simplified computation of LTP coefs
[opus.git] / silk / VQ_WMat_EC.c
1 /***********************************************************************
2 Copyright (c) 2006-2011, Skype Limited. All rights reserved.
3 Redistribution and use in source and binary forms, with or without
4 modification, are permitted provided that the following conditions
5 are met:
6 - Redistributions of source code must retain the above copyright notice,
7 this list of conditions and the following disclaimer.
8 - Redistributions in binary form must reproduce the above copyright
9 notice, this list of conditions and the following disclaimer in the
10 documentation and/or other materials provided with the distribution.
11 - Neither the name of Internet Society, IETF or IETF Trust, nor the
12 names of specific contributors, may be used to endorse or promote
13 products derived from this software without specific prior written
14 permission.
15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21 SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25 POSSIBILITY OF SUCH DAMAGE.
26 ***********************************************************************/
27
28 #ifdef HAVE_CONFIG_H
29 #include "config.h"
30 #endif
31
32 #include "main.h"
33
34 /* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */
35 void silk_VQ_WMat_EC_c(
36     opus_int8                   *ind,                           /* O    index of best codebook vector               */
37     opus_int32                  *res_nrg_Q15,                                   /* O    best residual energy                                            */
38     opus_int32                  *rate_dist_Q8,                  /* O    best total bitrate                                                      */
39     const opus_int32            *XX_Q17,                                                /* I    correlation matrix                          */
40     const opus_int32            *xX_Q17,                                                /* I    correlation vector                                                      */
41     const opus_int8             *cb_Q7,                         /* I    codebook                                    */
42     const opus_uint8            *cl_Q5,                         /* I    code length for each codebook vector        */
43     const opus_int              subfr_len,                                              /* I    number of samples per subframe                          */
44     const opus_int              L                               /* I    number of vectors in codebook               */
45 )
46 {
47     opus_int   k;
48     const opus_int8 *cb_row_Q7;
49         opus_int32 neg_xX_Q24[ 5 ];
50     opus_int32 sum1_Q15, sum2_Q24, sum1_best_Q15;
51         opus_int32 bits_res_Q8, bits_tot_Q8;
52
53         /* Negate and convert to new Q domain */
54         neg_xX_Q24[ 0 ] = -silk_LSHIFT32( xX_Q17[ 0 ], 7 );
55         neg_xX_Q24[ 1 ] = -silk_LSHIFT32( xX_Q17[ 1 ], 7 );
56         neg_xX_Q24[ 2 ] = -silk_LSHIFT32( xX_Q17[ 2 ], 7 );
57         neg_xX_Q24[ 3 ] = -silk_LSHIFT32( xX_Q17[ 3 ], 7 );
58         neg_xX_Q24[ 4 ] = -silk_LSHIFT32( xX_Q17[ 4 ], 7 );
59
60     /* Loop over codebook */
61     *rate_dist_Q8 = silk_int32_MAX;
62         *res_nrg_Q15 = silk_int32_MAX;
63         sum1_best_Q15 = silk_int32_MAX;
64     cb_row_Q7 = cb_Q7;
65     for( k = 0; k < L; k++ ) {
66         /* Weighted rate */
67                 /* Quantization error: 1 - 2* xX * cb + cb' * XX * cb */
68                 sum1_Q15 = SILK_FIX_CONST( 1.0001, 15 );
69
70         /* first row of XX_Q17 */
71         sum2_Q24 = silk_MLA( neg_xX_Q24[ 0 ], XX_Q17[  1 ], cb_row_Q7[ 1 ] );
72         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[  2 ], cb_row_Q7[ 2 ] );
73         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[  3 ], cb_row_Q7[ 3 ] );
74         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[  4 ], cb_row_Q7[ 4 ] );
75         sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
76         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[  0 ], cb_row_Q7[ 0 ] );
77         sum1_Q15 = silk_SMLAWB( sum1_Q15,        sum2_Q24,  cb_row_Q7[ 0 ] );
78
79         /* second row of XX_Q17 */
80         sum2_Q24 = silk_MLA( neg_xX_Q24[ 1 ], XX_Q17[  7 ], cb_row_Q7[ 2 ] );
81         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[  8 ], cb_row_Q7[ 3 ] );
82         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[  9 ], cb_row_Q7[ 4 ] );
83         sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
84         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[  6 ], cb_row_Q7[ 1 ] );
85         sum1_Q15 = silk_SMLAWB( sum1_Q15,        sum2_Q24,  cb_row_Q7[ 1 ] );
86
87         /* third row of XX_Q17 */
88         sum2_Q24 = silk_MLA( neg_xX_Q24[ 2 ], XX_Q17[ 13 ], cb_row_Q7[ 3 ] );
89         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[ 14 ], cb_row_Q7[ 4 ] );
90         sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
91         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[ 12 ], cb_row_Q7[ 2 ] );
92         sum1_Q15 = silk_SMLAWB( sum1_Q15,        sum2_Q24,  cb_row_Q7[ 2 ] );
93
94         /* fourth row of XX_Q17 */
95         sum2_Q24 = silk_MLA( neg_xX_Q24[ 3 ], XX_Q17[ 19 ], cb_row_Q7[ 4 ] );
96         sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
97         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[ 18 ], cb_row_Q7[ 3 ] );
98         sum1_Q15 = silk_SMLAWB( sum1_Q15,        sum2_Q24,  cb_row_Q7[ 3 ] );
99
100         /* last row of XX_Q17 */
101                 sum2_Q24 = silk_LSHIFT32( neg_xX_Q24[ 4 ], 1 );
102         sum2_Q24 = silk_MLA( sum2_Q24,        XX_Q17[ 24 ], cb_row_Q7[ 4 ] );
103         sum1_Q15 = silk_SMLAWB( sum1_Q15,        sum2_Q24,  cb_row_Q7[ 4 ] );
104
105                 /* If ever the following assert triggers, increase LTP_CORR_INV_MAX */
106         silk_assert( sum1_Q15 >= 0 );
107
108                 /* find best */
109                 if( sum1_Q15 <= sum1_best_Q15 ) {
110                         sum1_best_Q15 = sum1_Q15;
111                         /* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */
112                         bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 ) - (15 << 7) );
113                         bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 2 );
114                         if( bits_tot_Q8 <= *rate_dist_Q8 ) {
115                                 *rate_dist_Q8 = bits_tot_Q8;
116                                 *res_nrg_Q15 = sum1_Q15;
117                                 *ind = (opus_int8)k;
118                         }
119                 }
120
121         /* Go to next cbk vector */
122         cb_row_Q7 += LTP_ORDER;
123     }
124 }