decode_core.c (12216B)
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 #include "stack_alloc.h" 34 35 /**********************************************************/ 36 /* Core decoder. Performs inverse NSQ operation LTP + LPC */ 37 /**********************************************************/ 38 void silk_decode_core( 39 silk_decoder_state *psDec, /* I/O Decoder state */ 40 silk_decoder_control *psDecCtrl, /* I Decoder control */ 41 opus_int16 xq[], /* O Decoded speech */ 42 const opus_int16 pulses[ MAX_FRAME_LENGTH ], /* I Pulse signal */ 43 int arch /* I Run-time architecture */ 44 ) 45 { 46 opus_int i, k, lag = 0, start_idx, sLTP_buf_idx, NLSF_interpolation_flag, signalType; 47 opus_int16 *A_Q12, *B_Q14, *pxq, A_Q12_tmp[ MAX_LPC_ORDER ]; 48 VARDECL( opus_int16, sLTP ); 49 VARDECL( opus_int32, sLTP_Q15 ); 50 opus_int32 LTP_pred_Q13, LPC_pred_Q10, Gain_Q10, inv_gain_Q31, gain_adj_Q16, rand_seed, offset_Q10; 51 opus_int32 *pred_lag_ptr, *pexc_Q14, *pres_Q14; 52 VARDECL( opus_int32, res_Q14 ); 53 VARDECL( opus_int32, sLPC_Q14 ); 54 SAVE_STACK; 55 56 silk_assert( psDec->prev_gain_Q16 != 0 ); 57 58 ALLOC( sLTP, psDec->ltp_mem_length, opus_int16 ); 59 ALLOC( sLTP_Q15, psDec->ltp_mem_length + psDec->frame_length, opus_int32 ); 60 ALLOC( res_Q14, psDec->subfr_length, opus_int32 ); 61 /* Work around a clang bug (verified with clang 6.0 through clang 20.1.0) that causes the last 62 memset to be flagged as an invalid read by valgrind (not caught by asan). */ 63 #if defined(__clang__) && defined(VAR_ARRAYS) 64 ALLOC( sLPC_Q14, MAX_SUB_FRAME_LENGTH + MAX_LPC_ORDER, opus_int32 ); 65 #else 66 ALLOC( sLPC_Q14, psDec->subfr_length + MAX_LPC_ORDER, opus_int32 ); 67 #endif 68 69 offset_Q10 = silk_Quantization_Offsets_Q10[ psDec->indices.signalType >> 1 ][ psDec->indices.quantOffsetType ]; 70 71 if( psDec->indices.NLSFInterpCoef_Q2 < 1 << 2 ) { 72 NLSF_interpolation_flag = 1; 73 } else { 74 NLSF_interpolation_flag = 0; 75 } 76 77 /* Decode excitation */ 78 rand_seed = psDec->indices.Seed; 79 for( i = 0; i < psDec->frame_length; i++ ) { 80 rand_seed = silk_RAND( rand_seed ); 81 psDec->exc_Q14[ i ] = silk_LSHIFT( (opus_int32)pulses[ i ], 14 ); 82 if( psDec->exc_Q14[ i ] > 0 ) { 83 psDec->exc_Q14[ i ] -= QUANT_LEVEL_ADJUST_Q10 << 4; 84 } else 85 if( psDec->exc_Q14[ i ] < 0 ) { 86 psDec->exc_Q14[ i ] += QUANT_LEVEL_ADJUST_Q10 << 4; 87 } 88 psDec->exc_Q14[ i ] += offset_Q10 << 4; 89 if( rand_seed < 0 ) { 90 psDec->exc_Q14[ i ] = -psDec->exc_Q14[ i ]; 91 } 92 93 rand_seed = silk_ADD32_ovflw( rand_seed, pulses[ i ] ); 94 } 95 96 /* Copy LPC state */ 97 silk_memcpy( sLPC_Q14, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) ); 98 99 pexc_Q14 = psDec->exc_Q14; 100 pxq = xq; 101 sLTP_buf_idx = psDec->ltp_mem_length; 102 /* Loop over subframes */ 103 for( k = 0; k < psDec->nb_subfr; k++ ) { 104 pres_Q14 = res_Q14; 105 A_Q12 = psDecCtrl->PredCoef_Q12[ k >> 1 ]; 106 107 /* Preload LPC coefficients to array on stack. Gives small performance gain */ 108 silk_memcpy( A_Q12_tmp, A_Q12, psDec->LPC_order * sizeof( opus_int16 ) ); 109 B_Q14 = &psDecCtrl->LTPCoef_Q14[ k * LTP_ORDER ]; 110 signalType = psDec->indices.signalType; 111 112 Gain_Q10 = silk_RSHIFT( psDecCtrl->Gains_Q16[ k ], 6 ); 113 inv_gain_Q31 = silk_INVERSE32_varQ( psDecCtrl->Gains_Q16[ k ], 47 ); 114 115 /* Calculate gain adjustment factor */ 116 if( psDecCtrl->Gains_Q16[ k ] != psDec->prev_gain_Q16 ) { 117 gain_adj_Q16 = silk_DIV32_varQ( psDec->prev_gain_Q16, psDecCtrl->Gains_Q16[ k ], 16 ); 118 119 /* Scale short term state */ 120 for( i = 0; i < MAX_LPC_ORDER; i++ ) { 121 sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, sLPC_Q14[ i ] ); 122 } 123 } else { 124 gain_adj_Q16 = (opus_int32)1 << 16; 125 } 126 127 /* Save inv_gain */ 128 silk_assert( inv_gain_Q31 != 0 ); 129 psDec->prev_gain_Q16 = psDecCtrl->Gains_Q16[ k ]; 130 131 /* Avoid abrupt transition from voiced PLC to unvoiced normal decoding */ 132 if( psDec->lossCnt && psDec->prevSignalType == TYPE_VOICED && 133 psDec->indices.signalType != TYPE_VOICED && k < MAX_NB_SUBFR/2 ) { 134 135 silk_memset( B_Q14, 0, LTP_ORDER * sizeof( opus_int16 ) ); 136 B_Q14[ LTP_ORDER/2 ] = SILK_FIX_CONST( 0.25, 14 ); 137 138 signalType = TYPE_VOICED; 139 psDecCtrl->pitchL[ k ] = psDec->lagPrev; 140 } 141 142 if( signalType == TYPE_VOICED ) { 143 /* Voiced */ 144 lag = psDecCtrl->pitchL[ k ]; 145 146 /* Re-whitening */ 147 if( k == 0 || ( k == 2 && NLSF_interpolation_flag ) ) { 148 /* Rewhiten with new A coefs */ 149 start_idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2; 150 celt_assert( start_idx > 0 ); 151 152 if( k == 2 ) { 153 silk_memcpy( &psDec->outBuf[ psDec->ltp_mem_length ], xq, 2 * psDec->subfr_length * sizeof( opus_int16 ) ); 154 } 155 156 silk_LPC_analysis_filter( &sLTP[ start_idx ], &psDec->outBuf[ start_idx + k * psDec->subfr_length ], 157 A_Q12, psDec->ltp_mem_length - start_idx, psDec->LPC_order, arch ); 158 159 /* After rewhitening the LTP state is unscaled */ 160 if( k == 0 ) { 161 /* Do LTP downscaling to reduce inter-packet dependency */ 162 inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, psDecCtrl->LTP_scale_Q14 ), 2 ); 163 } 164 for( i = 0; i < lag + LTP_ORDER/2; i++ ) { 165 sLTP_Q15[ sLTP_buf_idx - i - 1 ] = silk_SMULWB( inv_gain_Q31, sLTP[ psDec->ltp_mem_length - i - 1 ] ); 166 } 167 } else { 168 /* Update LTP state when Gain changes */ 169 if( gain_adj_Q16 != (opus_int32)1 << 16 ) { 170 for( i = 0; i < lag + LTP_ORDER/2; i++ ) { 171 sLTP_Q15[ sLTP_buf_idx - i - 1 ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ sLTP_buf_idx - i - 1 ] ); 172 } 173 } 174 } 175 } 176 177 /* Long-term prediction */ 178 if( signalType == TYPE_VOICED ) { 179 /* Set up pointer */ 180 pred_lag_ptr = &sLTP_Q15[ sLTP_buf_idx - lag + LTP_ORDER / 2 ]; 181 for( i = 0; i < psDec->subfr_length; i++ ) { 182 /* Unrolled loop */ 183 /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ 184 LTP_pred_Q13 = 2; 185 LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); 186 LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -1 ], B_Q14[ 1 ] ); 187 LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -2 ], B_Q14[ 2 ] ); 188 LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -3 ], B_Q14[ 3 ] ); 189 LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -4 ], B_Q14[ 4 ] ); 190 pred_lag_ptr++; 191 192 /* Generate LPC excitation */ 193 pres_Q14[ i ] = silk_ADD_LSHIFT32( pexc_Q14[ i ], LTP_pred_Q13, 1 ); 194 195 /* Update states */ 196 sLTP_Q15[ sLTP_buf_idx ] = silk_LSHIFT( pres_Q14[ i ], 1 ); 197 sLTP_buf_idx++; 198 } 199 } else { 200 pres_Q14 = pexc_Q14; 201 } 202 203 for( i = 0; i < psDec->subfr_length; i++ ) { 204 /* Short-term prediction */ 205 celt_assert( psDec->LPC_order == 10 || psDec->LPC_order == 16 ); 206 /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ 207 LPC_pred_Q10 = silk_RSHIFT( psDec->LPC_order, 1 ); 208 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 1 ], A_Q12_tmp[ 0 ] ); 209 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 2 ], A_Q12_tmp[ 1 ] ); 210 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 3 ], A_Q12_tmp[ 2 ] ); 211 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 4 ], A_Q12_tmp[ 3 ] ); 212 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 5 ], A_Q12_tmp[ 4 ] ); 213 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 6 ], A_Q12_tmp[ 5 ] ); 214 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 7 ], A_Q12_tmp[ 6 ] ); 215 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 8 ], A_Q12_tmp[ 7 ] ); 216 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 9 ], A_Q12_tmp[ 8 ] ); 217 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 10 ], A_Q12_tmp[ 9 ] ); 218 if( psDec->LPC_order == 16 ) { 219 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 11 ], A_Q12_tmp[ 10 ] ); 220 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 12 ], A_Q12_tmp[ 11 ] ); 221 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 13 ], A_Q12_tmp[ 12 ] ); 222 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 14 ], A_Q12_tmp[ 13 ] ); 223 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 15 ], A_Q12_tmp[ 14 ] ); 224 LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 16 ], A_Q12_tmp[ 15 ] ); 225 } 226 227 /* Add prediction to LPC excitation */ 228 sLPC_Q14[ MAX_LPC_ORDER + i ] = silk_ADD_SAT32( pres_Q14[ i ], silk_LSHIFT_SAT32( LPC_pred_Q10, 4 ) ); 229 230 /* Scale with gain */ 231 pxq[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14[ MAX_LPC_ORDER + i ], Gain_Q10 ), 8 ) ); 232 } 233 234 /* Update LPC filter state */ 235 silk_memcpy( sLPC_Q14, &sLPC_Q14[ psDec->subfr_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); 236 pexc_Q14 += psDec->subfr_length; 237 pxq += psDec->subfr_length; 238 } 239 240 /* Save LPC state */ 241 silk_memcpy( psDec->sLPC_Q14_buf, sLPC_Q14, MAX_LPC_ORDER * sizeof( opus_int32 ) ); 242 RESTORE_STACK; 243 }