OLD | NEW |
1 /*********************************************************************** | 1 /*********************************************************************** |
2 Copyright (c) 2006-2011, Skype Limited. All rights reserved. | 2 Copyright (c) 2006-2011, Skype Limited. All rights reserved. |
3 Redistribution and use in source and binary forms, with or without | 3 Redistribution and use in source and binary forms, with or without |
4 modification, are permitted provided that the following conditions | 4 modification, are permitted provided that the following conditions |
5 are met: | 5 are met: |
6 - Redistributions of source code must retain the above copyright notice, | 6 - Redistributions of source code must retain the above copyright notice, |
7 this list of conditions and the following disclaimer. | 7 this list of conditions and the following disclaimer. |
8 - Redistributions in binary form must reproduce the above copyright | 8 - Redistributions in binary form must reproduce the above copyright |
9 notice, this list of conditions and the following disclaimer in the | 9 notice, this list of conditions and the following disclaimer in the |
10 documentation and/or other materials provided with the distribution. | 10 documentation and/or other materials provided with the distribution. |
(...skipping 16 matching lines...) Expand all Loading... |
27 | 27 |
28 #ifdef HAVE_CONFIG_H | 28 #ifdef HAVE_CONFIG_H |
29 #include "config.h" | 29 #include "config.h" |
30 #endif | 30 #endif |
31 | 31 |
32 #include "main.h" | 32 #include "main.h" |
33 | 33 |
34 /* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for
a single input data vector */ | 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( | 35 void silk_VQ_WMat_EC_c( |
36 opus_int8 *ind, /* O index of
best codebook vector */ | 36 opus_int8 *ind, /* O index of
best codebook vector */ |
37 opus_int32 *rate_dist_Q14, /* O best wei
ghted quant error + mu * rate */ | 37 opus_int32 *res_nrg_Q15, /* O best res
idual energy */ |
| 38 opus_int32 *rate_dist_Q8, /* O best tot
al bitrate */ |
38 opus_int *gain_Q7, /* O sum of a
bsolute LTP coefficients */ | 39 opus_int *gain_Q7, /* O sum of a
bsolute LTP coefficients */ |
39 const opus_int16 *in_Q14, /* I input ve
ctor to be quantized */ | 40 const opus_int32 *XX_Q17, /* I correlat
ion matrix */ |
40 const opus_int32 *W_Q18, /* I weightin
g matrix */ | 41 const opus_int32 *xX_Q17, /* I correlat
ion vector */ |
41 const opus_int8 *cb_Q7, /* I codebook
*/ | 42 const opus_int8 *cb_Q7, /* I codebook
*/ |
42 const opus_uint8 *cb_gain_Q7, /* I codebook
effective gain */ | 43 const opus_uint8 *cb_gain_Q7, /* I codebook
effective gain */ |
43 const opus_uint8 *cl_Q5, /* I code len
gth for each codebook vector */ | 44 const opus_uint8 *cl_Q5, /* I code len
gth for each codebook vector */ |
44 const opus_int mu_Q9, /* I tradeoff
betw. weighted error and rate */ | 45 const opus_int subfr_len, /* I number o
f samples per subframe */ |
45 const opus_int32 max_gain_Q7, /* I maximum
sum of absolute LTP coefficients */ | 46 const opus_int32 max_gain_Q7, /* I maximum
sum of absolute LTP coefficients */ |
46 opus_int L /* I number o
f vectors in codebook */ | 47 const opus_int L /* I number o
f vectors in codebook */ |
47 ) | 48 ) |
48 { | 49 { |
49 opus_int k, gain_tmp_Q7; | 50 opus_int k, gain_tmp_Q7; |
50 const opus_int8 *cb_row_Q7; | 51 const opus_int8 *cb_row_Q7; |
51 opus_int16 diff_Q14[ 5 ]; | 52 opus_int32 neg_xX_Q24[ 5 ]; |
52 opus_int32 sum1_Q14, sum2_Q16; | 53 opus_int32 sum1_Q15, sum2_Q24; |
| 54 opus_int32 bits_res_Q8, bits_tot_Q8; |
| 55 |
| 56 /* Negate and convert to new Q domain */ |
| 57 neg_xX_Q24[ 0 ] = -silk_LSHIFT32( xX_Q17[ 0 ], 7 ); |
| 58 neg_xX_Q24[ 1 ] = -silk_LSHIFT32( xX_Q17[ 1 ], 7 ); |
| 59 neg_xX_Q24[ 2 ] = -silk_LSHIFT32( xX_Q17[ 2 ], 7 ); |
| 60 neg_xX_Q24[ 3 ] = -silk_LSHIFT32( xX_Q17[ 3 ], 7 ); |
| 61 neg_xX_Q24[ 4 ] = -silk_LSHIFT32( xX_Q17[ 4 ], 7 ); |
53 | 62 |
54 /* Loop over codebook */ | 63 /* Loop over codebook */ |
55 *rate_dist_Q14 = silk_int32_MAX; | 64 *rate_dist_Q8 = silk_int32_MAX; |
| 65 *res_nrg_Q15 = silk_int32_MAX; |
56 cb_row_Q7 = cb_Q7; | 66 cb_row_Q7 = cb_Q7; |
| 67 /* In things go really bad, at least *ind is set to something safe. */ |
| 68 *ind = 0; |
57 for( k = 0; k < L; k++ ) { | 69 for( k = 0; k < L; k++ ) { |
| 70 opus_int32 penalty; |
58 gain_tmp_Q7 = cb_gain_Q7[k]; | 71 gain_tmp_Q7 = cb_gain_Q7[k]; |
59 | |
60 diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 ); | |
61 diff_Q14[ 1 ] = in_Q14[ 1 ] - silk_LSHIFT( cb_row_Q7[ 1 ], 7 ); | |
62 diff_Q14[ 2 ] = in_Q14[ 2 ] - silk_LSHIFT( cb_row_Q7[ 2 ], 7 ); | |
63 diff_Q14[ 3 ] = in_Q14[ 3 ] - silk_LSHIFT( cb_row_Q7[ 3 ], 7 ); | |
64 diff_Q14[ 4 ] = in_Q14[ 4 ] - silk_LSHIFT( cb_row_Q7[ 4 ], 7 ); | |
65 | |
66 /* Weighted rate */ | 72 /* Weighted rate */ |
67 sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] ); | 73 /* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */ |
| 74 sum1_Q15 = SILK_FIX_CONST( 1.001, 15 ); |
68 | 75 |
69 /* Penalty for too large gain */ | 76 /* Penalty for too large gain */ |
70 sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q
7, max_gain_Q7 ), 0 ), 10 ); | 77 penalty = silk_LSHIFT32( silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7
), 0 ), 11 ); |
71 | 78 |
72 silk_assert( sum1_Q14 >= 0 ); | 79 /* first row of XX_Q17 */ |
| 80 sum2_Q24 = silk_MLA( neg_xX_Q24[ 0 ], XX_Q17[ 1 ], cb_row_Q7[ 1 ] ); |
| 81 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 2 ], cb_row_Q7[ 2 ] ); |
| 82 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 3 ], cb_row_Q7[ 3 ] ); |
| 83 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 4 ], cb_row_Q7[ 4 ] ); |
| 84 sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); |
| 85 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 0 ], cb_row_Q7[ 0 ] ); |
| 86 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 0 ] ); |
73 | 87 |
74 /* first row of W_Q18 */ | 88 /* second row of XX_Q17 */ |
75 sum2_Q16 = silk_SMULWB( W_Q18[ 1 ], diff_Q14[ 1 ] ); | 89 sum2_Q24 = silk_MLA( neg_xX_Q24[ 1 ], XX_Q17[ 7 ], cb_row_Q7[ 2 ] ); |
76 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 2 ], diff_Q14[ 2 ] ); | 90 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 8 ], cb_row_Q7[ 3 ] ); |
77 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 3 ], diff_Q14[ 3 ] ); | 91 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 9 ], cb_row_Q7[ 4 ] ); |
78 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 4 ], diff_Q14[ 4 ] ); | 92 sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); |
79 sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); | 93 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 6 ], cb_row_Q7[ 1 ] ); |
80 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 0 ], diff_Q14[ 0 ] ); | 94 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 1 ] ); |
81 sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 0 ] ); | |
82 | 95 |
83 /* second row of W_Q18 */ | 96 /* third row of XX_Q17 */ |
84 sum2_Q16 = silk_SMULWB( W_Q18[ 7 ], diff_Q14[ 2 ] ); | 97 sum2_Q24 = silk_MLA( neg_xX_Q24[ 2 ], XX_Q17[ 13 ], cb_row_Q7[ 3 ] ); |
85 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 8 ], diff_Q14[ 3 ] ); | 98 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 14 ], cb_row_Q7[ 4 ] ); |
86 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 9 ], diff_Q14[ 4 ] ); | 99 sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); |
87 sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); | 100 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 12 ], cb_row_Q7[ 2 ] ); |
88 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 6 ], diff_Q14[ 1 ] ); | 101 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 2 ] ); |
89 sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 1 ] ); | |
90 | 102 |
91 /* third row of W_Q18 */ | 103 /* fourth row of XX_Q17 */ |
92 sum2_Q16 = silk_SMULWB( W_Q18[ 13 ], diff_Q14[ 3 ] ); | 104 sum2_Q24 = silk_MLA( neg_xX_Q24[ 3 ], XX_Q17[ 19 ], cb_row_Q7[ 4 ] ); |
93 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 14 ], diff_Q14[ 4 ] ); | 105 sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); |
94 sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); | 106 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 18 ], cb_row_Q7[ 3 ] ); |
95 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 12 ], diff_Q14[ 2 ] ); | 107 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 3 ] ); |
96 sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 2 ] ); | |
97 | 108 |
98 /* fourth row of W_Q18 */ | 109 /* last row of XX_Q17 */ |
99 sum2_Q16 = silk_SMULWB( W_Q18[ 19 ], diff_Q14[ 4 ] ); | 110 sum2_Q24 = silk_LSHIFT32( neg_xX_Q24[ 4 ], 1 ); |
100 sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); | 111 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 24 ], cb_row_Q7[ 4 ] ); |
101 sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 18 ], diff_Q14[ 3 ] ); | 112 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 4 ] ); |
102 sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 3 ] ); | |
103 | |
104 /* last row of W_Q18 */ | |
105 sum2_Q16 = silk_SMULWB( W_Q18[ 24 ], diff_Q14[ 4 ] ); | |
106 sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 4 ] ); | |
107 | |
108 silk_assert( sum1_Q14 >= 0 ); | |
109 | 113 |
110 /* find best */ | 114 /* find best */ |
111 if( sum1_Q14 < *rate_dist_Q14 ) { | 115 if( sum1_Q15 >= 0 ) { |
112 *rate_dist_Q14 = sum1_Q14; | 116 /* Translate residual energy to bits using high-rate assumption (6 d
B ==> 1 bit/sample) */ |
113 *ind = (opus_int8)k; | 117 bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 + penal
ty) - (15 << 7) ); |
114 *gain_Q7 = gain_tmp_Q7; | 118 /* In the following line we reduce the codelength component by half
("-1"); seems to slghtly improve quality */ |
| 119 bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 ); |
| 120 if( bits_tot_Q8 <= *rate_dist_Q8 ) { |
| 121 *rate_dist_Q8 = bits_tot_Q8; |
| 122 *res_nrg_Q15 = sum1_Q15 + penalty; |
| 123 *ind = (opus_int8)k; |
| 124 *gain_Q7 = gain_tmp_Q7; |
| 125 } |
115 } | 126 } |
116 | 127 |
117 /* Go to next cbk vector */ | 128 /* Go to next cbk vector */ |
118 cb_row_Q7 += LTP_ORDER; | 129 cb_row_Q7 += LTP_ORDER; |
119 } | 130 } |
120 } | 131 } |
OLD | NEW |