SKP_int16 SKP_Silk_int16_array_maxabs( /* O Maximum absolute value, max: 2^15-1 */
const SKP_int16 *vec, /* I Input vector [len] */
const SKP_int32 len /* I Length of input vector */
)
{
SKP_int32 max = 0, i, lvl = 0, ind;
if( len == 0 ) return 0;
ind = len - 1;
max = SKP_SMULBB( vec[ ind ], vec[ ind ] );
for( i = len - 2; i >= 0; i-- ) {
lvl = SKP_SMULBB( vec[ i ], vec[ i ] );
if( lvl > max ) {
max = lvl;
ind = i;
}
}
/* Do not return 32768, as it will not fit in an int16 so may lead to problems later on */
if( max >= 1073676289 ) { // (2^15-1)^2 = 1073676289
return( SKP_int16_MAX );
} else {
if( vec[ ind ] < 0 ) {
return( -vec[ ind ] );
} else {
return( vec[ ind ] );
}
}
}
/* Function that returns the maximum absolut value of the input vector */
int16_t SKP_Silk_int16_array_maxabs( /* O Maximum absolute value, max: 2^15-1 */
const int16_t * vec, /* I Input vector [len] */
const int32_t len /* I Length of input vector */
)
{
int32_t max = 0, i, lvl = 0, ind;
ind = len - 1;
max = SKP_SMULBB(vec[ind], vec[ind]);
for (i = len - 2; i >= 0; i--) {
lvl = SKP_SMULBB(vec[i], vec[i]);
if (lvl > max) {
max = lvl;
ind = i;
}
}
/* Do not return 32768, as it will not fit in an int16 so may lead to problems later on */
lvl = SKP_abs(vec[ind]);
if (lvl > int16_t_MAX) {
return (int16_t_MAX);
} else {
return ((int16_t) lvl);
}
}
SKP_int SKP_Silk_sigm_Q15( SKP_int in_Q5 )
{
SKP_int ind;
if( in_Q5 < 0 ) {
/* Negative input */
in_Q5 = -in_Q5;
if( in_Q5 >= 6 * 32 ) {
return 0; /* Clip */
} else {
/* Linear interpolation of look up table */
ind = SKP_RSHIFT( in_Q5, 5 );
return( sigm_LUT_neg_Q15[ ind ] - SKP_SMULBB( sigm_LUT_slope_Q10[ ind ], in_Q5 & 0x1F ) );
}
} else {
/* Positive input */
if( in_Q5 >= 6 * 32 ) {
return 32767; /* clip */
} else {
/* Linear interpolation of look up table */
ind = SKP_RSHIFT( in_Q5, 5 );
return( sigm_LUT_pos_Q15[ ind ] + SKP_SMULBB( sigm_LUT_slope_Q10[ ind ], in_Q5 & 0x1F ) );
}
}
}
/* Resamples by a factor 3/1 */
void SKP_Silk_resample_3_1(
SKP_int16 *out, /* O: Fs_high signal [inLen*3] */
SKP_int32 *S, /* I/O: State vector [7] */
const SKP_int16 *in, /* I: Fs_low signal [inLen] */
const SKP_int32 inLen /* I: Input length */
)
{
SKP_int k, LSubFrameIn, LSubFrameOut;
SKP_int32 out_tmp, idx, inLenTmp = inLen;
SKP_int32 scratch00[ IN_SUBFR_LEN_RESAMPLE_3_1 ];
SKP_int32 scratch0[ 3 * IN_SUBFR_LEN_RESAMPLE_3_1 ];
SKP_int32 scratch1[ 3 * IN_SUBFR_LEN_RESAMPLE_3_1 ];
/* Coefficients for 3-fold resampling */
const SKP_int16 A30[ 2 ] = { 1773, 17818 };
const SKP_int16 A31[ 2 ] = { 4942, 25677 };
const SKP_int16 A32[ 2 ] = { 11786, 29304 };
while( inLenTmp > 0 ) {
LSubFrameIn = SKP_min_int( IN_SUBFR_LEN_RESAMPLE_3_1, inLenTmp );
LSubFrameOut = SKP_SMULBB( 3, LSubFrameIn );
/* Convert Q15 -> Q25 */
for( k = 0; k < LSubFrameIn; k++ ) {
scratch00[k] = SKP_LSHIFT( (SKP_int32)in[ k ], 10 );
}
/* Allpass filtering */
/* Scratch size: 2 * 3* LSubFrame * sizeof(SKP_int32) */
SKP_Silk_allpass_int( scratch00, S + 1, A30[ 0 ], scratch1, LSubFrameIn );
SKP_Silk_allpass_int( scratch1, S + 2, A30[ 1 ], scratch0, LSubFrameIn );
SKP_Silk_allpass_int( scratch00, S + 3, A31[ 0 ], scratch1, LSubFrameIn );
SKP_Silk_allpass_int( scratch1, S + 4, A31[ 1 ], scratch0 + IN_SUBFR_LEN_RESAMPLE_3_1, LSubFrameIn );
SKP_Silk_allpass_int( scratch00, S + 5, A32[ 0 ], scratch1, LSubFrameIn );
SKP_Silk_allpass_int( scratch1, S + 6, A32[ 1 ], scratch0 + 2 * IN_SUBFR_LEN_RESAMPLE_3_1, LSubFrameIn );
/* Interleave three allpass outputs */
for( k = 0; k < LSubFrameIn; k++ ) {
idx = SKP_SMULBB( 3, k );
scratch1[ idx ] = scratch0[ k ];
scratch1[ idx + 1 ] = scratch0[ k + IN_SUBFR_LEN_RESAMPLE_3_1 ];
scratch1[ idx + 2 ] = scratch0[ k + 2 * IN_SUBFR_LEN_RESAMPLE_3_1 ];
}
/* Low-pass filtering */
SKP_Silk_lowpass_int( scratch1, S, scratch0, LSubFrameOut );
/* Saturate and convert to SKP_int16 */
for( k = 0; k < LSubFrameOut; k++ ) {
out_tmp = scratch0[ k ];
out[ k ] = (SKP_int16) SKP_SAT16( SKP_RSHIFT_ROUND( out_tmp, 10 ) );
}
in += LSubFrameIn;
inLenTmp -= LSubFrameIn;
out += LSubFrameOut;
}
}
/* Inverse filter of SKP_Silk_LPC_synthesis_filter */
void SKP_Silk_LPC_analysis_filter(
const SKP_int16 *in, /* I: Input signal */
const SKP_int16 *B, /* I: MA prediction coefficients, Q12 [order] */
SKP_int16 *S, /* I/O: State vector [order] */
SKP_int16 *out, /* O: Output signal */
const SKP_int32 len, /* I: Signal length */
const SKP_int32 Order /* I: Filter order */
)
{
SKP_int k, j, idx, Order_half = SKP_RSHIFT( Order, 1 );
SKP_int32 Btmp, B_align_Q12[ SigProc_MAX_ORDER_LPC >> 1 ], out32_Q12, out32;
SKP_int16 SA, SB;
/* Order must be even */
SKP_assert( 2 * Order_half == Order );
/* Combine two A_Q12 values and ensure 32-bit alignment */
for( k = 0; k < Order_half; k++ ) {
idx = SKP_SMULBB( 2, k );
B_align_Q12[ k ] = ( ( (SKP_int32)B[ idx ] ) & 0x0000ffff ) | SKP_LSHIFT( (SKP_int32)B[ idx + 1 ], 16 );
}
/* S[] values are in Q0 */
for( k = 0; k < len; k++ ) {
SA = S[ 0 ];
out32_Q12 = 0;
for( j = 0; j < ( Order_half - 1 ); j++ ) {
idx = SKP_SMULBB( 2, j ) + 1;
/* Multiply-add two prediction coefficients for each loop */
Btmp = B_align_Q12[ j ];
SB = S[ idx ];
S[ idx ] = SA;
out32_Q12 = SKP_SMLABB( out32_Q12, SA, Btmp );
out32_Q12 = SKP_SMLABT( out32_Q12, SB, Btmp );
SA = S[ idx + 1 ];
S[ idx + 1 ] = SB;
}
/* Unrolled loop: epilog */
Btmp = B_align_Q12[ Order_half - 1 ];
SB = S[ Order - 1 ];
S[ Order - 1 ] = SA;
out32_Q12 = SKP_SMLABB( out32_Q12, SA, Btmp );
out32_Q12 = SKP_SMLABT( out32_Q12, SB, Btmp );
/* Subtract prediction */
out32_Q12 = SKP_SUB_SAT32( SKP_LSHIFT( (SKP_int32)in[ k ], 12 ), out32_Q12 );
/* Scale to Q0 */
out32 = SKP_RSHIFT_ROUND( out32_Q12, 12 );
/* Saturate output */
out[ k ] = (SKP_int16)SKP_SAT16( out32 );
/* Move input line */
S[ 0 ] = in[ k ];
}
}
/* Compute weighted quantization errors for an LPC_order element input vector, over one codebook stage */
void SKP_Silk_NLSF_VQ_sum_error_FIX(int32_t * err_Q20, /* O Weighted quantization errors [N*K] */
const int *in_Q15, /* I Input vectors to be quantized [N*LPC_order] */
const int *w_Q6, /* I Weighting vectors [N*LPC_order] */
const int16_t * pCB_Q15, /* I Codebook vectors [K*LPC_order] */
const int N, /* I Number of input vectors */
const int K, /* I Number of codebook vectors */
const int LPC_order /* I Number of LPCs */
)
{
int i, n, m;
int32_t diff_Q15, sum_error, Wtmp_Q6;
int32_t Wcpy_Q6[MAX_LPC_ORDER / 2];
const int16_t *cb_vec_Q15;
assert(LPC_order <= 16);
assert((LPC_order & 1) == 0);
memzero(Wcpy_Q6, (MAX_LPC_ORDER / 2) * sizeof(int32_t));
/* Copy to local stack and pack two weights per int32 */
for (m = 0; m < SKP_RSHIFT(LPC_order, 1); m++) {
Wcpy_Q6[m] =
w_Q6[2 * m] | SKP_LSHIFT((int32_t) w_Q6[2 * m + 1], 16);
}
/* Loop over input vectors */
for (n = 0; n < N; n++) {
/* Loop over codebook */
cb_vec_Q15 = pCB_Q15;
for (i = 0; i < K; i++) {
sum_error = 0;
for (m = 0; m < LPC_order; m += 2) {
/* Get two weights packed in an int32 */
Wtmp_Q6 = Wcpy_Q6[SKP_RSHIFT(m, 1)];
/* Compute weighted squared quantization error for index m */
diff_Q15 = in_Q15[m] - *cb_vec_Q15++; // range: [ -32767 : 32767 ]
sum_error =
SKP_SMLAWB(sum_error,
SKP_SMULBB(diff_Q15, diff_Q15),
Wtmp_Q6);
/* Compute weighted squared quantization error for index m + 1 */
diff_Q15 = in_Q15[m + 1] - *cb_vec_Q15++; // range: [ -32767 : 32767 ]
sum_error =
SKP_SMLAWT(sum_error,
SKP_SMULBB(diff_Q15, diff_Q15),
Wtmp_Q6);
}
assert(sum_error >= 0);
err_Q20[i] = sum_error;
}
err_Q20 += K;
in_Q15 += LPC_order;
}
}
/* Set decoder sampling rate */
void SKP_Silk_decoder_set_fs(SKP_Silk_decoder_state * psDec, /* I/O Decoder state pointer */
int fs_kHz /* I Sampling frequency (kHz) */
)
{
if (psDec->fs_kHz != fs_kHz) {
psDec->fs_kHz = fs_kHz;
psDec->frame_length = SKP_SMULBB(FRAME_LENGTH_MS, fs_kHz);
psDec->subfr_length =
SKP_SMULBB(FRAME_LENGTH_MS / NB_SUBFR, fs_kHz);
if (psDec->fs_kHz == 8) {
psDec->LPC_order = MIN_LPC_ORDER;
psDec->psNLSF_CB[0] = &SKP_Silk_NLSF_CB0_10;
psDec->psNLSF_CB[1] = &SKP_Silk_NLSF_CB1_10;
} else {
psDec->LPC_order = MAX_LPC_ORDER;
psDec->psNLSF_CB[0] = &SKP_Silk_NLSF_CB0_16;
psDec->psNLSF_CB[1] = &SKP_Silk_NLSF_CB1_16;
}
/* Reset part of the decoder state */
SKP_memset(psDec->sLPC_Q14, 0, MAX_LPC_ORDER * sizeof(int32_t));
SKP_memset(psDec->outBuf, 0,
MAX_FRAME_LENGTH * sizeof(int16_t));
SKP_memset(psDec->prevNLSF_Q15, 0, MAX_LPC_ORDER * sizeof(int));
psDec->sLTP_buf_idx = 0;
psDec->lagPrev = 100;
psDec->LastGainIndex = 1;
psDec->prev_sigtype = 0;
psDec->first_frame_after_reset = 1;
if (fs_kHz == 24) {
psDec->HP_A = SKP_Silk_Dec_A_HP_24;
psDec->HP_B = SKP_Silk_Dec_B_HP_24;
} else if (fs_kHz == 16) {
psDec->HP_A = SKP_Silk_Dec_A_HP_16;
psDec->HP_B = SKP_Silk_Dec_B_HP_16;
} else if (fs_kHz == 12) {
psDec->HP_A = SKP_Silk_Dec_A_HP_12;
psDec->HP_B = SKP_Silk_Dec_B_HP_12;
} else if (fs_kHz == 8) {
psDec->HP_A = SKP_Silk_Dec_A_HP_8;
psDec->HP_B = SKP_Silk_Dec_B_HP_8;
} else {
/* unsupported sampling rate */
assert(0);
}
}
/* Check that settings are valid */
assert(psDec->frame_length > 0
&& psDec->frame_length <= MAX_FRAME_LENGTH);
}
/* Convert adaptive Mid/Side representation to Left/Right stereo signal */
void silk_stereo_MS_to_LR(
stereo_dec_state *state, /* I/O State */
opus_int16 x1[], /* I/O Left input signal, becomes mid signal */
opus_int16 x2[], /* I/O Right input signal, becomes side signal */
const opus_int32 pred_Q13[], /* I Predictors */
opus_int fs_kHz, /* I Samples rate (kHz) */
opus_int frame_length /* I Number of samples */
)
{
opus_int n, denom_Q16, delta0_Q13, delta1_Q13;
opus_int32 sum, diff, pred0_Q13, pred1_Q13;
/* Buffering */
SKP_memcpy( x1, state->sMid, 2 * sizeof( opus_int16 ) );
SKP_memcpy( x2, state->sSide, 2 * sizeof( opus_int16 ) );
SKP_memcpy( state->sMid, &x1[ frame_length ], 2 * sizeof( opus_int16 ) );
SKP_memcpy( state->sSide, &x2[ frame_length ], 2 * sizeof( opus_int16 ) );
/* Interpolate predictors and add prediction to side channel */
pred0_Q13 = state->pred_prev_Q13[ 0 ];
pred1_Q13 = state->pred_prev_Q13[ 1 ];
denom_Q16 = SKP_DIV32_16( 1 << 16, STEREO_INTERP_LEN_MS * fs_kHz );
delta0_Q13 = SKP_RSHIFT_ROUND( SKP_SMULBB( pred_Q13[ 0 ] - state->pred_prev_Q13[ 0 ], denom_Q16 ), 16 );
delta1_Q13 = SKP_RSHIFT_ROUND( SKP_SMULBB( pred_Q13[ 1 ] - state->pred_prev_Q13[ 1 ], denom_Q16 ), 16 );
for( n = 0; n < STEREO_INTERP_LEN_MS * fs_kHz; n++ ) {
pred0_Q13 += delta0_Q13;
pred1_Q13 += delta1_Q13;
sum = SKP_LSHIFT( SKP_ADD_LSHIFT( x1[ n ] + x1[ n + 2 ], x1[ n + 1 ], 1 ), 9 ); /* Q11 */
sum = SKP_SMLAWB( SKP_LSHIFT( ( opus_int32 )x2[ n + 1 ], 8 ), sum, pred0_Q13 ); /* Q8 */
sum = SKP_SMLAWB( sum, SKP_LSHIFT( ( opus_int32 )x1[ n + 1 ], 11 ), pred1_Q13 ); /* Q8 */
x2[ n + 1 ] = (opus_int16)SKP_SAT16( SKP_RSHIFT_ROUND( sum, 8 ) );
}
pred0_Q13 = pred_Q13[ 0 ];
pred1_Q13 = pred_Q13[ 1 ];
for( n = STEREO_INTERP_LEN_MS * fs_kHz; n < frame_length; n++ ) {
sum = SKP_LSHIFT( SKP_ADD_LSHIFT( x1[ n ] + x1[ n + 2 ], x1[ n + 1 ], 1 ), 9 ); /* Q11 */
sum = SKP_SMLAWB( SKP_LSHIFT( ( opus_int32 )x2[ n + 1 ], 8 ), sum, pred0_Q13 ); /* Q8 */
sum = SKP_SMLAWB( sum, SKP_LSHIFT( ( opus_int32 )x1[ n + 1 ], 11 ), pred1_Q13 ); /* Q8 */
x2[ n + 1 ] = (opus_int16)SKP_SAT16( SKP_RSHIFT_ROUND( sum, 8 ) );
}
state->pred_prev_Q13[ 0 ] = pred_Q13[ 0 ];
state->pred_prev_Q13[ 1 ] = pred_Q13[ 1 ];
/* Convert to left/right signals */
for( n = 0; n < frame_length; n++ ) {
sum = x1[ n + 1 ] + (opus_int32)x2[ n + 1 ];
diff = x1[ n + 1 ] - (opus_int32)x2[ n + 1 ];
x1[ n + 1 ] = (opus_int16)SKP_SAT16( sum );
x2[ n + 1 ] = (opus_int16)SKP_SAT16( diff );
}
}
/* 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;
}
}
/* Variable order MA filter */
void SKP_Silk_MA(
const SKP_int16 *in, /* I: input signal */
const SKP_int16 *B, /* I: MA coefficients, Q13 [order+1] */
SKP_int32 *S, /* I/O: state vector [order] */
SKP_int16 *out, /* O: output signal */
const SKP_int32 len, /* I: signal length */
const SKP_int32 order /* I: filter order */
)
{
SKP_int k, d, in16;
SKP_int32 out32;
for( k = 0; k < len; k++ ) {
in16 = in[ k ];
out32 = SKP_SMLABB( S[ 0 ], in16, B[ 0 ] );
out32 = SKP_RSHIFT_ROUND( out32, 13 );
for( d = 1; d < order; d++ ) {
S[ d - 1 ] = SKP_SMLABB( S[ d ], in16, B[ d ] );
}
S[ order - 1 ] = SKP_SMULBB( in16, B[ order ] );
/* Limit */
out[ k ] = (SKP_int16)SKP_SAT16( out32 );
}
}
/* Decodes signs of excitation */
void SKP_Silk_decode_signs(SKP_Silk_range_coder_state * sRC, /* I/O Range coder state */
int q[], /* I/O pulse signal */
const int length, /* I length of output */
const int sigtype, /* I Signal type */
const int QuantOffsetType, /* I Quantization offset type */
const int RateLevelIndex /* I Rate Level Index */
)
{
int i;
int data;
const uint16_t *cdf;
i = SKP_SMULBB(N_RATE_LEVELS - 1,
SKP_LSHIFT(sigtype,
1) + QuantOffsetType) + RateLevelIndex;
cdf = SKP_Silk_sign_CDF[i];
for (i = 0; i < length; i++) {
if (q[i] > 0) {
SKP_Silk_range_decoder(&data, sRC, cdf, 1);
/* attach sign */
/* implementation with shift, subtraction, multiplication */
q[i] *= SKP_dec_map(data);
}
}
}
/* Convert NLSF parameters to stable AR prediction filter coefficients */
void SKP_Silk_NLSF2A_stable(
SKP_int16 pAR_Q12[ MAX_LPC_ORDER ], /* O Stabilized AR coefs [LPC_order] */
const SKP_int pNLSF[ MAX_LPC_ORDER ], /* I NLSF vector [LPC_order] */
const SKP_int LPC_order /* I LPC/LSF order */
)
{
SKP_int i;
SKP_int32 invGain_Q30;
SKP_Silk_NLSF2A( pAR_Q12, pNLSF, LPC_order );
/* Ensure stable LPCs */
for( i = 0; i < MAX_LPC_STABILIZE_ITERATIONS; i++ ) {
if( SKP_Silk_LPC_inverse_pred_gain( &invGain_Q30, pAR_Q12, LPC_order ) == 1 ) {
SKP_Silk_bwexpander( pAR_Q12, LPC_order, 65536 - SKP_SMULBB( 10 + i, i ) ); /* 10_Q16 = 0.00015 */
} else {
break;
}
}
/* Reached the last iteration */
if( i == MAX_LPC_STABILIZE_ITERATIONS ) {
SKP_assert( 0 );
for( i = 0; i < LPC_order; i++ ) {
pAR_Q12[ i ] = 0;
}
}
}
/* Calculates correlation vector X'*t */
void SKP_Silk_corrVector_FIX(
const SKP_int16 *x, /* I x vector [L + order - 1] used to form data matrix X */
const SKP_int16 *t, /* I target vector [L] */
const SKP_int L, /* I Length of vectors */
const SKP_int order, /* I Max lag for correlation */
SKP_int32 *Xt, /* O Pointer to X'*t correlation vector [order] */
const SKP_int rshifts /* I Right shifts of correlations */
)
{
SKP_int lag, i;
const SKP_int16 *ptr1, *ptr2;
SKP_int32 inner_prod;
ptr1 = &x[ order - 1 ]; /* Points to first sample of column 0 of X: X[:,0] */
ptr2 = t;
/* Calculate X'*t */
if( rshifts > 0 ) {
/* Right shifting used */
for( lag = 0; lag < order; lag++ ) {
inner_prod = 0;
for( i = 0; i < L; i++ ) {
inner_prod += SKP_RSHIFT32( SKP_SMULBB( ptr1[ i ], ptr2[i] ), rshifts );
}
Xt[ lag ] = inner_prod; /* X[:,lag]'*t */
ptr1--; /* Go to next column of X */
}
} else {
SKP_assert( rshifts == 0 );
for( lag = 0; lag < order; lag++ ) {
Xt[ lag ] = SKP_Silk_inner_prod_aligned( ptr1, ptr2, L ); /* X[:,lag]'*t */
ptr1--; /* Go to next column of X */
}
}
}
/* Variable order MA prediction error filter */
void SKP_Silk_MA_Prediction(
const SKP_int16 *in, /* I: Input signal */
const SKP_int16 *B, /* I: MA prediction coefficients, Q12 [order] */
SKP_int32 *S, /* I/O: State vector [order] */
SKP_int16 *out, /* O: Output signal */
const SKP_int32 len, /* I: Signal length */
const SKP_int32 order /* I: Filter order */
)
{
SKP_int k, d, in16;
SKP_int32 out32;
for( k = 0; k < len; k++ ) {
in16 = in[ k ];
out32 = SKP_LSHIFT( in16, 12 ) - S[ 0 ];
out32 = SKP_RSHIFT_ROUND( out32, 12 );
for( d = 0; d < order - 1; d++ ) {
S[ d ] = SKP_SMLABB_ovflw( S[ d + 1 ], in16, B[ d ] );
}
S[ order - 1 ] = SKP_SMULBB( in16, B[ order - 1 ] );
/* Limit */
out[ k ] = (SKP_int16)SKP_SAT16( out32 );
}
}
/* Encodes signs of excitation */
void SKP_Silk_encode_signs(
SKP_Silk_range_coder_state *sRC, /* I/O Range coder state */
const SKP_int8 q[], /* I Pulse signal */
const SKP_int length, /* I Length of input */
const SKP_int sigtype, /* I Signal type */
const SKP_int QuantOffsetType, /* I Quantization offset type */
const SKP_int RateLevelIndex /* I Rate level index */
)
{
SKP_int i;
SKP_int inData;
SKP_uint16 cdf[ 3 ];
i = SKP_SMULBB( N_RATE_LEVELS - 1, SKP_LSHIFT( sigtype, 1 ) + QuantOffsetType ) + RateLevelIndex;
cdf[ 0 ] = 0;
cdf[ 1 ] = SKP_Silk_sign_CDF[ i ];
cdf[ 2 ] = 65535;
for( i = 0; i < length; i++ ) {
if( q[ i ] != 0 ) {
inData = SKP_enc_map( q[ i ] ); /* - = 0, + = 1 */
SKP_Silk_range_encoder( sRC, inData, cdf );
}
}
}
/* Compute quantization errors for an LPC_order element input vector for a VQ codebook */
void silk_NLSF_VQ(
SKP_int32 err_Q26[], /* O Quantization errors [K] */
const SKP_int16 in_Q15[], /* I Input vectors to be quantized [LPC_order] */
const SKP_uint8 pCB_Q8[], /* I Codebook vectors [K*LPC_order] */
const SKP_int K, /* I Number of codebook vectors */
const SKP_int LPC_order /* I Number of LPCs */
)
{
SKP_int i, m;
SKP_int32 diff_Q15, sum_error_Q30, sum_error_Q26;
SKP_assert( LPC_order <= 16 );
SKP_assert( ( LPC_order & 1 ) == 0 );
/* Loop over codebook */
for( i = 0; i < K; i++ ) {
sum_error_Q26 = 0;
for( m = 0; m < LPC_order; m += 2 ) {
/* Compute weighted squared quantization error for index m */
diff_Q15 = SKP_SUB_LSHIFT32( in_Q15[ m ], ( SKP_int32 )*pCB_Q8++, 7 ); // range: [ -32767 : 32767 ]
sum_error_Q30 = SKP_SMULBB( diff_Q15, diff_Q15 );
/* Compute weighted squared quantization error for index m + 1 */
diff_Q15 = SKP_SUB_LSHIFT32( in_Q15[m + 1], ( SKP_int32 )*pCB_Q8++, 7 ); // range: [ -32767 : 32767 ]
sum_error_Q30 = SKP_SMLABB( sum_error_Q30, diff_Q15, diff_Q15 );
sum_error_Q26 = SKP_ADD_RSHIFT32( sum_error_Q26, sum_error_Q30, 4 );
SKP_assert( sum_error_Q26 >= 0 );
SKP_assert( sum_error_Q30 >= 0 );
}
err_Q26[ i ] = sum_error_Q26;
}
}
/* 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 ) );
}
void SKP_Silk_LTP_analysis_filter_FIX(
SKP_int16 *LTP_res, /* O: LTP residual signal of length NB_SUBFR * ( pre_length + subfr_length ) */
const SKP_int16 *x, /* I: Pointer to input signal with at least max( pitchL ) preceeding samples */
const SKP_int16 LTPCoef_Q14[ LTP_ORDER * NB_SUBFR ],/* I: LTP_ORDER LTP coefficients for each NB_SUBFR subframe */
const SKP_int pitchL[ NB_SUBFR ], /* I: Pitch lag, one for each subframe */
const SKP_int32 invGains_Qxx[ NB_SUBFR ], /* I: Inverse quantization gains, one for each subframe */
const SKP_int Qxx, /* I: Inverse quantization gains Q domain */
const SKP_int subfr_length, /* I: Length of each subframe */
const SKP_int pre_length /* I: Length of the preceeding samples starting at &x[0] for each subframe */
)
{
const SKP_int16 *x_ptr, *x_lag_ptr;
SKP_int16 Btmp_Q14[ LTP_ORDER ];
SKP_int16 *LTP_res_ptr;
SKP_int k, i, j;
SKP_int32 LTP_est;
x_ptr = x;
LTP_res_ptr = LTP_res;
for( k = 0; k < NB_SUBFR; k++ ) {
x_lag_ptr = x_ptr - pitchL[ k ];
for( i = 0; i < LTP_ORDER; i++ ) {
Btmp_Q14[ i ] = LTPCoef_Q14[ k * LTP_ORDER + i ];
}
/* LTP analysis FIR filter */
for( i = 0; i < subfr_length + pre_length; i++ ) {
LTP_res_ptr[ i ] = x_ptr[ i ];
/* Long-term prediction */
LTP_est = SKP_SMULBB( x_lag_ptr[ LTP_ORDER / 2 ], Btmp_Q14[ 0 ] );
for( j = 1; j < LTP_ORDER; j++ ) {
LTP_est = SKP_SMLABB_ovflw( LTP_est, x_lag_ptr[ LTP_ORDER / 2 - j ], Btmp_Q14[ j ] );
}
LTP_est = SKP_RSHIFT_ROUND( LTP_est, 14 ); // round and -> Q0
/* Subtract long-term prediction */
LTP_res_ptr[ i ] = ( SKP_int16 )SKP_SAT16( ( SKP_int32 )x_ptr[ i ] - LTP_est );
/* Scale residual */
if( Qxx == 16 ) {
LTP_res_ptr[ i ] = SKP_SMULWB( invGains_Qxx[ k ], LTP_res_ptr[ i ] );
} else {
LTP_res_ptr[ i ] = ( SKP_int16 )SKP_CHECK_FIT16( SKP_RSHIFT64( SKP_SMULL( invGains_Qxx[ k ], LTP_res_ptr[ i ] ), Qxx ) );
}
x_lag_ptr++;
}
/* Update pointers */
LTP_res_ptr += subfr_length + pre_length;
x_ptr += subfr_length;
}
}
/* Variable order MA prediction error filter */
void SKP_Silk_MA_Prediction(
const SKP_int16 *in, /* I: Input signal */
const SKP_int16 *B, /* I: MA prediction coefficients, Q12 [order] */
SKP_int32 *S, /* I/O: State vector [order] */
SKP_int16 *out, /* O: Output signal */
const SKP_int32 len, /* I: Signal length */
const SKP_int32 order /* I: Filter order */
)
{
SKP_int k, d, in16;
SKP_int32 out32;
SKP_int32 B32;
if( ( order & 1 ) == 0 && (SKP_int32)( (SKP_int_ptr_size)B & 3 ) == 0 ) {
/* Even order and 4-byte aligned coefficient array */
/* NOTE: the code below loads two int16 values in an int32, and multiplies each using the */
/* SMLABB and SMLABT instructions. On a big-endian CPU the two int16 variables would be */
/* loaded in reverse order and the code will give the wrong result. In that case swapping */
/* the SMLABB and SMLABT instructions should solve the problem. */
for( k = 0; k < len; k++ ) {
in16 = in[ k ];
out32 = SKP_LSHIFT( in16, 12 ) - S[ 0 ];
out32 = SKP_RSHIFT_ROUND( out32, 12 );
for( d = 0; d < order - 2; d += 2 ) {
B32 = *( (SKP_int32*)&B[ d ] ); /* read two coefficients at once */
S[ d ] = SKP_SMLABB_ovflw( S[ d + 1 ], in16, B32 );
S[ d + 1 ] = SKP_SMLABT_ovflw( S[ d + 2 ], in16, B32 );
}
B32 = *( (SKP_int32*)&B[ d ] ); /* read two coefficients at once */
S[ order - 2 ] = SKP_SMLABB_ovflw( S[ order - 1 ], in16, B32 );
S[ order - 1 ] = SKP_SMULBT( in16, B32 );
/* Limit */
out[ k ] = (SKP_int16)SKP_SAT16( out32 );
}
} else {
/* Odd order or not 4-byte aligned coefficient array */
for( k = 0; k < len; k++ ) {
in16 = in[ k ];
out32 = SKP_LSHIFT( in16, 12 ) - S[ 0 ];
out32 = SKP_RSHIFT_ROUND( out32, 12 );
for( d = 0; d < order - 1; d++ ) {
S[ d ] = SKP_SMLABB_ovflw( S[ d + 1 ], in16, B[ d ] );
}
S[ order - 1 ] = SKP_SMULBB( in16, B[ order - 1 ] );
/* Limit */
out[ k ] = (SKP_int16)SKP_SAT16( out32 );
}
}
}
void SKP_Silk_noise_shape_analysis_FIX(
SKP_Silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */
SKP_Silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control FIX */
const SKP_int16 *pitch_res, /* I LPC residual from pitch analysis */
const SKP_int16 *x /* I Input signal [ frame_length + la_shape ] */
)
{
SKP_Silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
SKP_int k, i, nSamples, Qnrg, b_Q14, warping_Q16, scale = 0;
SKP_int32 SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32;
SKP_int32 nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7;
SKP_int32 delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8;
SKP_int32 auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ];
SKP_int32 refl_coef_Q16[ MAX_SHAPE_LPC_ORDER ];
SKP_int32 AR1_Q24[ MAX_SHAPE_LPC_ORDER ];
SKP_int32 AR2_Q24[ MAX_SHAPE_LPC_ORDER ];
SKP_int16 x_windowed[ SHAPE_LPC_WIN_MAX ];
SKP_int32 sqrt_nrg[ MAX_NB_SUBFR ], Qnrg_vec[ MAX_NB_SUBFR ];
const SKP_int16 *x_ptr, *pitch_res_ptr;
/* Point to start of first LPC analysis block */
x_ptr = x - psEnc->sCmn.la_shape;
/****************/
/* CONTROL SNR */
/****************/
/* Reduce SNR_dB values if recent bitstream has exceeded TargetRate */
psEncCtrl->current_SNR_dB_Q7 = psEnc->SNR_dB_Q7 - SKP_SMULBB( psEnc->BufferedInChannel_ms, SKP_FIX_CONST( 0.1, 7 ) );
/* Reduce SNR_dB because of any inband FEC used */
psEncCtrl->current_SNR_dB_Q7 -= psEnc->inBandFEC_SNR_comp_Q7;
/****************/
/* GAIN CONTROL */
/****************/
/* Input quality is the average of the quality in the lowest two VAD bands */
psEncCtrl->input_quality_Q14 = ( SKP_int )SKP_RSHIFT( ( SKP_int32 )psEnc->sCmn.input_quality_bands_Q15[ 0 ]
+ psEnc->sCmn.input_quality_bands_Q15[ 1 ], 2 );
/* Coding quality level, between 0.0_Q0 and 1.0_Q0, but in Q14 */
psEncCtrl->coding_quality_Q14 = SKP_RSHIFT( SKP_Silk_sigm_Q15( SKP_RSHIFT_ROUND( psEncCtrl->current_SNR_dB_Q7 -
SKP_FIX_CONST( 18.0, 7 ), 4 ) ), 1 );
/* Reduce coding SNR during low speech activity */
SNR_adj_dB_Q7 = psEncCtrl->current_SNR_dB_Q7;
if( psEnc->sCmn.useCBR == 0 ) {
b_Q8 = SKP_FIX_CONST( 1.0, 8 ) - psEnc->sCmn.speech_activity_Q8;
b_Q8 = SKP_SMULWB( SKP_LSHIFT( b_Q8, 8 ), b_Q8 );
SNR_adj_dB_Q7 = SKP_SMLAWB( SNR_adj_dB_Q7,
SKP_SMULBB( SKP_FIX_CONST( -BG_SNR_DECR_dB, 7 ) >> ( 4 + 1 ), b_Q8 ), // Q11
SKP_SMULWB( SKP_FIX_CONST( 1.0, 14 ) + psEncCtrl->input_quality_Q14, psEncCtrl->coding_quality_Q14 ) ); // Q12
}
/* SKP_Silk_prefilter. Prefilter for finding Quantizer input signal */
SKP_INLINE void SKP_Silk_prefilt_FIX(
SKP_Silk_prefilter_state_FIX *P, /* I/O state */
SKP_int32 st_res_Q12[], /* I short term residual signal */
SKP_int16 xw[], /* O prefiltered signal */
SKP_int32 HarmShapeFIRPacked_Q12, /* I Harmonic shaping coeficients */
SKP_int Tilt_Q14, /* I Tilt shaping coeficient */
SKP_int32 LF_shp_Q14, /* I Low-frequancy shaping coeficients*/
SKP_int lag, /* I Lag for harmonic shaping */
SKP_int length /* I Length of signals */
)
{
SKP_int i, idx, LTP_shp_buf_idx;
SKP_int32 n_LTP_Q12, n_Tilt_Q10, n_LF_Q10;
SKP_int32 sLF_MA_shp_Q12, sLF_AR_shp_Q12;
SKP_int16 *LTP_shp_buf;
/* To speed up use temp variables instead of using the struct */
LTP_shp_buf = P->sLTP_shp;
LTP_shp_buf_idx = P->sLTP_shp_buf_idx;
sLF_AR_shp_Q12 = P->sLF_AR_shp_Q12;
sLF_MA_shp_Q12 = P->sLF_MA_shp_Q12;
for( i = 0; i < length; i++ ) {
if( lag > 0 ) {
/* unrolled loop */
SKP_assert( HARM_SHAPE_FIR_TAPS == 3 );
idx = lag + LTP_shp_buf_idx;
n_LTP_Q12 = SKP_SMULBB( LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 - 1) & LTP_MASK ], HarmShapeFIRPacked_Q12 );
n_LTP_Q12 = SKP_SMLABT( n_LTP_Q12, LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 ) & LTP_MASK ], HarmShapeFIRPacked_Q12 );
n_LTP_Q12 = SKP_SMLABB( n_LTP_Q12, LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 + 1) & LTP_MASK ], HarmShapeFIRPacked_Q12 );
} else {
n_LTP_Q12 = 0;
}
n_Tilt_Q10 = SKP_SMULWB( sLF_AR_shp_Q12, Tilt_Q14 );
n_LF_Q10 = SKP_SMLAWB( SKP_SMULWT( sLF_AR_shp_Q12, LF_shp_Q14 ), sLF_MA_shp_Q12, LF_shp_Q14 );
sLF_AR_shp_Q12 = SKP_SUB32( st_res_Q12[ i ], SKP_LSHIFT( n_Tilt_Q10, 2 ) );
sLF_MA_shp_Q12 = SKP_SUB32( sLF_AR_shp_Q12, SKP_LSHIFT( n_LF_Q10, 2 ) );
LTP_shp_buf_idx = ( LTP_shp_buf_idx - 1 ) & LTP_MASK;
LTP_shp_buf[ LTP_shp_buf_idx ] = ( SKP_int16 )SKP_SAT16( SKP_RSHIFT_ROUND( sLF_MA_shp_Q12, 12 ) );
xw[i] = ( SKP_int16 )SKP_SAT16( SKP_RSHIFT_ROUND( SKP_SUB32( sLF_MA_shp_Q12, n_LTP_Q12 ), 12 ) );
}
/* Copy temp variable back to state */
P->sLF_AR_shp_Q12 = sLF_AR_shp_Q12;
P->sLF_MA_shp_Q12 = sLF_MA_shp_Q12;
P->sLTP_shp_buf_idx = LTP_shp_buf_idx;
}
opus_int32 silk_inner_prod_aligned_scale(
const opus_int16 *const inVec1, /* I input vector 1 */
const opus_int16 *const inVec2, /* I input vector 2 */
const opus_int scale, /* I number of bits to shift */
const opus_int len /* I vector lengths */
)
{
opus_int i;
opus_int32 sum = 0;
for( i = 0; i < len; i++ ) {
sum = SKP_ADD_RSHIFT32( sum, SKP_SMULBB( inVec1[ i ], inVec2[ i ] ), scale );
}
return sum;
}
SKP_int SKP_Silk_SDK_QueryEncoder(
const void *encState, /* I: State Vector */
SKP_Silk_EncodeControlStruct *encStatus /* O: Control Structure */
)
{
SKP_Silk_encoder_state_Fxx *psEnc;
SKP_int ret = SKP_SILK_NO_ERROR;
psEnc = ( SKP_Silk_encoder_state_Fxx* )encState;
encStatus->API_sampleRate = psEnc->sCmn.API_fs_Hz;
encStatus->maxInternalSampleRate = SKP_SMULBB( psEnc->sCmn.maxInternal_fs_kHz, 1000 );
encStatus->minInternalSampleRate = SKP_SMULBB( psEnc->sCmn.minInternal_fs_kHz, 1000 );
encStatus->payloadSize_ms = psEnc->sCmn.PacketSize_ms;
encStatus->bitRate = psEnc->sCmn.TargetRate_bps;
encStatus->packetLossPercentage = psEnc->sCmn.PacketLoss_perc;
encStatus->complexity = psEnc->sCmn.Complexity;
encStatus->useInBandFEC = psEnc->sCmn.useInBandFEC;
encStatus->useDTX = psEnc->sCmn.useDTX;
encStatus->useCBR = psEnc->sCmn.useCBR;
encStatus->internalSampleRate = SKP_SMULBB( psEnc->sCmn.fs_kHz, 1000 );
return ret;
}
/* Predict number of bytes used to encode q */
int SKP_Silk_pulses_to_bytes( /* O Return value, predicted number of bytes used to encode q */
SKP_Silk_encoder_state * psEncC, /* I/O Encoder State */
int q[] /* I Pulse signal */
)
{
int i, j, iter, *q_ptr;
int32_t sum_abs_val, nBytes, acc_nBytes;
/* Take the absolute value of the pulses */
iter = psEncC->frame_length / SHELL_CODEC_FRAME_LENGTH;
/* Calculate rate as a nonlinaer mapping of sum abs value of each Shell block */
q_ptr = q;
acc_nBytes = 0;
for (j = 0; j < iter; j++) {
sum_abs_val = 0;
for (i = 0; i < SHELL_CODEC_FRAME_LENGTH; i += 4) {
sum_abs_val += SKP_abs(q_ptr[i + 0]);
sum_abs_val += SKP_abs(q_ptr[i + 1]);
sum_abs_val += SKP_abs(q_ptr[i + 2]);
sum_abs_val += SKP_abs(q_ptr[i + 3]);
}
/* Calculate nBytes used for thi sshell frame */
nBytes = SKP_SMULWB(SKP_SMULBB(sum_abs_val, sum_abs_val), POLY_FIT_2_Q20); // Q4
nBytes = SKP_LSHIFT_SAT32(nBytes, 11); // Q15
nBytes += SKP_SMULBB(sum_abs_val, POLY_FIT_1_Q15); // Q15
nBytes += POLY_FIT_0_Q15; // Q15
acc_nBytes += nBytes;
q_ptr += SHELL_CODEC_FRAME_LENGTH; /* update pointer */
}
acc_nBytes = SKP_RSHIFT_ROUND(acc_nBytes, 15); // Q0
acc_nBytes = SKP_SAT16(acc_nBytes); // just to be sure // Q0
return ((int)acc_nBytes);
}
/* even order AR filter */
void SKP_Silk_LPC_synthesis_filter(
const SKP_int16 *in, /* I: excitation signal */
const SKP_int16 *A_Q12, /* I: AR coefficients [Order], between -8_Q0 and 8_Q0 */
const SKP_int32 Gain_Q26, /* I: gain */
SKP_int32 *S, /* I/O: state vector [Order] */
SKP_int16 *out, /* O: output signal */
const SKP_int32 len, /* I: signal length */
const SKP_int Order /* I: filter order, must be even */
)
{
SKP_int k, j, idx, Order_half = SKP_RSHIFT( Order, 1 );
SKP_int32 SA, SB, out32_Q10, out32;
/* Order must be even */
SKP_assert( 2 * Order_half == Order );
/* S[] values are in Q14 */
for( k = 0; k < len; k++ ) {
SA = S[ Order - 1 ];
out32_Q10 = 0;
for( j = 0; j < ( Order_half - 1 ); j++ ) {
idx = SKP_SMULBB( 2, j ) + 1;
SB = S[ Order - 1 - idx ];
S[ Order - 1 - idx ] = SA;
out32_Q10 = SKP_SMLAWB( out32_Q10, SA, A_Q12[ ( j << 1 ) ] );
out32_Q10 = SKP_SMLAWB( out32_Q10, SB, A_Q12[ ( j << 1 ) + 1 ] );
SA = S[ Order - 2 - idx ];
S[ Order - 2 - idx ] = SB;
}
/* unrolled loop: epilog */
SB = S[ 0 ];
S[ 0 ] = SA;
out32_Q10 = SKP_SMLAWB( out32_Q10, SA, A_Q12[ Order - 2 ] );
out32_Q10 = SKP_SMLAWB( out32_Q10, SB, A_Q12[ Order - 1 ] );
/* apply gain to excitation signal and add to prediction */
out32_Q10 = SKP_ADD_SAT32( out32_Q10, SKP_SMULWB( Gain_Q26, in[ k ] ) );
/* scale to Q0 */
out32 = SKP_RSHIFT_ROUND( out32_Q10, 10 );
/* saturate output */
out[ k ] = ( SKP_int16 )SKP_SAT16( out32 );
/* move result into delay line */
S[ Order - 1 ] = SKP_LSHIFT_SAT32( out32_Q10, 4 );
}
}
void SKP_Silk_detect_SWB_input(SKP_Silk_detect_SWB_state * psSWBdetect, /* (I/O) encoder state */
const int16_t samplesIn[], /* (I) input to encoder */
int nSamplesIn /* (I) length of input */
)
{
int HP_8_kHz_len, i;
int16_t in_HP_8_kHz[MAX_FRAME_LENGTH];
int32_t energy_32, shift;
/* High pass filter with cutoff at 8 khz */
HP_8_kHz_len = SKP_min_int(nSamplesIn, MAX_FRAME_LENGTH);
HP_8_kHz_len = SKP_max_int(HP_8_kHz_len, 0);
/* Cutoff around 9 khz */
/* A = conv(conv([8192,14613, 6868], [8192,12883, 7337]), [8192,11586, 7911]); */
/* B = conv(conv([575, -948, 575], [575, -221, 575]), [575, 104, 575]); */
SKP_Silk_biquad(samplesIn, SKP_Silk_SWB_detect_B_HP_Q13[0],
SKP_Silk_SWB_detect_A_HP_Q13[0],
psSWBdetect->S_HP_8_kHz[0], in_HP_8_kHz, HP_8_kHz_len);
for (i = 1; i < NB_SOS; i++) {
SKP_Silk_biquad(in_HP_8_kHz, SKP_Silk_SWB_detect_B_HP_Q13[i],
SKP_Silk_SWB_detect_A_HP_Q13[i],
psSWBdetect->S_HP_8_kHz[i], in_HP_8_kHz,
HP_8_kHz_len);
}
/* Calculate energy in HP signal */
SKP_Silk_sum_sqr_shift(&energy_32, &shift, in_HP_8_kHz, HP_8_kHz_len);
/* Count concecutive samples above threshold, after adjusting threshold for number of input samples and shift */
if (energy_32 >
SKP_RSHIFT(SKP_SMULBB(HP_8_KHZ_THRES, HP_8_kHz_len), shift)) {
psSWBdetect->ConsecSmplsAboveThres += nSamplesIn;
if (psSWBdetect->ConsecSmplsAboveThres > CONCEC_SWB_SMPLS_THRES) {
psSWBdetect->SWB_detected = 1;
}
} else {
psSWBdetect->ConsecSmplsAboveThres -= nSamplesIn;
psSWBdetect->ConsecSmplsAboveThres =
SKP_max(psSWBdetect->ConsecSmplsAboveThres, 0);
}
/* If sufficient speech activity and no SWB detected, we detect the signal as being WB */
if ((psSWBdetect->ActiveSpeech_ms > WB_DETECT_ACTIVE_SPEECH_MS_THRES)
&& (psSWBdetect->SWB_detected == 0)) {
psSWBdetect->WB_detected = 1;
}
}
/* Solve L^t*x = b, where L is lower triangular with ones on the diagonal */
SKP_INLINE void SKP_Silk_LS_SolveLast_FIX(
const SKP_int32 *L_Q16, /* I Pointer to Lower Triangular Matrix */
const SKP_int M, /* I Dim of Matrix equation */
const SKP_int32 *b, /* I b Vector */
SKP_int32 *x_Q16 /* O x Vector */
)
{
SKP_int i, j;
const SKP_int32 *ptr32;
SKP_int32 tmp_32;
for( i = M - 1; i >= 0; i-- ) {
ptr32 = matrix_adr( L_Q16, 0, i, M );
tmp_32 = 0;
for( j = M - 1; j > i; j-- ) {
tmp_32 = SKP_SMLAWW( tmp_32, ptr32[ SKP_SMULBB( j, M ) ], x_Q16[ j ] );
}
x_Q16[ i ] = SKP_SUB32( b[ i ], tmp_32 );
}
}
SKP_INLINE SKP_int16 *SKP_Silk_resampler_private_IIR_FIR_INTERPOL(
SKP_int16 * out, SKP_int16 * buf, SKP_int32 max_index_Q16 , SKP_int32 index_increment_Q16 ){
SKP_int32 index_Q16, res_Q15;
SKP_int16 *buf_ptr;
SKP_int32 table_index;
/* 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 ) );
}
return out;
}
/* 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;
}
}
/* Generic filter loop. Runs about 10..25% faster than the unrolled
C implementation. */
SKP_INLINE void FILTER_LOOP_ORDER_ANY(SKP_int32 * S, const SKP_int16 * B,
SKP_int16 in16, SKP_int32 order)
{
SKP_int d;
/* Unroll for 4 elements */
for(d = 0; d < order - 4; d += 4) {
S[d + 0] = SKP_SMLABB(S[d + 1], in16, B[d + 0]);
S[d + 1] = SKP_SMLABB(S[d + 2], in16, B[d + 1]);
S[d + 2] = SKP_SMLABB(S[d + 3], in16, B[d + 2]);
S[d + 3] = SKP_SMLABB(S[d + 4], in16, B[d + 3]);
}
if(d < order - 2) {
S[d + 0] = SKP_SMLABB(S[d + 1], in16, B[d + 0]);
S[d + 1] = SKP_SMLABB(S[d + 2], in16, B[d + 1]);
d += 2;
}
if(d == order - 2)
S[d] = SKP_SMLABB(S[d + 1], in16, B[d]);
S[order - 1] = SKP_SMULBB(in16, B[order - 1]);
}
