/* Gains scalar dequantization, uniform on log scale */
void silk_gains_dequant(
opus_int32 gain_Q16[ MAX_NB_SUBFR ], /* O quantized gains */
const opus_int8 ind[ MAX_NB_SUBFR ], /* I gain indices */
opus_int8 *prev_ind, /* I/O last index in previous frame */
const opus_int conditional, /* I first gain is delta coded if 1 */
const opus_int nb_subfr /* I number of subframes */
)
{
opus_int k, ind_tmp, double_step_size_threshold;
for( k = 0; k < nb_subfr; k++ ) {
if( k == 0 && conditional == 0 ) {
/* Gain index is not allowed to go down more than 16 steps (~21.8 dB) */
*prev_ind = silk_max_int( ind[ k ], *prev_ind - 16 );
} else {
/* Delta index */
ind_tmp = ind[ k ] + MIN_DELTA_GAIN_QUANT;
/* Accumulate deltas */
double_step_size_threshold = 2 * MAX_DELTA_GAIN_QUANT - N_LEVELS_QGAIN + *prev_ind;
if( ind_tmp > double_step_size_threshold ) {
*prev_ind += silk_LSHIFT( ind_tmp, 1 ) - double_step_size_threshold;
} else {
*prev_ind += ind_tmp;
}
}
*prev_ind = silk_LIMIT_int( *prev_ind, 0, N_LEVELS_QGAIN - 1 );
/* Scale and convert to linear scale */
gain_Q16[ k ] = silk_log2lin( silk_min_32( silk_SMULWB( INV_SCALE_Q16, *prev_ind ) + OFFSET, 3967 ) ); /* 3967 = 31 in Q7 */
}
}
/* Gain scalar quantization with hysteresis, uniform on log scale */
void silk_gains_quant(
opus_int8 ind[ MAX_NB_SUBFR ], /* O gain indices */
opus_int32 gain_Q16[ MAX_NB_SUBFR ], /* I/O gains (quantized out) */
opus_int8 *prev_ind, /* I/O last index in previous frame */
const opus_int conditional, /* I first gain is delta coded if 1 */
const opus_int nb_subfr /* I number of subframes */
)
{
opus_int k, double_step_size_threshold;
for( k = 0; k < nb_subfr; k++ ) {
/* Convert to log scale, scale, floor() */
ind[ k ] = silk_SMULWB( SCALE_Q16, silk_lin2log( gain_Q16[ k ] ) - OFFSET );
/* Round towards previous quantized gain (hysteresis) */
if( ind[ k ] < *prev_ind ) {
ind[ k ]++;
}
ind[ k ] = silk_LIMIT_int( ind[ k ], 0, N_LEVELS_QGAIN - 1 );
/* Compute delta indices and limit */
if( k == 0 && conditional == 0 ) {
/* Full index */
ind[ k ] = silk_LIMIT_int( ind[ k ], *prev_ind + MIN_DELTA_GAIN_QUANT, N_LEVELS_QGAIN - 1 );
*prev_ind = ind[ k ];
} else {
/* Delta index */
ind[ k ] = ind[ k ] - *prev_ind;
/* Double the quantization step size for large gain increases, so that the max gain level can be reached */
double_step_size_threshold = 2 * MAX_DELTA_GAIN_QUANT - N_LEVELS_QGAIN + *prev_ind;
if( ind[ k ] > double_step_size_threshold ) {
ind[ k ] = double_step_size_threshold + silk_RSHIFT( ind[ k ] - double_step_size_threshold + 1, 1 );
}
ind[ k ] = silk_LIMIT_int( ind[ k ], MIN_DELTA_GAIN_QUANT, MAX_DELTA_GAIN_QUANT );
/* Accumulate deltas */
if( ind[ k ] > double_step_size_threshold ) {
*prev_ind += silk_LSHIFT( ind[ k ], 1 ) - double_step_size_threshold;
*prev_ind = silk_min_int( *prev_ind, N_LEVELS_QGAIN - 1 );
} else {
*prev_ind += ind[ k ];
}
/* Shift to make non-negative */
ind[ k ] -= MIN_DELTA_GAIN_QUANT;
}
/* Scale and convert to linear scale */
gain_Q16[ k ] = silk_log2lin( silk_min_32( silk_SMULWB( INV_SCALE_Q16, *prev_ind ) + OFFSET, 3967 ) ); /* 3967 = 31 in Q7 */
}
}
/* Processing of gains */
void silk_process_gains_FIX(
silk_encoder_state_FIX *psEnc, /* I/O Encoder state */
silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control */
opus_int condCoding /* I The type of conditional coding to use */
)
{
silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
opus_int k;
opus_int32 s_Q16, InvMaxSqrVal_Q16, gain, gain_squared, ResNrg, ResNrgPart, quant_offset_Q10;
/* Gain reduction when LTP coding gain is high */
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
/*s = -0.5f * silk_sigmoid( 0.25f * ( psEncCtrl->LTPredCodGain - 12.0f ) ); */
s_Q16 = -silk_sigm_Q15( silk_RSHIFT_ROUND( psEncCtrl->LTPredCodGain_Q7 - SILK_FIX_CONST( 12.0, 7 ), 4 ) );
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
psEncCtrl->Gains_Q16[ k ] = silk_SMLAWB( psEncCtrl->Gains_Q16[ k ], psEncCtrl->Gains_Q16[ k ], s_Q16 );
}
}
/* Limit the quantized signal */
/* InvMaxSqrVal = pow( 2.0f, 0.33f * ( 21.0f - SNR_dB ) ) / subfr_length; */
InvMaxSqrVal_Q16 = silk_DIV32_16( silk_log2lin(
silk_SMULWB( SILK_FIX_CONST( 21 + 16 / 0.33, 7 ) - psEnc->sCmn.SNR_dB_Q7, SILK_FIX_CONST( 0.33, 16 ) ) ), psEnc->sCmn.subfr_length );
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
/* Soft limit on ratio residual energy and squared gains */
ResNrg = psEncCtrl->ResNrg[ k ];
ResNrgPart = silk_SMULWW( ResNrg, InvMaxSqrVal_Q16 );
if( psEncCtrl->ResNrgQ[ k ] > 0 ) {
ResNrgPart = silk_RSHIFT_ROUND( ResNrgPart, psEncCtrl->ResNrgQ[ k ] );
} else {
if( ResNrgPart >= silk_RSHIFT( silk_int32_MAX, -psEncCtrl->ResNrgQ[ k ] ) ) {
ResNrgPart = silk_int32_MAX;
} else {
ResNrgPart = silk_LSHIFT( ResNrgPart, -psEncCtrl->ResNrgQ[ k ] );
}
}
gain = psEncCtrl->Gains_Q16[ k ];
gain_squared = silk_ADD_SAT32( ResNrgPart, silk_SMMUL( gain, gain ) );
if( gain_squared < silk_int16_MAX ) {
/* recalculate with higher precision */
gain_squared = silk_SMLAWW( silk_LSHIFT( ResNrgPart, 16 ), gain, gain );
silk_assert( gain_squared > 0 );
gain = silk_SQRT_APPROX( gain_squared ); /* Q8 */
gain = silk_min( gain, silk_int32_MAX >> 8 );
psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT_SAT32( gain, 8 ); /* Q16 */
} else {
void silk_quant_LTP_gains(
opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (un)quantized LTP gains */
opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */
opus_int8 *periodicity_index, /* O Periodicity Index */
opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */
const opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Error Weights in Q18 */
opus_int mu_Q9, /* I Mu value (R/D tradeoff) */
opus_int lowComplexity, /* I Flag for low complexity */
const opus_int nb_subfr, /* I number of subframes */
int arch /* I Run-time architecture */
)
{
opus_int j, k, cbk_size;
opus_int8 temp_idx[ MAX_NB_SUBFR ];
const opus_uint8 *cl_ptr_Q5;
const opus_int8 *cbk_ptr_Q7;
const opus_uint8 *cbk_gain_ptr_Q7;
const opus_int16 *b_Q14_ptr;
const opus_int32 *W_Q18_ptr;
opus_int32 rate_dist_Q14_subfr, rate_dist_Q14, min_rate_dist_Q14;
opus_int32 sum_log_gain_tmp_Q7, best_sum_log_gain_Q7, max_gain_Q7, gain_Q7;
/***************************************************/
/* iterate over different codebooks with different */
/* rates/distortions, and choose best */
/***************************************************/
min_rate_dist_Q14 = silk_int32_MAX;
best_sum_log_gain_Q7 = 0;
for(k = 0; k < 3; k++) {
/* Safety margin for pitch gain control, to take into account factors
such as state rescaling/rewhitening. */
opus_int32 gain_safety = SILK_FIX_CONST(0.4, 7);
cl_ptr_Q5 = silk_LTP_gain_BITS_Q5_ptrs[ k ];
cbk_ptr_Q7 = silk_LTP_vq_ptrs_Q7[ k ];
cbk_gain_ptr_Q7 = silk_LTP_vq_gain_ptrs_Q7[ k ];
cbk_size = silk_LTP_vq_sizes[ k ];
/* Set up pointer to first subframe */
W_Q18_ptr = W_Q18;
b_Q14_ptr = B_Q14;
rate_dist_Q14 = 0;
sum_log_gain_tmp_Q7 = *sum_log_gain_Q7;
for(j = 0; j < nb_subfr; j++) {
max_gain_Q7 = silk_log2lin((SILK_FIX_CONST(MAX_SUM_LOG_GAIN_DB / 6.0, 7) - sum_log_gain_tmp_Q7)
+ SILK_FIX_CONST(7, 7)) - gain_safety;
silk_VQ_WMat_EC(
&temp_idx[ j ], /* O index of best codebook vector */
&rate_dist_Q14_subfr, /* O best weighted quantization error + mu * rate */
&gain_Q7, /* O sum of absolute LTP coefficients */
b_Q14_ptr, /* I input vector to be quantized */
W_Q18_ptr, /* I weighting matrix */
cbk_ptr_Q7, /* I codebook */
cbk_gain_ptr_Q7, /* I codebook effective gains */
cl_ptr_Q5, /* I code length for each codebook vector */
mu_Q9, /* I tradeoff between weighted error and rate */
max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
cbk_size, /* I number of vectors in codebook */
arch /* I Run-time architecture */
);
rate_dist_Q14 = silk_ADD_POS_SAT32(rate_dist_Q14, rate_dist_Q14_subfr);
sum_log_gain_tmp_Q7 = silk_max(0, sum_log_gain_tmp_Q7
+ silk_lin2log(gain_safety + gain_Q7) - SILK_FIX_CONST(7, 7));
b_Q14_ptr += LTP_ORDER;
W_Q18_ptr += LTP_ORDER * LTP_ORDER;
}
/* Avoid never finding a codebook */
rate_dist_Q14 = silk_min(silk_int32_MAX - 1, rate_dist_Q14);
if(rate_dist_Q14 < min_rate_dist_Q14) {
min_rate_dist_Q14 = rate_dist_Q14;
*periodicity_index = (opus_int8)k;
silk_memcpy(cbk_index, temp_idx, nb_subfr * sizeof(opus_int8));
best_sum_log_gain_Q7 = sum_log_gain_tmp_Q7;
}
/* Break early in low-complexity mode if rate distortion is below threshold */
if(lowComplexity && (rate_dist_Q14 < silk_LTP_gain_middle_avg_RD_Q14)) {
break;
}
}
cbk_ptr_Q7 = silk_LTP_vq_ptrs_Q7[ *periodicity_index ];
for(j = 0; j < nb_subfr; j++) {
for(k = 0; k < LTP_ORDER; k++) {
B_Q14[ j * LTP_ORDER + k ] = silk_LSHIFT(cbk_ptr_Q7[ cbk_index[ j ] * LTP_ORDER + k ], 7);
}
}
*sum_log_gain_Q7 = best_sum_log_gain_Q7;
}
opus_int silk_encode_frame_FLP(
silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */
opus_int32 *pnBytesOut, /* O Number of payload bytes; */
ec_enc *psRangeEnc, /* I/O compressor data structure */
opus_int condCoding, /* I The type of conditional coding to use */
opus_int maxBits, /* I If > 0: maximum number of output bits */
opus_int useCBR /* I Flag to force constant-bitrate operation */
)
{
silk_encoder_control_FLP sEncCtrl;
opus_int i, iter, maxIter, found_upper, found_lower, ret = 0;
silk_float *x_frame, *res_pitch_frame;
silk_float xfw[ MAX_FRAME_LENGTH ];
silk_float res_pitch[ 2 * MAX_FRAME_LENGTH + LA_PITCH_MAX ];
ec_enc sRangeEnc_copy, sRangeEnc_copy2;
silk_nsq_state sNSQ_copy, sNSQ_copy2;
opus_int32 seed_copy, nBits, nBits_lower, nBits_upper, gainMult_lower, gainMult_upper;
opus_int32 gainsID, gainsID_lower, gainsID_upper;
opus_int16 gainMult_Q8;
opus_int16 ec_prevLagIndex_copy;
opus_int ec_prevSignalType_copy;
opus_int8 LastGainIndex_copy2;
opus_int32 pGains_Q16[ MAX_NB_SUBFR ];
opus_uint8 ec_buf_copy[ 1275 ];
/* This is totally unnecessary but many compilers (including gcc) are too dumb to realise it */
LastGainIndex_copy2 = nBits_lower = nBits_upper = gainMult_lower = gainMult_upper = 0;
psEnc->sCmn.indices.Seed = psEnc->sCmn.frameCounter++ & 3;
/**************************************************************/
/* Set up Input Pointers, and insert frame in input buffer */
/**************************************************************/
/* pointers aligned with start of frame to encode */
x_frame = psEnc->x_buf + psEnc->sCmn.ltp_mem_length; /* start of frame to encode */
res_pitch_frame = res_pitch + psEnc->sCmn.ltp_mem_length; /* start of pitch LPC residual frame */
/***************************************/
/* Ensure smooth bandwidth transitions */
/***************************************/
silk_LP_variable_cutoff( &psEnc->sCmn.sLP, psEnc->sCmn.inputBuf + 1, psEnc->sCmn.frame_length );
/*******************************************/
/* Copy new frame to front of input buffer */
/*******************************************/
silk_short2float_array( x_frame + LA_SHAPE_MS * psEnc->sCmn.fs_kHz, psEnc->sCmn.inputBuf + 1, psEnc->sCmn.frame_length );
/* Add tiny signal to avoid high CPU load from denormalized floating point numbers */
for( i = 0; i < 8; i++ ) {
x_frame[ LA_SHAPE_MS * psEnc->sCmn.fs_kHz + i * ( psEnc->sCmn.frame_length >> 3 ) ] += ( 1 - ( i & 2 ) ) * 1e-6f;
}
if( !psEnc->sCmn.prefillFlag ) {
/*****************************************/
/* Find pitch lags, initial LPC analysis */
/*****************************************/
silk_find_pitch_lags_FLP( psEnc, &sEncCtrl, res_pitch, x_frame );
/************************/
/* Noise shape analysis */
/************************/
silk_noise_shape_analysis_FLP( psEnc, &sEncCtrl, res_pitch_frame, x_frame );
/***************************************************/
/* Find linear prediction coefficients (LPC + LTP) */
/***************************************************/
silk_find_pred_coefs_FLP( psEnc, &sEncCtrl, res_pitch, x_frame, condCoding );
/****************************************/
/* Process gains */
/****************************************/
silk_process_gains_FLP( psEnc, &sEncCtrl, condCoding );
/*****************************************/
/* Prefiltering for noise shaper */
/*****************************************/
silk_prefilter_FLP( psEnc, &sEncCtrl, xfw, x_frame );
/****************************************/
/* Low Bitrate Redundant Encoding */
/****************************************/
silk_LBRR_encode_FLP( psEnc, &sEncCtrl, xfw, condCoding );
/* Loop over quantizer and entroy coding to control bitrate */
maxIter = 6;
gainMult_Q8 = SILK_FIX_CONST( 1, 8 );
found_lower = 0;
found_upper = 0;
gainsID = silk_gains_ID( psEnc->sCmn.indices.GainsIndices, psEnc->sCmn.nb_subfr );
gainsID_lower = -1;
gainsID_upper = -1;
/* Copy part of the input state */
silk_memcpy( &sRangeEnc_copy, psRangeEnc, sizeof( ec_enc ) );
silk_memcpy( &sNSQ_copy, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) );
seed_copy = psEnc->sCmn.indices.Seed;
ec_prevLagIndex_copy = psEnc->sCmn.ec_prevLagIndex;
ec_prevSignalType_copy = psEnc->sCmn.ec_prevSignalType;
for( iter = 0; ; iter++ ) {
if( gainsID == gainsID_lower ) {
nBits = nBits_lower;
} else if( gainsID == gainsID_upper ) {
nBits = nBits_upper;
} else {
/* Restore part of the input state */
if( iter > 0 ) {
silk_memcpy( psRangeEnc, &sRangeEnc_copy, sizeof( ec_enc ) );
silk_memcpy( &psEnc->sCmn.sNSQ, &sNSQ_copy, sizeof( silk_nsq_state ) );
psEnc->sCmn.indices.Seed = seed_copy;
psEnc->sCmn.ec_prevLagIndex = ec_prevLagIndex_copy;
psEnc->sCmn.ec_prevSignalType = ec_prevSignalType_copy;
}
/*****************************************/
/* Noise shaping quantization */
/*****************************************/
silk_NSQ_wrapper_FLP( psEnc, &sEncCtrl, &psEnc->sCmn.indices, &psEnc->sCmn.sNSQ, psEnc->sCmn.pulses, xfw );
/****************************************/
/* Encode Parameters */
/****************************************/
silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding );
/****************************************/
/* Encode Excitation Signal */
/****************************************/
silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType,
psEnc->sCmn.pulses, psEnc->sCmn.frame_length );
nBits = ec_tell( psRangeEnc );
if( useCBR == 0 && iter == 0 && nBits <= maxBits ) {
break;
}
}
if( iter == maxIter ) {
if( found_lower && ( gainsID == gainsID_lower || nBits > maxBits ) ) {
/* Restore output state from earlier iteration that did meet the bitrate budget */
silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) );
silk_assert( sRangeEnc_copy2.offs <= 1275 );
silk_memcpy( psRangeEnc->buf, ec_buf_copy, sRangeEnc_copy2.offs );
silk_memcpy( &psEnc->sCmn.sNSQ, &sNSQ_copy2, sizeof( silk_nsq_state ) );
psEnc->sShape.LastGainIndex = LastGainIndex_copy2;
}
break;
}
if( nBits > maxBits ) {
if( found_lower == 0 && iter >= 2 ) {
/* Adjust the quantizer's rate/distortion tradeoff and discard previous "upper" results */
sEncCtrl.Lambda *= 1.5f;
found_upper = 0;
gainsID_upper = -1;
} else {
found_upper = 1;
nBits_upper = nBits;
gainMult_upper = gainMult_Q8;
gainsID_upper = gainsID;
}
} else if( nBits < maxBits - 5 ) {
found_lower = 1;
nBits_lower = nBits;
gainMult_lower = gainMult_Q8;
if( gainsID != gainsID_lower ) {
gainsID_lower = gainsID;
/* Copy part of the output state */
silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) );
silk_assert( psRangeEnc->offs <= 1275 );
silk_memcpy( ec_buf_copy, psRangeEnc->buf, psRangeEnc->offs );
silk_memcpy( &sNSQ_copy2, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) );
LastGainIndex_copy2 = psEnc->sShape.LastGainIndex;
}
} else {
/* Within 5 bits of budget: close enough */
break;
}
if( ( found_lower & found_upper ) == 0 ) {
/* Adjust gain according to high-rate rate/distortion curve */
opus_int32 gain_factor_Q16;
gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) );
gain_factor_Q16 = silk_min_32( gain_factor_Q16, SILK_FIX_CONST( 2, 16 ) );
if( nBits > maxBits ) {
gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) );
}
gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 );
} else {
/* Adjust gain by interpolating */
gainMult_Q8 = gainMult_lower + ( ( gainMult_upper - gainMult_lower ) * ( maxBits - nBits_lower ) ) / ( nBits_upper - nBits_lower );
/* New gain multplier must be between 25% and 75% of old range (note that gainMult_upper < gainMult_lower) */
if( gainMult_Q8 > silk_ADD_RSHIFT32( gainMult_lower, gainMult_upper - gainMult_lower, 2 ) ) {
gainMult_Q8 = silk_ADD_RSHIFT32( gainMult_lower, gainMult_upper - gainMult_lower, 2 );
} else
if( gainMult_Q8 < silk_SUB_RSHIFT32( gainMult_upper, gainMult_upper - gainMult_lower, 2 ) ) {
gainMult_Q8 = silk_SUB_RSHIFT32( gainMult_upper, gainMult_upper - gainMult_lower, 2 );
}
}
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
pGains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q8 ), 8 );
}
/* Quantize gains */
psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev;
silk_gains_quant( psEnc->sCmn.indices.GainsIndices, pGains_Q16,
&psEnc->sShape.LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr );
/* Unique identifier of gains vector */
gainsID = silk_gains_ID( psEnc->sCmn.indices.GainsIndices, psEnc->sCmn.nb_subfr );
/* Overwrite unquantized gains with quantized gains and convert back to Q0 from Q16 */
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
sEncCtrl.Gains[ i ] = pGains_Q16[ i ] / 65536.0f;
}
}
}
/* Update input buffer */
silk_memmove( psEnc->x_buf, &psEnc->x_buf[ psEnc->sCmn.frame_length ],
( psEnc->sCmn.ltp_mem_length + LA_SHAPE_MS * psEnc->sCmn.fs_kHz ) * sizeof( silk_float ) );
/* Parameters needed for next frame */
psEnc->sCmn.prevLag = sEncCtrl.pitchL[ psEnc->sCmn.nb_subfr - 1 ];
psEnc->sCmn.prevSignalType = psEnc->sCmn.indices.signalType;
/* Exit without entropy coding */
if( psEnc->sCmn.prefillFlag ) {
/* No payload */
*pnBytesOut = 0;
return ret;
}
/****************************************/
/* Finalize payload */
/****************************************/
psEnc->sCmn.first_frame_after_reset = 0;
/* Payload size */
*pnBytesOut = silk_RSHIFT( ec_tell( psRangeEnc ) + 7, 3 );
return ret;
}
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;
}
int silk_encode_frame_FIX(silk_encoder_state_FIX * psEnc, /* I/O Pointer to Silk FIX encoder state */
int32_t * pnBytesOut, /* O Pointer to number of payload bytes; */
ec_enc * psRangeEnc, /* I/O compressor data structure */
int condCoding, /* I The type of conditional coding to use */
int maxBits, /* I If > 0: maximum number of output bits */
int useCBR /* I Flag to force constant-bitrate operation */
) {
silk_encoder_control_FIX sEncCtrl;
int i, iter, maxIter, found_upper, found_lower, ret = 0;
int16_t *x_frame;
ec_enc sRangeEnc_copy, sRangeEnc_copy2;
silk_nsq_state sNSQ_copy, sNSQ_copy2;
int32_t seed_copy, nBits, nBits_lower, nBits_upper, gainMult_lower,
gainMult_upper;
int32_t gainsID, gainsID_lower, gainsID_upper;
int16_t gainMult_Q8;
int16_t ec_prevLagIndex_copy;
int ec_prevSignalType_copy;
int8_t LastGainIndex_copy2;
/* This is totally unnecessary but many compilers (including gcc) are too dumb to realise it */
LastGainIndex_copy2 = nBits_lower = nBits_upper = gainMult_lower =
gainMult_upper = 0;
psEnc->sCmn.indices.Seed = psEnc->sCmn.frameCounter++ & 3;
/**************************************************************/
/* Set up Input Pointers, and insert frame in input buffer */
/*************************************************************/
/* start of frame to encode */
x_frame = psEnc->x_buf + psEnc->sCmn.ltp_mem_length;
/***************************************/
/* Ensure smooth bandwidth transitions */
/***************************************/
silk_LP_variable_cutoff(&psEnc->sCmn.sLP, psEnc->sCmn.inputBuf + 1,
psEnc->sCmn.frame_length);
/*******************************************/
/* Copy new frame to front of input buffer */
/*******************************************/
memcpy(x_frame + LA_SHAPE_MS * psEnc->sCmn.fs_kHz,
psEnc->sCmn.inputBuf + 1,
psEnc->sCmn.frame_length * sizeof(int16_t));
if (!psEnc->sCmn.prefillFlag) {
int16_t *res_pitch_frame;
int16_t res_pitch[psEnc->sCmn.la_pitch +
psEnc->sCmn.frame_length +
psEnc->sCmn.ltp_mem_length];
/* start of pitch LPC residual frame */
res_pitch_frame = res_pitch + psEnc->sCmn.ltp_mem_length;
/*****************************************/
/* Find pitch lags, initial LPC analysis */
/*****************************************/
silk_find_pitch_lags_FIX(psEnc, &sEncCtrl, res_pitch, x_frame,
psEnc->sCmn.arch);
/************************/
/* Noise shape analysis */
/************************/
silk_noise_shape_analysis_FIX(psEnc, &sEncCtrl, res_pitch_frame,
x_frame, psEnc->sCmn.arch);
/***************************************************/
/* Find linear prediction coefficients (LPC + LTP) */
/***************************************************/
silk_find_pred_coefs_FIX(psEnc, &sEncCtrl, res_pitch, x_frame,
condCoding);
/****************************************/
/* Process gains */
/****************************************/
silk_process_gains_FIX(psEnc, &sEncCtrl, condCoding);
/*****************************************/
/* Prefiltering for noise shaper */
/*****************************************/
int32_t xfw_Q3[psEnc->sCmn.frame_length];
silk_prefilter_FIX(psEnc, &sEncCtrl, xfw_Q3, x_frame);
/****************************************/
/* Low Bitrate Redundant Encoding */
/****************************************/
silk_LBRR_encode_FIX(psEnc, &sEncCtrl, xfw_Q3, condCoding);
/* Loop over quantizer and entropy coding to control bitrate */
maxIter = 6;
gainMult_Q8 = SILK_FIX_CONST(1, 8);
found_lower = 0;
found_upper = 0;
gainsID =
silk_gains_ID(psEnc->sCmn.indices.GainsIndices,
psEnc->sCmn.nb_subfr);
gainsID_lower = -1;
gainsID_upper = -1;
/* Copy part of the input state */
memcpy(&sRangeEnc_copy, psRangeEnc, sizeof(ec_enc));
memcpy(&sNSQ_copy, &psEnc->sCmn.sNSQ,
sizeof(silk_nsq_state));
seed_copy = psEnc->sCmn.indices.Seed;
ec_prevLagIndex_copy = psEnc->sCmn.ec_prevLagIndex;
ec_prevSignalType_copy = psEnc->sCmn.ec_prevSignalType;
uint8_t ec_buf_copy[1275];
for (iter = 0;; iter++) {
if (gainsID == gainsID_lower) {
nBits = nBits_lower;
} else if (gainsID == gainsID_upper) {
nBits = nBits_upper;
} else {
/* Restore part of the input state */
if (iter > 0) {
memcpy(psRangeEnc, &sRangeEnc_copy,
sizeof(ec_enc));
memcpy(&psEnc->sCmn.sNSQ,
&sNSQ_copy,
sizeof(silk_nsq_state));
psEnc->sCmn.indices.Seed = seed_copy;
psEnc->sCmn.ec_prevLagIndex =
ec_prevLagIndex_copy;
psEnc->sCmn.ec_prevSignalType =
ec_prevSignalType_copy;
}
/*****************************************/
/* Noise shaping quantization */
/*****************************************/
if (psEnc->sCmn.nStatesDelayedDecision > 1
|| psEnc->sCmn.warping_Q16 > 0) {
silk_NSQ_del_dec(&psEnc->sCmn,
&psEnc->sCmn.sNSQ,
&psEnc->sCmn.indices,
xfw_Q3,
psEnc->sCmn.pulses,
sEncCtrl.
PredCoef_Q12[0],
sEncCtrl.LTPCoef_Q14,
sEncCtrl.AR2_Q13,
sEncCtrl.
HarmShapeGain_Q14,
sEncCtrl.Tilt_Q14,
sEncCtrl.LF_shp_Q14,
sEncCtrl.Gains_Q16,
sEncCtrl.pitchL,
sEncCtrl.Lambda_Q10,
sEncCtrl.
LTP_scale_Q14);
} else {
silk_NSQ(&psEnc->sCmn,
&psEnc->sCmn.sNSQ,
&psEnc->sCmn.indices, xfw_Q3,
psEnc->sCmn.pulses,
sEncCtrl.PredCoef_Q12[0],
sEncCtrl.LTPCoef_Q14,
sEncCtrl.AR2_Q13,
sEncCtrl.HarmShapeGain_Q14,
sEncCtrl.Tilt_Q14,
sEncCtrl.LF_shp_Q14,
sEncCtrl.Gains_Q16,
sEncCtrl.pitchL,
sEncCtrl.Lambda_Q10,
sEncCtrl.LTP_scale_Q14);
}
/****************************************/
/* Encode Parameters */
/****************************************/
silk_encode_indices(&psEnc->sCmn, psRangeEnc,
psEnc->sCmn.nFramesEncoded,
0, condCoding);
/****************************************/
/* Encode Excitation Signal */
/****************************************/
silk_encode_pulses(psRangeEnc,
psEnc->sCmn.indices.
signalType,
psEnc->sCmn.indices.
quantOffsetType,
psEnc->sCmn.pulses,
psEnc->sCmn.frame_length);
nBits = ec_tell(psRangeEnc);
if (useCBR == 0 && iter == 0
&& nBits <= maxBits) {
break;
}
}
if (iter == maxIter) {
if (found_lower
&& (gainsID == gainsID_lower
|| nBits > maxBits)) {
/* Restore output state from earlier iteration that did meet the bitrate budget */
memcpy(psRangeEnc,
&sRangeEnc_copy2,
sizeof(ec_enc));
assert(sRangeEnc_copy2.offs <=
1275);
memcpy(psRangeEnc->buf,
ec_buf_copy,
sRangeEnc_copy2.offs);
memcpy(&psEnc->sCmn.sNSQ,
&sNSQ_copy2,
sizeof(silk_nsq_state));
psEnc->sShape.LastGainIndex =
LastGainIndex_copy2;
}
break;
}
if (nBits > maxBits) {
if (found_lower == 0 && iter >= 2) {
/* Adjust the quantizer's rate/distortion tradeoff and discard previous "upper" results */
sEncCtrl.Lambda_Q10 =
silk_ADD_RSHIFT32(sEncCtrl.
Lambda_Q10,
sEncCtrl.
Lambda_Q10, 1);
found_upper = 0;
gainsID_upper = -1;
} else {
found_upper = 1;
nBits_upper = nBits;
gainMult_upper = gainMult_Q8;
gainsID_upper = gainsID;
}
} else if (nBits < maxBits - 5) {
found_lower = 1;
nBits_lower = nBits;
gainMult_lower = gainMult_Q8;
if (gainsID != gainsID_lower) {
gainsID_lower = gainsID;
/* Copy part of the output state */
memcpy(&sRangeEnc_copy2,
psRangeEnc, sizeof(ec_enc));
assert(psRangeEnc->offs <= 1275);
memcpy(ec_buf_copy,
psRangeEnc->buf,
psRangeEnc->offs);
memcpy(&sNSQ_copy2,
&psEnc->sCmn.sNSQ,
sizeof(silk_nsq_state));
LastGainIndex_copy2 =
psEnc->sShape.LastGainIndex;
}
} else {
/* Within 5 bits of budget: close enough */
break;
}
if ((found_lower & found_upper) == 0) {
/* Adjust gain according to high-rate rate/distortion curve */
int32_t gain_factor_Q16;
gain_factor_Q16 =
silk_log2lin(silk_LSHIFT(nBits - maxBits, 7)
/ psEnc->sCmn.frame_length +
SILK_FIX_CONST(16, 7));
gain_factor_Q16 =
silk_min_32(gain_factor_Q16,
SILK_FIX_CONST(2, 16));
if (nBits > maxBits) {
gain_factor_Q16 =
silk_max_32(gain_factor_Q16,
SILK_FIX_CONST(1.3,
16));
}
gainMult_Q8 =
silk_SMULWB(gain_factor_Q16, gainMult_Q8);
} else {
/* Adjust gain by interpolating */
assert(nBits_upper != nBits_lower);
gainMult_Q8 =
gainMult_lower +
silk_DIV32_16(silk_MUL
(gainMult_upper -
gainMult_lower,
maxBits - nBits_lower),
nBits_upper - nBits_lower);
/* New gain multplier must be between 25% and 75% of old range (note that gainMult_upper < gainMult_lower) */
if (gainMult_Q8 >
silk_ADD_RSHIFT32(gainMult_lower,
gainMult_upper -
gainMult_lower, 2)) {
gainMult_Q8 =
silk_ADD_RSHIFT32(gainMult_lower,
gainMult_upper -
gainMult_lower,
2);
} else if (gainMult_Q8 <
silk_SUB_RSHIFT32(gainMult_upper,
gainMult_upper -
gainMult_lower,
2)) {
gainMult_Q8 =
silk_SUB_RSHIFT32(gainMult_upper,
gainMult_upper -
gainMult_lower,
2);
}
}
for (i = 0; i < psEnc->sCmn.nb_subfr; i++) {
sEncCtrl.Gains_Q16[i] =
silk_LSHIFT_SAT32(silk_SMULWB
(sEncCtrl.GainsUnq_Q16[i],
gainMult_Q8), 8);
}
/* Quantize gains */
psEnc->sShape.LastGainIndex =
sEncCtrl.lastGainIndexPrev;
silk_gains_quant(psEnc->sCmn.indices.GainsIndices,
sEncCtrl.Gains_Q16,
&psEnc->sShape.LastGainIndex,
condCoding == CODE_CONDITIONALLY,
psEnc->sCmn.nb_subfr);
/* Unique identifier of gains vector */
gainsID =
silk_gains_ID(psEnc->sCmn.indices.GainsIndices,
psEnc->sCmn.nb_subfr);
}
}
/* Update input buffer */
memmove(psEnc->x_buf, &psEnc->x_buf[psEnc->sCmn.frame_length],
(psEnc->sCmn.ltp_mem_length + LA_SHAPE_MS * psEnc->sCmn.fs_kHz) * sizeof(int16_t));
/* Exit without entropy coding */
if (psEnc->sCmn.prefillFlag) {
/* No payload */
*pnBytesOut = 0;
return ret;
}
/* Parameters needed for next frame */
psEnc->sCmn.prevLag = sEncCtrl.pitchL[psEnc->sCmn.nb_subfr - 1];
psEnc->sCmn.prevSignalType = psEnc->sCmn.indices.signalType;
/****************************************/
/* Finalize payload */
/****************************************/
psEnc->sCmn.first_frame_after_reset = 0;
/* Payload size */
*pnBytesOut = silk_RSHIFT(ec_tell(psRangeEnc) + 7, 3);
return ret;
}
