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 "SigProc_FIX.h" 33 #include "define.h" 34 #include "tuning_parameters.h" 35 #include "pitch.h" 36 37 #define MAX_FRAME_SIZE 384 /* subfr_length * nb_subfr = ( 0.005 * 16000 + 16 ) * 4 = 384 */ 38 39 #define QA 25 40 #define N_BITS_HEAD_ROOM 2 41 #define MIN_RSHIFTS -16 42 #define MAX_RSHIFTS (32 - QA) 43 44 /* Compute reflection coefficients from input signal */ 45 void silk_burg_modified_c( 46 opus_int32 *res_nrg, /* O Residual energy */ 47 opus_int *res_nrg_Q, /* O Residual energy Q value */ 48 opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ 49 const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */ 50 const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ 51 const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ 52 const opus_int nb_subfr, /* I Number of subframes stacked in x */ 53 const opus_int D, /* I Order */ 54 int arch /* I Run-time architecture */ 55 ) 56 { 57 opus_int k, n, s, lz, rshifts, reached_max_gain; 58 opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; 59 const opus_int16 *x_ptr; 60 opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; 61 opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; 62 opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; 63 opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; 64 opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; 65 opus_int32 xcorr[ SILK_MAX_ORDER_LPC ]; 66 opus_int64 C0_64; 67 68 silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); 69 70 /* Compute autocorrelations, added over subframes */ 71 C0_64 = silk_inner_prod16_aligned_64( x, x, subfr_length*nb_subfr, arch ); 72 lz = silk_CLZ64(C0_64); 73 rshifts = 32 + 1 + N_BITS_HEAD_ROOM - lz; 74 if (rshifts > MAX_RSHIFTS) rshifts = MAX_RSHIFTS; 75 if (rshifts < MIN_RSHIFTS) rshifts = MIN_RSHIFTS; 76 77 if (rshifts > 0) { 78 C0 = (opus_int32)silk_RSHIFT64(C0_64, rshifts ); 79 } else { 80 C0 = silk_LSHIFT32((opus_int32)C0_64, -rshifts ); 81 } 82 83 CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ 84 silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); 85 if( rshifts > 0 ) { 86 for( s = 0; s < nb_subfr; s++ ) { 87 x_ptr = x + s * subfr_length; 88 for( n = 1; n < D + 1; n++ ) { 89 C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( 90 silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n, arch ), rshifts ); 91 } 92 } 93 } else { 94 for( s = 0; s < nb_subfr; s++ ) { 95 int i; 96 opus_int32 d; 97 x_ptr = x + s * subfr_length; 98 celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch ); 99 for( n = 1; n < D + 1; n++ ) { 100 for ( i = n + subfr_length - D, d = 0; i < subfr_length; i++ ) 101 d = MAC16_16( d, x_ptr[ i ], x_ptr[ i - n ] ); 102 xcorr[ n - 1 ] += d; 103 } 104 for( n = 1; n < D + 1; n++ ) { 105 C_first_row[ n - 1 ] += silk_LSHIFT32( xcorr[ n - 1 ], -rshifts ); 106 } 107 } 108 } 109 silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); 110 111 /* Initialize */ 112 CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ 113 114 invGain_Q30 = (opus_int32)1 << 30; 115 reached_max_gain = 0; 116 for( n = 0; n < D; n++ ) { 117 /* Update first row of correlation matrix (without first element) */ 118 /* Update last row of correlation matrix (without last element, stored in reversed order) */ 119 /* Update C * Af */ 120 /* Update C * flipud(Af) (stored in reversed order) */ 121 if( rshifts > -2 ) { 122 for( s = 0; s < nb_subfr; s++ ) { 123 x_ptr = x + s * subfr_length; 124 x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts) */ 125 x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts) */ 126 tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16) */ 127 tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16) */ 128 for( k = 0; k < n; k++ ) { 129 C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ 130 C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ 131 Atmp_QA = Af_QA[ k ]; 132 tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ 133 tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16) */ 134 } 135 tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts) */ 136 tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts) */ 137 for( k = 0; k <= n; k++ ) { 138 CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift ) */ 139 CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift ) */ 140 } 141 } 142 } else { 143 for( s = 0; s < nb_subfr; s++ ) { 144 x_ptr = x + s * subfr_length; 145 x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts ) */ 146 x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts ) */ 147 tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17 */ 148 tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17 */ 149 for( k = 0; k < n; k++ ) { 150 C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ 151 C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ 152 Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */ 153 /* We sometimes have get overflows in the multiplications (even beyond +/- 2^32), 154 but they cancel each other and the real result seems to always fit in a 32-bit 155 signed integer. This was determined experimentally, not theoretically (unfortunately). */ 156 tmp1 = silk_MLA_ovflw( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */ 157 tmp2 = silk_MLA_ovflw( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */ 158 } 159 tmp1 = -tmp1; /* Q17 */ 160 tmp2 = -tmp2; /* Q17 */ 161 for( k = 0; k <= n; k++ ) { 162 CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1, 163 silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift ) */ 164 CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2, 165 silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */ 166 } 167 } 168 } 169 170 /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ 171 tmp1 = C_first_row[ n ]; /* Q( -rshifts ) */ 172 tmp2 = C_last_row[ n ]; /* Q( -rshifts ) */ 173 num = 0; /* Q( -rshifts ) */ 174 nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts ) */ 175 for( k = 0; k < n; k++ ) { 176 Atmp_QA = Af_QA[ k ]; 177 lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1; 178 lz = silk_min( 32 - QA, lz ); 179 Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz ) */ 180 181 tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ 182 tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ 183 num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ 184 nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ), 185 Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts ) */ 186 } 187 CAf[ n + 1 ] = tmp1; /* Q( -rshifts ) */ 188 CAb[ n + 1 ] = tmp2; /* Q( -rshifts ) */ 189 num = silk_ADD32( num, tmp2 ); /* Q( -rshifts ) */ 190 num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts ) */ 191 192 /* Calculate the next order reflection (parcor) coefficient */ 193 if( silk_abs( num ) < nrg ) { 194 rc_Q31 = silk_DIV32_varQ( num, nrg, 31 ); 195 } else { 196 rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN; 197 } 198 199 /* Update inverse prediction gain */ 200 tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); 201 tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 ); 202 if( tmp1 <= minInvGain_Q30 ) { 203 /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ 204 tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */ 205 rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */ 206 if( rc_Q31 > 0 ) { 207 /* Newton-Raphson iteration */ 208 rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */ 209 rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */ 210 if( num < 0 ) { 211 /* Ensure adjusted reflection coefficients has the original sign */ 212 rc_Q31 = -rc_Q31; 213 } 214 } 215 invGain_Q30 = minInvGain_Q30; 216 reached_max_gain = 1; 217 } else { 218 invGain_Q30 = tmp1; 219 } 220 221 /* Update the AR coefficients */ 222 for( k = 0; k < (n + 1) >> 1; k++ ) { 223 tmp1 = Af_QA[ k ]; /* QA */ 224 tmp2 = Af_QA[ n - k - 1 ]; /* QA */ 225 Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA */ 226 Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA */ 227 } 228 Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA */ 229 230 if( reached_max_gain ) { 231 /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ 232 for( k = n + 1; k < D; k++ ) { 233 Af_QA[ k ] = 0; 234 } 235 break; 236 } 237 238 /* Update C * Af and C * Ab */ 239 for( k = 0; k <= n + 1; k++ ) { 240 tmp1 = CAf[ k ]; /* Q( -rshifts ) */ 241 tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts ) */ 242 CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts ) */ 243 CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts ) */ 244 } 245 } 246 247 if( reached_max_gain ) { 248 for( k = 0; k < D; k++ ) { 249 /* Scale coefficients */ 250 A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); 251 } 252 /* Subtract energy of preceding samples from C0 */ 253 if( rshifts > 0 ) { 254 for( s = 0; s < nb_subfr; s++ ) { 255 x_ptr = x + s * subfr_length; 256 C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D, arch ), rshifts ); 257 } 258 } else { 259 for( s = 0; s < nb_subfr; s++ ) { 260 x_ptr = x + s * subfr_length; 261 C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D, arch), -rshifts); 262 } 263 } 264 /* Approximate residual energy */ 265 *res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 ); 266 *res_nrg_Q = -rshifts; 267 } else { 268 /* Return residual energy */ 269 nrg = CAf[ 0 ]; /* Q( -rshifts ) */ 270 tmp1 = (opus_int32)1 << 16; /* Q16 */ 271 for( k = 0; k < D; k++ ) { 272 Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16 */ 273 nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts ) */ 274 tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16 */ 275 A_Q16[ k ] = -Atmp1; 276 } 277 *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ), -tmp1 );/* Q( -rshifts ) */ 278 *res_nrg_Q = -rshifts; 279 } 280 } 281
