1/***********************************************************************
2Copyright (c) 2006-2011, Skype Limited. All rights reserved.
3Redistribution and use in source and binary forms, with or without
4modification, are permitted provided that the following conditions
5are met:
6- Redistributions of source code must retain the above copyright notice,
7this list of conditions and the following disclaimer.
8- Redistributions in binary form must reproduce the above copyright
9notice, this list of conditions and the following disclaimer in the
10documentation and/or other materials provided with the distribution.
11- Neither the name of Internet Society, IETF or IETF Trust, nor the
12names of specific contributors, may be used to endorse or promote
13products derived from this software without specific prior written
14permission.
15THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25POSSIBILITY OF SUCH DAMAGE.
26***********************************************************************/
27
28#ifdef HAVE_CONFIG_H
29#include "config.h"
30#endif
31
32#include "main_FIX.h"
33#include "tuning_parameters.h"
34
35/* Head room for correlations */
36#define LTP_CORRS_HEAD_ROOM                             2
37
38void silk_fit_LTP(
39    opus_int32 LTP_coefs_Q16[ LTP_ORDER ],
40    opus_int16 LTP_coefs_Q14[ LTP_ORDER ]
41);
42
43void silk_find_LTP_FIX(
44    opus_int16                      b_Q14[ MAX_NB_SUBFR * LTP_ORDER ],      /* O    LTP coefs                                                                   */
45    opus_int32                      WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O    Weight for LTP quantization                                           */
46    opus_int                        *LTPredCodGain_Q7,                      /* O    LTP coding gain                                                             */
47    const opus_int16                r_lpc[],                                /* I    residual signal after LPC signal + state for first 10 ms                    */
48    const opus_int                  lag[ MAX_NB_SUBFR ],                    /* I    LTP lags                                                                    */
49    const opus_int32                Wght_Q15[ MAX_NB_SUBFR ],               /* I    weights                                                                     */
50    const opus_int                  subfr_length,                           /* I    subframe length                                                             */
51    const opus_int                  nb_subfr,                               /* I    number of subframes                                                         */
52    const opus_int                  mem_offset,                             /* I    number of samples in LTP memory                                             */
53    opus_int                        corr_rshifts[ MAX_NB_SUBFR ]            /* O    right shifts applied to correlations                                        */
54)
55{
56    opus_int   i, k, lshift;
57    const opus_int16 *r_ptr, *lag_ptr;
58    opus_int16 *b_Q14_ptr;
59
60    opus_int32 regu;
61    opus_int32 *WLTP_ptr;
62    opus_int32 b_Q16[ LTP_ORDER ], delta_b_Q14[ LTP_ORDER ], d_Q14[ MAX_NB_SUBFR ], nrg[ MAX_NB_SUBFR ], g_Q26;
63    opus_int32 w[ MAX_NB_SUBFR ], WLTP_max, max_abs_d_Q14, max_w_bits;
64
65    opus_int32 temp32, denom32;
66    opus_int   extra_shifts;
67    opus_int   rr_shifts, maxRshifts, maxRshifts_wxtra, LZs;
68    opus_int32 LPC_res_nrg, LPC_LTP_res_nrg, div_Q16;
69    opus_int32 Rr[ LTP_ORDER ], rr[ MAX_NB_SUBFR ];
70    opus_int32 wd, m_Q12;
71
72    b_Q14_ptr = b_Q14;
73    WLTP_ptr  = WLTP;
74    r_ptr     = &r_lpc[ mem_offset ];
75    for( k = 0; k < nb_subfr; k++ ) {
76        lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 );
77
78        silk_sum_sqr_shift( &rr[ k ], &rr_shifts, r_ptr, subfr_length ); /* rr[ k ] in Q( -rr_shifts ) */
79
80        /* Assure headroom */
81        LZs = silk_CLZ32( rr[k] );
82        if( LZs < LTP_CORRS_HEAD_ROOM ) {
83            rr[ k ] = silk_RSHIFT_ROUND( rr[ k ], LTP_CORRS_HEAD_ROOM - LZs );
84            rr_shifts += ( LTP_CORRS_HEAD_ROOM - LZs );
85        }
86        corr_rshifts[ k ] = rr_shifts;
87        silk_corrMatrix_FIX( lag_ptr, subfr_length, LTP_ORDER, LTP_CORRS_HEAD_ROOM, WLTP_ptr, &corr_rshifts[ k ] );  /* WLTP_fix_ptr in Q( -corr_rshifts[ k ] ) */
88
89        /* The correlation vector always has lower max abs value than rr and/or RR so head room is assured */
90        silk_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, Rr, corr_rshifts[ k ] );  /* Rr_fix_ptr   in Q( -corr_rshifts[ k ] ) */
91        if( corr_rshifts[ k ] > rr_shifts ) {
92            rr[ k ] = silk_RSHIFT( rr[ k ], corr_rshifts[ k ] - rr_shifts ); /* rr[ k ] in Q( -corr_rshifts[ k ] ) */
93        }
94        silk_assert( rr[ k ] >= 0 );
95
96        regu = 1;
97        regu = silk_SMLAWB( regu, rr[ k ], SILK_FIX_CONST( LTP_DAMPING/3, 16 ) );
98        regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, 0, 0, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) );
99        regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, LTP_ORDER-1, LTP_ORDER-1, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) );
100        silk_regularize_correlations_FIX( WLTP_ptr, &rr[k], regu, LTP_ORDER );
101
102        silk_solve_LDL_FIX( WLTP_ptr, LTP_ORDER, Rr, b_Q16 ); /* WLTP_fix_ptr and Rr_fix_ptr both in Q(-corr_rshifts[k]) */
103
104        /* Limit and store in Q14 */
105        silk_fit_LTP( b_Q16, b_Q14_ptr );
106
107        /* Calculate residual energy */
108        nrg[ k ] = silk_residual_energy16_covar_FIX( b_Q14_ptr, WLTP_ptr, Rr, rr[ k ], LTP_ORDER, 14 ); /* nrg_fix in Q( -corr_rshifts[ k ] ) */
109
110        /* temp = Wght[ k ] / ( nrg[ k ] * Wght[ k ] + 0.01f * subfr_length ); */
111        extra_shifts = silk_min_int( corr_rshifts[ k ], LTP_CORRS_HEAD_ROOM );
112        denom32 = silk_LSHIFT_SAT32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 + extra_shifts ) + /* Q( -corr_rshifts[ k ] + extra_shifts ) */
113            silk_RSHIFT( silk_SMULWB( (opus_int32)subfr_length, 655 ), corr_rshifts[ k ] - extra_shifts );    /* Q( -corr_rshifts[ k ] + extra_shifts ) */
114        denom32 = silk_max( denom32, 1 );
115        silk_assert( ((opus_int64)Wght_Q15[ k ] << 16 ) < silk_int32_MAX );                       /* Wght always < 0.5 in Q0 */
116        temp32 = silk_DIV32( silk_LSHIFT( (opus_int32)Wght_Q15[ k ], 16 ), denom32 );             /* Q( 15 + 16 + corr_rshifts[k] - extra_shifts ) */
117        temp32 = silk_RSHIFT( temp32, 31 + corr_rshifts[ k ] - extra_shifts - 26 );               /* Q26 */
118
119        /* Limit temp such that the below scaling never wraps around */
120        WLTP_max = 0;
121        for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) {
122            WLTP_max = silk_max( WLTP_ptr[ i ], WLTP_max );
123        }
124        lshift = silk_CLZ32( WLTP_max ) - 1 - 3; /* keep 3 bits free for vq_nearest_neighbor_fix */
125        silk_assert( 26 - 18 + lshift >= 0 );
126        if( 26 - 18 + lshift < 31 ) {
127            temp32 = silk_min_32( temp32, silk_LSHIFT( (opus_int32)1, 26 - 18 + lshift ) );
128        }
129
130        silk_scale_vector32_Q26_lshift_18( WLTP_ptr, temp32, LTP_ORDER * LTP_ORDER ); /* WLTP_ptr in Q( 18 - corr_rshifts[ k ] ) */
131
132        w[ k ] = matrix_ptr( WLTP_ptr, LTP_ORDER/2, LTP_ORDER/2, LTP_ORDER ); /* w in Q( 18 - corr_rshifts[ k ] ) */
133        silk_assert( w[k] >= 0 );
134
135        r_ptr     += subfr_length;
136        b_Q14_ptr += LTP_ORDER;
137        WLTP_ptr  += LTP_ORDER * LTP_ORDER;
138    }
139
140    maxRshifts = 0;
141    for( k = 0; k < nb_subfr; k++ ) {
142        maxRshifts = silk_max_int( corr_rshifts[ k ], maxRshifts );
143    }
144
145    /* Compute LTP coding gain */
146    if( LTPredCodGain_Q7 != NULL ) {
147        LPC_LTP_res_nrg = 0;
148        LPC_res_nrg     = 0;
149        silk_assert( LTP_CORRS_HEAD_ROOM >= 2 ); /* Check that no overflow will happen when adding */
150        for( k = 0; k < nb_subfr; k++ ) {
151            LPC_res_nrg     = silk_ADD32( LPC_res_nrg,     silk_RSHIFT( silk_ADD32( silk_SMULWB(  rr[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */
152            LPC_LTP_res_nrg = silk_ADD32( LPC_LTP_res_nrg, silk_RSHIFT( silk_ADD32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */
153        }
154        LPC_LTP_res_nrg = silk_max( LPC_LTP_res_nrg, 1 ); /* avoid division by zero */
155
156        div_Q16 = silk_DIV32_varQ( LPC_res_nrg, LPC_LTP_res_nrg, 16 );
157        *LTPredCodGain_Q7 = ( opus_int )silk_SMULBB( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) );
158
159        silk_assert( *LTPredCodGain_Q7 == ( opus_int )silk_SAT16( silk_MUL( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) ) ) );
160    }
161
162    /* smoothing */
163    /* d = sum( B, 1 ); */
164    b_Q14_ptr = b_Q14;
165    for( k = 0; k < nb_subfr; k++ ) {
166        d_Q14[ k ] = 0;
167        for( i = 0; i < LTP_ORDER; i++ ) {
168            d_Q14[ k ] += b_Q14_ptr[ i ];
169        }
170        b_Q14_ptr += LTP_ORDER;
171    }
172
173    /* m = ( w * d' ) / ( sum( w ) + 1e-3 ); */
174
175    /* Find maximum absolute value of d_Q14 and the bits used by w in Q0 */
176    max_abs_d_Q14 = 0;
177    max_w_bits    = 0;
178    for( k = 0; k < nb_subfr; k++ ) {
179        max_abs_d_Q14 = silk_max_32( max_abs_d_Q14, silk_abs( d_Q14[ k ] ) );
180        /* w[ k ] is in Q( 18 - corr_rshifts[ k ] ) */
181        /* Find bits needed in Q( 18 - maxRshifts ) */
182        max_w_bits = silk_max_32( max_w_bits, 32 - silk_CLZ32( w[ k ] ) + corr_rshifts[ k ] - maxRshifts );
183    }
184
185    /* max_abs_d_Q14 = (5 << 15); worst case, i.e. LTP_ORDER * -silk_int16_MIN */
186    silk_assert( max_abs_d_Q14 <= ( 5 << 15 ) );
187
188    /* How many bits is needed for w*d' in Q( 18 - maxRshifts ) in the worst case, of all d_Q14's being equal to max_abs_d_Q14 */
189    extra_shifts = max_w_bits + 32 - silk_CLZ32( max_abs_d_Q14 ) - 14;
190
191    /* Subtract what we got available; bits in output var plus maxRshifts */
192    extra_shifts -= ( 32 - 1 - 2 + maxRshifts ); /* Keep sign bit free as well as 2 bits for accumulation */
193    extra_shifts = silk_max_int( extra_shifts, 0 );
194
195    maxRshifts_wxtra = maxRshifts + extra_shifts;
196
197    temp32 = silk_RSHIFT( 262, maxRshifts + extra_shifts ) + 1; /* 1e-3f in Q( 18 - (maxRshifts + extra_shifts) ) */
198    wd = 0;
199    for( k = 0; k < nb_subfr; k++ ) {
200        /* w has at least 2 bits of headroom so no overflow should happen */
201        temp32 = silk_ADD32( temp32,                     silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ) );                      /* Q( 18 - maxRshifts_wxtra ) */
202        wd     = silk_ADD32( wd, silk_LSHIFT( silk_SMULWW( silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ), d_Q14[ k ] ), 2 ) ); /* Q( 18 - maxRshifts_wxtra ) */
203    }
204    m_Q12 = silk_DIV32_varQ( wd, temp32, 12 );
205
206    b_Q14_ptr = b_Q14;
207    for( k = 0; k < nb_subfr; k++ ) {
208        /* w_fix[ k ] from Q( 18 - corr_rshifts[ k ] ) to Q( 16 ) */
209        if( 2 - corr_rshifts[k] > 0 ) {
210            temp32 = silk_RSHIFT( w[ k ], 2 - corr_rshifts[ k ] );
211        } else {
212            temp32 = silk_LSHIFT_SAT32( w[ k ], corr_rshifts[ k ] - 2 );
213        }
214
215        g_Q26 = silk_MUL(
216            silk_DIV32(
217                SILK_FIX_CONST( LTP_SMOOTHING, 26 ),
218                silk_RSHIFT( SILK_FIX_CONST( LTP_SMOOTHING, 26 ), 10 ) + temp32 ),                          /* Q10 */
219            silk_LSHIFT_SAT32( silk_SUB_SAT32( (opus_int32)m_Q12, silk_RSHIFT( d_Q14[ k ], 2 ) ), 4 ) );    /* Q16 */
220
221        temp32 = 0;
222        for( i = 0; i < LTP_ORDER; i++ ) {
223            delta_b_Q14[ i ] = silk_max_16( b_Q14_ptr[ i ], 1638 );     /* 1638_Q14 = 0.1_Q0 */
224            temp32 += delta_b_Q14[ i ];                                 /* Q14 */
225        }
226        temp32 = silk_DIV32( g_Q26, temp32 );                           /* Q14 -> Q12 */
227        for( i = 0; i < LTP_ORDER; i++ ) {
228            b_Q14_ptr[ i ] = silk_LIMIT_32( (opus_int32)b_Q14_ptr[ i ] + silk_SMULWB( silk_LSHIFT_SAT32( temp32, 4 ), delta_b_Q14[ i ] ), -16000, 28000 );
229        }
230        b_Q14_ptr += LTP_ORDER;
231    }
232}
233
234void silk_fit_LTP(
235    opus_int32 LTP_coefs_Q16[ LTP_ORDER ],
236    opus_int16 LTP_coefs_Q14[ LTP_ORDER ]
237)
238{
239    opus_int i;
240
241    for( i = 0; i < LTP_ORDER; i++ ) {
242        LTP_coefs_Q14[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( LTP_coefs_Q16[ i ], 2 ) );
243    }
244}
245