHIFT32( C0, rshifts_extra ); } else { rshifts_extra = SKP_max( rshifts_extra, MIN_RSHIFTS - rshifts ); C0 = SKP_LSHIFT32( C0, -rshifts_extra ); } rshifts += rshifts_extra; } SKP_memset( C_first_row, 0, SKP_Silk_MAX_ORDER_LPC * sizeof( SKP_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 ] += (SKP_int32)SKP_RSHIFT64( SKP_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 ] += SKP_LSHIFT32( SKP_Silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts ); } } } SKP_memcpy( C_last_row, C_first_row, SKP_Silk_MAX_ORDER_LPC * sizeof( SKP_int32 ) ); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + SKP_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 = -SKP_LSHIFT32( (SKP_int32)x_ptr[ n ], 16 - rshifts ); // Q(16-rshifts) x2 = -SKP_LSHIFT32( (SKP_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); // Q(16-rshifts) tmp1 = SKP_LSHIFT32( (SKP_int32)x_ptr[ n ], QA - 16 ); // Q(QA-16) tmp2 = SKP_LSHIFT32( (SKP_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); // Q(QA-16) for( k = 0; k < n; k++ ) { C_first_row[ k ] = SKP_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); // Q( -rshifts ) C_last_row[ k ] = SKP_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); // Q( -rshifts ) Atmp_QA = Af_QA[ k ]; tmp1 = SKP_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); // Q(QA-16) tmp2 = SKP_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); // Q(QA-16) } tmp1 = SKP_LSHIFT32( -tmp1, 32 - QA - rshifts ); // Q(16-rshifts) tmp2 = SKP_LSHIFT32( -tmp2, 32 - QA - rshifts ); // Q(16-rshifts) for( k = 0; k <= n; k++ ) { CAf[ k ] = SKP_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); // Q( -rshift ) CAb[ k ] = SKP_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 = -SKP_LSHIFT32( (SKP_int32)x_ptr[ n ], -rshifts ); // Q( -rshifts ) x2 = -SKP_LSHIFT32( (SKP_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); // Q( -rshifts ) tmp1 = SKP_LSHIFT32( (SKP_int32)x_ptr[ n ], 17 ); // Q17 tmp2 = SKP_LSHIFT32( (SKP_int32)x_ptr[ subfr_length - n - 1 ], 17 ); // Q17 for( k = 0; k < n; k++ ) { C_first_row[ k ] = SKP_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); // Q( -rshifts ) C_last_row[ k ] = SKP_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); // Q( -rshifts ) Atmp1 = SKP_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); // Q17 tmp1 = SKP_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); // Q17 tmp2 = SKP_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); // Q17 } tmp1 = -tmp1; // Q17 tmp2 = -tmp2; // Q17 for( k = 0; k <= n; k++ ) { CAf[ k ] = SKP_SMLAWW( CAf[ k ], tmp1, SKP_LSHIFT32( (SKP_int32)x_ptr[ n - k ], -rshifts - 1 ) ); // Q( -rshift ) CAb[ k ] = SKP_SMLAWW( CAb[ k ], tmp2, SKP_LSHIFT32( (SKP_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 = SKP_ADD32( CAb[ 0 ], CAf[ 0 ] ); // Q( 1-rshifts ) for( k = 0; k < n; k++ ) { Atmp_QA = Af_QA[ k ]; lz = SKP_Silk_CLZ32( SKP_abs( Atmp_QA ) ) - 1; lz = SKP_min( 32 - QA, lz ); Atmp1 = SKP_LSHIFT32( Atmp_QA, lz ); // Q( QA + lz ) tmp1 = SKP_ADD_LSHIFT32( tmp1, SKP_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); // Q( -rshifts ) tmp2 = SKP_ADD_LSHIFT32( tmp2, SKP_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); // Q( -rshifts ) num = SKP_ADD_LSHIFT32( num, SKP_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); // Q( -rshifts ) nrg = SKP_ADD_LSHIFT32( nrg, SKP_SMMUL( SKP_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 = SKP_ADD32( num, tmp2 ); // Q( -rshifts ) num = SKP_LSHIFT32( -num, 1 ); // Q( 1-rshifts ) /* Calculate the next order reflection (parcor) coefficient */ if( SKP_abs( num ) < nrg ) { rc_Q31 = SKP_DIV32_varQ( num, nrg, 31 ); } else { /* Negative energy or ratio too high; set remaining coefficients to zero and exit loop */ SKP_memset( &Af_QA[ n ], 0, ( D - n ) * sizeof( SKP_int32 ) ); SKP_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 ] = SKP_ADD_LSHIFT32( tmp1, SKP_SMMUL( tmp2, rc_Q31 ), 1 ); // QA Af_QA[ n - k - 1 ] = SKP_ADD_LSHIFT32( tmp2, SKP_SMMUL( tmp1, rc_Q31 ), 1 ); // QA } Af_QA[ n ] = SKP_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 ] = SKP_ADD_LSHIFT32( tmp1, SKP_SMMUL( tmp2, rc_Q31 ), 1 ); // Q( -rshifts ) CAb[ n - k + 1 ] = SKP_ADD_LSHIFT32( tmp2, SKP_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 = SKP_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); // Q16 nrg = SKP_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); // Q( -rshifts ) tmp1 = SKP_SMLAWW( tmp1, Atmp1, Atmp1 ); // Q16 A_Q16[ k ] = -Atmp1; } *res_nrg = SKP_SMLAWW( nrg, SKP_SMMUL( WhiteNoiseFrac_Q32, C0 ), -tmp1 ); // Q( -rshifts ) *res_nrg_Q = -rshifts; }
/* Resample with a 2x downsampler (optional), a 2nd order AR filter followed by FIR interpolation */ void SKP_Silk_resampler_private_down_FIR( void *SS, /* I/O: Resampler state */ SKP_int16 out[], /* O: Output signal */ const SKP_int16 in[], /* I: Input signal */ SKP_int32 inLen /* I: Number of input samples */ ) { SKP_Silk_resampler_state_struct *S = (SKP_Silk_resampler_state_struct *)SS; SKP_int32 nSamplesIn, interpol_ind; SKP_int32 max_index_Q16, index_Q16, index_increment_Q16, res_Q6; SKP_int16 buf1[ RESAMPLER_MAX_BATCH_SIZE_IN / 2 ]; SKP_int32 buf2[ RESAMPLER_MAX_BATCH_SIZE_IN + RESAMPLER_DOWN_ORDER_FIR ]; SKP_int32 *buf_ptr; const SKP_int16 *interpol_ptr, *FIR_Coefs; /* Copy buffered samples to start of buffer */ SKP_memcpy( buf2, S->sFIR, RESAMPLER_DOWN_ORDER_FIR * sizeof( SKP_int32 ) ); FIR_Coefs = &S->Coefs[ 2 ]; /* Iterate over blocks of frameSizeIn input samples */ index_increment_Q16 = S->invRatio_Q16; while( 1 ) { nSamplesIn = SKP_min( inLen, S->batchSize ); if( S->input2x == 1 ) { /* Downsample 2x */ SKP_Silk_resampler_down2( S->sDown2, buf1, in, nSamplesIn ); nSamplesIn = SKP_RSHIFT32( nSamplesIn, 1 ); /* Second-order AR filter (output in Q8) */ SKP_Silk_resampler_private_AR2( S->sIIR, &buf2[ RESAMPLER_DOWN_ORDER_FIR ], buf1, S->Coefs, nSamplesIn ); } else { /* Second-order AR filter (output in Q8) */ SKP_Silk_resampler_private_AR2( S->sIIR, &buf2[ RESAMPLER_DOWN_ORDER_FIR ], in, S->Coefs, nSamplesIn ); } max_index_Q16 = SKP_LSHIFT32( nSamplesIn, 16 ); /* Interpolate filtered signal */ if( S->FIR_Fracs == 1 ) { for( index_Q16 = 0; index_Q16 < max_index_Q16; index_Q16 += index_increment_Q16 ) { /* Integer part gives pointer to buffered input */ buf_ptr = buf2 + SKP_RSHIFT( index_Q16, 16 ); /* Inner product */ res_Q6 = SKP_SMULWB( SKP_ADD32( buf_ptr[ 0 ], buf_ptr[ 11 ] ), FIR_Coefs[ 0 ] ); res_Q6 = SKP_SMLAWB( res_Q6, SKP_ADD32( buf_ptr[ 1 ], buf_ptr[ 10 ] ), FIR_Coefs[ 1 ] ); res_Q6 = SKP_SMLAWB( res_Q6, SKP_ADD32( buf_ptr[ 2 ], buf_ptr[ 9 ] ), FIR_Coefs[ 2 ] ); res_Q6 = SKP_SMLAWB( res_Q6, SKP_ADD32( buf_ptr[ 3 ], buf_ptr[ 8 ] ), FIR_Coefs[ 3 ] ); res_Q6 = SKP_SMLAWB( res_Q6, SKP_ADD32( buf_ptr[ 4 ], buf_ptr[ 7 ] ), FIR_Coefs[ 4 ] ); res_Q6 = SKP_SMLAWB( res_Q6, SKP_ADD32( buf_ptr[ 5 ], buf_ptr[ 6 ] ), FIR_Coefs[ 5 ] ); /* Scale down, saturate and store in output array */ *out++ = (SKP_int16)SKP_SAT16( SKP_RSHIFT_ROUND( res_Q6, 6 ) ); } } else { for( index_Q16 = 0; index_Q16 < max_index_Q16; index_Q16 += index_increment_Q16 ) { /* Integer part gives pointer to buffered input */ buf_ptr = buf2 + SKP_RSHIFT( index_Q16, 16 ); /* Fractional part gives interpolation coefficients */ interpol_ind = SKP_SMULWB( index_Q16 & 0xFFFF, S->FIR_Fracs ); /* Inner product */ interpol_ptr = &FIR_Coefs[ RESAMPLER_DOWN_ORDER_FIR / 2 * interpol_ind ]; res_Q6 = SKP_SMULWB( buf_ptr[ 0 ], interpol_ptr[ 0 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 1 ], interpol_ptr[ 1 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 2 ], interpol_ptr[ 2 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 3 ], interpol_ptr[ 3 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 4 ], interpol_ptr[ 4 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 5 ], interpol_ptr[ 5 ] ); interpol_ptr = &FIR_Coefs[ RESAMPLER_DOWN_ORDER_FIR / 2 * ( S->FIR_Fracs - 1 - interpol_ind ) ]; res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 11 ], interpol_ptr[ 0 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 10 ], interpol_ptr[ 1 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 9 ], interpol_ptr[ 2 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 8 ], interpol_ptr[ 3 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 7 ], interpol_ptr[ 4 ] ); res_Q6 = SKP_SMLAWB( res_Q6, buf_ptr[ 6 ], interpol_ptr[ 5 ] ); /* Scale down, saturate and store in output array */ *out++ = (SKP_int16)SKP_SAT16( SKP_RSHIFT_ROUND( res_Q6, 6 ) ); } } in += nSamplesIn << S->input2x; inLen -= nSamplesIn << S->input2x; if( inLen > S->input2x ) { /* More iterations to do; copy last part of filtered signal to beginning of buffer */ SKP_memcpy( buf2, &buf2[ nSamplesIn ], RESAMPLER_DOWN_ORDER_FIR * sizeof( SKP_int32 ) ); } else { break; } } /* Copy last part of filtered signal to the state for the next call */ SKP_memcpy( S->sFIR, &buf2[ nSamplesIn ], RESAMPLER_DOWN_ORDER_FIR * sizeof( SKP_int32 ) ); }
void SKP_Silk_quant_LTP_gains_FIX( SKP_int16 B_Q14[], /* I/O (un)quantized LTP gains */ SKP_int cbk_index[], /* O Codebook Index */ SKP_int *periodicity_index, /* O Periodicity Index */ const SKP_int32 W_Q18[], /* I Error Weights in Q18 */ SKP_int mu_Q8, /* I Mu value (R/D tradeoff) */ SKP_int lowComplexity /* I Flag for low complexity */ ) { SKP_int j, k, temp_idx[ NB_SUBFR ], cbk_size; const SKP_uint16 *cdf_ptr; const SKP_int16 *cl_ptr; const SKP_int16 *cbk_ptr_Q14; const SKP_int16 *b_Q14_ptr; const SKP_int32 *W_Q18_ptr; SKP_int32 rate_dist_subfr, rate_dist, min_rate_dist; /***************************************************/ /* iterate over different codebooks with different */ /* rates/distortions, and choose best */ /***************************************************/ min_rate_dist = SKP_int32_MAX; for( k = 0; k < 3; k++ ) { cdf_ptr = SKP_Silk_LTP_gain_CDF_ptrs[ k ]; cl_ptr = SKP_Silk_LTP_gain_BITS_Q6_ptrs[ k ]; cbk_ptr_Q14 = SKP_Silk_LTP_vq_ptrs_Q14[ k ]; cbk_size = SKP_Silk_LTP_vq_sizes[ k ]; /* Setup pointer to first subframe */ W_Q18_ptr = W_Q18; b_Q14_ptr = B_Q14; rate_dist = 0; for( j = 0; j < NB_SUBFR; j++ ) { SKP_Silk_VQ_WMat_EC_FIX( &temp_idx[ j ], /* O index of best codebook vector */ &rate_dist_subfr, /* O best weighted quantization error + mu * rate */ b_Q14_ptr, /* I input vector to be quantized */ W_Q18_ptr, /* I weighting matrix */ cbk_ptr_Q14, /* I codebook */ cl_ptr, /* I code length for each codebook vector */ mu_Q8, /* I tradeoff between weighted error and rate */ cbk_size /* I number of vectors in codebook */ ); rate_dist = SKP_ADD_POS_SAT32( rate_dist, rate_dist_subfr ); b_Q14_ptr += LTP_ORDER; W_Q18_ptr += LTP_ORDER * LTP_ORDER; } /* Avoid never finding a codebook */ rate_dist = SKP_min( SKP_int32_MAX - 1, rate_dist ); if( rate_dist < min_rate_dist ) { min_rate_dist = rate_dist; SKP_memcpy( cbk_index, temp_idx, NB_SUBFR * sizeof( SKP_int ) ); *periodicity_index = k; } /* Break early in low-complexity mode if rate distortion is below threshold */ if( lowComplexity && ( rate_dist < SKP_Silk_LTP_gain_middle_avg_RD_Q14 ) ) { break; } } cbk_ptr_Q14 = SKP_Silk_LTP_vq_ptrs_Q14[ *periodicity_index ]; for( j = 0; j < NB_SUBFR; j++ ) { for( k = 0; k < LTP_ORDER; k++ ) { B_Q14[ j * LTP_ORDER + k ] = cbk_ptr_Q14[ SKP_MLA( k, cbk_index[ j ], LTP_ORDER ) ]; } } }
/* Control internal sampling rate */ SKP_int SKP_Silk_control_audio_bandwidth( SKP_Silk_encoder_state *psEncC, /* I/O Pointer to Silk encoder state */ SKP_int32 TargetRate_bps /* I Target max bitrate (bps) */ ) { SKP_int fs_kHz; fs_kHz = psEncC->fs_kHz; /* Reduce bitrate for 10 ms modes in these calculations */ if( psEncC->nb_subfr == 2 ) { TargetRate_bps -= REDUCE_BITRATE_10_MS_BPS; } if( fs_kHz == 0 ) { /* Encoder has just been initialized */ if( TargetRate_bps >= WB2MB_BITRATE_BPS ) { fs_kHz = 16; } else if( TargetRate_bps >= MB2NB_BITRATE_BPS ) { fs_kHz = 12; } else { fs_kHz = 8; } /* Make sure internal rate is not higher than external rate or maximum allowed, or lower than minimum allowed */ fs_kHz = SKP_min( fs_kHz, SKP_DIV32_16( psEncC->API_fs_Hz, 1000 ) ); fs_kHz = SKP_min( fs_kHz, psEncC->maxInternal_fs_kHz ); fs_kHz = SKP_max( fs_kHz, psEncC->minInternal_fs_kHz ); } else if( SKP_SMULBB( fs_kHz, 1000 ) > psEncC->API_fs_Hz || fs_kHz > psEncC->maxInternal_fs_kHz || fs_kHz < psEncC->minInternal_fs_kHz ) { /* Make sure internal rate is not higher than external rate or maximum allowed, or lower than minimum allowed */ fs_kHz = SKP_DIV32_16( psEncC->API_fs_Hz, 1000 ); fs_kHz = SKP_min( fs_kHz, psEncC->maxInternal_fs_kHz ); fs_kHz = SKP_max( fs_kHz, psEncC->minInternal_fs_kHz ); } else { /* State machine for the internal sampling rate switching */ if( psEncC->API_fs_Hz > 8000 && psEncC->prevSignalType == TYPE_NO_VOICE_ACTIVITY ) { /* Low speech activity */ /* Check if we should switch down */ if( ( psEncC->fs_kHz == 12 && TargetRate_bps < MB2NB_BITRATE_BPS && psEncC->minInternal_fs_kHz <= 8 ) || ( psEncC->fs_kHz == 16 && TargetRate_bps < WB2MB_BITRATE_BPS && psEncC->minInternal_fs_kHz <= 12 ) ) { /* Switch down */ if( SWITCH_TRANSITION_FILTERING && psEncC->sLP.mode == 0 ) { /* New transition */ psEncC->sLP.transition_frame_no = TRANSITION_FRAMES; /* Reset transition filter state */ SKP_memset( psEncC->sLP.In_LP_State, 0, sizeof( psEncC->sLP.In_LP_State ) ); } if( psEncC->sLP.transition_frame_no <= 0 ) { /* Stop transition phase */ psEncC->sLP.mode = 0; /* Switch to a lower sample frequency */ fs_kHz = psEncC->fs_kHz == 16 ? 12 : 8; } else { /* Direction: down (at double speed) */ psEncC->sLP.mode = -2; } } else if( ( psEncC->fs_kHz == 8 && TargetRate_bps > NB2MB_BITRATE_BPS && psEncC->maxInternal_fs_kHz >= 12 && psEncC->API_fs_Hz >= 12000 ) || ( psEncC->fs_kHz == 12 && TargetRate_bps > MB2WB_BITRATE_BPS && psEncC->maxInternal_fs_kHz >= 16 && psEncC->API_fs_Hz >= 16000 ) ) { /* Switch up */ if( SWITCH_TRANSITION_FILTERING && psEncC->sLP.mode == 0 ) { /* Switch to a higher sample frequency */ fs_kHz = psEncC->fs_kHz == 8 ? 12 : 16; /* New transition */ psEncC->sLP.transition_frame_no = 0; } if( psEncC->sLP.transition_frame_no >= TRANSITION_FRAMES ) { /* Stop transition phase */ psEncC->sLP.mode = 0; } else { /* Direction: up */ psEncC->sLP.mode = 1; } } } } #ifdef FORCE_INTERNAL_FS_KHZ fs_kHz = FORCE_INTERNAL_FS_KHZ; #endif return fs_kHz; }
/* Control internal sampling rate */ SKP_int silk_control_audio_bandwidth( silk_encoder_state *psEncC /* I/O Pointer to Silk encoder state */ ) { SKP_int fs_kHz; SKP_int32 fs_Hz; fs_kHz = psEncC->fs_kHz; fs_Hz = SKP_SMULBB( fs_kHz, 1000 ); if( fs_Hz == 0 ) { /* Encoder has just been initialized */ fs_Hz = SKP_min( psEncC->desiredInternal_fs_Hz, psEncC->API_fs_Hz ); fs_kHz = SKP_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 = SKP_min( fs_Hz, psEncC->maxInternal_fs_Hz ); fs_Hz = SKP_max( fs_Hz, psEncC->minInternal_fs_Hz ); fs_kHz = SKP_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 ) { /* Check if we should switch down */ if( SKP_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 */ SKP_memset( psEncC->sLP.In_LP_State, 0, sizeof( psEncC->sLP.In_LP_State ) ); } if( psEncC->sLP.transition_frame_no <= 0 ) { /* Stop transition phase */ psEncC->sLP.mode = 0; /* Switch to a lower sample frequency */ fs_kHz = psEncC->fs_kHz == 16 ? 12 : 8; } else { /* Direction: down (at double speed) */ psEncC->sLP.mode = -2; } } else /* Check if we should switch up */ if( SKP_SMULBB( psEncC->fs_kHz, 1000 ) < psEncC->desiredInternal_fs_Hz ) { /* Switch up */ if( psEncC->sLP.mode == 0 ) { /* 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 */ SKP_memset( psEncC->sLP.In_LP_State, 0, sizeof( psEncC->sLP.In_LP_State ) ); } /* Direction: up */ psEncC->sLP.mode = 1; } } } #ifdef FORCE_INTERNAL_FS_KHZ fs_kHz = FORCE_INTERNAL_FS_KHZ; #endif return fs_kHz; }