Chris@69: /*********************************************************************** Chris@69: Copyright (c) 2006-2011, Skype Limited. All rights reserved. Chris@69: Redistribution and use in source and binary forms, with or without Chris@69: modification, are permitted provided that the following conditions Chris@69: are met: Chris@69: - Redistributions of source code must retain the above copyright notice, Chris@69: this list of conditions and the following disclaimer. Chris@69: - Redistributions in binary form must reproduce the above copyright Chris@69: notice, this list of conditions and the following disclaimer in the Chris@69: documentation and/or other materials provided with the distribution. Chris@69: - Neither the name of Internet Society, IETF or IETF Trust, nor the Chris@69: names of specific contributors, may be used to endorse or promote Chris@69: products derived from this software without specific prior written Chris@69: permission. Chris@69: THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" Chris@69: AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE Chris@69: IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE Chris@69: ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE Chris@69: LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR Chris@69: CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF Chris@69: SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS Chris@69: INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN Chris@69: CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) Chris@69: ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE Chris@69: POSSIBILITY OF SUCH DAMAGE. Chris@69: ***********************************************************************/ Chris@69: Chris@69: #ifdef HAVE_CONFIG_H Chris@69: #include "config.h" Chris@69: #endif Chris@69: Chris@69: #include "SigProc_FIX.h" Chris@69: #include "define.h" Chris@69: #include "tuning_parameters.h" Chris@69: #include "pitch.h" Chris@69: Chris@69: #define MAX_FRAME_SIZE 384 /* subfr_length * nb_subfr = ( 0.005 * 16000 + 16 ) * 4 = 384 */ Chris@69: Chris@69: #define QA 25 Chris@69: #define N_BITS_HEAD_ROOM 3 Chris@69: #define MIN_RSHIFTS -16 Chris@69: #define MAX_RSHIFTS (32 - QA) Chris@69: Chris@69: /* Compute reflection coefficients from input signal */ Chris@69: void silk_burg_modified_c( Chris@69: opus_int32 *res_nrg, /* O Residual energy */ Chris@69: opus_int *res_nrg_Q, /* O Residual energy Q value */ Chris@69: opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ Chris@69: const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */ Chris@69: const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ Chris@69: const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ Chris@69: const opus_int nb_subfr, /* I Number of subframes stacked in x */ Chris@69: const opus_int D, /* I Order */ Chris@69: int arch /* I Run-time architecture */ Chris@69: ) Chris@69: { Chris@69: opus_int k, n, s, lz, rshifts, reached_max_gain; Chris@69: opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; Chris@69: const opus_int16 *x_ptr; Chris@69: opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; Chris@69: opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; Chris@69: opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; Chris@69: opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; Chris@69: opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; Chris@69: opus_int32 xcorr[ SILK_MAX_ORDER_LPC ]; Chris@69: opus_int64 C0_64; Chris@69: Chris@69: celt_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); Chris@69: Chris@69: /* Compute autocorrelations, added over subframes */ Chris@69: C0_64 = silk_inner_prod16_aligned_64( x, x, subfr_length*nb_subfr, arch ); Chris@69: lz = silk_CLZ64(C0_64); Chris@69: rshifts = 32 + 1 + N_BITS_HEAD_ROOM - lz; Chris@69: if (rshifts > MAX_RSHIFTS) rshifts = MAX_RSHIFTS; Chris@69: if (rshifts < MIN_RSHIFTS) rshifts = MIN_RSHIFTS; Chris@69: Chris@69: if (rshifts > 0) { Chris@69: C0 = (opus_int32)silk_RSHIFT64(C0_64, rshifts ); Chris@69: } else { Chris@69: C0 = silk_LSHIFT32((opus_int32)C0_64, -rshifts ); Chris@69: } Chris@69: Chris@69: CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ Chris@69: silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); Chris@69: if( rshifts > 0 ) { Chris@69: for( s = 0; s < nb_subfr; s++ ) { Chris@69: x_ptr = x + s * subfr_length; Chris@69: for( n = 1; n < D + 1; n++ ) { Chris@69: C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( Chris@69: silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n, arch ), rshifts ); Chris@69: } Chris@69: } Chris@69: } else { Chris@69: for( s = 0; s < nb_subfr; s++ ) { Chris@69: int i; Chris@69: opus_int32 d; Chris@69: x_ptr = x + s * subfr_length; Chris@69: celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch ); Chris@69: for( n = 1; n < D + 1; n++ ) { Chris@69: for ( i = n + subfr_length - D, d = 0; i < subfr_length; i++ ) Chris@69: d = MAC16_16( d, x_ptr[ i ], x_ptr[ i - n ] ); Chris@69: xcorr[ n - 1 ] += d; Chris@69: } Chris@69: for( n = 1; n < D + 1; n++ ) { Chris@69: C_first_row[ n - 1 ] += silk_LSHIFT32( xcorr[ n - 1 ], -rshifts ); Chris@69: } Chris@69: } Chris@69: } Chris@69: silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); Chris@69: Chris@69: /* Initialize */ Chris@69: CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ Chris@69: Chris@69: invGain_Q30 = (opus_int32)1 << 30; Chris@69: reached_max_gain = 0; Chris@69: for( n = 0; n < D; n++ ) { Chris@69: /* Update first row of correlation matrix (without first element) */ Chris@69: /* Update last row of correlation matrix (without last element, stored in reversed order) */ Chris@69: /* Update C * Af */ Chris@69: /* Update C * flipud(Af) (stored in reversed order) */ Chris@69: if( rshifts > -2 ) { Chris@69: for( s = 0; s < nb_subfr; s++ ) { Chris@69: x_ptr = x + s * subfr_length; Chris@69: x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts) */ Chris@69: x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts) */ Chris@69: tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16) */ Chris@69: tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16) */ Chris@69: for( k = 0; k < n; k++ ) { Chris@69: C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ Chris@69: C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ Chris@69: Atmp_QA = Af_QA[ k ]; Chris@69: tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ Chris@69: tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16) */ Chris@69: } Chris@69: tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts) */ Chris@69: tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts) */ Chris@69: for( k = 0; k <= n; k++ ) { Chris@69: CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift ) */ Chris@69: CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift ) */ Chris@69: } Chris@69: } Chris@69: } else { Chris@69: for( s = 0; s < nb_subfr; s++ ) { Chris@69: x_ptr = x + s * subfr_length; Chris@69: x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts ) */ Chris@69: x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts ) */ Chris@69: tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17 */ Chris@69: tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17 */ Chris@69: for( k = 0; k < n; k++ ) { Chris@69: C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ Chris@69: C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ Chris@69: Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */ Chris@69: /* We sometimes have get overflows in the multiplications (even beyond +/- 2^32), Chris@69: but they cancel each other and the real result seems to always fit in a 32-bit Chris@69: signed integer. This was determined experimentally, not theoretically (unfortunately). */ Chris@69: tmp1 = silk_MLA_ovflw( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */ Chris@69: tmp2 = silk_MLA_ovflw( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */ Chris@69: } Chris@69: tmp1 = -tmp1; /* Q17 */ Chris@69: tmp2 = -tmp2; /* Q17 */ Chris@69: for( k = 0; k <= n; k++ ) { Chris@69: CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1, Chris@69: silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift ) */ Chris@69: CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2, Chris@69: silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */ Chris@69: } Chris@69: } Chris@69: } Chris@69: Chris@69: /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ Chris@69: tmp1 = C_first_row[ n ]; /* Q( -rshifts ) */ Chris@69: tmp2 = C_last_row[ n ]; /* Q( -rshifts ) */ Chris@69: num = 0; /* Q( -rshifts ) */ Chris@69: nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts ) */ Chris@69: for( k = 0; k < n; k++ ) { Chris@69: Atmp_QA = Af_QA[ k ]; Chris@69: lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1; Chris@69: lz = silk_min( 32 - QA, lz ); Chris@69: Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz ) */ Chris@69: Chris@69: tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ Chris@69: tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ Chris@69: num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ Chris@69: nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ), Chris@69: Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts ) */ Chris@69: } Chris@69: CAf[ n + 1 ] = tmp1; /* Q( -rshifts ) */ Chris@69: CAb[ n + 1 ] = tmp2; /* Q( -rshifts ) */ Chris@69: num = silk_ADD32( num, tmp2 ); /* Q( -rshifts ) */ Chris@69: num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts ) */ Chris@69: Chris@69: /* Calculate the next order reflection (parcor) coefficient */ Chris@69: if( silk_abs( num ) < nrg ) { Chris@69: rc_Q31 = silk_DIV32_varQ( num, nrg, 31 ); Chris@69: } else { Chris@69: rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN; Chris@69: } Chris@69: Chris@69: /* Update inverse prediction gain */ Chris@69: tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); Chris@69: tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 ); Chris@69: if( tmp1 <= minInvGain_Q30 ) { Chris@69: /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ Chris@69: tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */ Chris@69: rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */ Chris@69: if( rc_Q31 > 0 ) { Chris@69: /* Newton-Raphson iteration */ Chris@69: rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */ Chris@69: rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */ Chris@69: if( num < 0 ) { Chris@69: /* Ensure adjusted reflection coefficients has the original sign */ Chris@69: rc_Q31 = -rc_Q31; Chris@69: } Chris@69: } Chris@69: invGain_Q30 = minInvGain_Q30; Chris@69: reached_max_gain = 1; Chris@69: } else { Chris@69: invGain_Q30 = tmp1; Chris@69: } Chris@69: Chris@69: /* Update the AR coefficients */ Chris@69: for( k = 0; k < (n + 1) >> 1; k++ ) { Chris@69: tmp1 = Af_QA[ k ]; /* QA */ Chris@69: tmp2 = Af_QA[ n - k - 1 ]; /* QA */ Chris@69: Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA */ Chris@69: Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA */ Chris@69: } Chris@69: Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA */ Chris@69: Chris@69: if( reached_max_gain ) { Chris@69: /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ Chris@69: for( k = n + 1; k < D; k++ ) { Chris@69: Af_QA[ k ] = 0; Chris@69: } Chris@69: break; Chris@69: } Chris@69: Chris@69: /* Update C * Af and C * Ab */ Chris@69: for( k = 0; k <= n + 1; k++ ) { Chris@69: tmp1 = CAf[ k ]; /* Q( -rshifts ) */ Chris@69: tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts ) */ Chris@69: CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts ) */ Chris@69: CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts ) */ Chris@69: } Chris@69: } Chris@69: Chris@69: if( reached_max_gain ) { Chris@69: for( k = 0; k < D; k++ ) { Chris@69: /* Scale coefficients */ Chris@69: A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); Chris@69: } Chris@69: /* Subtract energy of preceding samples from C0 */ Chris@69: if( rshifts > 0 ) { Chris@69: for( s = 0; s < nb_subfr; s++ ) { Chris@69: x_ptr = x + s * subfr_length; Chris@69: C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D, arch ), rshifts ); Chris@69: } Chris@69: } else { Chris@69: for( s = 0; s < nb_subfr; s++ ) { Chris@69: x_ptr = x + s * subfr_length; Chris@69: C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D, arch), -rshifts); Chris@69: } Chris@69: } Chris@69: /* Approximate residual energy */ Chris@69: *res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 ); Chris@69: *res_nrg_Q = -rshifts; Chris@69: } else { Chris@69: /* Return residual energy */ Chris@69: nrg = CAf[ 0 ]; /* Q( -rshifts ) */ Chris@69: tmp1 = (opus_int32)1 << 16; /* Q16 */ Chris@69: for( k = 0; k < D; k++ ) { Chris@69: Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16 */ Chris@69: nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts ) */ Chris@69: tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16 */ Chris@69: A_Q16[ k ] = -Atmp1; Chris@69: } Chris@69: *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ), -tmp1 );/* Q( -rshifts ) */ Chris@69: *res_nrg_Q = -rshifts; Chris@69: } Chris@69: }