diff options
| author | Felicia Lim <flim@google.com> | 2017-07-05 17:36:56 -0700 |
|---|---|---|
| committer | Felicia Lim <flim@google.com> | 2017-07-07 09:42:55 -0700 |
| commit | 0c2090c324e4f2ba2a8621c8b083559bab74c7c5 (patch) | |
| tree | fea1ab0038bc4102569d1ab4ee57a0f973895570 /silk | |
| parent | a7703b70699299f078a189e19b6915120cded732 (diff) | |
| download | libopus-0c2090c324e4f2ba2a8621c8b083559bab74c7c5.tar.gz | |
Opus 1.2.1android-o-iot-preview-5o-iot-preview-5
Change-Id: I551f1de5c5e121ac1275334e67c7e0f96ab18114
Test: - verified builds for arm*/mips*/x86*
- checked functionality using an emulator and stagefright
Diffstat (limited to 'silk')
101 files changed, 3633 insertions, 2464 deletions
diff --git a/silk/A2NLSF.c b/silk/A2NLSF.c index b6e9e5ff..b487686f 100644 --- a/silk/A2NLSF.c +++ b/silk/A2NLSF.c @@ -40,7 +40,7 @@ POSSIBILITY OF SUCH DAMAGE. /* Number of binary divisions, when not in low complexity mode */ #define BIN_DIV_STEPS_A2NLSF_FIX 3 /* must be no higher than 16 - log2( LSF_COS_TAB_SZ_FIX ) */ -#define MAX_ITERATIONS_A2NLSF_FIX 30 +#define MAX_ITERATIONS_A2NLSF_FIX 16 /* Helper function for A2NLSF(..) */ /* Transforms polynomials from cos(n*f) to cos(f)^n */ @@ -130,7 +130,7 @@ void silk_A2NLSF( const opus_int d /* I Filter order (must be even) */ ) { - opus_int i, k, m, dd, root_ix, ffrac; + opus_int i, k, m, dd, root_ix, ffrac; opus_int32 xlo, xhi, xmid; opus_int32 ylo, yhi, ymid, thr; opus_int32 nom, den; @@ -239,13 +239,13 @@ void silk_A2NLSF( /* Set NLSFs to white spectrum and exit */ NLSF[ 0 ] = (opus_int16)silk_DIV32_16( 1 << 15, d + 1 ); for( k = 1; k < d; k++ ) { - NLSF[ k ] = (opus_int16)silk_SMULBB( k + 1, NLSF[ 0 ] ); + NLSF[ k ] = (opus_int16)silk_ADD16( NLSF[ k-1 ], NLSF[ 0 ] ); } return; } /* Error: Apply progressively more bandwidth expansion and run again */ - silk_bwexpander_32( a_Q16, d, 65536 - silk_SMULBB( 10 + i, i ) ); /* 10_Q16 = 0.00015*/ + silk_bwexpander_32( a_Q16, d, 65536 - silk_LSHIFT( 1, i ) ); silk_A2NLSF_init( a_Q16, P, Q, dd ); p = P; /* Pointer to polynomial */ @@ -138,11 +138,11 @@ void silk_CNG( gain_Q16 = silk_LSHIFT32( silk_SQRT_APPROX( gain_Q16 ), 8 ); } gain_Q10 = silk_RSHIFT( gain_Q16, 6 ); - + silk_CNG_exc( CNG_sig_Q14 + MAX_LPC_ORDER, psCNG->CNG_exc_buf_Q14, length, &psCNG->rand_seed ); /* Convert CNG NLSF to filter representation */ - silk_NLSF2A( A_Q12, psCNG->CNG_smth_NLSF_Q15, psDec->LPC_order ); + silk_NLSF2A( A_Q12, psCNG->CNG_smth_NLSF_Q15, psDec->LPC_order, psDec->arch ); /* Generate CNG signal, by synthesis filtering */ silk_memcpy( CNG_sig_Q14, psCNG->CNG_synth_state, MAX_LPC_ORDER * sizeof( opus_int32 ) ); @@ -170,11 +170,11 @@ void silk_CNG( } /* Update states */ - CNG_sig_Q14[ MAX_LPC_ORDER + i ] = silk_ADD_LSHIFT( CNG_sig_Q14[ MAX_LPC_ORDER + i ], LPC_pred_Q10, 4 ); - + CNG_sig_Q14[ MAX_LPC_ORDER + i ] = silk_ADD_SAT32( CNG_sig_Q14[ MAX_LPC_ORDER + i ], silk_LSHIFT_SAT32( LPC_pred_Q10, 4 ) ); + /* Scale with Gain and add to input signal */ frame[ i ] = (opus_int16)silk_ADD_SAT16( frame[ i ], silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( CNG_sig_Q14[ MAX_LPC_ORDER + i ], gain_Q10 ), 8 ) ) ); - + } silk_memcpy( psCNG->CNG_synth_state, &CNG_sig_Q14[ length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); } else { diff --git a/silk/LPC_analysis_filter.c b/silk/LPC_analysis_filter.c index 20906673..7715f70f 100644 --- a/silk/LPC_analysis_filter.c +++ b/silk/LPC_analysis_filter.c @@ -39,6 +39,13 @@ POSSIBILITY OF SUCH DAMAGE. /* first d output samples are set to zero */ /*******************************************/ +/* OPT: Using celt_fir() for this function should be faster, but it may cause + integer overflows in intermediate values (not final results), which the + current implementation silences by casting to unsigned. Enabling + this should be safe in pretty much all cases, even though it is not technically + C89-compliant. */ +#define USE_CELT_FIR 0 + void silk_LPC_analysis_filter( opus_int16 *out, /* O Output signal */ const opus_int16 *in, /* I Input signal */ @@ -49,8 +56,7 @@ void silk_LPC_analysis_filter( ) { opus_int j; -#ifdef FIXED_POINT - opus_int16 mem[SILK_MAX_ORDER_LPC]; +#if defined(FIXED_POINT) && USE_CELT_FIR opus_int16 num[SILK_MAX_ORDER_LPC]; #else int ix; @@ -62,15 +68,12 @@ void silk_LPC_analysis_filter( silk_assert( (d & 1) == 0 ); silk_assert( d <= len ); -#ifdef FIXED_POINT +#if defined(FIXED_POINT) && USE_CELT_FIR silk_assert( d <= SILK_MAX_ORDER_LPC ); for ( j = 0; j < d; j++ ) { num[ j ] = -B[ j ]; } - for (j=0;j<d;j++) { - mem[ j ] = in[ d - j - 1 ]; - } - celt_fir( in + d, num, out + d, len - d, d, mem, arch ); + celt_fir( in + d, num, out + d, len - d, d, arch ); for ( j = 0; j < d; j++ ) { out[ j ] = 0; } diff --git a/silk/LPC_fit.c b/silk/LPC_fit.c new file mode 100644 index 00000000..cdea4f3a --- /dev/null +++ b/silk/LPC_fit.c @@ -0,0 +1,81 @@ +/*********************************************************************** +Copyright (c) 2013, Koen Vos. All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "SigProc_FIX.h" + +/* Convert int32 coefficients to int16 coefs and make sure there's no wrap-around */ +void silk_LPC_fit( + opus_int16 *a_QOUT, /* O Output signal */ + opus_int32 *a_QIN, /* I/O Input signal */ + const opus_int QOUT, /* I Input Q domain */ + const opus_int QIN, /* I Input Q domain */ + const opus_int d /* I Filter order */ +) +{ + opus_int i, k, idx = 0; + opus_int32 maxabs, absval, chirp_Q16; + + /* Limit the maximum absolute value of the prediction coefficients, so that they'll fit in int16 */ + for( i = 0; i < 10; i++ ) { + /* Find maximum absolute value and its index */ + maxabs = 0; + for( k = 0; k < d; k++ ) { + absval = silk_abs( a_QIN[k] ); + if( absval > maxabs ) { + maxabs = absval; + idx = k; + } + } + maxabs = silk_RSHIFT_ROUND( maxabs, QIN - QOUT ); + + if( maxabs > silk_int16_MAX ) { + /* Reduce magnitude of prediction coefficients */ + maxabs = silk_min( maxabs, 163838 ); /* ( silk_int32_MAX >> 14 ) + silk_int16_MAX = 163838 */ + chirp_Q16 = SILK_FIX_CONST( 0.999, 16 ) - silk_DIV32( silk_LSHIFT( maxabs - silk_int16_MAX, 14 ), + silk_RSHIFT32( silk_MUL( maxabs, idx + 1), 2 ) ); + silk_bwexpander_32( a_QIN, d, chirp_Q16 ); + } else { + break; + } + } + + if( i == 10 ) { + /* Reached the last iteration, clip the coefficients */ + for( k = 0; k < d; k++ ) { + a_QOUT[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT ) ); + a_QIN[ k ] = silk_LSHIFT( (opus_int32)a_QOUT[ k ], QIN - QOUT ); + } + } else { + for( k = 0; k < d; k++ ) { + a_QOUT[ k ] = (opus_int16)silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT ); + } + } +} diff --git a/silk/LPC_inv_pred_gain.c b/silk/LPC_inv_pred_gain.c index 4af89aa5..a3746a6e 100644 --- a/silk/LPC_inv_pred_gain.c +++ b/silk/LPC_inv_pred_gain.c @@ -30,6 +30,7 @@ POSSIBILITY OF SUCH DAMAGE. #endif #include "SigProc_FIX.h" +#include "define.h" #define QA 24 #define A_LIMIT SILK_FIX_CONST( 0.99975, QA ) @@ -38,117 +39,103 @@ POSSIBILITY OF SUCH DAMAGE. /* Compute inverse of LPC prediction gain, and */ /* test if LPC coefficients are stable (all poles within unit circle) */ -static opus_int32 LPC_inverse_pred_gain_QA( /* O Returns inverse prediction gain in energy domain, Q30 */ - opus_int32 A_QA[ 2 ][ SILK_MAX_ORDER_LPC ], /* I Prediction coefficients */ +static opus_int32 LPC_inverse_pred_gain_QA_c( /* O Returns inverse prediction gain in energy domain, Q30 */ + opus_int32 A_QA[ SILK_MAX_ORDER_LPC ], /* I Prediction coefficients */ const opus_int order /* I Prediction order */ ) { opus_int k, n, mult2Q; - opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp_QA; - opus_int32 *Aold_QA, *Anew_QA; + opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp1, tmp2; - Anew_QA = A_QA[ order & 1 ]; - - invGain_Q30 = (opus_int32)1 << 30; + invGain_Q30 = SILK_FIX_CONST( 1, 30 ); for( k = order - 1; k > 0; k-- ) { /* Check for stability */ - if( ( Anew_QA[ k ] > A_LIMIT ) || ( Anew_QA[ k ] < -A_LIMIT ) ) { + if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) { return 0; } /* Set RC equal to negated AR coef */ - rc_Q31 = -silk_LSHIFT( Anew_QA[ k ], 31 - QA ); + rc_Q31 = -silk_LSHIFT( A_QA[ k ], 31 - QA ); /* rc_mult1_Q30 range: [ 1 : 2^30 ] */ - rc_mult1_Q30 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); + rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) ); silk_assert( rc_mult1_Q30 > ( 1 << 15 ) ); /* reduce A_LIMIT if fails */ silk_assert( rc_mult1_Q30 <= ( 1 << 30 ) ); - /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */ - mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) ); - rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 ); - /* Update inverse gain */ /* invGain_Q30 range: [ 0 : 2^30 ] */ invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 ); silk_assert( invGain_Q30 >= 0 ); silk_assert( invGain_Q30 <= ( 1 << 30 ) ); + if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) { + return 0; + } - /* Swap pointers */ - Aold_QA = Anew_QA; - Anew_QA = A_QA[ k & 1 ]; + /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */ + mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) ); + rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 ); /* Update AR coefficient */ - for( n = 0; n < k; n++ ) { - tmp_QA = Aold_QA[ n ] - MUL32_FRAC_Q( Aold_QA[ k - n - 1 ], rc_Q31, 31 ); - Anew_QA[ n ] = MUL32_FRAC_Q( tmp_QA, rc_mult2 , mult2Q ); + for( n = 0; n < (k + 1) >> 1; n++ ) { + opus_int64 tmp64; + tmp1 = A_QA[ n ]; + tmp2 = A_QA[ k - n - 1 ]; + tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp1, + MUL32_FRAC_Q( tmp2, rc_Q31, 31 ) ), rc_mult2 ), mult2Q); + if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) { + return 0; + } + A_QA[ n ] = ( opus_int32 )tmp64; + tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp2, + MUL32_FRAC_Q( tmp1, rc_Q31, 31 ) ), rc_mult2), mult2Q); + if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) { + return 0; + } + A_QA[ k - n - 1 ] = ( opus_int32 )tmp64; } } /* Check for stability */ - if( ( Anew_QA[ 0 ] > A_LIMIT ) || ( Anew_QA[ 0 ] < -A_LIMIT ) ) { + if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) { return 0; } /* Set RC equal to negated AR coef */ - rc_Q31 = -silk_LSHIFT( Anew_QA[ 0 ], 31 - QA ); + rc_Q31 = -silk_LSHIFT( A_QA[ 0 ], 31 - QA ); /* Range: [ 1 : 2^30 ] */ - rc_mult1_Q30 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); + rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) ); /* Update inverse gain */ /* Range: [ 0 : 2^30 ] */ invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 ); - silk_assert( invGain_Q30 >= 0 ); - silk_assert( invGain_Q30 <= 1<<30 ); + silk_assert( invGain_Q30 >= 0 ); + silk_assert( invGain_Q30 <= ( 1 << 30 ) ); + if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) { + return 0; + } return invGain_Q30; } /* For input in Q12 domain */ -opus_int32 silk_LPC_inverse_pred_gain( /* O Returns inverse prediction gain in energy domain, Q30 */ +opus_int32 silk_LPC_inverse_pred_gain_c( /* O Returns inverse prediction gain in energy domain, Q30 */ const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ const opus_int order /* I Prediction order */ ) { opus_int k; - opus_int32 Atmp_QA[ 2 ][ SILK_MAX_ORDER_LPC ]; - opus_int32 *Anew_QA; + opus_int32 Atmp_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 DC_resp = 0; - Anew_QA = Atmp_QA[ order & 1 ]; - /* Increase Q domain of the AR coefficients */ for( k = 0; k < order; k++ ) { DC_resp += (opus_int32)A_Q12[ k ]; - Anew_QA[ k ] = silk_LSHIFT32( (opus_int32)A_Q12[ k ], QA - 12 ); + Atmp_QA[ k ] = silk_LSHIFT32( (opus_int32)A_Q12[ k ], QA - 12 ); } /* If the DC is unstable, we don't even need to do the full calculations */ if( DC_resp >= 4096 ) { return 0; } - return LPC_inverse_pred_gain_QA( Atmp_QA, order ); + return LPC_inverse_pred_gain_QA_c( Atmp_QA, order ); } - -#ifdef FIXED_POINT - -/* For input in Q24 domain */ -opus_int32 silk_LPC_inverse_pred_gain_Q24( /* O Returns inverse prediction gain in energy domain, Q30 */ - const opus_int32 *A_Q24, /* I Prediction coefficients [order] */ - const opus_int order /* I Prediction order */ -) -{ - opus_int k; - opus_int32 Atmp_QA[ 2 ][ SILK_MAX_ORDER_LPC ]; - opus_int32 *Anew_QA; - - Anew_QA = Atmp_QA[ order & 1 ]; - - /* Increase Q domain of the AR coefficients */ - for( k = 0; k < order; k++ ) { - Anew_QA[ k ] = silk_RSHIFT32( A_Q24[ k ], 24 - QA ); - } - - return LPC_inverse_pred_gain_QA( Atmp_QA, order ); -} -#endif diff --git a/silk/LP_variable_cutoff.c b/silk/LP_variable_cutoff.c index f639e1f8..79112ad3 100644 --- a/silk/LP_variable_cutoff.c +++ b/silk/LP_variable_cutoff.c @@ -130,6 +130,6 @@ void silk_LP_variable_cutoff( /* ARMA low-pass filtering */ silk_assert( TRANSITION_NB == 3 && TRANSITION_NA == 2 ); - silk_biquad_alt( frame, B_Q28, A_Q28, psLP->In_LP_State, frame, frame_length, 1); + silk_biquad_alt_stride1( frame, B_Q28, A_Q28, psLP->In_LP_State, frame, frame_length); } } diff --git a/silk/MacroCount.h b/silk/MacroCount.h index 834817d0..78100ffe 100644 --- a/silk/MacroCount.h +++ b/silk/MacroCount.h @@ -319,14 +319,6 @@ static OPUS_INLINE opus_int32 silk_ADD_POS_SAT32(opus_int64 a, opus_int64 b){ return(tmp); } -#undef silk_ADD_POS_SAT64 -static OPUS_INLINE opus_int64 silk_ADD_POS_SAT64(opus_int64 a, opus_int64 b){ - opus_int64 tmp; - ops_count += 1; - tmp = ((((a)+(b)) & 0x8000000000000000LL) ? silk_int64_MAX : ((a)+(b))); - return(tmp); -} - #undef silk_LSHIFT8 static OPUS_INLINE opus_int8 silk_LSHIFT8(opus_int8 a, opus_int32 shift){ opus_int8 ret; @@ -699,7 +691,7 @@ return(ret); #undef silk_LIMIT_32 -static OPUS_INLINE opus_int silk_LIMIT_32(opus_int32 a, opus_int32 limit1, opus_int32 limit2) +static OPUS_INLINE opus_int32 silk_LIMIT_32(opus_int32 a, opus_int32 limit1, opus_int32 limit2) { opus_int32 ret; ops_count += 6; diff --git a/silk/MacroDebug.h b/silk/MacroDebug.h index 35aedc5c..8dd4ce2e 100644 --- a/silk/MacroDebug.h +++ b/silk/MacroDebug.h @@ -539,8 +539,7 @@ static OPUS_INLINE opus_int32 silk_DIV32_16_(opus_int32 a32, opus_int32 b32, cha no checking needed for silk_POS_SAT32 no checking needed for silk_ADD_POS_SAT8 no checking needed for silk_ADD_POS_SAT16 - no checking needed for silk_ADD_POS_SAT32 - no checking needed for silk_ADD_POS_SAT64 */ + no checking needed for silk_ADD_POS_SAT32 */ #undef silk_LSHIFT8 #define silk_LSHIFT8(a,b) silk_LSHIFT8_((a), (b), __FILE__, __LINE__) diff --git a/silk/NLSF2A.c b/silk/NLSF2A.c index b1c559ea..116b465b 100644 --- a/silk/NLSF2A.c +++ b/silk/NLSF2A.c @@ -66,7 +66,8 @@ static OPUS_INLINE void silk_NLSF2A_find_poly( void silk_NLSF2A( opus_int16 *a_Q12, /* O monic whitening filter coefficients in Q12, [ d ] */ const opus_int16 *NLSF, /* I normalized line spectral frequencies in Q15, [ d ] */ - const opus_int d /* I filter order (should be even) */ + const opus_int d, /* I filter order (should be even) */ + int arch /* I Run-time architecture */ ) { /* This ordering was found to maximize quality. It improves numerical accuracy of @@ -83,15 +84,14 @@ void silk_NLSF2A( opus_int32 P[ SILK_MAX_ORDER_LPC / 2 + 1 ], Q[ SILK_MAX_ORDER_LPC / 2 + 1 ]; opus_int32 Ptmp, Qtmp, f_int, f_frac, cos_val, delta; opus_int32 a32_QA1[ SILK_MAX_ORDER_LPC ]; - opus_int32 maxabs, absval, idx=0, sc_Q16; silk_assert( LSF_COS_TAB_SZ_FIX == 128 ); - silk_assert( d==10||d==16 ); + silk_assert( d==10 || d==16 ); /* convert LSFs to 2*cos(LSF), using piecewise linear curve from table */ ordering = d == 16 ? ordering16 : ordering10; for( k = 0; k < d; k++ ) { - silk_assert(NLSF[k] >= 0 ); + silk_assert( NLSF[k] >= 0 ); /* f_int on a scale 0-127 (rounded down) */ f_int = silk_RSHIFT( NLSF[k], 15 - 7 ); @@ -126,52 +126,15 @@ void silk_NLSF2A( a32_QA1[ d-k-1 ] = Qtmp - Ptmp; /* QA+1 */ } - /* Limit the maximum absolute value of the prediction coefficients, so that they'll fit in int16 */ - for( i = 0; i < 10; i++ ) { - /* Find maximum absolute value and its index */ - maxabs = 0; - for( k = 0; k < d; k++ ) { - absval = silk_abs( a32_QA1[k] ); - if( absval > maxabs ) { - maxabs = absval; - idx = k; - } - } - maxabs = silk_RSHIFT_ROUND( maxabs, QA + 1 - 12 ); /* QA+1 -> Q12 */ - - if( maxabs > silk_int16_MAX ) { - /* Reduce magnitude of prediction coefficients */ - maxabs = silk_min( maxabs, 163838 ); /* ( silk_int32_MAX >> 14 ) + silk_int16_MAX = 163838 */ - sc_Q16 = SILK_FIX_CONST( 0.999, 16 ) - silk_DIV32( silk_LSHIFT( maxabs - silk_int16_MAX, 14 ), - silk_RSHIFT32( silk_MUL( maxabs, idx + 1), 2 ) ); - silk_bwexpander_32( a32_QA1, d, sc_Q16 ); - } else { - break; - } - } + /* Convert int32 coefficients to Q12 int16 coefs */ + silk_LPC_fit( a_Q12, a32_QA1, 12, QA + 1, d ); - if( i == 10 ) { - /* Reached the last iteration, clip the coefficients */ + for( i = 0; silk_LPC_inverse_pred_gain( a_Q12, d, arch ) == 0 && i < MAX_LPC_STABILIZE_ITERATIONS; i++ ) { + /* Prediction coefficients are (too close to) unstable; apply bandwidth expansion */ + /* on the unscaled coefficients, convert to Q12 and measure again */ + silk_bwexpander_32( a32_QA1, d, 65536 - silk_LSHIFT( 2, i ) ); for( k = 0; k < d; k++ ) { - a_Q12[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ) ); /* QA+1 -> Q12 */ - a32_QA1[ k ] = silk_LSHIFT( (opus_int32)a_Q12[ k ], QA + 1 - 12 ); - } - } else { - for( k = 0; k < d; k++ ) { - a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */ - } - } - - for( i = 0; i < MAX_LPC_STABILIZE_ITERATIONS; i++ ) { - if( silk_LPC_inverse_pred_gain( a_Q12, d ) < SILK_FIX_CONST( 1.0 / MAX_PREDICTION_POWER_GAIN, 30 ) ) { - /* Prediction coefficients are (too close to) unstable; apply bandwidth expansion */ - /* on the unscaled coefficients, convert to Q12 and measure again */ - silk_bwexpander_32( a32_QA1, d, 65536 - silk_LSHIFT( 2, i ) ); - for( k = 0; k < d; k++ ) { - a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */ - } - } else { - break; + a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */ } } } diff --git a/silk/NLSF_VQ.c b/silk/NLSF_VQ.c index 69b6e22e..452f3dcb 100644 --- a/silk/NLSF_VQ.c +++ b/silk/NLSF_VQ.c @@ -33,36 +33,44 @@ POSSIBILITY OF SUCH DAMAGE. /* Compute quantization errors for an LPC_order element input vector for a VQ codebook */ void silk_NLSF_VQ( - opus_int32 err_Q26[], /* O Quantization errors [K] */ + opus_int32 err_Q24[], /* O Quantization errors [K] */ const opus_int16 in_Q15[], /* I Input vectors to be quantized [LPC_order] */ const opus_uint8 pCB_Q8[], /* I Codebook vectors [K*LPC_order] */ + const opus_int16 pWght_Q9[], /* I Codebook weights [K*LPC_order] */ const opus_int K, /* I Number of codebook vectors */ const opus_int LPC_order /* I Number of LPCs */ ) { - opus_int i, m; - opus_int32 diff_Q15, sum_error_Q30, sum_error_Q26; + opus_int i, m; + opus_int32 diff_Q15, diffw_Q24, sum_error_Q24, pred_Q24; + const opus_int16 *w_Q9_ptr; + const opus_uint8 *cb_Q8_ptr; - silk_assert( LPC_order <= 16 ); silk_assert( ( LPC_order & 1 ) == 0 ); /* Loop over codebook */ + cb_Q8_ptr = pCB_Q8; + w_Q9_ptr = pWght_Q9; for( i = 0; i < K; i++ ) { - sum_error_Q26 = 0; - for( m = 0; m < LPC_order; m += 2 ) { - /* Compute weighted squared quantization error for index m */ - diff_Q15 = silk_SUB_LSHIFT32( in_Q15[ m ], (opus_int32)*pCB_Q8++, 7 ); /* range: [ -32767 : 32767 ]*/ - sum_error_Q30 = silk_SMULBB( diff_Q15, diff_Q15 ); + sum_error_Q24 = 0; + pred_Q24 = 0; + for( m = LPC_order-2; m >= 0; m -= 2 ) { + /* Compute weighted absolute predictive quantization error for index m + 1 */ + diff_Q15 = silk_SUB_LSHIFT32( in_Q15[ m + 1 ], (opus_int32)cb_Q8_ptr[ m + 1 ], 7 ); /* range: [ -32767 : 32767 ]*/ + diffw_Q24 = silk_SMULBB( diff_Q15, w_Q9_ptr[ m + 1 ] ); + sum_error_Q24 = silk_ADD32( sum_error_Q24, silk_abs( silk_SUB_RSHIFT32( diffw_Q24, pred_Q24, 1 ) ) ); + pred_Q24 = diffw_Q24; - /* Compute weighted squared quantization error for index m + 1 */ - diff_Q15 = silk_SUB_LSHIFT32( in_Q15[m + 1], (opus_int32)*pCB_Q8++, 7 ); /* range: [ -32767 : 32767 ]*/ - sum_error_Q30 = silk_SMLABB( sum_error_Q30, diff_Q15, diff_Q15 ); + /* Compute weighted absolute predictive quantization error for index m */ + diff_Q15 = silk_SUB_LSHIFT32( in_Q15[ m ], (opus_int32)cb_Q8_ptr[ m ], 7 ); /* range: [ -32767 : 32767 ]*/ + diffw_Q24 = silk_SMULBB( diff_Q15, w_Q9_ptr[ m ] ); + sum_error_Q24 = silk_ADD32( sum_error_Q24, silk_abs( silk_SUB_RSHIFT32( diffw_Q24, pred_Q24, 1 ) ) ); + pred_Q24 = diffw_Q24; - sum_error_Q26 = silk_ADD_RSHIFT32( sum_error_Q26, sum_error_Q30, 4 ); - - silk_assert( sum_error_Q26 >= 0 ); - silk_assert( sum_error_Q30 >= 0 ); + silk_assert( sum_error_Q24 >= 0 ); } - err_Q26[ i ] = sum_error_Q26; + err_Q24[ i ] = sum_error_Q24; + cb_Q8_ptr += LPC_order; + w_Q9_ptr += LPC_order; } } diff --git a/silk/NLSF_decode.c b/silk/NLSF_decode.c index 9f715060..eeb0ba8c 100644 --- a/silk/NLSF_decode.c +++ b/silk/NLSF_decode.c @@ -32,7 +32,7 @@ POSSIBILITY OF SUCH DAMAGE. #include "main.h" /* Predictive dequantizer for NLSF residuals */ -static OPUS_INLINE void silk_NLSF_residual_dequant( /* O Returns RD value in Q30 */ +static OPUS_INLINE void silk_NLSF_residual_dequant( /* O Returns RD value in Q30 */ opus_int16 x_Q10[], /* O Output [ order ] */ const opus_int8 indices[], /* I Quantization indices [ order ] */ const opus_uint8 pred_coef_Q8[], /* I Backward predictor coefs [ order ] */ @@ -70,15 +70,9 @@ void silk_NLSF_decode( opus_uint8 pred_Q8[ MAX_LPC_ORDER ]; opus_int16 ec_ix[ MAX_LPC_ORDER ]; opus_int16 res_Q10[ MAX_LPC_ORDER ]; - opus_int16 W_tmp_QW[ MAX_LPC_ORDER ]; - opus_int32 W_tmp_Q9, NLSF_Q15_tmp; + opus_int32 NLSF_Q15_tmp; const opus_uint8 *pCB_element; - - /* Decode first stage */ - pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ NLSFIndices[ 0 ] * psNLSF_CB->order ]; - for( i = 0; i < psNLSF_CB->order; i++ ) { - pNLSF_Q15[ i ] = silk_LSHIFT( (opus_int16)pCB_element[ i ], 7 ); - } + const opus_int16 *pCB_Wght_Q9; /* Unpack entropy table indices and predictor for current CB1 index */ silk_NLSF_unpack( ec_ix, pred_Q8, psNLSF_CB, NLSFIndices[ 0 ] ); @@ -86,13 +80,11 @@ void silk_NLSF_decode( /* Predictive residual dequantizer */ silk_NLSF_residual_dequant( res_Q10, &NLSFIndices[ 1 ], pred_Q8, psNLSF_CB->quantStepSize_Q16, psNLSF_CB->order ); - /* Weights from codebook vector */ - silk_NLSF_VQ_weights_laroia( W_tmp_QW, pNLSF_Q15, psNLSF_CB->order ); - - /* Apply inverse square-rooted weights and add to output */ + /* Apply inverse square-rooted weights to first stage and add to output */ + pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ NLSFIndices[ 0 ] * psNLSF_CB->order ]; + pCB_Wght_Q9 = &psNLSF_CB->CB1_Wght_Q9[ NLSFIndices[ 0 ] * psNLSF_CB->order ]; for( i = 0; i < psNLSF_CB->order; i++ ) { - W_tmp_Q9 = silk_SQRT_APPROX( silk_LSHIFT( (opus_int32)W_tmp_QW[ i ], 18 - NLSF_W_Q ) ); - NLSF_Q15_tmp = silk_ADD32( pNLSF_Q15[ i ], silk_DIV32_16( silk_LSHIFT( (opus_int32)res_Q10[ i ], 14 ), W_tmp_Q9 ) ); + NLSF_Q15_tmp = silk_ADD_LSHIFT32( silk_DIV32_16( silk_LSHIFT( (opus_int32)res_Q10[ i ], 14 ), pCB_Wght_Q9[ i ] ), (opus_int16)pCB_element[ i ], 7 ); pNLSF_Q15[ i ] = (opus_int16)silk_LIMIT( NLSF_Q15_tmp, 0, 32767 ); } diff --git a/silk/NLSF_del_dec_quant.c b/silk/NLSF_del_dec_quant.c index de88fee0..44a16acd 100644 --- a/silk/NLSF_del_dec_quant.c +++ b/silk/NLSF_del_dec_quant.c @@ -84,7 +84,7 @@ opus_int32 silk_NLSF_del_dec_quant( /* O Returns nStates = 1; RD_Q25[ 0 ] = 0; prev_out_Q10[ 0 ] = 0; - for( i = order - 1; ; i-- ) { + for( i = order - 1; i >= 0; i-- ) { rates_Q5 = &ec_rates_Q5[ ec_ix[ i ] ]; in_Q10 = x_Q10[ i ]; for( j = 0; j < nStates; j++ ) { @@ -131,7 +131,7 @@ opus_int32 silk_NLSF_del_dec_quant( /* O Returns RD_Q25[ j + nStates ] = silk_SMLABB( silk_MLA( RD_tmp_Q25, silk_SMULBB( diff_Q10, diff_Q10 ), w_Q5[ i ] ), mu_Q20, rate1_Q5 ); } - if( nStates <= ( NLSF_QUANT_DEL_DEC_STATES >> 1 ) ) { + if( nStates <= NLSF_QUANT_DEL_DEC_STATES/2 ) { /* double number of states and copy */ for( j = 0; j < nStates; j++ ) { ind[ j + nStates ][ i ] = ind[ j ][ i ] + 1; @@ -140,7 +140,7 @@ opus_int32 silk_NLSF_del_dec_quant( /* O Returns for( j = nStates; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { ind[ j ][ i ] = ind[ j - nStates ][ i ]; } - } else if( i > 0 ) { + } else { /* sort lower and upper half of RD_Q25, pairwise */ for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { if( RD_Q25[ j ] > RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ] ) { @@ -191,8 +191,6 @@ opus_int32 silk_NLSF_del_dec_quant( /* O Returns for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { ind[ j ][ i ] += silk_RSHIFT( ind_sort[ j ], NLSF_QUANT_DEL_DEC_STATES_LOG2 ); } - } else { /* i == 0 */ - break; } } diff --git a/silk/NLSF_encode.c b/silk/NLSF_encode.c index f03c3f1c..268b9a19 100644 --- a/silk/NLSF_encode.c +++ b/silk/NLSF_encode.c @@ -37,9 +37,9 @@ POSSIBILITY OF SUCH DAMAGE. /***********************/ opus_int32 silk_NLSF_encode( /* O Returns RD value in Q25 */ opus_int8 *NLSFIndices, /* I Codebook path vector [ LPC_ORDER + 1 ] */ - opus_int16 *pNLSF_Q15, /* I/O Quantized NLSF vector [ LPC_ORDER ] */ + opus_int16 *pNLSF_Q15, /* I/O (Un)quantized NLSF vector [ LPC_ORDER ] */ const silk_NLSF_CB_struct *psNLSF_CB, /* I Codebook object */ - const opus_int16 *pW_QW, /* I NLSF weight vector [ LPC_ORDER ] */ + const opus_int16 *pW_Q2, /* I NLSF weight vector [ LPC_ORDER ] */ const opus_int NLSF_mu_Q20, /* I Rate weight for the RD optimization */ const opus_int nSurvivors, /* I Max survivors after first stage */ const opus_int signalType /* I Signal type: 0/1/2 */ @@ -47,21 +47,19 @@ opus_int32 silk_NLSF_encode( /* O Returns { opus_int i, s, ind1, bestIndex, prob_Q8, bits_q7; opus_int32 W_tmp_Q9, ret; - VARDECL( opus_int32, err_Q26 ); + VARDECL( opus_int32, err_Q24 ); VARDECL( opus_int32, RD_Q25 ); VARDECL( opus_int, tempIndices1 ); VARDECL( opus_int8, tempIndices2 ); - opus_int16 res_Q15[ MAX_LPC_ORDER ]; opus_int16 res_Q10[ MAX_LPC_ORDER ]; opus_int16 NLSF_tmp_Q15[ MAX_LPC_ORDER ]; - opus_int16 W_tmp_QW[ MAX_LPC_ORDER ]; opus_int16 W_adj_Q5[ MAX_LPC_ORDER ]; opus_uint8 pred_Q8[ MAX_LPC_ORDER ]; opus_int16 ec_ix[ MAX_LPC_ORDER ]; const opus_uint8 *pCB_element, *iCDF_ptr; + const opus_int16 *pCB_Wght_Q9; SAVE_STACK; - silk_assert( nSurvivors <= NLSF_VQ_MAX_SURVIVORS ); silk_assert( signalType >= 0 && signalType <= 2 ); silk_assert( NLSF_mu_Q20 <= 32767 && NLSF_mu_Q20 >= 0 ); @@ -69,12 +67,12 @@ opus_int32 silk_NLSF_encode( /* O Returns silk_NLSF_stabilize( pNLSF_Q15, psNLSF_CB->deltaMin_Q15, psNLSF_CB->order ); /* First stage: VQ */ - ALLOC( err_Q26, psNLSF_CB->nVectors, opus_int32 ); - silk_NLSF_VQ( err_Q26, pNLSF_Q15, psNLSF_CB->CB1_NLSF_Q8, psNLSF_CB->nVectors, psNLSF_CB->order ); + ALLOC( err_Q24, psNLSF_CB->nVectors, opus_int32 ); + silk_NLSF_VQ( err_Q24, pNLSF_Q15, psNLSF_CB->CB1_NLSF_Q8, psNLSF_CB->CB1_Wght_Q9, psNLSF_CB->nVectors, psNLSF_CB->order ); /* Sort the quantization errors */ ALLOC( tempIndices1, nSurvivors, opus_int ); - silk_insertion_sort_increasing( err_Q26, tempIndices1, psNLSF_CB->nVectors, nSurvivors ); + silk_insertion_sort_increasing( err_Q24, tempIndices1, psNLSF_CB->nVectors, nSurvivors ); ALLOC( RD_Q25, nSurvivors, opus_int32 ); ALLOC( tempIndices2, nSurvivors * MAX_LPC_ORDER, opus_int8 ); @@ -85,23 +83,12 @@ opus_int32 silk_NLSF_encode( /* O Returns /* Residual after first stage */ pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ ind1 * psNLSF_CB->order ]; + pCB_Wght_Q9 = &psNLSF_CB->CB1_Wght_Q9[ ind1 * psNLSF_CB->order ]; for( i = 0; i < psNLSF_CB->order; i++ ) { NLSF_tmp_Q15[ i ] = silk_LSHIFT16( (opus_int16)pCB_element[ i ], 7 ); - res_Q15[ i ] = pNLSF_Q15[ i ] - NLSF_tmp_Q15[ i ]; - } - - /* Weights from codebook vector */ - silk_NLSF_VQ_weights_laroia( W_tmp_QW, NLSF_tmp_Q15, psNLSF_CB->order ); - - /* Apply square-rooted weights */ - for( i = 0; i < psNLSF_CB->order; i++ ) { - W_tmp_Q9 = silk_SQRT_APPROX( silk_LSHIFT( (opus_int32)W_tmp_QW[ i ], 18 - NLSF_W_Q ) ); - res_Q10[ i ] = (opus_int16)silk_RSHIFT( silk_SMULBB( res_Q15[ i ], W_tmp_Q9 ), 14 ); - } - - /* Modify input weights accordingly */ - for( i = 0; i < psNLSF_CB->order; i++ ) { - W_adj_Q5[ i ] = silk_DIV32_16( silk_LSHIFT( (opus_int32)pW_QW[ i ], 5 ), W_tmp_QW[ i ] ); + W_tmp_Q9 = pCB_Wght_Q9[ i ]; + res_Q10[ i ] = (opus_int16)silk_RSHIFT( silk_SMULBB( pNLSF_Q15[ i ] - NLSF_tmp_Q15[ i ], W_tmp_Q9 ), 14 ); + W_adj_Q5[ i ] = silk_DIV32_varQ( (opus_int32)pW_Q2[ i ], silk_SMULBB( W_tmp_Q9, W_tmp_Q9 ), 21 ); } /* Unpack entropy table indices and predictor for current CB1 index */ @@ -37,7 +37,7 @@ POSSIBILITY OF SUCH DAMAGE. static OPUS_INLINE void silk_nsq_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ - const opus_int32 x_Q3[], /* I input in Q3 */ + const opus_int16 x16[], /* I input */ opus_int32 x_sc_Q10[], /* O input scaled with 1/Gain */ const opus_int16 sLTP[], /* I re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -75,14 +75,14 @@ static OPUS_INLINE void silk_noise_shape_quantizer( void silk_NSQ_c ( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q3[], /* I Prefiltered input signal */ + const opus_int16 x16[], /* I Input */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ - const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ @@ -117,8 +117,7 @@ void silk_NSQ_c LSF_interpolation_flag = 1; } - ALLOC( sLTP_Q15, - psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); + ALLOC( sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 ); ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 ); /* Set up pointers to start of sub frame */ @@ -128,7 +127,7 @@ void silk_NSQ_c for( k = 0; k < psEncC->nb_subfr; k++ ) { A_Q12 = &PredCoef_Q12[ (( k >> 1 ) | ( 1 - LSF_interpolation_flag )) * MAX_LPC_ORDER ]; B_Q14 = <PCoef_Q14[ k * LTP_ORDER ]; - AR_shp_Q13 = &AR2_Q13[ k * MAX_SHAPE_LPC_ORDER ]; + AR_shp_Q13 = &AR_Q13[ k * MAX_SHAPE_LPC_ORDER ]; /* Noise shape parameters */ silk_assert( HarmShapeGain_Q14[ k ] >= 0 ); @@ -154,13 +153,13 @@ void silk_NSQ_c } } - silk_nsq_scale_states( psEncC, NSQ, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType ); + silk_nsq_scale_states( psEncC, NSQ, x16, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType ); silk_noise_shape_quantizer( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->arch ); - x_Q3 += psEncC->subfr_length; + x16 += psEncC->subfr_length; pulses += psEncC->subfr_length; pxq += psEncC->subfr_length; } @@ -169,7 +168,6 @@ void silk_NSQ_c NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; /* Save quantized speech and noise shaping signals */ - /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[ psEncC->ltp_mem_length ], psEncC->frame_length * sizeof( opus_int16 ) ) */ silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); RESTORE_STACK; @@ -250,7 +248,7 @@ void silk_noise_shape_quantizer( /* Noise shape feedback */ silk_assert( ( shapingLPCOrder & 1 ) == 0 ); /* check that order is even */ - n_AR_Q12 = silk_NSQ_noise_shape_feedback_loop(psLPC_Q14, NSQ->sAR2_Q14, AR_shp_Q13, shapingLPCOrder, arch); + n_AR_Q12 = silk_NSQ_noise_shape_feedback_loop(&NSQ->sDiff_shp_Q14, NSQ->sAR2_Q14, AR_shp_Q13, shapingLPCOrder, arch); n_AR_Q12 = silk_SMLAWB( n_AR_Q12, NSQ->sLF_AR_shp_Q14, Tilt_Q14 ); @@ -279,14 +277,27 @@ void silk_noise_shape_quantizer( r_Q10 = silk_SUB32( x_sc_Q10[ i ], tmp1 ); /* residual error Q10 */ /* Flip sign depending on dither */ - if ( NSQ->rand_seed < 0 ) { - r_Q10 = -r_Q10; + if( NSQ->rand_seed < 0 ) { + r_Q10 = -r_Q10; } r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 ); /* Find two quantization level candidates and measure their rate-distortion */ q1_Q10 = silk_SUB32( r_Q10, offset_Q10 ); q1_Q0 = silk_RSHIFT( q1_Q10, 10 ); + if (Lambda_Q10 > 2048) { + /* For aggressive RDO, the bias becomes more than one pulse. */ + int rdo_offset = Lambda_Q10/2 - 512; + if (q1_Q10 > rdo_offset) { + q1_Q0 = silk_RSHIFT( q1_Q10 - rdo_offset, 10 ); + } else if (q1_Q10 < -rdo_offset) { + q1_Q0 = silk_RSHIFT( q1_Q10 + rdo_offset, 10 ); + } else if (q1_Q10 < 0) { + q1_Q0 = -1; + } else { + q1_Q0 = 0; + } + } if( q1_Q0 > 0 ) { q1_Q10 = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 ); q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 ); @@ -337,7 +348,8 @@ void silk_noise_shape_quantizer( /* Update states */ psLPC_Q14++; *psLPC_Q14 = xq_Q14; - sLF_AR_shp_Q14 = silk_SUB_LSHIFT32( xq_Q14, n_AR_Q12, 2 ); + NSQ->sDiff_shp_Q14 = silk_SUB_LSHIFT32( xq_Q14, x_sc_Q10[ i ], 4 ); + sLF_AR_shp_Q14 = silk_SUB_LSHIFT32( NSQ->sDiff_shp_Q14, n_AR_Q12, 2 ); NSQ->sLF_AR_shp_Q14 = sLF_AR_shp_Q14; NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx ] = silk_SUB_LSHIFT32( sLF_AR_shp_Q14, n_LF_Q12, 2 ); @@ -356,7 +368,7 @@ void silk_noise_shape_quantizer( static OPUS_INLINE void silk_nsq_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ - const opus_int32 x_Q3[], /* I input in Q3 */ + const opus_int16 x16[], /* I input */ opus_int32 x_sc_Q10[], /* O input scaled with 1/Gain */ const opus_int16 sLTP[], /* I re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -368,28 +380,18 @@ static OPUS_INLINE void silk_nsq_scale_states( ) { opus_int i, lag; - opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23; + opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q26; lag = pitchL[ subfr ]; inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); silk_assert( inv_gain_Q31 != 0 ); - /* Calculate gain adjustment factor */ - if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { - gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); - } else { - gain_adj_Q16 = (opus_int32)1 << 16; - } - /* Scale input */ - inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 ); + inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 ); for( i = 0; i < psEncC->subfr_length; i++ ) { - x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 ); + x_sc_Q10[ i ] = silk_SMULWW( x16[ i ], inv_gain_Q26 ); } - /* Save inverse gain */ - NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; - /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ if( NSQ->rewhite_flag ) { if( subfr == 0 ) { @@ -403,7 +405,9 @@ static OPUS_INLINE void silk_nsq_scale_states( } /* Adjust for changing gain */ - if( gain_adj_Q16 != (opus_int32)1 << 16 ) { + if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { + gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); + /* Scale long-term shaping state */ for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) { NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); @@ -417,6 +421,7 @@ static OPUS_INLINE void silk_nsq_scale_states( } NSQ->sLF_AR_shp_Q14 = silk_SMULWW( gain_adj_Q16, NSQ->sLF_AR_shp_Q14 ); + NSQ->sDiff_shp_Q14 = silk_SMULWW( gain_adj_Q16, NSQ->sDiff_shp_Q14 ); /* Scale short-term prediction and shaping states */ for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { @@ -425,5 +430,8 @@ static OPUS_INLINE void silk_nsq_scale_states( for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) { NSQ->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sAR2_Q14[ i ] ); } + + /* Save inverse gain */ + NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; } } diff --git a/silk/NSQ_del_dec.c b/silk/NSQ_del_dec.c index ab6feeac..1cd29d96 100644 --- a/silk/NSQ_del_dec.c +++ b/silk/NSQ_del_dec.c @@ -43,6 +43,7 @@ typedef struct { opus_int32 Shape_Q14[ DECISION_DELAY ]; opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ]; opus_int32 LF_AR_Q14; + opus_int32 Diff_Q14; opus_int32 Seed; opus_int32 SeedInit; opus_int32 RD_Q10; @@ -53,6 +54,7 @@ typedef struct { opus_int32 RD_Q10; opus_int32 xq_Q14; opus_int32 LF_AR_Q14; + opus_int32 Diff_Q14; opus_int32 sLTP_shp_Q14; opus_int32 LPC_exc_Q14; } NSQ_sample_struct; @@ -66,7 +68,7 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ - const opus_int32 x_Q3[], /* I Input in Q3 */ + const opus_int16 x16[], /* I Input */ opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -107,20 +109,20 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay, /* I */ int arch /* I */ ); void silk_NSQ_del_dec_c( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q3[], /* I Prefiltered input signal */ + const opus_int16 x16[], /* I Input */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ - const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ @@ -159,6 +161,7 @@ void silk_NSQ_del_dec_c( psDD->SeedInit = psDD->Seed; psDD->RD_Q10 = 0; psDD->LF_AR_Q14 = NSQ->sLF_AR_shp_Q14; + psDD->Diff_Q14 = NSQ->sDiff_shp_Q14; psDD->Shape_Q14[ 0 ] = NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ]; silk_memcpy( psDD->sLPC_Q14, NSQ->sLPC_Q14, NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) ); silk_memcpy( psDD->sAR2_Q14, NSQ->sAR2_Q14, sizeof( NSQ->sAR2_Q14 ) ); @@ -186,8 +189,7 @@ void silk_NSQ_del_dec_c( LSF_interpolation_flag = 1; } - ALLOC( sLTP_Q15, - psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); + ALLOC( sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 ); ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 ); ALLOC( delayedGain_Q10, DECISION_DELAY, opus_int32 ); @@ -199,7 +201,7 @@ void silk_NSQ_del_dec_c( for( k = 0; k < psEncC->nb_subfr; k++ ) { A_Q12 = &PredCoef_Q12[ ( ( k >> 1 ) | ( 1 - LSF_interpolation_flag ) ) * MAX_LPC_ORDER ]; B_Q14 = <PCoef_Q14[ k * LTP_ORDER ]; - AR_shp_Q13 = &AR2_Q13[ k * MAX_SHAPE_LPC_ORDER ]; + AR_shp_Q13 = &AR_Q13[ k * MAX_SHAPE_LPC_ORDER ]; /* Noise shape parameters */ silk_assert( HarmShapeGain_Q14[ k ] >= 0 ); @@ -235,7 +237,8 @@ void silk_NSQ_del_dec_c( psDD = &psDelDec[ Winner_ind ]; last_smple_idx = smpl_buf_idx + decisionDelay; for( i = 0; i < decisionDelay; i++ ) { - last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; + last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY; + if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY; pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) ); @@ -257,7 +260,7 @@ void silk_NSQ_del_dec_c( } } - silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, + silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x16, x_sc_Q10, sLTP, sLTP_Q15, k, psEncC->nStatesDelayedDecision, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay ); silk_noise_shape_quantizer_del_dec( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, @@ -265,7 +268,7 @@ void silk_NSQ_del_dec_c( Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay, psEncC->arch ); - x_Q3 += psEncC->subfr_length; + x16 += psEncC->subfr_length; pulses += psEncC->subfr_length; pxq += psEncC->subfr_length; } @@ -286,7 +289,9 @@ void silk_NSQ_del_dec_c( last_smple_idx = smpl_buf_idx + decisionDelay; Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 ); for( i = 0; i < decisionDelay; i++ ) { - last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; + last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY; + if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY; + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) ); @@ -297,10 +302,10 @@ void silk_NSQ_del_dec_c( /* Update states */ NSQ->sLF_AR_shp_Q14 = psDD->LF_AR_Q14; + NSQ->sDiff_shp_Q14 = psDD->Diff_Q14; NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; /* Save quantized speech signal */ - /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[psEncC->ltp_mem_length], psEncC->frame_length * sizeof( opus_int16 ) ) */ silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); RESTORE_STACK; @@ -335,7 +340,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay, /* I */ int arch /* I */ ) @@ -416,7 +421,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( /* Noise shape feedback */ silk_assert( ( shapingLPCOrder & 1 ) == 0 ); /* check that order is even */ /* Output of lowpass section */ - tmp2 = silk_SMLAWB( psLPC_Q14[ 0 ], psDD->sAR2_Q14[ 0 ], warping_Q16 ); + tmp2 = silk_SMLAWB( psDD->Diff_Q14, psDD->sAR2_Q14[ 0 ], warping_Q16 ); /* Output of allpass section */ tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ 0 ], psDD->sAR2_Q14[ 1 ] - tmp2, warping_Q16 ); psDD->sAR2_Q14[ 0 ] = tmp2; @@ -462,6 +467,19 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( /* Find two quantization level candidates and measure their rate-distortion */ q1_Q10 = silk_SUB32( r_Q10, offset_Q10 ); q1_Q0 = silk_RSHIFT( q1_Q10, 10 ); + if (Lambda_Q10 > 2048) { + /* For aggressive RDO, the bias becomes more than one pulse. */ + int rdo_offset = Lambda_Q10/2 - 512; + if (q1_Q10 > rdo_offset) { + q1_Q0 = silk_RSHIFT( q1_Q10 - rdo_offset, 10 ); + } else if (q1_Q10 < -rdo_offset) { + q1_Q0 = silk_RSHIFT( q1_Q10 + rdo_offset, 10 ); + } else if (q1_Q10 < 0) { + q1_Q0 = -1; + } else { + q1_Q0 = 0; + } + } if( q1_Q0 > 0 ) { q1_Q10 = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 ); q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 ); @@ -515,7 +533,8 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 ); /* Update states */ - sLF_AR_shp_Q14 = silk_SUB32( xq_Q14, n_AR_Q14 ); + psSS[ 0 ].Diff_Q14 = silk_SUB_LSHIFT32( xq_Q14, x_Q10[ i ], 4 ); + sLF_AR_shp_Q14 = silk_SUB32( psSS[ 0 ].Diff_Q14, n_AR_Q14 ); psSS[ 0 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 ); psSS[ 0 ].LF_AR_Q14 = sLF_AR_shp_Q14; psSS[ 0 ].LPC_exc_Q14 = LPC_exc_Q14; @@ -529,21 +548,22 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( exc_Q14 = -exc_Q14; } - /* Add predictions */ LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 ); xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 ); /* Update states */ - sLF_AR_shp_Q14 = silk_SUB32( xq_Q14, n_AR_Q14 ); + psSS[ 1 ].Diff_Q14 = silk_SUB_LSHIFT32( xq_Q14, x_Q10[ i ], 4 ); + sLF_AR_shp_Q14 = silk_SUB32( psSS[ 1 ].Diff_Q14, n_AR_Q14 ); psSS[ 1 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 ); psSS[ 1 ].LF_AR_Q14 = sLF_AR_shp_Q14; psSS[ 1 ].LPC_exc_Q14 = LPC_exc_Q14; psSS[ 1 ].xq_Q14 = xq_Q14; } - *smpl_buf_idx = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK; /* Index to newest samples */ - last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK; /* Index to decisionDelay old samples */ + *smpl_buf_idx = ( *smpl_buf_idx - 1 ) % DECISION_DELAY; + if( *smpl_buf_idx < 0 ) *smpl_buf_idx += DECISION_DELAY; + last_smple_idx = ( *smpl_buf_idx + decisionDelay ) % DECISION_DELAY; /* Find winner */ RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10; @@ -607,6 +627,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( psDD = &psDelDec[ k ]; psSS = &psSampleState[ k ][ 0 ]; psDD->LF_AR_Q14 = psSS->LF_AR_Q14; + psDD->Diff_Q14 = psSS->Diff_Q14; psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ] = psSS->xq_Q14; psDD->Xq_Q14[ *smpl_buf_idx ] = psSS->xq_Q14; psDD->Q_Q10[ *smpl_buf_idx ] = psSS->Q_Q10; @@ -631,7 +652,7 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ - const opus_int32 x_Q3[], /* I Input in Q3 */ + const opus_int16 x16[], /* I Input */ opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -645,29 +666,19 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( ) { opus_int i, k, lag; - opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23; + opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q26; NSQ_del_dec_struct *psDD; lag = pitchL[ subfr ]; inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); silk_assert( inv_gain_Q31 != 0 ); - /* Calculate gain adjustment factor */ - if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { - gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); - } else { - gain_adj_Q16 = (opus_int32)1 << 16; - } - /* Scale input */ - inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 ); + inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 ); for( i = 0; i < psEncC->subfr_length; i++ ) { - x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 ); + x_sc_Q10[ i ] = silk_SMULWW( x16[ i ], inv_gain_Q26 ); } - /* Save inverse gain */ - NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; - /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ if( NSQ->rewhite_flag ) { if( subfr == 0 ) { @@ -681,7 +692,9 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( } /* Adjust for changing gain */ - if( gain_adj_Q16 != (opus_int32)1 << 16 ) { + if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { + gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); + /* Scale long-term shaping state */ for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) { NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); @@ -699,6 +712,7 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( /* Scale scalar states */ psDD->LF_AR_Q14 = silk_SMULWW( gain_adj_Q16, psDD->LF_AR_Q14 ); + psDD->Diff_Q14 = silk_SMULWW( gain_adj_Q16, psDD->Diff_Q14 ); /* Scale short-term prediction and shaping states */ for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { @@ -712,5 +726,8 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( psDD->Shape_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Shape_Q14[ i ] ); } } + + /* Save inverse gain */ + NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; } } @@ -275,7 +275,7 @@ static OPUS_INLINE void silk_PLC_conceal( /* Reduce random noise for unvoiced frames with high LPC gain */ opus_int32 invGain_Q30, down_scale_Q30; - invGain_Q30 = silk_LPC_inverse_pred_gain( psPLC->prevLPC_Q12, psDec->LPC_order ); + invGain_Q30 = silk_LPC_inverse_pred_gain( psPLC->prevLPC_Q12, psDec->LPC_order, arch ); down_scale_Q30 = silk_min_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 ); down_scale_Q30 = silk_max_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 ); @@ -328,8 +328,10 @@ static OPUS_INLINE void silk_PLC_conceal( for( j = 0; j < LTP_ORDER; j++ ) { B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 ); } - /* Gradually reduce excitation gain */ - rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 ); + if ( psDec->indices.signalType != TYPE_NO_VOICE_ACTIVITY ) { + /* Gradually reduce excitation gain */ + rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 ); + } /* Slowly increase pitch lag */ psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 ); diff --git a/silk/SigProc_FIX.h b/silk/SigProc_FIX.h index b6329944..f9ae3263 100644 --- a/silk/SigProc_FIX.h +++ b/silk/SigProc_FIX.h @@ -35,7 +35,7 @@ extern "C" /*#define silk_MACRO_COUNT */ /* Used to enable WMOPS counting */ -#define SILK_MAX_ORDER_LPC 16 /* max order of the LPC analysis in schur() and k2a() */ +#define SILK_MAX_ORDER_LPC 24 /* max order of the LPC analysis in schur() and k2a() */ #include <string.h> /* for memset(), memcpy(), memmove() */ #include "typedef.h" @@ -47,6 +47,11 @@ extern "C" #include "x86/SigProc_FIX_sse.h" #endif +#if (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) +#include "arm/biquad_alt_arm.h" +#include "arm/LPC_inv_pred_gain_arm.h" +#endif + /********************************************************************/ /* SIGNAL PROCESSING FUNCTIONS */ /********************************************************************/ @@ -96,14 +101,22 @@ void silk_resampler_down2_3( * slower than biquad() but uses more precise coefficients * can handle (slowly) varying coefficients */ -void silk_biquad_alt( +void silk_biquad_alt_stride1( const opus_int16 *in, /* I input signal */ const opus_int32 *B_Q28, /* I MA coefficients [3] */ const opus_int32 *A_Q28, /* I AR coefficients [2] */ opus_int32 *S, /* I/O State vector [2] */ opus_int16 *out, /* O output signal */ - const opus_int32 len, /* I signal length (must be even) */ - opus_int stride /* I Operate on interleaved signal if > 1 */ + const opus_int32 len /* I signal length (must be even) */ +); + +void silk_biquad_alt_stride2_c( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ ); /* Variable order MA prediction error filter. */ @@ -132,17 +145,11 @@ void silk_bwexpander_32( /* Compute inverse of LPC prediction gain, and */ /* test if LPC coefficients are stable (all poles within unit circle) */ -opus_int32 silk_LPC_inverse_pred_gain( /* O Returns inverse prediction gain in energy domain, Q30 */ +opus_int32 silk_LPC_inverse_pred_gain_c( /* O Returns inverse prediction gain in energy domain, Q30 */ const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ const opus_int order /* I Prediction order */ ); -/* For input in Q24 domain */ -opus_int32 silk_LPC_inverse_pred_gain_Q24( /* O Returns inverse prediction gain in energy domain, Q30 */ - const opus_int32 *A_Q24, /* I Prediction coefficients [order] */ - const opus_int order /* I Prediction order */ -); - /* Split signal in two decimated bands using first-order allpass filters */ void silk_ana_filt_bank_1( const opus_int16 *in, /* I Input signal [N] */ @@ -152,6 +159,14 @@ void silk_ana_filt_bank_1( const opus_int32 N /* I Number of input samples */ ); +#if !defined(OVERRIDE_silk_biquad_alt_stride2) +#define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((void)(arch), silk_biquad_alt_stride2_c(in, B_Q28, A_Q28, S, out, len)) +#endif + +#if !defined(OVERRIDE_silk_LPC_inverse_pred_gain) +#define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((void)(arch), silk_LPC_inverse_pred_gain_c(A_Q12, order)) +#endif + /********************************************************************/ /* SCALAR FUNCTIONS */ /********************************************************************/ @@ -271,7 +286,17 @@ void silk_A2NLSF( void silk_NLSF2A( opus_int16 *a_Q12, /* O monic whitening filter coefficients in Q12, [ d ] */ const opus_int16 *NLSF, /* I normalized line spectral frequencies in Q15, [ d ] */ - const opus_int d /* I filter order (should be even) */ + const opus_int d, /* I filter order (should be even) */ + int arch /* I Run-time architecture */ +); + +/* Convert int32 coefficients to int16 coefs and make sure there's no wrap-around */ +void silk_LPC_fit( + opus_int16 *a_QOUT, /* O Output signal */ + opus_int32 *a_QIN, /* I/O Input signal */ + const opus_int QOUT, /* I Input Q domain */ + const opus_int QIN, /* I Input Q domain */ + const opus_int d /* I Filter order */ ); void silk_insertion_sort_increasing( @@ -471,8 +496,7 @@ static OPUS_INLINE opus_int32 silk_ROR32( opus_int32 a32, opus_int rot ) /* Add with saturation for positive input values */ #define silk_ADD_POS_SAT8(a, b) ((((a)+(b)) & 0x80) ? silk_int8_MAX : ((a)+(b))) #define silk_ADD_POS_SAT16(a, b) ((((a)+(b)) & 0x8000) ? silk_int16_MAX : ((a)+(b))) -#define silk_ADD_POS_SAT32(a, b) ((((a)+(b)) & 0x80000000) ? silk_int32_MAX : ((a)+(b))) -#define silk_ADD_POS_SAT64(a, b) ((((a)+(b)) & 0x8000000000000000LL) ? silk_int64_MAX : ((a)+(b))) +#define silk_ADD_POS_SAT32(a, b) ((((opus_uint32)(a)+(opus_uint32)(b)) & 0x80000000) ? silk_int32_MAX : ((a)+(b))) #define silk_LSHIFT8(a, shift) ((opus_int8)((opus_uint8)(a)<<(shift))) /* shift >= 0, shift < 8 */ #define silk_LSHIFT16(a, shift) ((opus_int16)((opus_uint16)(a)<<(shift))) /* shift >= 0, shift < 16 */ @@ -572,7 +596,9 @@ static OPUS_INLINE opus_int64 silk_max_64(opus_int64 a, opus_int64 b) /* Make sure to store the result as the seed for the next call (also in between */ /* frames), otherwise result won't be random at all. When only using some of the */ /* bits, take the most significant bits by right-shifting. */ -#define silk_RAND(seed) (silk_MLA_ovflw(907633515, (seed), 196314165)) +#define RAND_MULTIPLIER 196314165 +#define RAND_INCREMENT 907633515 +#define silk_RAND(seed) (silk_MLA_ovflw((RAND_INCREMENT), (seed), (RAND_MULTIPLIER))) /* Add some multiplication functions that can be easily mapped to ARM. */ diff --git a/silk/VQ_WMat_EC.c b/silk/VQ_WMat_EC.c index 7983f1db..0f3d545c 100644 --- a/silk/VQ_WMat_EC.c +++ b/silk/VQ_WMat_EC.c @@ -34,84 +34,95 @@ POSSIBILITY OF SUCH DAMAGE. /* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */ void silk_VQ_WMat_EC_c( opus_int8 *ind, /* O index of best codebook vector */ - opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */ + opus_int32 *res_nrg_Q15, /* O best residual energy */ + opus_int32 *rate_dist_Q8, /* O best total bitrate */ opus_int *gain_Q7, /* O sum of absolute LTP coefficients */ - const opus_int16 *in_Q14, /* I input vector to be quantized */ - const opus_int32 *W_Q18, /* I weighting matrix */ + const opus_int32 *XX_Q17, /* I correlation matrix */ + const opus_int32 *xX_Q17, /* I correlation vector */ const opus_int8 *cb_Q7, /* I codebook */ const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */ const opus_uint8 *cl_Q5, /* I code length for each codebook vector */ - const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */ + const opus_int subfr_len, /* I number of samples per subframe */ const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ - opus_int L /* I number of vectors in codebook */ + const opus_int L /* I number of vectors in codebook */ ) { opus_int k, gain_tmp_Q7; const opus_int8 *cb_row_Q7; - opus_int16 diff_Q14[ 5 ]; - opus_int32 sum1_Q14, sum2_Q16; + opus_int32 neg_xX_Q24[ 5 ]; + opus_int32 sum1_Q15, sum2_Q24; + opus_int32 bits_res_Q8, bits_tot_Q8; + + /* Negate and convert to new Q domain */ + neg_xX_Q24[ 0 ] = -silk_LSHIFT32( xX_Q17[ 0 ], 7 ); + neg_xX_Q24[ 1 ] = -silk_LSHIFT32( xX_Q17[ 1 ], 7 ); + neg_xX_Q24[ 2 ] = -silk_LSHIFT32( xX_Q17[ 2 ], 7 ); + neg_xX_Q24[ 3 ] = -silk_LSHIFT32( xX_Q17[ 3 ], 7 ); + neg_xX_Q24[ 4 ] = -silk_LSHIFT32( xX_Q17[ 4 ], 7 ); /* Loop over codebook */ - *rate_dist_Q14 = silk_int32_MAX; + *rate_dist_Q8 = silk_int32_MAX; + *res_nrg_Q15 = silk_int32_MAX; cb_row_Q7 = cb_Q7; + /* In things go really bad, at least *ind is set to something safe. */ + *ind = 0; for( k = 0; k < L; k++ ) { + opus_int32 penalty; gain_tmp_Q7 = cb_gain_Q7[k]; - - diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 ); - diff_Q14[ 1 ] = in_Q14[ 1 ] - silk_LSHIFT( cb_row_Q7[ 1 ], 7 ); - diff_Q14[ 2 ] = in_Q14[ 2 ] - silk_LSHIFT( cb_row_Q7[ 2 ], 7 ); - diff_Q14[ 3 ] = in_Q14[ 3 ] - silk_LSHIFT( cb_row_Q7[ 3 ], 7 ); - diff_Q14[ 4 ] = in_Q14[ 4 ] - silk_LSHIFT( cb_row_Q7[ 4 ], 7 ); - /* Weighted rate */ - sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] ); + /* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */ + sum1_Q15 = SILK_FIX_CONST( 1.001, 15 ); /* Penalty for too large gain */ - sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 ); - - silk_assert( sum1_Q14 >= 0 ); - - /* first row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 1 ], diff_Q14[ 1 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 2 ], diff_Q14[ 2 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 3 ], diff_Q14[ 3 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 4 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 0 ], diff_Q14[ 0 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 0 ] ); - - /* second row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 7 ], diff_Q14[ 2 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 8 ], diff_Q14[ 3 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 9 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 6 ], diff_Q14[ 1 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 1 ] ); - - /* third row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 13 ], diff_Q14[ 3 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 14 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 12 ], diff_Q14[ 2 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 2 ] ); - - /* fourth row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 19 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 18 ], diff_Q14[ 3 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 3 ] ); - - /* last row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 24 ], diff_Q14[ 4 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 4 ] ); - - silk_assert( sum1_Q14 >= 0 ); + penalty = silk_LSHIFT32( silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 11 ); + + /* first row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 0 ], XX_Q17[ 1 ], cb_row_Q7[ 1 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 2 ], cb_row_Q7[ 2 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 3 ], cb_row_Q7[ 3 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 4 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 0 ], cb_row_Q7[ 0 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 0 ] ); + + /* second row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 1 ], XX_Q17[ 7 ], cb_row_Q7[ 2 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 8 ], cb_row_Q7[ 3 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 9 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 6 ], cb_row_Q7[ 1 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 1 ] ); + + /* third row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 2 ], XX_Q17[ 13 ], cb_row_Q7[ 3 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 14 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 12 ], cb_row_Q7[ 2 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 2 ] ); + + /* fourth row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 3 ], XX_Q17[ 19 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 18 ], cb_row_Q7[ 3 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 3 ] ); + + /* last row of XX_Q17 */ + sum2_Q24 = silk_LSHIFT32( neg_xX_Q24[ 4 ], 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 24 ], cb_row_Q7[ 4 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 4 ] ); /* find best */ - if( sum1_Q14 < *rate_dist_Q14 ) { - *rate_dist_Q14 = sum1_Q14; - *ind = (opus_int8)k; - *gain_Q7 = gain_tmp_Q7; + if( sum1_Q15 >= 0 ) { + /* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */ + bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 + penalty) - (15 << 7) ); + /* In the following line we reduce the codelength component by half ("-1"); seems to slghtly improve quality */ + bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 ); + if( bits_tot_Q8 <= *rate_dist_Q8 ) { + *rate_dist_Q8 = bits_tot_Q8; + *res_nrg_Q15 = sum1_Q15 + penalty; + *ind = (opus_int8)k; + *gain_Q7 = gain_tmp_Q7; + } } /* Go to next cbk vector */ diff --git a/silk/arm/LPC_inv_pred_gain_arm.h b/silk/arm/LPC_inv_pred_gain_arm.h new file mode 100644 index 00000000..9895b555 --- /dev/null +++ b/silk/arm/LPC_inv_pred_gain_arm.h @@ -0,0 +1,57 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifndef SILK_LPC_INV_PRED_GAIN_ARM_H +# define SILK_LPC_INV_PRED_GAIN_ARM_H + +# include "celt/arm/armcpu.h" + +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +opus_int32 silk_LPC_inverse_pred_gain_neon( /* O Returns inverse prediction gain in energy domain, Q30 */ + const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ + const opus_int order /* I Prediction order */ +); + +# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON) +# define OVERRIDE_silk_LPC_inverse_pred_gain (1) +# define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((void)(arch), PRESUME_NEON(silk_LPC_inverse_pred_gain)(A_Q12, order)) +# endif +# endif + +# if !defined(OVERRIDE_silk_LPC_inverse_pred_gain) +/*Is run-time CPU detection enabled on this platform?*/ +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern opus_int32 (*const SILK_LPC_INVERSE_PRED_GAIN_IMPL[OPUS_ARCHMASK+1])(const opus_int16 *A_Q12, const opus_int order); +# define OVERRIDE_silk_LPC_inverse_pred_gain (1) +# define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((*SILK_LPC_INVERSE_PRED_GAIN_IMPL[(arch)&OPUS_ARCHMASK])(A_Q12, order)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_silk_LPC_inverse_pred_gain (1) +# define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((void)(arch), silk_LPC_inverse_pred_gain_neon(A_Q12, order)) +# endif +# endif + +#endif /* end SILK_LPC_INV_PRED_GAIN_ARM_H */ diff --git a/silk/arm/LPC_inv_pred_gain_neon_intr.c b/silk/arm/LPC_inv_pred_gain_neon_intr.c new file mode 100644 index 00000000..27142f34 --- /dev/null +++ b/silk/arm/LPC_inv_pred_gain_neon_intr.c @@ -0,0 +1,280 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include <arm_neon.h> +#include "SigProc_FIX.h" +#include "define.h" + +#define QA 24 +#define A_LIMIT SILK_FIX_CONST( 0.99975, QA ) + +#define MUL32_FRAC_Q(a32, b32, Q) ((opus_int32)(silk_RSHIFT_ROUND64(silk_SMULL(a32, b32), Q))) + +/* The difficulty is how to judge a 64-bit signed integer tmp64 is 32-bit overflowed, + * since NEON has no 64-bit min, max or comparison instructions. + * A failed idea is to compare the results of vmovn(tmp64) and vqmovn(tmp64) whether they are equal or not. + * However, this idea fails when the tmp64 is something like 0xFFFFFFF980000000. + * Here we know that mult2Q >= 1, so the highest bit (bit 63, sign bit) of tmp64 must equal to bit 62. + * tmp64 was shifted left by 1 and we got tmp64'. If high_half(tmp64') != 0 and high_half(tmp64') != -1, + * then we know that bit 31 to bit 63 of tmp64 can not all be the sign bit, and therefore tmp64 is 32-bit overflowed. + * That is, we judge if tmp64' > 0x00000000FFFFFFFF, or tmp64' <= 0xFFFFFFFF00000000. + * We use narrowing shift right 31 bits to tmp32' to save data bandwidth and instructions. + * That is, we judge if tmp32' > 0x00000000, or tmp32' <= 0xFFFFFFFF. + */ + +/* Compute inverse of LPC prediction gain, and */ +/* test if LPC coefficients are stable (all poles within unit circle) */ +static OPUS_INLINE opus_int32 LPC_inverse_pred_gain_QA_neon( /* O Returns inverse prediction gain in energy domain, Q30 */ + opus_int32 A_QA[ SILK_MAX_ORDER_LPC ], /* I Prediction coefficients */ + const opus_int order /* I Prediction order */ +) +{ + opus_int k, n, mult2Q; + opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp1, tmp2; + opus_int32 max, min; + int32x4_t max_s32x4, min_s32x4; + int32x2_t max_s32x2, min_s32x2; + + max_s32x4 = vdupq_n_s32( silk_int32_MIN ); + min_s32x4 = vdupq_n_s32( silk_int32_MAX ); + invGain_Q30 = SILK_FIX_CONST( 1, 30 ); + for( k = order - 1; k > 0; k-- ) { + int32x2_t rc_Q31_s32x2, rc_mult2_s32x2; + int64x2_t mult2Q_s64x2; + + /* Check for stability */ + if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) { + return 0; + } + + /* Set RC equal to negated AR coef */ + rc_Q31 = -silk_LSHIFT( A_QA[ k ], 31 - QA ); + + /* rc_mult1_Q30 range: [ 1 : 2^30 ] */ + rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) ); + silk_assert( rc_mult1_Q30 > ( 1 << 15 ) ); /* reduce A_LIMIT if fails */ + silk_assert( rc_mult1_Q30 <= ( 1 << 30 ) ); + + /* Update inverse gain */ + /* invGain_Q30 range: [ 0 : 2^30 ] */ + invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 ); + silk_assert( invGain_Q30 >= 0 ); + silk_assert( invGain_Q30 <= ( 1 << 30 ) ); + if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) { + return 0; + } + + /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */ + mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) ); + rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 ); + + /* Update AR coefficient */ + rc_Q31_s32x2 = vdup_n_s32( rc_Q31 ); + mult2Q_s64x2 = vdupq_n_s64( -mult2Q ); + rc_mult2_s32x2 = vdup_n_s32( rc_mult2 ); + + for( n = 0; n < ( ( k + 1 ) >> 1 ) - 3; n += 4 ) { + /* We always calculate extra elements of A_QA buffer when ( k % 4 ) != 0, to take the advantage of SIMD parallelization. */ + int32x4_t tmp1_s32x4, tmp2_s32x4, t0_s32x4, t1_s32x4, s0_s32x4, s1_s32x4, t_QA0_s32x4, t_QA1_s32x4; + int64x2_t t0_s64x2, t1_s64x2, t2_s64x2, t3_s64x2; + tmp1_s32x4 = vld1q_s32( A_QA + n ); + tmp2_s32x4 = vld1q_s32( A_QA + k - n - 4 ); + tmp2_s32x4 = vrev64q_s32( tmp2_s32x4 ); + tmp2_s32x4 = vcombine_s32( vget_high_s32( tmp2_s32x4 ), vget_low_s32( tmp2_s32x4 ) ); + t0_s32x4 = vqrdmulhq_lane_s32( tmp2_s32x4, rc_Q31_s32x2, 0 ); + t1_s32x4 = vqrdmulhq_lane_s32( tmp1_s32x4, rc_Q31_s32x2, 0 ); + t_QA0_s32x4 = vqsubq_s32( tmp1_s32x4, t0_s32x4 ); + t_QA1_s32x4 = vqsubq_s32( tmp2_s32x4, t1_s32x4 ); + t0_s64x2 = vmull_s32( vget_low_s32 ( t_QA0_s32x4 ), rc_mult2_s32x2 ); + t1_s64x2 = vmull_s32( vget_high_s32( t_QA0_s32x4 ), rc_mult2_s32x2 ); + t2_s64x2 = vmull_s32( vget_low_s32 ( t_QA1_s32x4 ), rc_mult2_s32x2 ); + t3_s64x2 = vmull_s32( vget_high_s32( t_QA1_s32x4 ), rc_mult2_s32x2 ); + t0_s64x2 = vrshlq_s64( t0_s64x2, mult2Q_s64x2 ); + t1_s64x2 = vrshlq_s64( t1_s64x2, mult2Q_s64x2 ); + t2_s64x2 = vrshlq_s64( t2_s64x2, mult2Q_s64x2 ); + t3_s64x2 = vrshlq_s64( t3_s64x2, mult2Q_s64x2 ); + t0_s32x4 = vcombine_s32( vmovn_s64( t0_s64x2 ), vmovn_s64( t1_s64x2 ) ); + t1_s32x4 = vcombine_s32( vmovn_s64( t2_s64x2 ), vmovn_s64( t3_s64x2 ) ); + s0_s32x4 = vcombine_s32( vshrn_n_s64( t0_s64x2, 31 ), vshrn_n_s64( t1_s64x2, 31 ) ); + s1_s32x4 = vcombine_s32( vshrn_n_s64( t2_s64x2, 31 ), vshrn_n_s64( t3_s64x2, 31 ) ); + max_s32x4 = vmaxq_s32( max_s32x4, s0_s32x4 ); + min_s32x4 = vminq_s32( min_s32x4, s0_s32x4 ); + max_s32x4 = vmaxq_s32( max_s32x4, s1_s32x4 ); + min_s32x4 = vminq_s32( min_s32x4, s1_s32x4 ); + t1_s32x4 = vrev64q_s32( t1_s32x4 ); + t1_s32x4 = vcombine_s32( vget_high_s32( t1_s32x4 ), vget_low_s32( t1_s32x4 ) ); + vst1q_s32( A_QA + n, t0_s32x4 ); + vst1q_s32( A_QA + k - n - 4, t1_s32x4 ); + } + for( ; n < (k + 1) >> 1; n++ ) { + opus_int64 tmp64; + tmp1 = A_QA[ n ]; + tmp2 = A_QA[ k - n - 1 ]; + tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp1, + MUL32_FRAC_Q( tmp2, rc_Q31, 31 ) ), rc_mult2 ), mult2Q); + if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) { + return 0; + } + A_QA[ n ] = ( opus_int32 )tmp64; + tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp2, + MUL32_FRAC_Q( tmp1, rc_Q31, 31 ) ), rc_mult2), mult2Q); + if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) { + return 0; + } + A_QA[ k - n - 1 ] = ( opus_int32 )tmp64; + } + } + + /* Check for stability */ + if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) { + return 0; + } + + max_s32x2 = vmax_s32( vget_low_s32( max_s32x4 ), vget_high_s32( max_s32x4 ) ); + min_s32x2 = vmin_s32( vget_low_s32( min_s32x4 ), vget_high_s32( min_s32x4 ) ); + max_s32x2 = vmax_s32( max_s32x2, vreinterpret_s32_s64( vshr_n_s64( vreinterpret_s64_s32( max_s32x2 ), 32 ) ) ); + min_s32x2 = vmin_s32( min_s32x2, vreinterpret_s32_s64( vshr_n_s64( vreinterpret_s64_s32( min_s32x2 ), 32 ) ) ); + max = vget_lane_s32( max_s32x2, 0 ); + min = vget_lane_s32( min_s32x2, 0 ); + if( ( max > 0 ) || ( min < -1 ) ) { + return 0; + } + + /* Set RC equal to negated AR coef */ + rc_Q31 = -silk_LSHIFT( A_QA[ 0 ], 31 - QA ); + + /* Range: [ 1 : 2^30 ] */ + rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) ); + + /* Update inverse gain */ + /* Range: [ 0 : 2^30 ] */ + invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 ); + silk_assert( invGain_Q30 >= 0 ); + silk_assert( invGain_Q30 <= ( 1 << 30 ) ); + if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) { + return 0; + } + + return invGain_Q30; +} + +/* For input in Q12 domain */ +opus_int32 silk_LPC_inverse_pred_gain_neon( /* O Returns inverse prediction gain in energy domain, Q30 */ + const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ + const opus_int order /* I Prediction order */ +) +{ +#ifdef OPUS_CHECK_ASM + const opus_int32 invGain_Q30_c = silk_LPC_inverse_pred_gain_c( A_Q12, order ); +#endif + + opus_int32 invGain_Q30; + if( ( SILK_MAX_ORDER_LPC != 24 ) || ( order & 1 )) { + invGain_Q30 = silk_LPC_inverse_pred_gain_c( A_Q12, order ); + } + else { + opus_int32 Atmp_QA[ SILK_MAX_ORDER_LPC ]; + opus_int32 DC_resp; + int16x8_t t0_s16x8, t1_s16x8, t2_s16x8; + int32x4_t t0_s32x4; + const opus_int leftover = order & 7; + + /* Increase Q domain of the AR coefficients */ + t0_s16x8 = vld1q_s16( A_Q12 + 0 ); + t1_s16x8 = vld1q_s16( A_Q12 + 8 ); + t2_s16x8 = vld1q_s16( A_Q12 + 16 ); + t0_s32x4 = vpaddlq_s16( t0_s16x8 ); + + switch( order - leftover ) + { + case 24: + t0_s32x4 = vpadalq_s16( t0_s32x4, t2_s16x8 ); + /* Intend to fall through */ + + case 16: + t0_s32x4 = vpadalq_s16( t0_s32x4, t1_s16x8 ); + vst1q_s32( Atmp_QA + 16, vshll_n_s16( vget_low_s16 ( t2_s16x8 ), QA - 12 ) ); + vst1q_s32( Atmp_QA + 20, vshll_n_s16( vget_high_s16( t2_s16x8 ), QA - 12 ) ); + /* Intend to fall through */ + + case 8: + { + const int32x2_t t_s32x2 = vpadd_s32( vget_low_s32( t0_s32x4 ), vget_high_s32( t0_s32x4 ) ); + const int64x1_t t_s64x1 = vpaddl_s32( t_s32x2 ); + DC_resp = vget_lane_s32( vreinterpret_s32_s64( t_s64x1 ), 0 ); + vst1q_s32( Atmp_QA + 8, vshll_n_s16( vget_low_s16 ( t1_s16x8 ), QA - 12 ) ); + vst1q_s32( Atmp_QA + 12, vshll_n_s16( vget_high_s16( t1_s16x8 ), QA - 12 ) ); + } + break; + + default: + DC_resp = 0; + break; + } + A_Q12 += order - leftover; + + switch( leftover ) + { + case 6: + DC_resp += (opus_int32)A_Q12[ 5 ]; + DC_resp += (opus_int32)A_Q12[ 4 ]; + /* Intend to fall through */ + + case 4: + DC_resp += (opus_int32)A_Q12[ 3 ]; + DC_resp += (opus_int32)A_Q12[ 2 ]; + /* Intend to fall through */ + + case 2: + DC_resp += (opus_int32)A_Q12[ 1 ]; + DC_resp += (opus_int32)A_Q12[ 0 ]; + /* Intend to fall through */ + + default: + break; + } + + /* If the DC is unstable, we don't even need to do the full calculations */ + if( DC_resp >= 4096 ) { + invGain_Q30 = 0; + } else { + vst1q_s32( Atmp_QA + 0, vshll_n_s16( vget_low_s16 ( t0_s16x8 ), QA - 12 ) ); + vst1q_s32( Atmp_QA + 4, vshll_n_s16( vget_high_s16( t0_s16x8 ), QA - 12 ) ); + invGain_Q30 = LPC_inverse_pred_gain_QA_neon( Atmp_QA, order ); + } + } + +#ifdef OPUS_CHECK_ASM + silk_assert( invGain_Q30_c == invGain_Q30 ); +#endif + + return invGain_Q30; +} diff --git a/silk/arm/NSQ_del_dec_arm.h b/silk/arm/NSQ_del_dec_arm.h new file mode 100644 index 00000000..9e76e169 --- /dev/null +++ b/silk/arm/NSQ_del_dec_arm.h @@ -0,0 +1,100 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifndef SILK_NSQ_DEL_DEC_ARM_H +#define SILK_NSQ_DEL_DEC_ARM_H + +#include "celt/arm/armcpu.h" + +#if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +void silk_NSQ_del_dec_neon( + const silk_encoder_state *psEncC, silk_nsq_state *NSQ, + SideInfoIndices *psIndices, const opus_int16 x16[], opus_int8 pulses[], + const opus_int16 PredCoef_Q12[2 * MAX_LPC_ORDER], + const opus_int16 LTPCoef_Q14[LTP_ORDER * MAX_NB_SUBFR], + const opus_int16 AR_Q13[MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER], + const opus_int HarmShapeGain_Q14[MAX_NB_SUBFR], + const opus_int Tilt_Q14[MAX_NB_SUBFR], + const opus_int32 LF_shp_Q14[MAX_NB_SUBFR], + const opus_int32 Gains_Q16[MAX_NB_SUBFR], + const opus_int pitchL[MAX_NB_SUBFR], const opus_int Lambda_Q10, + const opus_int LTP_scale_Q14); + +#if !defined(OPUS_HAVE_RTCD) +#define OVERRIDE_silk_NSQ_del_dec (1) +#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \ + LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \ + LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \ + LTP_scale_Q14, arch) \ + ((void)(arch), \ + PRESUME_NEON(silk_NSQ_del_dec)( \ + psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, \ + AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, \ + Lambda_Q10, LTP_scale_Q14)) +#endif +#endif + +#if !defined(OVERRIDE_silk_NSQ_del_dec) +/*Is run-time CPU detection enabled on this platform?*/ +#if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && \ + !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])( + const silk_encoder_state *psEncC, silk_nsq_state *NSQ, + SideInfoIndices *psIndices, const opus_int16 x16[], opus_int8 pulses[], + const opus_int16 PredCoef_Q12[2 * MAX_LPC_ORDER], + const opus_int16 LTPCoef_Q14[LTP_ORDER * MAX_NB_SUBFR], + const opus_int16 AR_Q13[MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER], + const opus_int HarmShapeGain_Q14[MAX_NB_SUBFR], + const opus_int Tilt_Q14[MAX_NB_SUBFR], + const opus_int32 LF_shp_Q14[MAX_NB_SUBFR], + const opus_int32 Gains_Q16[MAX_NB_SUBFR], + const opus_int pitchL[MAX_NB_SUBFR], const opus_int Lambda_Q10, + const opus_int LTP_scale_Q14); +#define OVERRIDE_silk_NSQ_del_dec (1) +#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \ + LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \ + LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \ + LTP_scale_Q14, arch) \ + ((*SILK_NSQ_DEL_DEC_IMPL[(arch)&OPUS_ARCHMASK])( \ + psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, \ + AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, \ + Lambda_Q10, LTP_scale_Q14)) +#elif defined(OPUS_ARM_PRESUME_NEON_INTR) +#define OVERRIDE_silk_NSQ_del_dec (1) +#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \ + LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \ + LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \ + LTP_scale_Q14, arch) \ + ((void)(arch), \ + silk_NSQ_del_dec_neon(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \ + LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \ + LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \ + LTP_scale_Q14)) +#endif +#endif + +#endif /* end SILK_NSQ_DEL_DEC_ARM_H */ diff --git a/silk/arm/NSQ_del_dec_neon_intr.c b/silk/arm/NSQ_del_dec_neon_intr.c new file mode 100644 index 00000000..212410f3 --- /dev/null +++ b/silk/arm/NSQ_del_dec_neon_intr.c @@ -0,0 +1,1124 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include <arm_neon.h> +#ifdef OPUS_CHECK_ASM +# include <string.h> +#endif +#include "main.h" +#include "stack_alloc.h" + +/* NEON intrinsics optimization now can only parallelize up to 4 delay decision states. */ +/* If there are more states, C function is called, and this optimization must be expanded. */ +#define NEON_MAX_DEL_DEC_STATES 4 + +typedef struct { + opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 RandState[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Q_Q10[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Xq_Q14[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Pred_Q15[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Shape_Q14[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 LF_AR_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Diff_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Seed[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 SeedInit[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 RD_Q10[ NEON_MAX_DEL_DEC_STATES ]; +} NSQ_del_decs_struct; + +typedef struct { + opus_int32 Q_Q10[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 RD_Q10[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 xq_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 LF_AR_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Diff_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 sLTP_shp_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 LPC_exc_Q14[ NEON_MAX_DEL_DEC_STATES ]; +} NSQ_samples_struct; + +static OPUS_INLINE void silk_nsq_del_dec_scale_states_neon( + const silk_encoder_state *psEncC, /* I Encoder State */ + silk_nsq_state *NSQ, /* I/O NSQ state */ + NSQ_del_decs_struct psDelDec[], /* I/O Delayed decision states */ + const opus_int16 x16[], /* I Input */ + opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ + const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ + opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ + opus_int subfr, /* I Subframe number */ + const opus_int LTP_scale_Q14, /* I LTP state scaling */ + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */ + const opus_int signal_type, /* I Signal type */ + const opus_int decisionDelay /* I Decision delay */ +); + +/******************************************/ +/* Noise shape quantizer for one subframe */ +/******************************************/ +static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_neon( + silk_nsq_state *NSQ, /* I/O NSQ state */ + NSQ_del_decs_struct psDelDec[], /* I/O Delayed decision states */ + opus_int signalType, /* I Signal type */ + const opus_int32 x_Q10[], /* I */ + opus_int8 pulses[], /* O */ + opus_int16 xq[], /* O */ + opus_int32 sLTP_Q15[], /* I/O LTP filter state */ + opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */ + const opus_int16 a_Q12[], /* I Short term prediction coefs */ + const opus_int16 b_Q14[], /* I Long term prediction coefs */ + const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */ + opus_int lag, /* I Pitch lag */ + opus_int32 HarmShapeFIRPacked_Q14, /* I */ + opus_int Tilt_Q14, /* I Spectral tilt */ + opus_int32 LF_shp_Q14, /* I */ + opus_int32 Gain_Q16, /* I */ + opus_int Lambda_Q10, /* I */ + opus_int offset_Q10, /* I */ + opus_int length, /* I Input length */ + opus_int subfr, /* I Subframe number */ + opus_int shapingLPCOrder, /* I Shaping LPC filter order */ + opus_int predictLPCOrder, /* I Prediction filter order */ + opus_int warping_Q16, /* I */ + opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ + opus_int decisionDelay /* I */ +); + +static OPUS_INLINE void copy_winner_state_kernel( + const NSQ_del_decs_struct *psDelDec, + const opus_int offset, + const opus_int last_smple_idx, + const opus_int Winner_ind, + const int32x2_t gain_lo_s32x2, + const int32x2_t gain_hi_s32x2, + const int32x4_t shift_s32x4, + int32x4_t t0_s32x4, + int32x4_t t1_s32x4, + opus_int8 *const pulses, + opus_int16 *pxq, + silk_nsq_state *NSQ +) +{ + int16x8_t t_s16x8; + int32x4_t o0_s32x4, o1_s32x4; + + t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 0 ][ Winner_ind ], t0_s32x4, 0 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 1 ][ Winner_ind ], t0_s32x4, 1 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 2 ][ Winner_ind ], t0_s32x4, 2 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 3 ][ Winner_ind ], t0_s32x4, 3 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 4 ][ Winner_ind ], t1_s32x4, 0 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 5 ][ Winner_ind ], t1_s32x4, 1 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 6 ][ Winner_ind ], t1_s32x4, 2 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 7 ][ Winner_ind ], t1_s32x4, 3 ); + t_s16x8 = vcombine_s16( vrshrn_n_s32( t0_s32x4, 10 ), vrshrn_n_s32( t1_s32x4, 10 ) ); + vst1_s8( &pulses[ offset ], vmovn_s16( t_s16x8 ) ); + + t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 0 ][ Winner_ind ], t0_s32x4, 0 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 1 ][ Winner_ind ], t0_s32x4, 1 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 2 ][ Winner_ind ], t0_s32x4, 2 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 3 ][ Winner_ind ], t0_s32x4, 3 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 4 ][ Winner_ind ], t1_s32x4, 0 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 5 ][ Winner_ind ], t1_s32x4, 1 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 6 ][ Winner_ind ], t1_s32x4, 2 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 7 ][ Winner_ind ], t1_s32x4, 3 ); + o0_s32x4 = vqdmulhq_lane_s32( t0_s32x4, gain_lo_s32x2, 0 ); + o1_s32x4 = vqdmulhq_lane_s32( t1_s32x4, gain_lo_s32x2, 0 ); + o0_s32x4 = vmlaq_lane_s32( o0_s32x4, t0_s32x4, gain_hi_s32x2, 0 ); + o1_s32x4 = vmlaq_lane_s32( o1_s32x4, t1_s32x4, gain_hi_s32x2, 0 ); + o0_s32x4 = vrshlq_s32( o0_s32x4, shift_s32x4 ); + o1_s32x4 = vrshlq_s32( o1_s32x4, shift_s32x4 ); + vst1_s16( &pxq[ offset + 0 ], vqmovn_s32( o0_s32x4 ) ); + vst1_s16( &pxq[ offset + 4 ], vqmovn_s32( o1_s32x4 ) ); + + t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 0 ][ Winner_ind ], t0_s32x4, 0 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 1 ][ Winner_ind ], t0_s32x4, 1 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 2 ][ Winner_ind ], t0_s32x4, 2 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 3 ][ Winner_ind ], t0_s32x4, 3 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 4 ][ Winner_ind ], t1_s32x4, 0 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 5 ][ Winner_ind ], t1_s32x4, 1 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 6 ][ Winner_ind ], t1_s32x4, 2 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 7 ][ Winner_ind ], t1_s32x4, 3 ); + vst1q_s32( &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx + offset + 0 ], t0_s32x4 ); + vst1q_s32( &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx + offset + 4 ], t1_s32x4 ); +} + +static OPUS_INLINE void copy_winner_state( + const NSQ_del_decs_struct *psDelDec, + const opus_int decisionDelay, + const opus_int smpl_buf_idx, + const opus_int Winner_ind, + const opus_int32 gain, + const opus_int32 shift, + opus_int8 *const pulses, + opus_int16 *pxq, + silk_nsq_state *NSQ +) +{ + opus_int i, last_smple_idx; + const int32x2_t gain_lo_s32x2 = vdup_n_s32( silk_LSHIFT32( gain & 0x0000FFFF, 15 ) ); + const int32x2_t gain_hi_s32x2 = vdup_n_s32( gain >> 16 ); + const int32x4_t shift_s32x4 = vdupq_n_s32( -shift ); + int32x4_t t0_s32x4, t1_s32x4; + + t0_s32x4 = t1_s32x4 = vdupq_n_s32( 0 ); /* initialization */ + last_smple_idx = smpl_buf_idx + decisionDelay - 1 + DECISION_DELAY; + if( last_smple_idx >= DECISION_DELAY ) last_smple_idx -= DECISION_DELAY; + if( last_smple_idx >= DECISION_DELAY ) last_smple_idx -= DECISION_DELAY; + + for( i = 0; ( i < ( decisionDelay - 7 ) ) && ( last_smple_idx >= 7 ); i += 8, last_smple_idx -= 8 ) { + copy_winner_state_kernel( psDelDec, i - decisionDelay, last_smple_idx, Winner_ind, gain_lo_s32x2, gain_hi_s32x2, shift_s32x4, t0_s32x4, t1_s32x4, pulses, pxq, NSQ ); + } + for( ; ( i < decisionDelay ) && ( last_smple_idx >= 0 ); i++, last_smple_idx-- ) { + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDelDec->Q_Q10[ last_smple_idx ][ Winner_ind ], 10 ); + pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDelDec->Xq_Q14[ last_smple_idx ][ Winner_ind ], gain ), shift ) ); + NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDelDec->Shape_Q14[ last_smple_idx ][ Winner_ind ]; + } + + last_smple_idx += DECISION_DELAY; + for( ; i < ( decisionDelay - 7 ); i++, last_smple_idx-- ) { + copy_winner_state_kernel( psDelDec, i - decisionDelay, last_smple_idx, Winner_ind, gain_lo_s32x2, gain_hi_s32x2, shift_s32x4, t0_s32x4, t1_s32x4, pulses, pxq, NSQ ); + } + for( ; i < decisionDelay; i++, last_smple_idx-- ) { + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDelDec->Q_Q10[ last_smple_idx ][ Winner_ind ], 10 ); + pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDelDec->Xq_Q14[ last_smple_idx ][ Winner_ind ], gain ), shift ) ); + NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDelDec->Shape_Q14[ last_smple_idx ][ Winner_ind ]; + } +} + +void silk_NSQ_del_dec_neon( + const silk_encoder_state *psEncC, /* I Encoder State */ + silk_nsq_state *NSQ, /* I/O NSQ state */ + SideInfoIndices *psIndices, /* I/O Quantization Indices */ + const opus_int16 x16[], /* I Input */ + opus_int8 pulses[], /* O Quantized pulse signal */ + const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ + const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ + const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ + const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */ + const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */ + const opus_int LTP_scale_Q14 /* I LTP state scaling */ +) +{ +#ifdef OPUS_CHECK_ASM + silk_nsq_state NSQ_c; + SideInfoIndices psIndices_c; + opus_int8 pulses_c[ MAX_FRAME_LENGTH ]; + const opus_int8 *const pulses_a = pulses; + + ( void )pulses_a; + silk_memcpy( &NSQ_c, NSQ, sizeof( NSQ_c ) ); + silk_memcpy( &psIndices_c, psIndices, sizeof( psIndices_c ) ); + silk_memcpy( pulses_c, pulses, sizeof( pulses_c ) ); + silk_NSQ_del_dec_c( psEncC, &NSQ_c, &psIndices_c, x16, pulses_c, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, + pitchL, Lambda_Q10, LTP_scale_Q14 ); +#endif + + /* The optimization parallelizes the different delay decision states. */ + if(( psEncC->nStatesDelayedDecision > NEON_MAX_DEL_DEC_STATES ) || ( psEncC->nStatesDelayedDecision <= 2 )) { + /* NEON intrinsics optimization now can only parallelize up to 4 delay decision states. */ + /* If there are more states, C function is called, and this optimization must be expanded. */ + /* When the number of delay decision states is less than 3, there are penalties using this */ + /* optimization, and C function is called. */ + /* When the number of delay decision states is 2, it's better to specialize another */ + /* structure NSQ_del_dec2_struct and optimize with shorter NEON registers. (Low priority) */ + silk_NSQ_del_dec_c( psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, + Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14 ); + } else { + opus_int i, k, lag, start_idx, LSF_interpolation_flag, Winner_ind, subfr; + opus_int smpl_buf_idx, decisionDelay; + const opus_int16 *A_Q12, *B_Q14, *AR_shp_Q13; + opus_int16 *pxq; + VARDECL( opus_int32, sLTP_Q15 ); + VARDECL( opus_int16, sLTP ); + opus_int32 HarmShapeFIRPacked_Q14; + opus_int offset_Q10; + opus_int32 RDmin_Q10, Gain_Q10; + VARDECL( opus_int32, x_sc_Q10 ); + VARDECL( opus_int32, delayedGain_Q10 ); + VARDECL( NSQ_del_decs_struct, psDelDec ); + int32x4_t t_s32x4; + SAVE_STACK; + + /* Set unvoiced lag to the previous one, overwrite later for voiced */ + lag = NSQ->lagPrev; + + silk_assert( NSQ->prev_gain_Q16 != 0 ); + + /* Initialize delayed decision states */ + ALLOC( psDelDec, 1, NSQ_del_decs_struct ); + /* Only RandState and RD_Q10 need to be initialized to 0. */ + silk_memset( psDelDec->RandState, 0, sizeof( psDelDec->RandState ) ); + vst1q_s32( psDelDec->RD_Q10, vdupq_n_s32( 0 ) ); + + for( k = 0; k < psEncC->nStatesDelayedDecision; k++ ) { + psDelDec->SeedInit[ k ] = psDelDec->Seed[ k ] = ( k + psIndices->Seed ) & 3; + } + vst1q_s32( psDelDec->LF_AR_Q14, vld1q_dup_s32( &NSQ->sLF_AR_shp_Q14 ) ); + vst1q_s32( psDelDec->Diff_Q14, vld1q_dup_s32( &NSQ->sDiff_shp_Q14 ) ); + vst1q_s32( psDelDec->Shape_Q14[ 0 ], vld1q_dup_s32( &NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ] ) ); + for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { + vst1q_s32( psDelDec->sLPC_Q14[ i ], vld1q_dup_s32( &NSQ->sLPC_Q14[ i ] ) ); + } + for( i = 0; i < (opus_int)( sizeof( NSQ->sAR2_Q14 ) / sizeof( NSQ->sAR2_Q14[ 0 ] ) ); i++ ) { + vst1q_s32( psDelDec->sAR2_Q14[ i ], vld1q_dup_s32( &NSQ->sAR2_Q14[ i ] ) ); + } + + offset_Q10 = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ]; + smpl_buf_idx = 0; /* index of oldest samples */ + + decisionDelay = silk_min_int( DECISION_DELAY, psEncC->subfr_length ); + + /* For voiced frames limit the decision delay to lower than the pitch lag */ + if( psIndices->signalType == TYPE_VOICED ) { + opus_int pitch_min = pitchL[ 0 ]; + for( k = 1; k < psEncC->nb_subfr; k++ ) { + pitch_min = silk_min_int( pitch_min, pitchL[ k ] ); + } + decisionDelay = silk_min_int( decisionDelay, pitch_min - LTP_ORDER / 2 - 1 ); + } else { + if( lag > 0 ) { + decisionDelay = silk_min_int( decisionDelay, lag - LTP_ORDER / 2 - 1 ); + } + } + + if( psIndices->NLSFInterpCoef_Q2 == 4 ) { + LSF_interpolation_flag = 0; + } else { + LSF_interpolation_flag = 1; + } + + ALLOC( sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); + ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 ); + ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 ); + ALLOC( delayedGain_Q10, DECISION_DELAY, opus_int32 ); + /* Set up pointers to start of sub frame */ + pxq = &NSQ->xq[ psEncC->ltp_mem_length ]; + NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length; + NSQ->sLTP_buf_idx = psEncC->ltp_mem_length; + subfr = 0; + for( k = 0; k < psEncC->nb_subfr; k++ ) { + A_Q12 = &PredCoef_Q12[ ( ( k >> 1 ) | ( 1 - LSF_interpolation_flag ) ) * MAX_LPC_ORDER ]; + B_Q14 = <PCoef_Q14[ k * LTP_ORDER ]; + AR_shp_Q13 = &AR_Q13[ k * MAX_SHAPE_LPC_ORDER ]; + + /* Noise shape parameters */ + silk_assert( HarmShapeGain_Q14[ k ] >= 0 ); + HarmShapeFIRPacked_Q14 = silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 ); + HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 ); + + NSQ->rewhite_flag = 0; + if( psIndices->signalType == TYPE_VOICED ) { + /* Voiced */ + lag = pitchL[ k ]; + + /* Re-whitening */ + if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) { + if( k == 2 ) { + /* RESET DELAYED DECISIONS */ + /* Find winner */ + int32x4_t RD_Q10_s32x4; + RDmin_Q10 = psDelDec->RD_Q10[ 0 ]; + Winner_ind = 0; + for( i = 1; i < psEncC->nStatesDelayedDecision; i++ ) { + if( psDelDec->RD_Q10[ i ] < RDmin_Q10 ) { + RDmin_Q10 = psDelDec->RD_Q10[ i ]; + Winner_ind = i; + } + } + psDelDec->RD_Q10[ Winner_ind ] -= ( silk_int32_MAX >> 4 ); + RD_Q10_s32x4 = vld1q_s32( psDelDec->RD_Q10 ); + RD_Q10_s32x4 = vaddq_s32( RD_Q10_s32x4, vdupq_n_s32( silk_int32_MAX >> 4 ) ); + vst1q_s32( psDelDec->RD_Q10, RD_Q10_s32x4 ); + + /* Copy final part of signals from winner state to output and long-term filter states */ + copy_winner_state( psDelDec, decisionDelay, smpl_buf_idx, Winner_ind, Gains_Q16[ 1 ], 14, pulses, pxq, NSQ ); + + subfr = 0; + } + + /* Rewhiten with new A coefs */ + start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2; + silk_assert( start_idx > 0 ); + + silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ], + A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch ); + + NSQ->sLTP_buf_idx = psEncC->ltp_mem_length; + NSQ->rewhite_flag = 1; + } + } + + silk_nsq_del_dec_scale_states_neon( psEncC, NSQ, psDelDec, x16, x_sc_Q10, sLTP, sLTP_Q15, k, + LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay ); + + silk_noise_shape_quantizer_del_dec_neon( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, + delayedGain_Q10, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], + Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder, + psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay ); + + x16 += psEncC->subfr_length; + pulses += psEncC->subfr_length; + pxq += psEncC->subfr_length; + } + + /* Find winner */ + RDmin_Q10 = psDelDec->RD_Q10[ 0 ]; + Winner_ind = 0; + for( k = 1; k < psEncC->nStatesDelayedDecision; k++ ) { + if( psDelDec->RD_Q10[ k ] < RDmin_Q10 ) { + RDmin_Q10 = psDelDec->RD_Q10[ k ]; + Winner_ind = k; + } + } + + /* Copy final part of signals from winner state to output and long-term filter states */ + psIndices->Seed = psDelDec->SeedInit[ Winner_ind ]; + Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 ); + copy_winner_state( psDelDec, decisionDelay, smpl_buf_idx, Winner_ind, Gain_Q10, 8, pulses, pxq, NSQ ); + + t_s32x4 = vdupq_n_s32( 0 ); /* initialization */ + for( i = 0; i < ( NSQ_LPC_BUF_LENGTH - 3 ); i += 4 ) { + t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 0 ][ Winner_ind ], t_s32x4, 0 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 1 ][ Winner_ind ], t_s32x4, 1 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 2 ][ Winner_ind ], t_s32x4, 2 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 3 ][ Winner_ind ], t_s32x4, 3 ); + vst1q_s32( &NSQ->sLPC_Q14[ i ], t_s32x4 ); + } + + for( ; i < NSQ_LPC_BUF_LENGTH; i++ ) { + NSQ->sLPC_Q14[ i ] = psDelDec->sLPC_Q14[ i ][ Winner_ind ]; + } + + for( i = 0; i < (opus_int)( sizeof( NSQ->sAR2_Q14 ) / sizeof( NSQ->sAR2_Q14[ 0 ] ) - 3 ); i += 4 ) { + t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 0 ][ Winner_ind ], t_s32x4, 0 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 1 ][ Winner_ind ], t_s32x4, 1 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 2 ][ Winner_ind ], t_s32x4, 2 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 3 ][ Winner_ind ], t_s32x4, 3 ); + vst1q_s32( &NSQ->sAR2_Q14[ i ], t_s32x4 ); + } + + for( ; i < (opus_int)( sizeof( NSQ->sAR2_Q14 ) / sizeof( NSQ->sAR2_Q14[ 0 ] ) ); i++ ) { + NSQ->sAR2_Q14[ i ] = psDelDec->sAR2_Q14[ i ][ Winner_ind ]; + } + + /* Update states */ + NSQ->sLF_AR_shp_Q14 = psDelDec->LF_AR_Q14[ Winner_ind ]; + NSQ->sDiff_shp_Q14 = psDelDec->Diff_Q14[ Winner_ind ]; + NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; + + /* Save quantized speech signal */ + silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); + silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); + RESTORE_STACK; + } + +#ifdef OPUS_CHECK_ASM + silk_assert( !memcmp( &NSQ_c, NSQ, sizeof( NSQ_c ) ) ); + silk_assert( !memcmp( &psIndices_c, psIndices, sizeof( psIndices_c ) ) ); + silk_assert( !memcmp( pulses_c, pulses_a, sizeof( pulses_c ) ) ); +#endif +} + +/******************************************/ +/* Noise shape quantizer for one subframe */ +/******************************************/ +/* Note: Function silk_short_prediction_create_arch_coef_neon() defined in NSQ_neon.h is actually a hacking C function. */ +/* Therefore here we append "_local" to the NEON function name to avoid confusion. */ +static OPUS_INLINE void silk_short_prediction_create_arch_coef_neon_local(opus_int32 *out, const opus_int16 *in, opus_int order) +{ + int16x8_t t_s16x8; + int32x4_t t0_s32x4, t1_s32x4, t2_s32x4, t3_s32x4; + silk_assert( order == 10 || order == 16 ); + + t_s16x8 = vld1q_s16( in + 0 ); /* 7 6 5 4 3 2 1 0 */ + t_s16x8 = vrev64q_s16( t_s16x8 ); /* 4 5 6 7 0 1 2 3 */ + t2_s32x4 = vshll_n_s16( vget_high_s16( t_s16x8 ), 15 ); /* 4 5 6 7 */ + t3_s32x4 = vshll_n_s16( vget_low_s16( t_s16x8 ), 15 ); /* 0 1 2 3 */ + + if( order == 16 ) { + t_s16x8 = vld1q_s16( in + 8 ); /* F E D C B A 9 8 */ + t_s16x8 = vrev64q_s16( t_s16x8 ); /* C D E F 8 9 A B */ + t0_s32x4 = vshll_n_s16( vget_high_s16( t_s16x8 ), 15 ); /* C D E F */ + t1_s32x4 = vshll_n_s16( vget_low_s16( t_s16x8 ), 15 ); /* 8 9 A B */ + } else { + int16x4_t t_s16x4; + + t0_s32x4 = vdupq_n_s32( 0 ); /* zero zero zero zero */ + t_s16x4 = vld1_s16( in + 6 ); /* 9 8 7 6 */ + t_s16x4 = vrev64_s16( t_s16x4 ); /* 6 7 8 9 */ + t1_s32x4 = vshll_n_s16( t_s16x4, 15 ); + t1_s32x4 = vcombine_s32( vget_low_s32(t0_s32x4), vget_low_s32( t1_s32x4 ) ); /* 8 9 zero zero */ + } + vst1q_s32( out + 0, t0_s32x4 ); + vst1q_s32( out + 4, t1_s32x4 ); + vst1q_s32( out + 8, t2_s32x4 ); + vst1q_s32( out + 12, t3_s32x4 ); +} + +static OPUS_INLINE int32x4_t silk_SMLAWB_lane0_neon( + const int32x4_t out_s32x4, + const int32x4_t in_s32x4, + const int32x2_t coef_s32x2 +) +{ + return vaddq_s32( out_s32x4, vqdmulhq_lane_s32( in_s32x4, coef_s32x2, 0 ) ); +} + +static OPUS_INLINE int32x4_t silk_SMLAWB_lane1_neon( + const int32x4_t out_s32x4, + const int32x4_t in_s32x4, + const int32x2_t coef_s32x2 +) +{ + return vaddq_s32( out_s32x4, vqdmulhq_lane_s32( in_s32x4, coef_s32x2, 1 ) ); +} + +/* Note: This function has different return value than silk_noise_shape_quantizer_short_prediction_neon(). */ +/* Therefore here we append "_local" to the function name to avoid confusion. */ +static OPUS_INLINE int32x4_t silk_noise_shape_quantizer_short_prediction_neon_local(const opus_int32 *buf32, const opus_int32 *a_Q12_arch, opus_int order) +{ + const int32x4_t a_Q12_arch0_s32x4 = vld1q_s32( a_Q12_arch + 0 ); + const int32x4_t a_Q12_arch1_s32x4 = vld1q_s32( a_Q12_arch + 4 ); + const int32x4_t a_Q12_arch2_s32x4 = vld1q_s32( a_Q12_arch + 8 ); + const int32x4_t a_Q12_arch3_s32x4 = vld1q_s32( a_Q12_arch + 12 ); + int32x4_t LPC_pred_Q14_s32x4; + + silk_assert( order == 10 || order == 16 ); + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + LPC_pred_Q14_s32x4 = vdupq_n_s32( silk_RSHIFT( order, 1 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 0 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch0_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 1 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch0_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 2 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch0_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 3 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch0_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 4 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch1_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 5 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch1_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 6 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch1_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 7 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch1_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 8 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch2_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 9 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch2_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 10 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch2_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 11 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch2_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 12 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch3_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 13 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch3_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 14 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch3_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 15 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch3_s32x4 ) ); + + return LPC_pred_Q14_s32x4; +} + +static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_neon( + silk_nsq_state *NSQ, /* I/O NSQ state */ + NSQ_del_decs_struct psDelDec[], /* I/O Delayed decision states */ + opus_int signalType, /* I Signal type */ + const opus_int32 x_Q10[], /* I */ + opus_int8 pulses[], /* O */ + opus_int16 xq[], /* O */ + opus_int32 sLTP_Q15[], /* I/O LTP filter state */ + opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */ + const opus_int16 a_Q12[], /* I Short term prediction coefs */ + const opus_int16 b_Q14[], /* I Long term prediction coefs */ + const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */ + opus_int lag, /* I Pitch lag */ + opus_int32 HarmShapeFIRPacked_Q14, /* I */ + opus_int Tilt_Q14, /* I Spectral tilt */ + opus_int32 LF_shp_Q14, /* I */ + opus_int32 Gain_Q16, /* I */ + opus_int Lambda_Q10, /* I */ + opus_int offset_Q10, /* I */ + opus_int length, /* I Input length */ + opus_int subfr, /* I Subframe number */ + opus_int shapingLPCOrder, /* I Shaping LPC filter order */ + opus_int predictLPCOrder, /* I Prediction filter order */ + opus_int warping_Q16, /* I */ + opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ + opus_int decisionDelay /* I */ +) +{ + opus_int i, j, k, Winner_ind, RDmin_ind, RDmax_ind, last_smple_idx; + opus_int32 Winner_rand_state; + opus_int32 LTP_pred_Q14, n_LTP_Q14; + opus_int32 RDmin_Q10, RDmax_Q10; + opus_int32 Gain_Q10; + opus_int32 *pred_lag_ptr, *shp_lag_ptr; + opus_int32 a_Q12_arch[MAX_LPC_ORDER]; + const int32x2_t warping_Q16_s32x2 = vdup_n_s32( silk_LSHIFT32( warping_Q16, 16 ) >> 1 ); + const opus_int32 LF_shp_Q29 = silk_LSHIFT32( LF_shp_Q14, 16 ) >> 1; + opus_int32 AR_shp_Q28[ MAX_SHAPE_LPC_ORDER ]; + const uint32x4_t rand_multiplier_u32x4 = vdupq_n_u32( RAND_MULTIPLIER ); + const uint32x4_t rand_increment_u32x4 = vdupq_n_u32( RAND_INCREMENT ); + + VARDECL( NSQ_samples_struct, psSampleState ); + SAVE_STACK; + + silk_assert( nStatesDelayedDecision > 0 ); + silk_assert( ( shapingLPCOrder & 1 ) == 0 ); /* check that order is even */ + ALLOC( psSampleState, 2, NSQ_samples_struct ); + + shp_lag_ptr = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ]; + pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ]; + Gain_Q10 = silk_RSHIFT( Gain_Q16, 6 ); + + for( i = 0; i < ( MAX_SHAPE_LPC_ORDER - 7 ); i += 8 ) { + const int16x8_t t_s16x8 = vld1q_s16( AR_shp_Q13 + i ); + vst1q_s32( AR_shp_Q28 + i + 0, vshll_n_s16( vget_low_s16( t_s16x8 ), 15 ) ); + vst1q_s32( AR_shp_Q28 + i + 4, vshll_n_s16( vget_high_s16( t_s16x8 ), 15 ) ); + } + + for( ; i < MAX_SHAPE_LPC_ORDER; i++ ) { + AR_shp_Q28[i] = silk_LSHIFT32( AR_shp_Q13[i], 15 ); + } + + silk_short_prediction_create_arch_coef_neon_local( a_Q12_arch, a_Q12, predictLPCOrder ); + + for( i = 0; i < length; i++ ) { + int32x4_t Seed_s32x4, LPC_pred_Q14_s32x4; + int32x4_t sign_s32x4, tmp1_s32x4, tmp2_s32x4; + int32x4_t n_AR_Q14_s32x4, n_LF_Q14_s32x4; + int32x2_t AR_shp_Q28_s32x2; + int16x4_t r_Q10_s16x4, rr_Q10_s16x4; + + /* Perform common calculations used in all states */ + + /* Long-term prediction */ + if( signalType == TYPE_VOICED ) { + /* Unrolled loop */ + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + LTP_pred_Q14 = 2; + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ 0 ], b_Q14[ 0 ] ); + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -1 ], b_Q14[ 1 ] ); + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -2 ], b_Q14[ 2 ] ); + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -3 ], b_Q14[ 3 ] ); + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -4 ], b_Q14[ 4 ] ); + LTP_pred_Q14 = silk_LSHIFT( LTP_pred_Q14, 1 ); /* Q13 -> Q14 */ + pred_lag_ptr++; + } else { + LTP_pred_Q14 = 0; + } + + /* Long-term shaping */ + if( lag > 0 ) { + /* Symmetric, packed FIR coefficients */ + n_LTP_Q14 = silk_SMULWB( silk_ADD32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 ); + n_LTP_Q14 = silk_SMLAWT( n_LTP_Q14, shp_lag_ptr[ -1 ], HarmShapeFIRPacked_Q14 ); + n_LTP_Q14 = silk_SUB_LSHIFT32( LTP_pred_Q14, n_LTP_Q14, 2 ); /* Q12 -> Q14 */ + shp_lag_ptr++; + } else { + n_LTP_Q14 = 0; + } + + /* Generate dither */ + Seed_s32x4 = vld1q_s32( psDelDec->Seed ); + Seed_s32x4 = vreinterpretq_s32_u32( vmlaq_u32( rand_increment_u32x4, vreinterpretq_u32_s32( Seed_s32x4 ), rand_multiplier_u32x4 ) ); + vst1q_s32( psDelDec->Seed, Seed_s32x4 ); + + /* Short-term prediction */ + LPC_pred_Q14_s32x4 = silk_noise_shape_quantizer_short_prediction_neon_local(psDelDec->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 16 + i ], a_Q12_arch, predictLPCOrder); + LPC_pred_Q14_s32x4 = vshlq_n_s32( LPC_pred_Q14_s32x4, 4 ); /* Q10 -> Q14 */ + + /* Noise shape feedback */ + /* Output of lowpass section */ + tmp2_s32x4 = silk_SMLAWB_lane0_neon( vld1q_s32( psDelDec->Diff_Q14 ), vld1q_s32( psDelDec->sAR2_Q14[ 0 ] ), warping_Q16_s32x2 ); + /* Output of allpass section */ + tmp1_s32x4 = vsubq_s32( vld1q_s32( psDelDec->sAR2_Q14[ 1 ] ), tmp2_s32x4 ); + tmp1_s32x4 = silk_SMLAWB_lane0_neon( vld1q_s32( psDelDec->sAR2_Q14[ 0 ] ), tmp1_s32x4, warping_Q16_s32x2 ); + vst1q_s32( psDelDec->sAR2_Q14[ 0 ], tmp2_s32x4 ); + AR_shp_Q28_s32x2 = vld1_s32( AR_shp_Q28 ); + n_AR_Q14_s32x4 = vaddq_s32( vdupq_n_s32( silk_RSHIFT( shapingLPCOrder, 1 ) ), vqdmulhq_lane_s32( tmp2_s32x4, AR_shp_Q28_s32x2, 0 ) ); + + /* Loop over allpass sections */ + for( j = 2; j < shapingLPCOrder; j += 2 ) { + /* Output of allpass section */ + tmp2_s32x4 = vsubq_s32( vld1q_s32( psDelDec->sAR2_Q14[ j + 0 ] ), tmp1_s32x4 ); + tmp2_s32x4 = silk_SMLAWB_lane0_neon( vld1q_s32( psDelDec->sAR2_Q14[ j - 1 ] ), tmp2_s32x4, warping_Q16_s32x2 ); + vst1q_s32( psDelDec->sAR2_Q14[ j - 1 ], tmp1_s32x4 ); + n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_lane_s32( tmp1_s32x4, AR_shp_Q28_s32x2, 1 ) ); + /* Output of allpass section */ + tmp1_s32x4 = vsubq_s32( vld1q_s32( psDelDec->sAR2_Q14[ j + 1 ] ), tmp2_s32x4 ); + tmp1_s32x4 = silk_SMLAWB_lane0_neon( vld1q_s32( psDelDec->sAR2_Q14[ j + 0 ] ), tmp1_s32x4, warping_Q16_s32x2 ); + vst1q_s32( psDelDec->sAR2_Q14[ j + 0 ], tmp2_s32x4 ); + AR_shp_Q28_s32x2 = vld1_s32( &AR_shp_Q28[ j ] ); + n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_lane_s32( tmp2_s32x4, AR_shp_Q28_s32x2, 0 ) ); + } + vst1q_s32( psDelDec->sAR2_Q14[ shapingLPCOrder - 1 ], tmp1_s32x4 ); + n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_lane_s32( tmp1_s32x4, AR_shp_Q28_s32x2, 1 ) ); + n_AR_Q14_s32x4 = vshlq_n_s32( n_AR_Q14_s32x4, 1 ); /* Q11 -> Q12 */ + n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_n_s32( vld1q_s32( psDelDec->LF_AR_Q14 ), silk_LSHIFT32( Tilt_Q14, 16 ) >> 1 ) ); /* Q12 */ + n_AR_Q14_s32x4 = vshlq_n_s32( n_AR_Q14_s32x4, 2 ); /* Q12 -> Q14 */ + n_LF_Q14_s32x4 = vqdmulhq_n_s32( vld1q_s32( psDelDec->Shape_Q14[ *smpl_buf_idx ] ), LF_shp_Q29 ); /* Q12 */ + n_LF_Q14_s32x4 = vaddq_s32( n_LF_Q14_s32x4, vqdmulhq_n_s32( vld1q_s32( psDelDec->LF_AR_Q14 ), silk_LSHIFT32( LF_shp_Q14 >> 16 , 15 ) ) ); /* Q12 */ + n_LF_Q14_s32x4 = vshlq_n_s32( n_LF_Q14_s32x4, 2 ); /* Q12 -> Q14 */ + + /* Input minus prediction plus noise feedback */ + /* r = x[ i ] - LTP_pred - LPC_pred + n_AR + n_Tilt + n_LF + n_LTP */ + tmp1_s32x4 = vaddq_s32( n_AR_Q14_s32x4, n_LF_Q14_s32x4 ); /* Q14 */ + tmp2_s32x4 = vaddq_s32( vdupq_n_s32( n_LTP_Q14 ), LPC_pred_Q14_s32x4 ); /* Q13 */ + tmp1_s32x4 = vsubq_s32( tmp2_s32x4, tmp1_s32x4 ); /* Q13 */ + tmp1_s32x4 = vrshrq_n_s32( tmp1_s32x4, 4 ); /* Q10 */ + tmp1_s32x4 = vsubq_s32( vdupq_n_s32( x_Q10[ i ] ), tmp1_s32x4 ); /* residual error Q10 */ + + /* Flip sign depending on dither */ + sign_s32x4 = vreinterpretq_s32_u32( vcltq_s32( Seed_s32x4, vdupq_n_s32( 0 ) ) ); + tmp1_s32x4 = veorq_s32( tmp1_s32x4, sign_s32x4 ); + tmp1_s32x4 = vsubq_s32( tmp1_s32x4, sign_s32x4 ); + tmp1_s32x4 = vmaxq_s32( tmp1_s32x4, vdupq_n_s32( -( 31 << 10 ) ) ); + tmp1_s32x4 = vminq_s32( tmp1_s32x4, vdupq_n_s32( 30 << 10 ) ); + r_Q10_s16x4 = vmovn_s32( tmp1_s32x4 ); + + /* Find two quantization level candidates and measure their rate-distortion */ + { + int16x4_t q1_Q10_s16x4 = vsub_s16( r_Q10_s16x4, vdup_n_s16( offset_Q10 ) ); + int16x4_t q1_Q0_s16x4 = vshr_n_s16( q1_Q10_s16x4, 10 ); + int16x4_t q2_Q10_s16x4; + int32x4_t rd1_Q10_s32x4, rd2_Q10_s32x4; + uint32x4_t t_u32x4; + + if( Lambda_Q10 > 2048 ) { + /* For aggressive RDO, the bias becomes more than one pulse. */ + const int rdo_offset = Lambda_Q10/2 - 512; + const uint16x4_t greaterThanRdo = vcgt_s16( q1_Q10_s16x4, vdup_n_s16( rdo_offset ) ); + const uint16x4_t lessThanMinusRdo = vclt_s16( q1_Q10_s16x4, vdup_n_s16( -rdo_offset ) ); + /* If Lambda_Q10 > 32767, then q1_Q0, q1_Q10 and q2_Q10 must change to 32-bit. */ + silk_assert( Lambda_Q10 <= 32767 ); + + q1_Q0_s16x4 = vreinterpret_s16_u16( vclt_s16( q1_Q10_s16x4, vdup_n_s16( 0 ) ) ); + q1_Q0_s16x4 = vbsl_s16( greaterThanRdo, vsub_s16( q1_Q10_s16x4, vdup_n_s16( rdo_offset ) ), q1_Q0_s16x4 ); + q1_Q0_s16x4 = vbsl_s16( lessThanMinusRdo, vadd_s16( q1_Q10_s16x4, vdup_n_s16( rdo_offset ) ), q1_Q0_s16x4 ); + q1_Q0_s16x4 = vshr_n_s16( q1_Q0_s16x4, 10 ); + } + { + const uint16x4_t equal0_u16x4 = vceq_s16( q1_Q0_s16x4, vdup_n_s16( 0 ) ); + const uint16x4_t equalMinus1_u16x4 = vceq_s16( q1_Q0_s16x4, vdup_n_s16( -1 ) ); + const uint16x4_t lessThanMinus1_u16x4 = vclt_s16( q1_Q0_s16x4, vdup_n_s16( -1 ) ); + int16x4_t tmp1_s16x4, tmp2_s16x4; + + q1_Q10_s16x4 = vshl_n_s16( q1_Q0_s16x4, 10 ); + tmp1_s16x4 = vadd_s16( q1_Q10_s16x4, vdup_n_s16( offset_Q10 - QUANT_LEVEL_ADJUST_Q10 ) ); + q1_Q10_s16x4 = vadd_s16( q1_Q10_s16x4, vdup_n_s16( offset_Q10 + QUANT_LEVEL_ADJUST_Q10 ) ); + q1_Q10_s16x4 = vbsl_s16( lessThanMinus1_u16x4, q1_Q10_s16x4, tmp1_s16x4 ); + q1_Q10_s16x4 = vbsl_s16( equal0_u16x4, vdup_n_s16( offset_Q10 ), q1_Q10_s16x4 ); + q1_Q10_s16x4 = vbsl_s16( equalMinus1_u16x4, vdup_n_s16( offset_Q10 - ( 1024 - QUANT_LEVEL_ADJUST_Q10 ) ), q1_Q10_s16x4 ); + q2_Q10_s16x4 = vadd_s16( q1_Q10_s16x4, vdup_n_s16( 1024 ) ); + q2_Q10_s16x4 = vbsl_s16( equal0_u16x4, vdup_n_s16( offset_Q10 + 1024 - QUANT_LEVEL_ADJUST_Q10 ), q2_Q10_s16x4 ); + q2_Q10_s16x4 = vbsl_s16( equalMinus1_u16x4, vdup_n_s16( offset_Q10 ), q2_Q10_s16x4 ); + tmp1_s16x4 = q1_Q10_s16x4; + tmp2_s16x4 = q2_Q10_s16x4; + tmp1_s16x4 = vbsl_s16( vorr_u16( equalMinus1_u16x4, lessThanMinus1_u16x4 ), vneg_s16( tmp1_s16x4 ), tmp1_s16x4 ); + tmp2_s16x4 = vbsl_s16( lessThanMinus1_u16x4, vneg_s16( tmp2_s16x4 ), tmp2_s16x4 ); + rd1_Q10_s32x4 = vmull_s16( tmp1_s16x4, vdup_n_s16( Lambda_Q10 ) ); + rd2_Q10_s32x4 = vmull_s16( tmp2_s16x4, vdup_n_s16( Lambda_Q10 ) ); + } + + rr_Q10_s16x4 = vsub_s16( r_Q10_s16x4, q1_Q10_s16x4 ); + rd1_Q10_s32x4 = vmlal_s16( rd1_Q10_s32x4, rr_Q10_s16x4, rr_Q10_s16x4 ); + rd1_Q10_s32x4 = vshrq_n_s32( rd1_Q10_s32x4, 10 ); + + rr_Q10_s16x4 = vsub_s16( r_Q10_s16x4, q2_Q10_s16x4 ); + rd2_Q10_s32x4 = vmlal_s16( rd2_Q10_s32x4, rr_Q10_s16x4, rr_Q10_s16x4 ); + rd2_Q10_s32x4 = vshrq_n_s32( rd2_Q10_s32x4, 10 ); + + tmp2_s32x4 = vld1q_s32( psDelDec->RD_Q10 ); + tmp1_s32x4 = vaddq_s32( tmp2_s32x4, vminq_s32( rd1_Q10_s32x4, rd2_Q10_s32x4 ) ); + tmp2_s32x4 = vaddq_s32( tmp2_s32x4, vmaxq_s32( rd1_Q10_s32x4, rd2_Q10_s32x4 ) ); + vst1q_s32( psSampleState[ 0 ].RD_Q10, tmp1_s32x4 ); + vst1q_s32( psSampleState[ 1 ].RD_Q10, tmp2_s32x4 ); + t_u32x4 = vcltq_s32( rd1_Q10_s32x4, rd2_Q10_s32x4 ); + tmp1_s32x4 = vbslq_s32( t_u32x4, vmovl_s16( q1_Q10_s16x4 ), vmovl_s16( q2_Q10_s16x4 ) ); + tmp2_s32x4 = vbslq_s32( t_u32x4, vmovl_s16( q2_Q10_s16x4 ), vmovl_s16( q1_Q10_s16x4 ) ); + vst1q_s32( psSampleState[ 0 ].Q_Q10, tmp1_s32x4 ); + vst1q_s32( psSampleState[ 1 ].Q_Q10, tmp2_s32x4 ); + } + + { + /* Update states for best quantization */ + int32x4_t exc_Q14_s32x4, LPC_exc_Q14_s32x4, xq_Q14_s32x4, sLF_AR_shp_Q14_s32x4; + + /* Quantized excitation */ + exc_Q14_s32x4 = vshlq_n_s32( tmp1_s32x4, 4 ); + exc_Q14_s32x4 = veorq_s32( exc_Q14_s32x4, sign_s32x4 ); + exc_Q14_s32x4 = vsubq_s32( exc_Q14_s32x4, sign_s32x4 ); + + /* Add predictions */ + LPC_exc_Q14_s32x4 = vaddq_s32( exc_Q14_s32x4, vdupq_n_s32( LTP_pred_Q14 ) ); + xq_Q14_s32x4 = vaddq_s32( LPC_exc_Q14_s32x4, LPC_pred_Q14_s32x4 ); + + /* Update states */ + tmp1_s32x4 = vsubq_s32( xq_Q14_s32x4, vshlq_n_s32( vdupq_n_s32( x_Q10[ i ] ), 4 ) ); + vst1q_s32( psSampleState[ 0 ].Diff_Q14, tmp1_s32x4 ); + sLF_AR_shp_Q14_s32x4 = vsubq_s32( tmp1_s32x4, n_AR_Q14_s32x4 ); + vst1q_s32( psSampleState[ 0 ].sLTP_shp_Q14, vsubq_s32( sLF_AR_shp_Q14_s32x4, n_LF_Q14_s32x4 ) ); + vst1q_s32( psSampleState[ 0 ].LF_AR_Q14, sLF_AR_shp_Q14_s32x4 ); + vst1q_s32( psSampleState[ 0 ].LPC_exc_Q14, LPC_exc_Q14_s32x4 ); + vst1q_s32( psSampleState[ 0 ].xq_Q14, xq_Q14_s32x4 ); + + /* Quantized excitation */ + exc_Q14_s32x4 = vshlq_n_s32( tmp2_s32x4, 4 ); + exc_Q14_s32x4 = veorq_s32( exc_Q14_s32x4, sign_s32x4 ); + exc_Q14_s32x4 = vsubq_s32( exc_Q14_s32x4, sign_s32x4 ); + + /* Add predictions */ + LPC_exc_Q14_s32x4 = vaddq_s32( exc_Q14_s32x4, vdupq_n_s32( LTP_pred_Q14 ) ); + xq_Q14_s32x4 = vaddq_s32( LPC_exc_Q14_s32x4, LPC_pred_Q14_s32x4 ); + + /* Update states */ + tmp1_s32x4 = vsubq_s32( xq_Q14_s32x4, vshlq_n_s32( vdupq_n_s32( x_Q10[ i ] ), 4 ) ); + vst1q_s32( psSampleState[ 1 ].Diff_Q14, tmp1_s32x4 ); + sLF_AR_shp_Q14_s32x4 = vsubq_s32( tmp1_s32x4, n_AR_Q14_s32x4 ); + vst1q_s32( psSampleState[ 1 ].sLTP_shp_Q14, vsubq_s32( sLF_AR_shp_Q14_s32x4, n_LF_Q14_s32x4 ) ); + vst1q_s32( psSampleState[ 1 ].LF_AR_Q14, sLF_AR_shp_Q14_s32x4 ); + vst1q_s32( psSampleState[ 1 ].LPC_exc_Q14, LPC_exc_Q14_s32x4 ); + vst1q_s32( psSampleState[ 1 ].xq_Q14, xq_Q14_s32x4 ); + } + + *smpl_buf_idx = *smpl_buf_idx ? ( *smpl_buf_idx - 1 ) : ( DECISION_DELAY - 1); + last_smple_idx = *smpl_buf_idx + decisionDelay + DECISION_DELAY; + if( last_smple_idx >= DECISION_DELAY ) last_smple_idx -= DECISION_DELAY; + if( last_smple_idx >= DECISION_DELAY ) last_smple_idx -= DECISION_DELAY; + + /* Find winner */ + RDmin_Q10 = psSampleState[ 0 ].RD_Q10[ 0 ]; + Winner_ind = 0; + for( k = 1; k < nStatesDelayedDecision; k++ ) { + if( psSampleState[ 0 ].RD_Q10[ k ] < RDmin_Q10 ) { + RDmin_Q10 = psSampleState[ 0 ].RD_Q10[ k ]; + Winner_ind = k; + } + } + + /* Increase RD values of expired states */ + { + uint32x4_t t_u32x4; + Winner_rand_state = psDelDec->RandState[ last_smple_idx ][ Winner_ind ]; + t_u32x4 = vceqq_s32( vld1q_s32( psDelDec->RandState[ last_smple_idx ] ), vdupq_n_s32( Winner_rand_state ) ); + t_u32x4 = vmvnq_u32( t_u32x4 ); + t_u32x4 = vshrq_n_u32( t_u32x4, 5 ); + tmp1_s32x4 = vld1q_s32( psSampleState[ 0 ].RD_Q10 ); + tmp2_s32x4 = vld1q_s32( psSampleState[ 1 ].RD_Q10 ); + tmp1_s32x4 = vaddq_s32( tmp1_s32x4, vreinterpretq_s32_u32( t_u32x4 ) ); + tmp2_s32x4 = vaddq_s32( tmp2_s32x4, vreinterpretq_s32_u32( t_u32x4 ) ); + vst1q_s32( psSampleState[ 0 ].RD_Q10, tmp1_s32x4 ); + vst1q_s32( psSampleState[ 1 ].RD_Q10, tmp2_s32x4 ); + + /* Find worst in first set and best in second set */ + RDmax_Q10 = psSampleState[ 0 ].RD_Q10[ 0 ]; + RDmin_Q10 = psSampleState[ 1 ].RD_Q10[ 0 ]; + RDmax_ind = 0; + RDmin_ind = 0; + for( k = 1; k < nStatesDelayedDecision; k++ ) { + /* find worst in first set */ + if( psSampleState[ 0 ].RD_Q10[ k ] > RDmax_Q10 ) { + RDmax_Q10 = psSampleState[ 0 ].RD_Q10[ k ]; + RDmax_ind = k; + } + /* find best in second set */ + if( psSampleState[ 1 ].RD_Q10[ k ] < RDmin_Q10 ) { + RDmin_Q10 = psSampleState[ 1 ].RD_Q10[ k ]; + RDmin_ind = k; + } + } + } + + /* Replace a state if best from second set outperforms worst in first set */ + if( RDmin_Q10 < RDmax_Q10 ) { + opus_int32 (*ptr)[NEON_MAX_DEL_DEC_STATES] = psDelDec->RandState; + const int numOthers = (int)( ( sizeof( NSQ_del_decs_struct ) - sizeof( ( (NSQ_del_decs_struct *)0 )->sLPC_Q14 ) ) + / ( NEON_MAX_DEL_DEC_STATES * sizeof( opus_int32 ) ) ); + /* Only ( predictLPCOrder - 1 ) of sLPC_Q14 buffer need to be updated, though the first several */ + /* useless sLPC_Q14[] will be different comparing with C when predictLPCOrder < NSQ_LPC_BUF_LENGTH. */ + /* Here just update constant ( NSQ_LPC_BUF_LENGTH - 1 ) for simplicity. */ + for( j = i + 1; j < i + NSQ_LPC_BUF_LENGTH; j++ ) { + psDelDec->sLPC_Q14[ j ][ RDmax_ind ] = psDelDec->sLPC_Q14[ j ][ RDmin_ind ]; + } + for( j = 0; j < numOthers; j++ ) { + ptr[ j ][ RDmax_ind ] = ptr[ j ][ RDmin_ind ]; + } + + psSampleState[ 0 ].Q_Q10[ RDmax_ind ] = psSampleState[ 1 ].Q_Q10[ RDmin_ind ]; + psSampleState[ 0 ].RD_Q10[ RDmax_ind ] = psSampleState[ 1 ].RD_Q10[ RDmin_ind ]; + psSampleState[ 0 ].xq_Q14[ RDmax_ind ] = psSampleState[ 1 ].xq_Q14[ RDmin_ind ]; + psSampleState[ 0 ].LF_AR_Q14[ RDmax_ind ] = psSampleState[ 1 ].LF_AR_Q14[ RDmin_ind ]; + psSampleState[ 0 ].Diff_Q14[ RDmax_ind ] = psSampleState[ 1 ].Diff_Q14[ RDmin_ind ]; + psSampleState[ 0 ].sLTP_shp_Q14[ RDmax_ind ] = psSampleState[ 1 ].sLTP_shp_Q14[ RDmin_ind ]; + psSampleState[ 0 ].LPC_exc_Q14[ RDmax_ind ] = psSampleState[ 1 ].LPC_exc_Q14[ RDmin_ind ]; + } + + /* Write samples from winner to output and long-term filter states */ + if( subfr > 0 || i >= decisionDelay ) { + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDelDec->Q_Q10[ last_smple_idx ][ Winner_ind ], 10 ); + xq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( + silk_SMULWW( psDelDec->Xq_Q14[ last_smple_idx ][ Winner_ind ], delayedGain_Q10[ last_smple_idx ] ), 8 ) ); + NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay ] = psDelDec->Shape_Q14[ last_smple_idx ][ Winner_ind ]; + sLTP_Q15[ NSQ->sLTP_buf_idx - decisionDelay ] = psDelDec->Pred_Q15[ last_smple_idx ][ Winner_ind ]; + } + NSQ->sLTP_shp_buf_idx++; + NSQ->sLTP_buf_idx++; + + /* Update states */ + vst1q_s32( psDelDec->LF_AR_Q14, vld1q_s32( psSampleState[ 0 ].LF_AR_Q14 ) ); + vst1q_s32( psDelDec->Diff_Q14, vld1q_s32( psSampleState[ 0 ].Diff_Q14 ) ); + vst1q_s32( psDelDec->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ], vld1q_s32( psSampleState[ 0 ].xq_Q14 ) ); + vst1q_s32( psDelDec->Xq_Q14[ *smpl_buf_idx ], vld1q_s32( psSampleState[ 0 ].xq_Q14 ) ); + tmp1_s32x4 = vld1q_s32( psSampleState[ 0 ].Q_Q10 ); + vst1q_s32( psDelDec->Q_Q10[ *smpl_buf_idx ], tmp1_s32x4 ); + vst1q_s32( psDelDec->Pred_Q15[ *smpl_buf_idx ], vshlq_n_s32( vld1q_s32( psSampleState[ 0 ].LPC_exc_Q14 ), 1 ) ); + vst1q_s32( psDelDec->Shape_Q14[ *smpl_buf_idx ], vld1q_s32( psSampleState[ 0 ].sLTP_shp_Q14 ) ); + tmp1_s32x4 = vrshrq_n_s32( tmp1_s32x4, 10 ); + tmp1_s32x4 = vaddq_s32( vld1q_s32( psDelDec->Seed ), tmp1_s32x4 ); + vst1q_s32( psDelDec->Seed, tmp1_s32x4 ); + vst1q_s32( psDelDec->RandState[ *smpl_buf_idx ], tmp1_s32x4 ); + vst1q_s32( psDelDec->RD_Q10, vld1q_s32( psSampleState[ 0 ].RD_Q10 ) ); + delayedGain_Q10[ *smpl_buf_idx ] = Gain_Q10; + } + /* Update LPC states */ + silk_memcpy( psDelDec->sLPC_Q14[ 0 ], psDelDec->sLPC_Q14[ length ], NEON_MAX_DEL_DEC_STATES * NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) ); + + RESTORE_STACK; +} + +static OPUS_INLINE void silk_SMULWB_8_neon( + const opus_int16 *a, + const int32x2_t b, + opus_int32 *o +) +{ + const int16x8_t a_s16x8 = vld1q_s16( a ); + int32x4_t o0_s32x4, o1_s32x4; + + o0_s32x4 = vshll_n_s16( vget_low_s16( a_s16x8 ), 15 ); + o1_s32x4 = vshll_n_s16( vget_high_s16( a_s16x8 ), 15 ); + o0_s32x4 = vqdmulhq_lane_s32( o0_s32x4, b, 0 ); + o1_s32x4 = vqdmulhq_lane_s32( o1_s32x4, b, 0 ); + vst1q_s32( o, o0_s32x4 ); + vst1q_s32( o + 4, o1_s32x4 ); +} + +/* Only works when ( b >= -65536 ) && ( b < 65536 ). */ +static OPUS_INLINE void silk_SMULWW_small_b_4_neon( + opus_int32 *a, + const int32x2_t b_s32x2) +{ + int32x4_t o_s32x4; + + o_s32x4 = vld1q_s32( a ); + o_s32x4 = vqdmulhq_lane_s32( o_s32x4, b_s32x2, 0 ); + vst1q_s32( a, o_s32x4 ); +} + +/* Only works when ( b >= -65536 ) && ( b < 65536 ). */ +static OPUS_INLINE void silk_SMULWW_small_b_8_neon( + opus_int32 *a, + const int32x2_t b_s32x2 +) +{ + int32x4_t o0_s32x4, o1_s32x4; + + o0_s32x4 = vld1q_s32( a ); + o1_s32x4 = vld1q_s32( a + 4 ); + o0_s32x4 = vqdmulhq_lane_s32( o0_s32x4, b_s32x2, 0 ); + o1_s32x4 = vqdmulhq_lane_s32( o1_s32x4, b_s32x2, 0 ); + vst1q_s32( a, o0_s32x4 ); + vst1q_s32( a + 4, o1_s32x4 ); +} + +static OPUS_INLINE void silk_SMULWW_4_neon( + opus_int32 *a, + const int32x2_t b_s32x2) +{ + int32x4_t a_s32x4, o_s32x4; + + a_s32x4 = vld1q_s32( a ); + o_s32x4 = vqdmulhq_lane_s32( a_s32x4, b_s32x2, 0 ); + o_s32x4 = vmlaq_lane_s32( o_s32x4, a_s32x4, b_s32x2, 1 ); + vst1q_s32( a, o_s32x4 ); +} + +static OPUS_INLINE void silk_SMULWW_8_neon( + opus_int32 *a, + const int32x2_t b_s32x2 +) +{ + int32x4_t a0_s32x4, a1_s32x4, o0_s32x4, o1_s32x4; + + a0_s32x4 = vld1q_s32( a ); + a1_s32x4 = vld1q_s32( a + 4 ); + o0_s32x4 = vqdmulhq_lane_s32( a0_s32x4, b_s32x2, 0 ); + o1_s32x4 = vqdmulhq_lane_s32( a1_s32x4, b_s32x2, 0 ); + o0_s32x4 = vmlaq_lane_s32( o0_s32x4, a0_s32x4, b_s32x2, 1 ); + o1_s32x4 = vmlaq_lane_s32( o1_s32x4, a1_s32x4, b_s32x2, 1 ); + vst1q_s32( a, o0_s32x4 ); + vst1q_s32( a + 4, o1_s32x4 ); +} + +static OPUS_INLINE void silk_SMULWW_loop_neon( + const opus_int16 *a, + const opus_int32 b, + opus_int32 *o, + const opus_int loop_num +) +{ + opus_int i; + int32x2_t b_s32x2; + + b_s32x2 = vdup_n_s32( b ); + for( i = 0; i < loop_num - 7; i += 8 ) { + silk_SMULWB_8_neon( a + i, b_s32x2, o + i ); + } + for( ; i < loop_num; i++ ) { + o[ i ] = silk_SMULWW( a[ i ], b ); + } +} + +static OPUS_INLINE void silk_nsq_del_dec_scale_states_neon( + const silk_encoder_state *psEncC, /* I Encoder State */ + silk_nsq_state *NSQ, /* I/O NSQ state */ + NSQ_del_decs_struct psDelDec[], /* I/O Delayed decision states */ + const opus_int16 x16[], /* I Input */ + opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ + const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ + opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ + opus_int subfr, /* I Subframe number */ + const opus_int LTP_scale_Q14, /* I LTP state scaling */ + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */ + const opus_int signal_type, /* I Signal type */ + const opus_int decisionDelay /* I Decision delay */ +) +{ + opus_int i, lag; + opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q26; + + lag = pitchL[ subfr ]; + inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); + silk_assert( inv_gain_Q31 != 0 ); + + /* Scale input */ + inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 ); + silk_SMULWW_loop_neon( x16, inv_gain_Q26, x_sc_Q10, psEncC->subfr_length ); + + /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ + if( NSQ->rewhite_flag ) { + if( subfr == 0 ) { + /* Do LTP downscaling */ + inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 ); + } + silk_SMULWW_loop_neon( sLTP + NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2, inv_gain_Q31, sLTP_Q15 + NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2, lag + LTP_ORDER / 2 ); + } + + /* Adjust for changing gain */ + if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { + int32x2_t gain_adj_Q16_s32x2; + gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); + + /* Scale long-term shaping state */ + if( ( gain_adj_Q16 >= -65536 ) && ( gain_adj_Q16 < 65536 ) ) { + gain_adj_Q16_s32x2 = vdup_n_s32( silk_LSHIFT32( gain_adj_Q16, 15 ) ); + for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 7; i += 8 ) { + silk_SMULWW_small_b_8_neon( NSQ->sLTP_shp_Q14 + i, gain_adj_Q16_s32x2 ); + } + for( ; i < NSQ->sLTP_shp_buf_idx; i++ ) { + NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); + } + + /* Scale long-term prediction state */ + if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) { + for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay - 7; i += 8 ) { + silk_SMULWW_small_b_8_neon( sLTP_Q15 + i, gain_adj_Q16_s32x2 ); + } + for( ; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) { + sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] ); + } + } + + /* Scale scalar states */ + silk_SMULWW_small_b_4_neon( psDelDec->LF_AR_Q14, gain_adj_Q16_s32x2 ); + silk_SMULWW_small_b_4_neon( psDelDec->Diff_Q14, gain_adj_Q16_s32x2 ); + + /* Scale short-term prediction and shaping states */ + for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { + silk_SMULWW_small_b_4_neon( psDelDec->sLPC_Q14[ i ], gain_adj_Q16_s32x2 ); + } + + for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) { + silk_SMULWW_small_b_4_neon( psDelDec->sAR2_Q14[ i ], gain_adj_Q16_s32x2 ); + } + + for( i = 0; i < DECISION_DELAY; i++ ) { + silk_SMULWW_small_b_4_neon( psDelDec->Pred_Q15[ i ], gain_adj_Q16_s32x2 ); + silk_SMULWW_small_b_4_neon( psDelDec->Shape_Q14[ i ], gain_adj_Q16_s32x2 ); + } + } else { + gain_adj_Q16_s32x2 = vdup_n_s32( silk_LSHIFT32( gain_adj_Q16 & 0x0000FFFF, 15 ) ); + gain_adj_Q16_s32x2 = vset_lane_s32( gain_adj_Q16 >> 16, gain_adj_Q16_s32x2, 1 ); + for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 7; i += 8 ) { + silk_SMULWW_8_neon( NSQ->sLTP_shp_Q14 + i, gain_adj_Q16_s32x2 ); + } + for( ; i < NSQ->sLTP_shp_buf_idx; i++ ) { + NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); + } + + /* Scale long-term prediction state */ + if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) { + for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay - 7; i += 8 ) { + silk_SMULWW_8_neon( sLTP_Q15 + i, gain_adj_Q16_s32x2 ); + } + for( ; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) { + sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] ); + } + } + + /* Scale scalar states */ + silk_SMULWW_4_neon( psDelDec->LF_AR_Q14, gain_adj_Q16_s32x2 ); + silk_SMULWW_4_neon( psDelDec->Diff_Q14, gain_adj_Q16_s32x2 ); + + /* Scale short-term prediction and shaping states */ + for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { + silk_SMULWW_4_neon( psDelDec->sLPC_Q14[ i ], gain_adj_Q16_s32x2 ); + } + + for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) { + silk_SMULWW_4_neon( psDelDec->sAR2_Q14[ i ], gain_adj_Q16_s32x2 ); + } + + for( i = 0; i < DECISION_DELAY; i++ ) { + silk_SMULWW_4_neon( psDelDec->Pred_Q15[ i ], gain_adj_Q16_s32x2 ); + silk_SMULWW_4_neon( psDelDec->Shape_Q14[ i ], gain_adj_Q16_s32x2 ); + } + } + + /* Save inverse gain */ + NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; + } +} diff --git a/silk/arm/NSQ_neon.h b/silk/arm/NSQ_neon.h index 77c946af..b31d9442 100644 --- a/silk/arm/NSQ_neon.h +++ b/silk/arm/NSQ_neon.h @@ -28,30 +28,31 @@ POSSIBILITY OF SUCH DAMAGE. #define SILK_NSQ_NEON_H #include "cpu_support.h" +#include "SigProc_FIX.h" #undef silk_short_prediction_create_arch_coef /* For vectorized calc, reverse a_Q12 coefs, convert to 32-bit, and shift for vqdmulhq_s32. */ static OPUS_INLINE void silk_short_prediction_create_arch_coef_neon(opus_int32 *out, const opus_int16 *in, opus_int order) { - out[15] = in[0] << 15; - out[14] = in[1] << 15; - out[13] = in[2] << 15; - out[12] = in[3] << 15; - out[11] = in[4] << 15; - out[10] = in[5] << 15; - out[9] = in[6] << 15; - out[8] = in[7] << 15; - out[7] = in[8] << 15; - out[6] = in[9] << 15; + out[15] = silk_LSHIFT32(in[0], 15); + out[14] = silk_LSHIFT32(in[1], 15); + out[13] = silk_LSHIFT32(in[2], 15); + out[12] = silk_LSHIFT32(in[3], 15); + out[11] = silk_LSHIFT32(in[4], 15); + out[10] = silk_LSHIFT32(in[5], 15); + out[9] = silk_LSHIFT32(in[6], 15); + out[8] = silk_LSHIFT32(in[7], 15); + out[7] = silk_LSHIFT32(in[8], 15); + out[6] = silk_LSHIFT32(in[9], 15); if (order == 16) { - out[5] = in[10] << 15; - out[4] = in[11] << 15; - out[3] = in[12] << 15; - out[2] = in[13] << 15; - out[1] = in[14] << 15; - out[0] = in[15] << 15; + out[5] = silk_LSHIFT32(in[10], 15); + out[4] = silk_LSHIFT32(in[11], 15); + out[3] = silk_LSHIFT32(in[12], 15); + out[2] = silk_LSHIFT32(in[13], 15); + out[1] = silk_LSHIFT32(in[14], 15); + out[0] = silk_LSHIFT32(in[15], 15); } else { diff --git a/silk/arm/arm_silk_map.c b/silk/arm/arm_silk_map.c index 9bd86a7b..0b9bfec2 100644 --- a/silk/arm/arm_silk_map.c +++ b/silk/arm/arm_silk_map.c @@ -28,13 +28,62 @@ POSSIBILITY OF SUCH DAMAGE. # include "config.h" #endif +#include "main_FIX.h" #include "NSQ.h" +#include "SigProc_FIX.h" #if defined(OPUS_HAVE_RTCD) # if (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && \ !defined(OPUS_ARM_PRESUME_NEON_INTR)) +void (*const SILK_BIQUAD_ALT_STRIDE2_IMPL[OPUS_ARCHMASK + 1])( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ +) = { + silk_biquad_alt_stride2_c, /* ARMv4 */ + silk_biquad_alt_stride2_c, /* EDSP */ + silk_biquad_alt_stride2_c, /* Media */ + silk_biquad_alt_stride2_neon, /* Neon */ +}; + +opus_int32 (*const SILK_LPC_INVERSE_PRED_GAIN_IMPL[OPUS_ARCHMASK + 1])( /* O Returns inverse prediction gain in energy domain, Q30 */ + const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ + const opus_int order /* I Prediction order */ +) = { + silk_LPC_inverse_pred_gain_c, /* ARMv4 */ + silk_LPC_inverse_pred_gain_c, /* EDSP */ + silk_LPC_inverse_pred_gain_c, /* Media */ + silk_LPC_inverse_pred_gain_neon, /* Neon */ +}; + +void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])( + const silk_encoder_state *psEncC, /* I Encoder State */ + silk_nsq_state *NSQ, /* I/O NSQ state */ + SideInfoIndices *psIndices, /* I/O Quantization Indices */ + const opus_int16 x16[], /* I Input */ + opus_int8 pulses[], /* O Quantized pulse signal */ + const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ + const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ + const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ + const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */ + const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */ + const opus_int LTP_scale_Q14 /* I LTP state scaling */ +) = { + silk_NSQ_del_dec_c, /* ARMv4 */ + silk_NSQ_del_dec_c, /* EDSP */ + silk_NSQ_del_dec_c, /* Media */ + silk_NSQ_del_dec_neon, /* Neon */ +}; + /*There is no table for silk_noise_shape_quantizer_short_prediction because the NEON version takes different parameters than the C version. Instead RTCD is done via if statements at the call sites. @@ -52,4 +101,23 @@ opus_int32 # endif +# if defined(FIXED_POINT) && \ + defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR) + +void (*const SILK_WARPED_AUTOCORRELATION_FIX_IMPL[OPUS_ARCHMASK + 1])( + opus_int32 *corr, /* O Result [order + 1] */ + opus_int *scale, /* O Scaling of the correlation vector */ + const opus_int16 *input, /* I Input data to correlate */ + const opus_int warping_Q16, /* I Warping coefficient */ + const opus_int length, /* I Length of input */ + const opus_int order /* I Correlation order (even) */ +) = { + silk_warped_autocorrelation_FIX_c, /* ARMv4 */ + silk_warped_autocorrelation_FIX_c, /* EDSP */ + silk_warped_autocorrelation_FIX_c, /* Media */ + silk_warped_autocorrelation_FIX_neon, /* Neon */ +}; + +# endif + #endif /* OPUS_HAVE_RTCD */ diff --git a/silk/arm/biquad_alt_arm.h b/silk/arm/biquad_alt_arm.h new file mode 100644 index 00000000..66ea9f43 --- /dev/null +++ b/silk/arm/biquad_alt_arm.h @@ -0,0 +1,68 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifndef SILK_BIQUAD_ALT_ARM_H +# define SILK_BIQUAD_ALT_ARM_H + +# include "celt/arm/armcpu.h" + +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +void silk_biquad_alt_stride2_neon( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ +); + +# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON) +# define OVERRIDE_silk_biquad_alt_stride2 (1) +# define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((void)(arch), PRESUME_NEON(silk_biquad_alt_stride2)(in, B_Q28, A_Q28, S, out, len)) +# endif +# endif + +# if !defined(OVERRIDE_silk_biquad_alt_stride2) +/*Is run-time CPU detection enabled on this platform?*/ +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern void (*const SILK_BIQUAD_ALT_STRIDE2_IMPL[OPUS_ARCHMASK+1])( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ + ); +# define OVERRIDE_silk_biquad_alt_stride2 (1) +# define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((*SILK_BIQUAD_ALT_STRIDE2_IMPL[(arch)&OPUS_ARCHMASK])(in, B_Q28, A_Q28, S, out, len)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_silk_biquad_alt_stride2 (1) +# define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((void)(arch), silk_biquad_alt_stride2_neon(in, B_Q28, A_Q28, S, out, len)) +# endif +# endif + +#endif /* end SILK_BIQUAD_ALT_ARM_H */ diff --git a/silk/arm/biquad_alt_neon_intr.c b/silk/arm/biquad_alt_neon_intr.c new file mode 100644 index 00000000..97157331 --- /dev/null +++ b/silk/arm/biquad_alt_neon_intr.c @@ -0,0 +1,156 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include <arm_neon.h> +#ifdef OPUS_CHECK_ASM +# include <string.h> +# include "stack_alloc.h" +#endif +#include "SigProc_FIX.h" + +static inline void silk_biquad_alt_stride2_kernel( const int32x4_t A_L_s32x4, const int32x4_t A_U_s32x4, const int32x4_t B_Q28_s32x4, const int32x2_t t_s32x2, const int32x4_t in_s32x4, int32x4_t *S_s32x4, int32x2_t *out32_Q14_s32x2 ) +{ + int32x4_t t_s32x4, out32_Q14_s32x4; + + *out32_Q14_s32x2 = vadd_s32( vget_low_s32( *S_s32x4 ), t_s32x2 ); /* silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} ) */ + *S_s32x4 = vcombine_s32( vget_high_s32( *S_s32x4 ), vdup_n_s32( 0 ) ); /* S{0,1} = S{2,3}; S{2,3} = 0; */ + *out32_Q14_s32x2 = vshl_n_s32( *out32_Q14_s32x2, 2 ); /* out32_Q14_{0,1} = silk_LSHIFT( silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} ), 2 ); */ + out32_Q14_s32x4 = vcombine_s32( *out32_Q14_s32x2, *out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} */ + t_s32x4 = vqdmulhq_s32( out32_Q14_s32x4, A_L_s32x4 ); /* silk_SMULWB( out32_Q14_{0,1,0,1}, A{0,0,1,1}_L_Q28 ) */ + *S_s32x4 = vrsraq_n_s32( *S_s32x4, t_s32x4, 14 ); /* S{0,1} = S{2,3} + silk_RSHIFT_ROUND(); S{2,3} = silk_RSHIFT_ROUND(); */ + t_s32x4 = vqdmulhq_s32( out32_Q14_s32x4, A_U_s32x4 ); /* silk_SMULWB( out32_Q14_{0,1,0,1}, A{0,0,1,1}_U_Q28 ) */ + *S_s32x4 = vaddq_s32( *S_s32x4, t_s32x4 ); /* S0 = silk_SMLAWB( S{0,1,2,3}, out32_Q14_{0,1,0,1}, A{0,0,1,1}_U_Q28 ); */ + t_s32x4 = vqdmulhq_s32( in_s32x4, B_Q28_s32x4 ); /* silk_SMULWB( B_Q28[ {1,1,2,2} ], in{0,1,0,1} ) */ + *S_s32x4 = vaddq_s32( *S_s32x4, t_s32x4 ); /* S0 = silk_SMLAWB( S0, B_Q28[ {1,1,2,2} ], in{0,1,0,1} ); */ +} + +void silk_biquad_alt_stride2_neon( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ +) +{ + /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */ + opus_int k = 0; + const int32x2_t offset_s32x2 = vdup_n_s32( (1<<14) - 1 ); + const int32x4_t offset_s32x4 = vcombine_s32( offset_s32x2, offset_s32x2 ); + int16x4_t in_s16x4 = vdup_n_s16( 0 ); + int16x4_t out_s16x4; + int32x2_t A_Q28_s32x2, A_L_s32x2, A_U_s32x2, B_Q28_s32x2, t_s32x2; + int32x4_t A_L_s32x4, A_U_s32x4, B_Q28_s32x4, S_s32x4, out32_Q14_s32x4; + int32x2x2_t t0_s32x2x2, t1_s32x2x2, t2_s32x2x2, S_s32x2x2; + +#ifdef OPUS_CHECK_ASM + opus_int32 S_c[ 4 ]; + VARDECL( opus_int16, out_c ); + SAVE_STACK; + ALLOC( out_c, 2 * len, opus_int16 ); + + silk_memcpy( &S_c, S, sizeof( S_c ) ); + silk_biquad_alt_stride2_c( in, B_Q28, A_Q28, S_c, out_c, len ); +#endif + + /* Negate A_Q28 values and split in two parts */ + A_Q28_s32x2 = vld1_s32( A_Q28 ); + A_Q28_s32x2 = vneg_s32( A_Q28_s32x2 ); + A_L_s32x2 = vshl_n_s32( A_Q28_s32x2, 18 ); /* ( -A_Q28[] & 0x00003FFF ) << 18 */ + A_L_s32x2 = vreinterpret_s32_u32( vshr_n_u32( vreinterpret_u32_s32( A_L_s32x2 ), 3 ) ); /* ( -A_Q28[] & 0x00003FFF ) << 15 */ + A_U_s32x2 = vshr_n_s32( A_Q28_s32x2, 14 ); /* silk_RSHIFT( -A_Q28[], 14 ) */ + A_U_s32x2 = vshl_n_s32( A_U_s32x2, 16 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 16 (Clip two leading bits to conform to C function.) */ + A_U_s32x2 = vshr_n_s32( A_U_s32x2, 1 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 15 */ + + B_Q28_s32x2 = vld1_s32( B_Q28 ); + t_s32x2 = vld1_s32( B_Q28 + 1 ); + t0_s32x2x2 = vzip_s32( A_L_s32x2, A_L_s32x2 ); + t1_s32x2x2 = vzip_s32( A_U_s32x2, A_U_s32x2 ); + t2_s32x2x2 = vzip_s32( t_s32x2, t_s32x2 ); + A_L_s32x4 = vcombine_s32( t0_s32x2x2.val[ 0 ], t0_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_L_Q28 */ + A_U_s32x4 = vcombine_s32( t1_s32x2x2.val[ 0 ], t1_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_U_Q28 */ + B_Q28_s32x4 = vcombine_s32( t2_s32x2x2.val[ 0 ], t2_s32x2x2.val[ 1 ] ); /* B_Q28[ {1,1,2,2} ] */ + S_s32x4 = vld1q_s32( S ); /* S0 = S[ 0 ]; S3 = S[ 3 ]; */ + S_s32x2x2 = vtrn_s32( vget_low_s32( S_s32x4 ), vget_high_s32( S_s32x4 ) ); /* S2 = S[ 1 ]; S1 = S[ 2 ]; */ + S_s32x4 = vcombine_s32( S_s32x2x2.val[ 0 ], S_s32x2x2.val[ 1 ] ); + + for( ; k < len - 1; k += 2 ) { + int32x4_t in_s32x4[ 2 ], t_s32x4; + int32x2_t out32_Q14_s32x2[ 2 ]; + + /* S[ 2 * i + 0 ], S[ 2 * i + 1 ], S[ 2 * i + 2 ], S[ 2 * i + 3 ]: Q12 */ + in_s16x4 = vld1_s16( &in[ 2 * k ] ); /* in{0,1,2,3} = in[ 2 * k + {0,1,2,3} ]; */ + in_s32x4[ 0 ] = vshll_n_s16( in_s16x4, 15 ); /* in{0,1,2,3} << 15 */ + t_s32x4 = vqdmulhq_lane_s32( in_s32x4[ 0 ], B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1,2,3} ) */ + in_s32x4[ 1 ] = vcombine_s32( vget_high_s32( in_s32x4[ 0 ] ), vget_high_s32( in_s32x4[ 0 ] ) ); /* in{2,3,2,3} << 15 */ + in_s32x4[ 0 ] = vcombine_s32( vget_low_s32 ( in_s32x4[ 0 ] ), vget_low_s32 ( in_s32x4[ 0 ] ) ); /* in{0,1,0,1} << 15 */ + silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_low_s32 ( t_s32x4 ), in_s32x4[ 0 ], &S_s32x4, &out32_Q14_s32x2[ 0 ] ); + silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_high_s32( t_s32x4 ), in_s32x4[ 1 ], &S_s32x4, &out32_Q14_s32x2[ 1 ] ); + + /* Scale back to Q0 and saturate */ + out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2[ 0 ], out32_Q14_s32x2[ 1 ] ); /* out32_Q14_{0,1,2,3} */ + out32_Q14_s32x4 = vaddq_s32( out32_Q14_s32x4, offset_s32x4 ); /* out32_Q14_{0,1,2,3} + (1<<14) - 1 */ + out_s16x4 = vqshrn_n_s32( out32_Q14_s32x4, 14 ); /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) ) */ + vst1_s16( &out[ 2 * k ], out_s16x4 ); /* out[ 2 * k + {0,1,2,3} ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) ); */ + } + + /* Process leftover. */ + if( k < len ) { + int32x4_t in_s32x4; + int32x2_t out32_Q14_s32x2; + + /* S[ 2 * i + 0 ], S[ 2 * i + 1 ]: Q12 */ + in_s16x4 = vld1_lane_s16( &in[ 2 * k + 0 ], in_s16x4, 0 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */ + in_s16x4 = vld1_lane_s16( &in[ 2 * k + 1 ], in_s16x4, 1 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */ + in_s32x4 = vshll_n_s16( in_s16x4, 15 ); /* in{0,1} << 15 */ + t_s32x2 = vqdmulh_lane_s32( vget_low_s32( in_s32x4 ), B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1} ) */ + in_s32x4 = vcombine_s32( vget_low_s32( in_s32x4 ), vget_low_s32( in_s32x4 ) ); /* in{0,1,0,1} << 15 */ + silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, t_s32x2, in_s32x4, &S_s32x4, &out32_Q14_s32x2 ); + + /* Scale back to Q0 and saturate */ + out32_Q14_s32x2 = vadd_s32( out32_Q14_s32x2, offset_s32x2 ); /* out32_Q14_{0,1} + (1<<14) - 1 */ + out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2, out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} + (1<<14) - 1 */ + out_s16x4 = vqshrn_n_s32( out32_Q14_s32x4, 14 ); /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,0,1} + (1<<14) - 1, 14 ) ) */ + vst1_lane_s16( &out[ 2 * k + 0 ], out_s16x4, 0 ); /* out[ 2 * k + 0 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_0 + (1<<14) - 1, 14 ) ); */ + vst1_lane_s16( &out[ 2 * k + 1 ], out_s16x4, 1 ); /* out[ 2 * k + 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_1 + (1<<14) - 1, 14 ) ); */ + } + + vst1q_lane_s32( &S[ 0 ], S_s32x4, 0 ); /* S[ 0 ] = S0; */ + vst1q_lane_s32( &S[ 1 ], S_s32x4, 2 ); /* S[ 1 ] = S2; */ + vst1q_lane_s32( &S[ 2 ], S_s32x4, 1 ); /* S[ 2 ] = S1; */ + vst1q_lane_s32( &S[ 3 ], S_s32x4, 3 ); /* S[ 3 ] = S3; */ + +#ifdef OPUS_CHECK_ASM + silk_assert( !memcmp( S_c, S, sizeof( S_c ) ) ); + silk_assert( !memcmp( out_c, out, 2 * len * sizeof( opus_int16 ) ) ); + RESTORE_STACK; +#endif +} diff --git a/silk/arm/macros_armv4.h b/silk/arm/macros_armv4.h index 3f30e972..877eb18d 100644 --- a/silk/arm/macros_armv4.h +++ b/silk/arm/macros_armv4.h @@ -28,6 +28,11 @@ POSSIBILITY OF SUCH DAMAGE. #ifndef SILK_MACROS_ARMv4_H #define SILK_MACROS_ARMv4_H +/* This macro only avoids the undefined behaviour from a left shift of + a negative value. It should only be used in macros that can't include + SigProc_FIX.h. In other cases, use silk_LSHIFT32(). */ +#define SAFE_SHL(a,b) ((opus_int32)((opus_uint32)(a) << (b))) + /* (a32 * (opus_int32)((opus_int16)(b32))) >> 16 output have to be 32bit int */ #undef silk_SMULWB static OPUS_INLINE opus_int32 silk_SMULWB_armv4(opus_int32 a, opus_int16 b) @@ -38,7 +43,7 @@ static OPUS_INLINE opus_int32 silk_SMULWB_armv4(opus_int32 a, opus_int16 b) "#silk_SMULWB\n\t" "smull %0, %1, %2, %3\n\t" : "=&r"(rd_lo), "=&r"(rd_hi) - : "%r"(a), "r"(b<<16) + : "%r"(a), "r"(SAFE_SHL(b,16)) ); return rd_hi; } @@ -80,7 +85,7 @@ static OPUS_INLINE opus_int32 silk_SMULWW_armv4(opus_int32 a, opus_int32 b) : "=&r"(rd_lo), "=&r"(rd_hi) : "%r"(a), "r"(b) ); - return (rd_hi<<16)+(rd_lo>>16); + return SAFE_SHL(rd_hi,16)+(rd_lo>>16); } #define silk_SMULWW(a, b) (silk_SMULWW_armv4(a, b)) @@ -96,8 +101,10 @@ static OPUS_INLINE opus_int32 silk_SMLAWW_armv4(opus_int32 a, opus_int32 b, : "=&r"(rd_lo), "=&r"(rd_hi) : "%r"(b), "r"(c) ); - return a+(rd_hi<<16)+(rd_lo>>16); + return a+SAFE_SHL(rd_hi,16)+(rd_lo>>16); } #define silk_SMLAWW(a, b, c) (silk_SMLAWW_armv4(a, b, c)) +#undef SAFE_SHL + #endif /* SILK_MACROS_ARMv4_H */ diff --git a/silk/arm/macros_armv5e.h b/silk/arm/macros_armv5e.h index aad4117e..b14ec65d 100644 --- a/silk/arm/macros_armv5e.h +++ b/silk/arm/macros_armv5e.h @@ -29,6 +29,11 @@ POSSIBILITY OF SUCH DAMAGE. #ifndef SILK_MACROS_ARMv5E_H #define SILK_MACROS_ARMv5E_H +/* This macro only avoids the undefined behaviour from a left shift of + a negative value. It should only be used in macros that can't include + SigProc_FIX.h. In other cases, use silk_LSHIFT32(). */ +#define SAFE_SHL(a,b) ((opus_int32)((opus_uint32)(a) << (b))) + /* (a32 * (opus_int32)((opus_int16)(b32))) >> 16 output have to be 32bit int */ #undef silk_SMULWB static OPUS_INLINE opus_int32 silk_SMULWB_armv5e(opus_int32 a, opus_int16 b) @@ -190,7 +195,7 @@ static OPUS_INLINE opus_int32 silk_CLZ16_armv5(opus_int16 in16) "#silk_CLZ16\n\t" "clz %0, %1;\n" : "=r"(res) - : "r"(in16<<16|0x8000) + : "r"(SAFE_SHL(in16,16)|0x8000) ); return res; } @@ -210,4 +215,6 @@ static OPUS_INLINE opus_int32 silk_CLZ32_armv5(opus_int32 in32) } #define silk_CLZ32(in32) (silk_CLZ32_armv5(in32)) +#undef SAFE_SHL + #endif /* SILK_MACROS_ARMv5E_H */ diff --git a/silk/biquad_alt.c b/silk/biquad_alt.c index d55f5ee9..54566a43 100644 --- a/silk/biquad_alt.c +++ b/silk/biquad_alt.c @@ -39,14 +39,13 @@ POSSIBILITY OF SUCH DAMAGE. #include "SigProc_FIX.h" /* Second order ARMA filter, alternative implementation */ -void silk_biquad_alt( +void silk_biquad_alt_stride1( const opus_int16 *in, /* I input signal */ const opus_int32 *B_Q28, /* I MA coefficients [3] */ const opus_int32 *A_Q28, /* I AR coefficients [2] */ opus_int32 *S, /* I/O State vector [2] */ opus_int16 *out, /* O output signal */ - const opus_int32 len, /* I signal length (must be even) */ - opus_int stride /* I Operate on interleaved signal if > 1 */ + const opus_int32 len /* I signal length (must be even) */ ) { /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */ @@ -61,7 +60,7 @@ void silk_biquad_alt( for( k = 0; k < len; k++ ) { /* S[ 0 ], S[ 1 ]: Q12 */ - inval = in[ k * stride ]; + inval = in[ k ]; out32_Q14 = silk_LSHIFT( silk_SMLAWB( S[ 0 ], B_Q28[ 0 ], inval ), 2 ); S[ 0 ] = S[1] + silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14, A0_L_Q28 ), 14 ); @@ -73,6 +72,50 @@ void silk_biquad_alt( S[ 1 ] = silk_SMLAWB( S[ 1 ], B_Q28[ 2 ], inval ); /* Scale back to Q0 and saturate */ - out[ k * stride ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14 + (1<<14) - 1, 14 ) ); + out[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14 + (1<<14) - 1, 14 ) ); + } +} + +void silk_biquad_alt_stride2_c( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ +) +{ + /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */ + opus_int k; + opus_int32 A0_U_Q28, A0_L_Q28, A1_U_Q28, A1_L_Q28, out32_Q14[ 2 ]; + + /* Negate A_Q28 values and split in two parts */ + A0_L_Q28 = ( -A_Q28[ 0 ] ) & 0x00003FFF; /* lower part */ + A0_U_Q28 = silk_RSHIFT( -A_Q28[ 0 ], 14 ); /* upper part */ + A1_L_Q28 = ( -A_Q28[ 1 ] ) & 0x00003FFF; /* lower part */ + A1_U_Q28 = silk_RSHIFT( -A_Q28[ 1 ], 14 ); /* upper part */ + + for( k = 0; k < len; k++ ) { + /* S[ 0 ], S[ 1 ], S[ 2 ], S[ 3 ]: Q12 */ + out32_Q14[ 0 ] = silk_LSHIFT( silk_SMLAWB( S[ 0 ], B_Q28[ 0 ], in[ 2 * k + 0 ] ), 2 ); + out32_Q14[ 1 ] = silk_LSHIFT( silk_SMLAWB( S[ 2 ], B_Q28[ 0 ], in[ 2 * k + 1 ] ), 2 ); + + S[ 0 ] = S[ 1 ] + silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14[ 0 ], A0_L_Q28 ), 14 ); + S[ 2 ] = S[ 3 ] + silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14[ 1 ], A0_L_Q28 ), 14 ); + S[ 0 ] = silk_SMLAWB( S[ 0 ], out32_Q14[ 0 ], A0_U_Q28 ); + S[ 2 ] = silk_SMLAWB( S[ 2 ], out32_Q14[ 1 ], A0_U_Q28 ); + S[ 0 ] = silk_SMLAWB( S[ 0 ], B_Q28[ 1 ], in[ 2 * k + 0 ] ); + S[ 2 ] = silk_SMLAWB( S[ 2 ], B_Q28[ 1 ], in[ 2 * k + 1 ] ); + + S[ 1 ] = silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14[ 0 ], A1_L_Q28 ), 14 ); + S[ 3 ] = silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14[ 1 ], A1_L_Q28 ), 14 ); + S[ 1 ] = silk_SMLAWB( S[ 1 ], out32_Q14[ 0 ], A1_U_Q28 ); + S[ 3 ] = silk_SMLAWB( S[ 3 ], out32_Q14[ 1 ], A1_U_Q28 ); + S[ 1 ] = silk_SMLAWB( S[ 1 ], B_Q28[ 2 ], in[ 2 * k + 0 ] ); + S[ 3 ] = silk_SMLAWB( S[ 3 ], B_Q28[ 2 ], in[ 2 * k + 1 ] ); + + /* Scale back to Q0 and saturate */ + out[ 2 * k + 0 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14[ 0 ] + (1<<14) - 1, 14 ) ); + out[ 2 * k + 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14[ 1 ] + (1<<14) - 1, 14 ) ); } } diff --git a/silk/bwexpander.c b/silk/bwexpander.c index 2eb44566..afa97907 100644 --- a/silk/bwexpander.c +++ b/silk/bwexpander.c @@ -45,7 +45,7 @@ void silk_bwexpander( /* Bias in silk_SMULWB can lead to unstable filters */ for( i = 0; i < d - 1; i++ ) { ar[ i ] = (opus_int16)silk_RSHIFT_ROUND( silk_MUL( chirp_Q16, ar[ i ] ), 16 ); - chirp_Q16 += silk_RSHIFT_ROUND( silk_MUL( chirp_Q16, chirp_minus_one_Q16 ), 16 ); + chirp_Q16 += silk_RSHIFT_ROUND( silk_MUL( chirp_Q16, chirp_minus_one_Q16 ), 16 ); } ar[ d - 1 ] = (opus_int16)silk_RSHIFT_ROUND( silk_MUL( chirp_Q16, ar[ d - 1 ] ), 16 ); } diff --git a/silk/control.h b/silk/control.h index 747e5426..b76ec33c 100644 --- a/silk/control.h +++ b/silk/control.h @@ -77,6 +77,9 @@ typedef struct { /* I: Flag to enable in-band Forward Error Correction (FEC); 0/1 */ opus_int useInBandFEC; + /* I: Flag to actually code in-band Forward Error Correction (FEC) in the current packet; 0/1 */ + opus_int LBRR_coded; + /* I: Flag to enable discontinuous transmission (DTX); 0/1 */ opus_int useDTX; @@ -110,6 +113,11 @@ typedef struct { /* O: Tells the Opus encoder we're ready to switch */ opus_int switchReady; + /* O: SILK Signal type */ + opus_int signalType; + + /* O: SILK offset (dithering) */ + opus_int offset; } silk_EncControlStruct; /**************************************************************************/ diff --git a/silk/control_SNR.c b/silk/control_SNR.c index cee87eb0..464c1acf 100644 --- a/silk/control_SNR.c +++ b/silk/control_SNR.c @@ -64,8 +64,7 @@ opus_int silk_control_SNR( /* Find bitrate interval in table and interpolate */ for( k = 1; k < TARGET_RATE_TAB_SZ; k++ ) { if( TargetRate_bps <= rateTable[ k ] ) { - frac_Q6 = silk_DIV32( silk_LSHIFT( TargetRate_bps - rateTable[ k - 1 ], 6 ), - rateTable[ k ] - rateTable[ k - 1 ] ); + frac_Q6 = silk_DIV32( silk_LSHIFT( TargetRate_bps - rateTable[ k - 1 ], 6 ), rateTable[ k ] - rateTable[ k - 1 ] ); psEncC->SNR_dB_Q7 = silk_LSHIFT( silk_SNR_table_Q1[ k - 1 ], 6 ) + silk_MUL( frac_Q6, silk_SNR_table_Q1[ k ] - silk_SNR_table_Q1[ k - 1 ] ); break; } diff --git a/silk/control_codec.c b/silk/control_codec.c index 044eea3f..9350fd4b 100644 --- a/silk/control_codec.c +++ b/silk/control_codec.c @@ -57,7 +57,7 @@ static opus_int silk_setup_complexity( static OPUS_INLINE opus_int silk_setup_LBRR( silk_encoder_state *psEncC, /* I/O */ - const opus_int32 TargetRate_bps /* I */ + const silk_EncControlStruct *encControl /* I */ ); @@ -65,7 +65,6 @@ static OPUS_INLINE opus_int silk_setup_LBRR( opus_int silk_control_encoder( silk_encoder_state_Fxx *psEnc, /* I/O Pointer to Silk encoder state */ silk_EncControlStruct *encControl, /* I Control structure */ - const opus_int32 TargetRate_bps, /* I Target max bitrate (bps) */ const opus_int allow_bw_switch, /* I Flag to allow switching audio bandwidth */ const opus_int channelNb, /* I Channel number */ const opus_int force_fs_kHz @@ -125,7 +124,7 @@ opus_int silk_control_encoder( /********************************************/ /* Set LBRR usage */ /********************************************/ - ret += silk_setup_LBRR( &psEnc->sCmn, TargetRate_bps ); + ret += silk_setup_LBRR( &psEnc->sCmn, encControl ); psEnc->sCmn.controlled_since_last_payload = 1; @@ -244,7 +243,6 @@ static opus_int silk_setup_fs( if( psEnc->sCmn.fs_kHz != fs_kHz ) { /* reset part of the state */ silk_memset( &psEnc->sShape, 0, sizeof( psEnc->sShape ) ); - silk_memset( &psEnc->sPrefilt, 0, sizeof( psEnc->sPrefilt ) ); silk_memset( &psEnc->sCmn.sNSQ, 0, sizeof( psEnc->sCmn.sNSQ ) ); silk_memset( psEnc->sCmn.prev_NLSFq_Q15, 0, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); silk_memset( &psEnc->sCmn.sLP.In_LP_State, 0, sizeof( psEnc->sCmn.sLP.In_LP_State ) ); @@ -255,7 +253,6 @@ static opus_int silk_setup_fs( /* Initialize non-zero parameters */ psEnc->sCmn.prevLag = 100; psEnc->sCmn.first_frame_after_reset = 1; - psEnc->sPrefilt.lagPrev = 100; psEnc->sShape.LastGainIndex = 10; psEnc->sCmn.sNSQ.lagPrev = 100; psEnc->sCmn.sNSQ.prev_gain_Q16 = 65536; @@ -293,13 +290,10 @@ static opus_int silk_setup_fs( psEnc->sCmn.pitch_LPC_win_length = silk_SMULBB( FIND_PITCH_LPC_WIN_MS_2_SF, fs_kHz ); } if( psEnc->sCmn.fs_kHz == 16 ) { - psEnc->sCmn.mu_LTP_Q9 = SILK_FIX_CONST( MU_LTP_QUANT_WB, 9 ); psEnc->sCmn.pitch_lag_low_bits_iCDF = silk_uniform8_iCDF; } else if( psEnc->sCmn.fs_kHz == 12 ) { - psEnc->sCmn.mu_LTP_Q9 = SILK_FIX_CONST( MU_LTP_QUANT_MB, 9 ); psEnc->sCmn.pitch_lag_low_bits_iCDF = silk_uniform6_iCDF; } else { - psEnc->sCmn.mu_LTP_Q9 = SILK_FIX_CONST( MU_LTP_QUANT_NB, 9 ); psEnc->sCmn.pitch_lag_low_bits_iCDF = silk_uniform4_iCDF; } } @@ -319,60 +313,75 @@ static opus_int silk_setup_complexity( /* Set encoding complexity */ silk_assert( Complexity >= 0 && Complexity <= 10 ); - if( Complexity < 2 ) { + if( Complexity < 1 ) { psEncC->pitchEstimationComplexity = SILK_PE_MIN_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.8, 16 ); psEncC->pitchEstimationLPCOrder = 6; - psEncC->shapingLPCOrder = 8; + psEncC->shapingLPCOrder = 12; psEncC->la_shape = 3 * psEncC->fs_kHz; psEncC->nStatesDelayedDecision = 1; psEncC->useInterpolatedNLSFs = 0; - psEncC->LTPQuantLowComplexity = 1; psEncC->NLSF_MSVQ_Survivors = 2; psEncC->warping_Q16 = 0; - } else if( Complexity < 4 ) { + } else if( Complexity < 2 ) { psEncC->pitchEstimationComplexity = SILK_PE_MID_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.76, 16 ); psEncC->pitchEstimationLPCOrder = 8; - psEncC->shapingLPCOrder = 10; + psEncC->shapingLPCOrder = 14; psEncC->la_shape = 5 * psEncC->fs_kHz; psEncC->nStatesDelayedDecision = 1; psEncC->useInterpolatedNLSFs = 0; - psEncC->LTPQuantLowComplexity = 0; + psEncC->NLSF_MSVQ_Survivors = 3; + psEncC->warping_Q16 = 0; + } else if( Complexity < 3 ) { + psEncC->pitchEstimationComplexity = SILK_PE_MIN_COMPLEX; + psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.8, 16 ); + psEncC->pitchEstimationLPCOrder = 6; + psEncC->shapingLPCOrder = 12; + psEncC->la_shape = 3 * psEncC->fs_kHz; + psEncC->nStatesDelayedDecision = 2; + psEncC->useInterpolatedNLSFs = 0; + psEncC->NLSF_MSVQ_Survivors = 2; + psEncC->warping_Q16 = 0; + } else if( Complexity < 4 ) { + psEncC->pitchEstimationComplexity = SILK_PE_MID_COMPLEX; + psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.76, 16 ); + psEncC->pitchEstimationLPCOrder = 8; + psEncC->shapingLPCOrder = 14; + psEncC->la_shape = 5 * psEncC->fs_kHz; + psEncC->nStatesDelayedDecision = 2; + psEncC->useInterpolatedNLSFs = 0; psEncC->NLSF_MSVQ_Survivors = 4; psEncC->warping_Q16 = 0; } else if( Complexity < 6 ) { psEncC->pitchEstimationComplexity = SILK_PE_MID_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.74, 16 ); psEncC->pitchEstimationLPCOrder = 10; - psEncC->shapingLPCOrder = 12; + psEncC->shapingLPCOrder = 16; psEncC->la_shape = 5 * psEncC->fs_kHz; psEncC->nStatesDelayedDecision = 2; psEncC->useInterpolatedNLSFs = 1; - psEncC->LTPQuantLowComplexity = 0; - psEncC->NLSF_MSVQ_Survivors = 8; + psEncC->NLSF_MSVQ_Survivors = 6; psEncC->warping_Q16 = psEncC->fs_kHz * SILK_FIX_CONST( WARPING_MULTIPLIER, 16 ); } else if( Complexity < 8 ) { psEncC->pitchEstimationComplexity = SILK_PE_MID_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.72, 16 ); psEncC->pitchEstimationLPCOrder = 12; - psEncC->shapingLPCOrder = 14; + psEncC->shapingLPCOrder = 20; psEncC->la_shape = 5 * psEncC->fs_kHz; psEncC->nStatesDelayedDecision = 3; psEncC->useInterpolatedNLSFs = 1; - psEncC->LTPQuantLowComplexity = 0; - psEncC->NLSF_MSVQ_Survivors = 16; + psEncC->NLSF_MSVQ_Survivors = 8; psEncC->warping_Q16 = psEncC->fs_kHz * SILK_FIX_CONST( WARPING_MULTIPLIER, 16 ); } else { psEncC->pitchEstimationComplexity = SILK_PE_MAX_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.7, 16 ); psEncC->pitchEstimationLPCOrder = 16; - psEncC->shapingLPCOrder = 16; + psEncC->shapingLPCOrder = 24; psEncC->la_shape = 5 * psEncC->fs_kHz; psEncC->nStatesDelayedDecision = MAX_DEL_DEC_STATES; psEncC->useInterpolatedNLSFs = 1; - psEncC->LTPQuantLowComplexity = 0; - psEncC->NLSF_MSVQ_Survivors = 32; + psEncC->NLSF_MSVQ_Survivors = 16; psEncC->warping_Q16 = psEncC->fs_kHz * SILK_FIX_CONST( WARPING_MULTIPLIER, 16 ); } @@ -387,40 +396,26 @@ static opus_int silk_setup_complexity( silk_assert( psEncC->warping_Q16 <= 32767 ); silk_assert( psEncC->la_shape <= LA_SHAPE_MAX ); silk_assert( psEncC->shapeWinLength <= SHAPE_LPC_WIN_MAX ); - silk_assert( psEncC->NLSF_MSVQ_Survivors <= NLSF_VQ_MAX_SURVIVORS ); return ret; } static OPUS_INLINE opus_int silk_setup_LBRR( silk_encoder_state *psEncC, /* I/O */ - const opus_int32 TargetRate_bps /* I */ + const silk_EncControlStruct *encControl /* I */ ) { opus_int LBRR_in_previous_packet, ret = SILK_NO_ERROR; - opus_int32 LBRR_rate_thres_bps; LBRR_in_previous_packet = psEncC->LBRR_enabled; - psEncC->LBRR_enabled = 0; - if( psEncC->useInBandFEC && psEncC->PacketLoss_perc > 0 ) { - if( psEncC->fs_kHz == 8 ) { - LBRR_rate_thres_bps = LBRR_NB_MIN_RATE_BPS; - } else if( psEncC->fs_kHz == 12 ) { - LBRR_rate_thres_bps = LBRR_MB_MIN_RATE_BPS; + psEncC->LBRR_enabled = encControl->LBRR_coded; + if( psEncC->LBRR_enabled ) { + /* Set gain increase for coding LBRR excitation */ + if( LBRR_in_previous_packet == 0 ) { + /* Previous packet did not have LBRR, and was therefore coded at a higher bitrate */ + psEncC->LBRR_GainIncreases = 7; } else { - LBRR_rate_thres_bps = LBRR_WB_MIN_RATE_BPS; - } - LBRR_rate_thres_bps = silk_SMULWB( silk_MUL( LBRR_rate_thres_bps, 125 - silk_min( psEncC->PacketLoss_perc, 25 ) ), SILK_FIX_CONST( 0.01, 16 ) ); - - if( TargetRate_bps > LBRR_rate_thres_bps ) { - /* Set gain increase for coding LBRR excitation */ - if( LBRR_in_previous_packet == 0 ) { - /* Previous packet did not have LBRR, and was therefore coded at a higher bitrate */ - psEncC->LBRR_GainIncreases = 7; - } else { - psEncC->LBRR_GainIncreases = silk_max_int( 7 - silk_SMULWB( (opus_int32)psEncC->PacketLoss_perc, SILK_FIX_CONST( 0.4, 16 ) ), 2 ); - } - psEncC->LBRR_enabled = 1; + psEncC->LBRR_GainIncreases = silk_max_int( 7 - silk_SMULWB( (opus_int32)psEncC->PacketLoss_perc, SILK_FIX_CONST( 0.4, 16 ) ), 2 ); } } diff --git a/silk/debug.h b/silk/debug.h index efb6d3e9..6f68c1ca 100644 --- a/silk/debug.h +++ b/silk/debug.h @@ -39,23 +39,10 @@ extern "C" unsigned long GetHighResolutionTime(void); /* O time in usec*/ -/* make SILK_DEBUG dependent on compiler's _DEBUG */ -#if defined _WIN32 - #ifdef _DEBUG - #define SILK_DEBUG 1 - #else - #define SILK_DEBUG 0 - #endif - - /* overrule the above */ - #if 0 - /* #define NO_ASSERTS*/ - #undef SILK_DEBUG - #define SILK_DEBUG 1 - #endif -#else - #define SILK_DEBUG 0 -#endif +/* Set to 1 to enable DEBUG_STORE_DATA() macros for dumping + * intermediate signals from the codec. + */ +#define SILK_DEBUG 0 /* Flag for using timers */ #define SILK_TIC_TOC 0 diff --git a/silk/decode_core.c b/silk/decode_core.c index e569c0e7..fd4e857a 100644 --- a/silk/decode_core.c +++ b/silk/decode_core.c @@ -225,8 +225,6 @@ void silk_decode_core( pxq[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14[ MAX_LPC_ORDER + i ], Gain_Q10 ), 8 ) ); } - /* DEBUG_STORE_DATA( dec.pcm, pxq, psDec->subfr_length * sizeof( opus_int16 ) ) */ - /* Update LPC filter state */ silk_memcpy( sLPC_Q14, &sLPC_Q14[ psDec->subfr_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); pexc_Q14 += psDec->subfr_length; diff --git a/silk/decode_frame.c b/silk/decode_frame.c index a605d95a..dfa73c4a 100644 --- a/silk/decode_frame.c +++ b/silk/decode_frame.c @@ -97,6 +97,7 @@ opus_int silk_decode_frame( psDec->first_frame_after_reset = 0; } else { /* Handle packet loss by extrapolation */ + psDec->indices.signalType = psDec->prevSignalType; silk_PLC( psDec, psDecCtrl, pOut, 1, arch ); } diff --git a/silk/decode_parameters.c b/silk/decode_parameters.c index e345b1dc..a56a4098 100644 --- a/silk/decode_parameters.c +++ b/silk/decode_parameters.c @@ -52,7 +52,7 @@ void silk_decode_parameters( silk_NLSF_decode( pNLSF_Q15, psDec->indices.NLSFIndices, psDec->psNLSF_CB ); /* Convert NLSF parameters to AR prediction filter coefficients */ - silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 1 ], pNLSF_Q15, psDec->LPC_order ); + silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 1 ], pNLSF_Q15, psDec->LPC_order, psDec->arch ); /* If just reset, e.g., because internal Fs changed, do not allow interpolation */ /* improves the case of packet loss in the first frame after a switch */ @@ -69,7 +69,7 @@ void silk_decode_parameters( } /* Convert NLSF parameters to AR prediction filter coefficients */ - silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 0 ], pNLSF0_Q15, psDec->LPC_order ); + silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 0 ], pNLSF0_Q15, psDec->LPC_order, psDec->arch ); } else { /* Copy LPC coefficients for first half from second half */ silk_memcpy( psDecCtrl->PredCoef_Q12[ 0 ], psDecCtrl->PredCoef_Q12[ 1 ], psDec->LPC_order * sizeof( opus_int16 ) ); diff --git a/silk/define.h b/silk/define.h index 19c9b00e..1286048e 100644 --- a/silk/define.h +++ b/silk/define.h @@ -56,6 +56,7 @@ extern "C" /* DTX settings */ #define NB_SPEECH_FRAMES_BEFORE_DTX 10 /* eq 200 ms */ #define MAX_CONSECUTIVE_DTX 20 /* eq 400 ms */ +#define DTX_ACTIVITY_THRESHOLD 0.1f /* Maximum sampling frequency */ #define MAX_FS_KHZ 16 @@ -147,7 +148,7 @@ extern "C" #define USE_HARM_SHAPING 1 /* Max LPC order of noise shaping filters */ -#define MAX_SHAPE_LPC_ORDER 16 +#define MAX_SHAPE_LPC_ORDER 24 #define HARM_SHAPE_FIR_TAPS 3 @@ -157,8 +158,7 @@ extern "C" #define LTP_BUF_LENGTH 512 #define LTP_MASK ( LTP_BUF_LENGTH - 1 ) -#define DECISION_DELAY 32 -#define DECISION_DELAY_MASK ( DECISION_DELAY - 1 ) +#define DECISION_DELAY 40 /* Number of subframes for excitation entropy coding */ #define SHELL_CODEC_FRAME_LENGTH 16 @@ -173,11 +173,7 @@ extern "C" #define MAX_MATRIX_SIZE MAX_LPC_ORDER /* Max of LPC Order and LTP order */ -#if( MAX_LPC_ORDER > DECISION_DELAY ) # define NSQ_LPC_BUF_LENGTH MAX_LPC_ORDER -#else -# define NSQ_LPC_BUF_LENGTH DECISION_DELAY -#endif /***************************/ /* Voice activity detector */ @@ -205,7 +201,6 @@ extern "C" /******************/ #define NLSF_W_Q 2 #define NLSF_VQ_MAX_VECTORS 32 -#define NLSF_VQ_MAX_SURVIVORS 32 #define NLSF_QUANT_MAX_AMPLITUDE 4 #define NLSF_QUANT_MAX_AMPLITUDE_EXT 10 #define NLSF_QUANT_LEVEL_ADJ 0.1 diff --git a/silk/enc_API.c b/silk/enc_API.c index f8060286..701c2905 100644 --- a/silk/enc_API.c +++ b/silk/enc_API.c @@ -233,11 +233,10 @@ opus_int silk_Encode( /* O Returns error co } } - TargetRate_bps = silk_RSHIFT32( encControl->bitRate, encControl->nChannelsInternal - 1 ); for( n = 0; n < encControl->nChannelsInternal; n++ ) { /* Force the side channel to the same rate as the mid */ opus_int force_fs_kHz = (n==1) ? psEnc->state_Fxx[0].sCmn.fs_kHz : 0; - if( ( ret = silk_control_encoder( &psEnc->state_Fxx[ n ], encControl, TargetRate_bps, psEnc->allowBandwidthSwitch, n, force_fs_kHz ) ) != 0 ) { + if( ( ret = silk_control_encoder( &psEnc->state_Fxx[ n ], encControl, psEnc->allowBandwidthSwitch, n, force_fs_kHz ) ) != 0 ) { silk_assert( 0 ); RESTORE_STACK; return ret; @@ -416,7 +415,6 @@ opus_int silk_Encode( /* O Returns error co /* Reset side channel encoder memory for first frame with side coding */ if( psEnc->prev_decode_only_middle == 1 ) { silk_memset( &psEnc->state_Fxx[ 1 ].sShape, 0, sizeof( psEnc->state_Fxx[ 1 ].sShape ) ); - silk_memset( &psEnc->state_Fxx[ 1 ].sPrefilt, 0, sizeof( psEnc->state_Fxx[ 1 ].sPrefilt ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sNSQ, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sNSQ ) ); silk_memset( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15 ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State ) ); @@ -557,6 +555,10 @@ opus_int silk_Encode( /* O Returns error co } } + encControl->signalType = psEnc->state_Fxx[0].sCmn.indices.signalType; + encControl->offset = silk_Quantization_Offsets_Q10 + [ psEnc->state_Fxx[0].sCmn.indices.signalType >> 1 ] + [ psEnc->state_Fxx[0].sCmn.indices.quantOffsetType ]; RESTORE_STACK; return ret; } diff --git a/silk/fixed/arm/warped_autocorrelation_FIX_arm.h b/silk/fixed/arm/warped_autocorrelation_FIX_arm.h new file mode 100644 index 00000000..1992e432 --- /dev/null +++ b/silk/fixed/arm/warped_autocorrelation_FIX_arm.h @@ -0,0 +1,68 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifndef SILK_WARPED_AUTOCORRELATION_FIX_ARM_H +# define SILK_WARPED_AUTOCORRELATION_FIX_ARM_H + +# include "celt/arm/armcpu.h" + +# if defined(FIXED_POINT) + +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +void silk_warped_autocorrelation_FIX_neon( + opus_int32 *corr, /* O Result [order + 1] */ + opus_int *scale, /* O Scaling of the correlation vector */ + const opus_int16 *input, /* I Input data to correlate */ + const opus_int warping_Q16, /* I Warping coefficient */ + const opus_int length, /* I Length of input */ + const opus_int order /* I Correlation order (even) */ +); + +# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON) +# define OVERRIDE_silk_warped_autocorrelation_FIX (1) +# define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \ + ((void)(arch), PRESUME_NEON(silk_warped_autocorrelation_FIX)(corr, scale, input, warping_Q16, length, order)) +# endif +# endif + +# if !defined(OVERRIDE_silk_warped_autocorrelation_FIX) +/*Is run-time CPU detection enabled on this platform?*/ +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern void (*const SILK_WARPED_AUTOCORRELATION_FIX_IMPL[OPUS_ARCHMASK+1])(opus_int32*, opus_int*, const opus_int16*, const opus_int, const opus_int, const opus_int); +# define OVERRIDE_silk_warped_autocorrelation_FIX (1) +# define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \ + ((*SILK_WARPED_AUTOCORRELATION_FIX_IMPL[(arch)&OPUS_ARCHMASK])(corr, scale, input, warping_Q16, length, order)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_silk_warped_autocorrelation_FIX (1) +# define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \ + ((void)(arch), silk_warped_autocorrelation_FIX_neon(corr, scale, input, warping_Q16, length, order)) +# endif +# endif + +# endif /* end FIXED_POINT */ + +#endif /* end SILK_WARPED_AUTOCORRELATION_FIX_ARM_H */ diff --git a/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c b/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c new file mode 100644 index 00000000..00a70cb5 --- /dev/null +++ b/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c @@ -0,0 +1,260 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc., Jean-Marc Valin +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include <arm_neon.h> +#ifdef OPUS_CHECK_ASM +# include <string.h> +#endif +#include "stack_alloc.h" +#include "main_FIX.h" + +static OPUS_INLINE void calc_corr( const opus_int32 *const input_QS, opus_int64 *const corr_QC, const opus_int offset, const int32x4_t state_QS_s32x4 ) +{ + int64x2_t corr_QC_s64x2[ 2 ], t_s64x2[ 2 ]; + const int32x4_t input_QS_s32x4 = vld1q_s32( input_QS + offset ); + corr_QC_s64x2[ 0 ] = vld1q_s64( corr_QC + offset + 0 ); + corr_QC_s64x2[ 1 ] = vld1q_s64( corr_QC + offset + 2 ); + t_s64x2[ 0 ] = vmull_s32( vget_low_s32( state_QS_s32x4 ), vget_low_s32( input_QS_s32x4 ) ); + t_s64x2[ 1 ] = vmull_s32( vget_high_s32( state_QS_s32x4 ), vget_high_s32( input_QS_s32x4 ) ); + corr_QC_s64x2[ 0 ] = vsraq_n_s64( corr_QC_s64x2[ 0 ], t_s64x2[ 0 ], 2 * QS - QC ); + corr_QC_s64x2[ 1 ] = vsraq_n_s64( corr_QC_s64x2[ 1 ], t_s64x2[ 1 ], 2 * QS - QC ); + vst1q_s64( corr_QC + offset + 0, corr_QC_s64x2[ 0 ] ); + vst1q_s64( corr_QC + offset + 2, corr_QC_s64x2[ 1 ] ); +} + +static OPUS_INLINE int32x4_t calc_state( const int32x4_t state_QS0_s32x4, const int32x4_t state_QS0_1_s32x4, const int32x4_t state_QS1_1_s32x4, const int32x4_t warping_Q16_s32x4 ) +{ + int32x4_t t_s32x4 = vsubq_s32( state_QS0_s32x4, state_QS0_1_s32x4 ); + t_s32x4 = vqdmulhq_s32( t_s32x4, warping_Q16_s32x4 ); + return vaddq_s32( state_QS1_1_s32x4, t_s32x4 ); +} + +void silk_warped_autocorrelation_FIX_neon( + opus_int32 *corr, /* O Result [order + 1] */ + opus_int *scale, /* O Scaling of the correlation vector */ + const opus_int16 *input, /* I Input data to correlate */ + const opus_int warping_Q16, /* I Warping coefficient */ + const opus_int length, /* I Length of input */ + const opus_int order /* I Correlation order (even) */ +) +{ + if( ( MAX_SHAPE_LPC_ORDER > 24 ) || ( order < 6 ) ) { + silk_warped_autocorrelation_FIX_c( corr, scale, input, warping_Q16, length, order ); + } else { + opus_int n, i, lsh; + opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; /* In reverse order */ + opus_int64 corr_QC_orderT; + int64x2_t lsh_s64x2; + const opus_int orderT = ( order + 3 ) & ~3; + opus_int64 *corr_QCT; + opus_int32 *input_QS; + VARDECL( opus_int32, input_QST ); + VARDECL( opus_int32, state ); + SAVE_STACK; + + /* Order must be even */ + silk_assert( ( order & 1 ) == 0 ); + silk_assert( 2 * QS - QC >= 0 ); + + ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER, opus_int32 ); + + input_QS = input_QST; + /* input_QS has zero paddings in the beginning and end. */ + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + + /* Loop over samples */ + for( n = 0; n < length - 7; n += 8, input_QS += 8 ) { + const int16x8_t t0_s16x4 = vld1q_s16( input + n ); + vst1q_s32( input_QS + 0, vshll_n_s16( vget_low_s16( t0_s16x4 ), QS ) ); + vst1q_s32( input_QS + 4, vshll_n_s16( vget_high_s16( t0_s16x4 ), QS ) ); + } + for( ; n < length; n++, input_QS++ ) { + input_QS[ 0 ] = silk_LSHIFT32( (opus_int32)input[ n ], QS ); + } + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS = input_QST + MAX_SHAPE_LPC_ORDER - orderT; + + /* The following loop runs ( length + order ) times, with ( order ) extra epilogues. */ + /* The zero paddings in input_QS guarantee corr_QC's correctness even with the extra epilogues. */ + /* The values of state_QS will be polluted by the extra epilogues, however they are temporary values. */ + + /* Keep the C code here to help understand the intrinsics optimization. */ + /* + { + opus_int32 state_QS[ 2 ][ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; + opus_int32 *state_QST[ 3 ]; + state_QST[ 0 ] = state_QS[ 0 ]; + state_QST[ 1 ] = state_QS[ 1 ]; + for( n = 0; n < length + order; n++, input_QS++ ) { + state_QST[ 0 ][ orderT ] = input_QS[ orderT ]; + for( i = 0; i < orderT; i++ ) { + corr_QC[ i ] += silk_RSHIFT64( silk_SMULL( state_QST[ 0 ][ i ], input_QS[ i ] ), 2 * QS - QC ); + state_QST[ 1 ][ i ] = silk_SMLAWB( state_QST[ 1 ][ i + 1 ], state_QST[ 0 ][ i ] - state_QST[ 0 ][ i + 1 ], warping_Q16 ); + } + state_QST[ 2 ] = state_QST[ 0 ]; + state_QST[ 0 ] = state_QST[ 1 ]; + state_QST[ 1 ] = state_QST[ 2 ]; + } + } + */ + + { + const int32x4_t warping_Q16_s32x4 = vdupq_n_s32( warping_Q16 << 15 ); + const opus_int32 *in = input_QS + orderT; + opus_int o = orderT; + int32x4_t state_QS_s32x4[ 3 ][ 2 ]; + + ALLOC( state, length + orderT, opus_int32 ); + state_QS_s32x4[ 2 ][ 1 ] = vdupq_n_s32( 0 ); + + /* Calculate 8 taps of all inputs in each loop. */ + do { + state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 0 ][ 1 ] = + state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 1 ][ 1 ] = vdupq_n_s32( 0 ); + n = 0; + do { + calc_corr( input_QS + n, corr_QC, o - 8, state_QS_s32x4[ 0 ][ 0 ] ); + calc_corr( input_QS + n, corr_QC, o - 4, state_QS_s32x4[ 0 ][ 1 ] ); + state_QS_s32x4[ 2 ][ 1 ] = vld1q_s32( in + n ); + vst1q_lane_s32( state + n, state_QS_s32x4[ 0 ][ 0 ], 0 ); + state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 0 ][ 1 ], 1 ); + state_QS_s32x4[ 2 ][ 1 ] = vextq_s32( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], 1 ); + state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 ); + state_QS_s32x4[ 0 ][ 1 ] = calc_state( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], state_QS_s32x4[ 1 ][ 1 ], warping_Q16_s32x4 ); + state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ]; + state_QS_s32x4[ 1 ][ 1 ] = state_QS_s32x4[ 2 ][ 1 ]; + } while( ++n < ( length + order ) ); + in = state; + o -= 8; + } while( o > 4 ); + + if( o ) { + /* Calculate the last 4 taps of all inputs. */ + opus_int32 *stateT = state; + silk_assert( o == 4 ); + state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 1 ][ 0 ] = vdupq_n_s32( 0 ); + n = length + order; + do { + calc_corr( input_QS, corr_QC, 0, state_QS_s32x4[ 0 ][ 0 ] ); + state_QS_s32x4[ 2 ][ 0 ] = vld1q_s32( stateT ); + vst1q_lane_s32( stateT, state_QS_s32x4[ 0 ][ 0 ], 0 ); + state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], 1 ); + state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 ); + state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ]; + input_QS++; + stateT++; + } while( --n ); + } + } + + { + const opus_int16 *inputT = input; + int32x4_t t_s32x4; + int64x1_t t_s64x1; + int64x2_t t_s64x2 = vdupq_n_s64( 0 ); + for( n = 0; n <= length - 8; n += 8 ) { + int16x8_t input_s16x8 = vld1q_s16( inputT ); + t_s32x4 = vmull_s16( vget_low_s16( input_s16x8 ), vget_low_s16( input_s16x8 ) ); + t_s32x4 = vmlal_s16( t_s32x4, vget_high_s16( input_s16x8 ), vget_high_s16( input_s16x8 ) ); + t_s64x2 = vaddw_s32( t_s64x2, vget_low_s32( t_s32x4 ) ); + t_s64x2 = vaddw_s32( t_s64x2, vget_high_s32( t_s32x4 ) ); + inputT += 8; + } + t_s64x1 = vadd_s64( vget_low_s64( t_s64x2 ), vget_high_s64( t_s64x2 ) ); + corr_QC_orderT = vget_lane_s64( t_s64x1, 0 ); + for( ; n < length; n++ ) { + corr_QC_orderT += silk_SMULL( input[ n ], input[ n ] ); + } + corr_QC_orderT = silk_LSHIFT64( corr_QC_orderT, QC ); + corr_QC[ orderT ] = corr_QC_orderT; + } + + corr_QCT = corr_QC + orderT - order; + lsh = silk_CLZ64( corr_QC_orderT ) - 35; + lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC ); + *scale = -( QC + lsh ); + silk_assert( *scale >= -30 && *scale <= 12 ); + lsh_s64x2 = vdupq_n_s64( lsh ); + for( i = 0; i <= order - 3; i += 4 ) { + int32x4_t corr_s32x4; + int64x2_t corr_QC0_s64x2, corr_QC1_s64x2; + corr_QC0_s64x2 = vld1q_s64( corr_QCT + i ); + corr_QC1_s64x2 = vld1q_s64( corr_QCT + i + 2 ); + corr_QC0_s64x2 = vshlq_s64( corr_QC0_s64x2, lsh_s64x2 ); + corr_QC1_s64x2 = vshlq_s64( corr_QC1_s64x2, lsh_s64x2 ); + corr_s32x4 = vcombine_s32( vmovn_s64( corr_QC1_s64x2 ), vmovn_s64( corr_QC0_s64x2 ) ); + corr_s32x4 = vrev64q_s32( corr_s32x4 ); + vst1q_s32( corr + order - i - 3, corr_s32x4 ); + } + if( lsh >= 0 ) { + for( ; i < order + 1; i++ ) { + corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QCT[ i ], lsh ) ); + } + } else { + for( ; i < order + 1; i++ ) { + corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QCT[ i ], -lsh ) ); + } + } + silk_assert( corr_QCT[ order ] >= 0 ); /* If breaking, decrease QC*/ + RESTORE_STACK; + } + +#ifdef OPUS_CHECK_ASM + { + opus_int32 corr_c[ MAX_SHAPE_LPC_ORDER + 1 ]; + opus_int scale_c; + silk_warped_autocorrelation_FIX_c( corr_c, &scale_c, input, warping_Q16, length, order ); + silk_assert( !memcmp( corr_c, corr, sizeof( corr_c[ 0 ] ) * ( order + 1 ) ) ); + silk_assert( scale_c == *scale ); + } +#endif +} diff --git a/silk/fixed/burg_modified_FIX.c b/silk/fixed/burg_modified_FIX.c index 17d0e099..80edbdfa 100644 --- a/silk/fixed/burg_modified_FIX.c +++ b/silk/fixed/burg_modified_FIX.c @@ -37,7 +37,7 @@ POSSIBILITY OF SUCH DAMAGE. #define MAX_FRAME_SIZE 384 /* subfr_length * nb_subfr = ( 0.005 * 16000 + 16 ) * 4 = 384 */ #define QA 25 -#define N_BITS_HEAD_ROOM 2 +#define N_BITS_HEAD_ROOM 3 #define MIN_RSHIFTS -16 #define MAX_RSHIFTS (32 - QA) diff --git a/silk/fixed/corrMatrix_FIX.c b/silk/fixed/corrMatrix_FIX.c index c1d437c7..1b4a29c2 100644 --- a/silk/fixed/corrMatrix_FIX.c +++ b/silk/fixed/corrMatrix_FIX.c @@ -58,7 +58,7 @@ void silk_corrVector_FIX( for( lag = 0; lag < order; lag++ ) { inner_prod = 0; for( i = 0; i < L; i++ ) { - inner_prod += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts ); + inner_prod = silk_ADD_RSHIFT32( inner_prod, silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts ); } Xt[ lag ] = inner_prod; /* X[:,lag]'*t */ ptr1--; /* Go to next column of X */ @@ -77,61 +77,54 @@ void silk_corrMatrix_FIX( const opus_int16 *x, /* I x vector [L + order - 1] used to form data matrix X */ const opus_int L, /* I Length of vectors */ const opus_int order, /* I Max lag for correlation */ - const opus_int head_room, /* I Desired headroom */ opus_int32 *XX, /* O Pointer to X'*X correlation matrix [ order x order ] */ - opus_int *rshifts, /* I/O Right shifts of correlations */ + opus_int32 *nrg, /* O Energy of x vector */ + opus_int *rshifts, /* O Right shifts of correlations and energy */ int arch /* I Run-time architecture */ ) { - opus_int i, j, lag, rshifts_local, head_room_rshifts; + opus_int i, j, lag; opus_int32 energy; const opus_int16 *ptr1, *ptr2; /* Calculate energy to find shift used to fit in 32 bits */ - silk_sum_sqr_shift( &energy, &rshifts_local, x, L + order - 1 ); - /* Add shifts to get the desired head room */ - head_room_rshifts = silk_max( head_room - silk_CLZ32( energy ), 0 ); - - energy = silk_RSHIFT32( energy, head_room_rshifts ); - rshifts_local += head_room_rshifts; + silk_sum_sqr_shift( nrg, rshifts, x, L + order - 1 ); + energy = *nrg; /* Calculate energy of first column (0) of X: X[:,0]'*X[:,0] */ /* Remove contribution of first order - 1 samples */ for( i = 0; i < order - 1; i++ ) { - energy -= silk_RSHIFT32( silk_SMULBB( x[ i ], x[ i ] ), rshifts_local ); - } - if( rshifts_local < *rshifts ) { - /* Adjust energy */ - energy = silk_RSHIFT32( energy, *rshifts - rshifts_local ); - rshifts_local = *rshifts; + energy -= silk_RSHIFT32( silk_SMULBB( x[ i ], x[ i ] ), *rshifts ); } /* Calculate energy of remaining columns of X: X[:,j]'*X[:,j] */ /* Fill out the diagonal of the correlation matrix */ matrix_ptr( XX, 0, 0, order ) = energy; + silk_assert( energy >= 0 ); ptr1 = &x[ order - 1 ]; /* First sample of column 0 of X */ for( j = 1; j < order; j++ ) { - energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr1[ L - j ] ), rshifts_local ) ); - energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr1[ -j ] ), rshifts_local ) ); + energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr1[ L - j ] ), *rshifts ) ); + energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr1[ -j ] ), *rshifts ) ); matrix_ptr( XX, j, j, order ) = energy; + silk_assert( energy >= 0 ); } ptr2 = &x[ order - 2 ]; /* First sample of column 1 of X */ /* Calculate the remaining elements of the correlation matrix */ - if( rshifts_local > 0 ) { + if( *rshifts > 0 ) { /* Right shifting used */ for( lag = 1; lag < order; lag++ ) { /* Inner product of column 0 and column lag: X[:,0]'*X[:,lag] */ energy = 0; for( i = 0; i < L; i++ ) { - energy += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts_local ); + energy += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), *rshifts ); } /* Calculate remaining off diagonal: X[:,j]'*X[:,j + lag] */ matrix_ptr( XX, lag, 0, order ) = energy; matrix_ptr( XX, 0, lag, order ) = energy; for( j = 1; j < ( order - lag ); j++ ) { - energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr2[ L - j ] ), rshifts_local ) ); - energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr2[ -j ] ), rshifts_local ) ); + energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr2[ L - j ] ), *rshifts ) ); + energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr2[ -j ] ), *rshifts ) ); matrix_ptr( XX, lag + j, j, order ) = energy; matrix_ptr( XX, j, lag + j, order ) = energy; } @@ -153,6 +146,5 @@ void silk_corrMatrix_FIX( ptr2--;/* Update pointer to first sample of next column (lag) in X */ } } - *rshifts = rshifts_local; } diff --git a/silk/fixed/encode_frame_FIX.c b/silk/fixed/encode_frame_FIX.c index 5ef44b03..4f9e0861 100644 --- a/silk/fixed/encode_frame_FIX.c +++ b/silk/fixed/encode_frame_FIX.c @@ -29,6 +29,7 @@ POSSIBILITY OF SUCH DAMAGE. #include "config.h" #endif +#include <stdlib.h> #include "main_FIX.h" #include "stack_alloc.h" #include "tuning_parameters.h" @@ -37,7 +38,7 @@ POSSIBILITY OF SUCH DAMAGE. static OPUS_INLINE void silk_LBRR_encode_FIX( silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O Pointer to Silk FIX encoder control struct */ - const opus_int32 xfw_Q3[], /* I Input signal */ + const opus_int16 x16[], /* I Input signal */ opus_int condCoding /* I The type of conditional coding used so far for this frame */ ); @@ -94,6 +95,9 @@ opus_int silk_encode_frame_FIX( opus_int16 ec_prevLagIndex_copy; opus_int ec_prevSignalType_copy; opus_int8 LastGainIndex_copy2; + opus_int gain_lock[ MAX_NB_SUBFR ] = {0}; + opus_int16 best_gain_mult[ MAX_NB_SUBFR ]; + opus_int best_sum[ MAX_NB_SUBFR ]; SAVE_STACK; /* This is totally unnecessary but many compilers (including gcc) are too dumb to realise it */ @@ -118,7 +122,6 @@ opus_int silk_encode_frame_FIX( silk_memcpy( x_frame + LA_SHAPE_MS * psEnc->sCmn.fs_kHz, psEnc->sCmn.inputBuf + 1, psEnc->sCmn.frame_length * sizeof( opus_int16 ) ); if( !psEnc->sCmn.prefillFlag ) { - VARDECL( opus_int32, xfw_Q3 ); VARDECL( opus_int16, res_pitch ); VARDECL( opus_uint8, ec_buf_copy ); opus_int16 *res_pitch_frame; @@ -132,7 +135,7 @@ opus_int silk_encode_frame_FIX( /*****************************************/ /* Find pitch lags, initial LPC analysis */ /*****************************************/ - silk_find_pitch_lags_FIX( psEnc, &sEncCtrl, res_pitch, x_frame, psEnc->sCmn.arch ); + silk_find_pitch_lags_FIX( psEnc, &sEncCtrl, res_pitch, x_frame - psEnc->sCmn.ltp_mem_length, psEnc->sCmn.arch ); /************************/ /* Noise shape analysis */ @@ -142,23 +145,17 @@ opus_int silk_encode_frame_FIX( /***************************************************/ /* Find linear prediction coefficients (LPC + LTP) */ /***************************************************/ - silk_find_pred_coefs_FIX( psEnc, &sEncCtrl, res_pitch, x_frame, condCoding ); + silk_find_pred_coefs_FIX( psEnc, &sEncCtrl, res_pitch_frame, x_frame, condCoding ); /****************************************/ /* Process gains */ /****************************************/ silk_process_gains_FIX( psEnc, &sEncCtrl, condCoding ); - /*****************************************/ - /* Prefiltering for noise shaper */ - /*****************************************/ - ALLOC( xfw_Q3, psEnc->sCmn.frame_length, opus_int32 ); - silk_prefilter_FIX( psEnc, &sEncCtrl, xfw_Q3, x_frame ); - /****************************************/ /* Low Bitrate Redundant Encoding */ /****************************************/ - silk_LBRR_encode_FIX( psEnc, &sEncCtrl, xfw_Q3, condCoding ); + silk_LBRR_encode_FIX( psEnc, &sEncCtrl, x_frame, condCoding ); /* Loop over quantizer and entropy coding to control bitrate */ maxIter = 6; @@ -194,17 +191,21 @@ opus_int silk_encode_frame_FIX( /* Noise shaping quantization */ /*****************************************/ if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { - silk_NSQ_del_dec( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, xfw_Q3, psEnc->sCmn.pulses, - sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR2_Q13, sEncCtrl.HarmShapeGain_Q14, + silk_NSQ_del_dec( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, x_frame, psEnc->sCmn.pulses, + sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR_Q13, sEncCtrl.HarmShapeGain_Q14, sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14, psEnc->sCmn.arch ); } else { - silk_NSQ( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, xfw_Q3, psEnc->sCmn.pulses, - sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR2_Q13, sEncCtrl.HarmShapeGain_Q14, + silk_NSQ( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, x_frame, psEnc->sCmn.pulses, + sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR_Q13, sEncCtrl.HarmShapeGain_Q14, sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14, psEnc->sCmn.arch); } + if ( iter == maxIter && !found_lower ) { + silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) ); + } + /****************************************/ /* Encode Parameters */ /****************************************/ @@ -218,6 +219,33 @@ opus_int silk_encode_frame_FIX( nBits = ec_tell( psRangeEnc ); + /* If we still bust after the last iteration, do some damage control. */ + if ( iter == maxIter && !found_lower && nBits > maxBits ) { + silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) ); + + /* Keep gains the same as the last frame. */ + psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev; + for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + psEnc->sCmn.indices.GainsIndices[ i ] = 4; + } + if (condCoding != CODE_CONDITIONALLY) { + psEnc->sCmn.indices.GainsIndices[ 0 ] = sEncCtrl.lastGainIndexPrev; + } + psEnc->sCmn.ec_prevLagIndex = ec_prevLagIndex_copy; + psEnc->sCmn.ec_prevSignalType = ec_prevSignalType_copy; + /* Clear all pulses. */ + for ( i = 0; i < psEnc->sCmn.frame_length; i++ ) { + psEnc->sCmn.pulses[ i ] = 0; + } + + silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding ); + + silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType, + psEnc->sCmn.pulses, psEnc->sCmn.frame_length ); + + nBits = ec_tell( psRangeEnc ); + } + if( useCBR == 0 && iter == 0 && nBits <= maxBits ) { break; } @@ -265,15 +293,35 @@ opus_int silk_encode_frame_FIX( break; } + if ( !found_lower && nBits > maxBits ) { + int j; + for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + int sum=0; + for ( j = i*psEnc->sCmn.subfr_length; j < (i+1)*psEnc->sCmn.subfr_length; j++ ) { + sum += abs( psEnc->sCmn.pulses[j] ); + } + if ( iter == 0 || (sum < best_sum[i] && !gain_lock[i]) ) { + best_sum[i] = sum; + best_gain_mult[i] = gainMult_Q8; + } else { + gain_lock[i] = 1; + } + } + } if( ( found_lower & found_upper ) == 0 ) { /* Adjust gain according to high-rate rate/distortion curve */ - opus_int32 gain_factor_Q16; - gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) ); - gain_factor_Q16 = silk_min_32( gain_factor_Q16, SILK_FIX_CONST( 2, 16 ) ); if( nBits > maxBits ) { - gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) ); + if (gainMult_Q8 < 16384) { + gainMult_Q8 *= 2; + } else { + gainMult_Q8 = 32767; + } + } else { + opus_int32 gain_factor_Q16; + gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) ); + gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 ); } - gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 ); + } else { /* Adjust gain by interpolating */ gainMult_Q8 = gainMult_lower + silk_DIV32_16( silk_MUL( gainMult_upper - gainMult_lower, maxBits - nBits_lower ), nBits_upper - nBits_lower ); @@ -287,7 +335,13 @@ opus_int silk_encode_frame_FIX( } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { - sEncCtrl.Gains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q8 ), 8 ); + opus_int16 tmp; + if ( gain_lock[i] ) { + tmp = best_gain_mult[i]; + } else { + tmp = gainMult_Q8; + } + sEncCtrl.Gains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], tmp ), 8 ); } /* Quantize gains */ @@ -331,7 +385,7 @@ opus_int silk_encode_frame_FIX( static OPUS_INLINE void silk_LBRR_encode_FIX( silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O Pointer to Silk FIX encoder control struct */ - const opus_int32 xfw_Q3[], /* I Input signal */ + const opus_int16 x16[], /* I Input signal */ opus_int condCoding /* I The type of conditional coding used so far for this frame */ ) { @@ -370,14 +424,14 @@ static OPUS_INLINE void silk_LBRR_encode_FIX( /* Noise shaping quantization */ /*****************************************/ if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { - silk_NSQ_del_dec( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q3, + silk_NSQ_del_dec( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, x16, psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], psEncCtrl->PredCoef_Q12[ 0 ], psEncCtrl->LTPCoef_Q14, - psEncCtrl->AR2_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, + psEncCtrl->AR_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, psEncCtrl->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14, psEnc->sCmn.arch ); } else { - silk_NSQ( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q3, + silk_NSQ( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, x16, psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], psEncCtrl->PredCoef_Q12[ 0 ], psEncCtrl->LTPCoef_Q14, - psEncCtrl->AR2_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, + psEncCtrl->AR_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, psEncCtrl->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14, psEnc->sCmn.arch ); } diff --git a/silk/fixed/find_LPC_FIX.c b/silk/fixed/find_LPC_FIX.c index e11cdc86..e55b63ac 100644 --- a/silk/fixed/find_LPC_FIX.c +++ b/silk/fixed/find_LPC_FIX.c @@ -92,7 +92,7 @@ void silk_find_LPC_FIX( silk_interpolate( NLSF0_Q15, psEncC->prev_NLSFq_Q15, NLSF_Q15, k, psEncC->predictLPCOrder ); /* Convert to LPC for residual energy evaluation */ - silk_NLSF2A( a_tmp_Q12, NLSF0_Q15, psEncC->predictLPCOrder ); + silk_NLSF2A( a_tmp_Q12, NLSF0_Q15, psEncC->predictLPCOrder, psEncC->arch ); /* Calculate residual energy with NLSF interpolation */ silk_LPC_analysis_filter( LPC_res, x, a_tmp_Q12, 2 * subfr_length, psEncC->predictLPCOrder, psEncC->arch ); diff --git a/silk/fixed/find_LTP_FIX.c b/silk/fixed/find_LTP_FIX.c index 1314a281..62d4afb2 100644 --- a/silk/fixed/find_LTP_FIX.c +++ b/silk/fixed/find_LTP_FIX.c @@ -32,214 +32,68 @@ POSSIBILITY OF SUCH DAMAGE. #include "main_FIX.h" #include "tuning_parameters.h" -/* Head room for correlations */ -#define LTP_CORRS_HEAD_ROOM 2 - -void silk_fit_LTP( - opus_int32 LTP_coefs_Q16[ LTP_ORDER ], - opus_int16 LTP_coefs_Q14[ LTP_ORDER ] -); - void silk_find_LTP_FIX( - opus_int16 b_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */ - opus_int32 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ - opus_int *LTPredCodGain_Q7, /* O LTP coding gain */ - const opus_int16 r_lpc[], /* I residual signal after LPC signal + state for first 10 ms */ + opus_int32 XXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Correlation matrix */ + opus_int32 xXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER ], /* O Correlation vector */ + const opus_int16 r_ptr[], /* I Residual signal after LPC */ const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ - const opus_int32 Wght_Q15[ MAX_NB_SUBFR ], /* I weights */ - const opus_int subfr_length, /* I subframe length */ - const opus_int nb_subfr, /* I number of subframes */ - const opus_int mem_offset, /* I number of samples in LTP memory */ - opus_int corr_rshifts[ MAX_NB_SUBFR ], /* O right shifts applied to correlations */ + const opus_int subfr_length, /* I Subframe length */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ) { - opus_int i, k, lshift; - const opus_int16 *r_ptr, *lag_ptr; - opus_int16 *b_Q14_ptr; - - opus_int32 regu; - opus_int32 *WLTP_ptr; - opus_int32 b_Q16[ LTP_ORDER ], delta_b_Q14[ LTP_ORDER ], d_Q14[ MAX_NB_SUBFR ], nrg[ MAX_NB_SUBFR ], g_Q26; - opus_int32 w[ MAX_NB_SUBFR ], WLTP_max, max_abs_d_Q14, max_w_bits; - - opus_int32 temp32, denom32; - opus_int extra_shifts; - opus_int rr_shifts, maxRshifts, maxRshifts_wxtra, LZs; - opus_int32 LPC_res_nrg, LPC_LTP_res_nrg, div_Q16; - opus_int32 Rr[ LTP_ORDER ], rr[ MAX_NB_SUBFR ]; - opus_int32 wd, m_Q12; - - b_Q14_ptr = b_Q14; - WLTP_ptr = WLTP; - r_ptr = &r_lpc[ mem_offset ]; + opus_int i, k, extra_shifts; + opus_int xx_shifts, xX_shifts, XX_shifts; + const opus_int16 *lag_ptr; + opus_int32 *XXLTP_Q17_ptr, *xXLTP_Q17_ptr; + opus_int32 xx, nrg, temp; + + xXLTP_Q17_ptr = xXLTP_Q17; + XXLTP_Q17_ptr = XXLTP_Q17; for( k = 0; k < nb_subfr; k++ ) { lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 ); - silk_sum_sqr_shift( &rr[ k ], &rr_shifts, r_ptr, subfr_length ); /* rr[ k ] in Q( -rr_shifts ) */ - - /* Assure headroom */ - LZs = silk_CLZ32( rr[k] ); - if( LZs < LTP_CORRS_HEAD_ROOM ) { - rr[ k ] = silk_RSHIFT_ROUND( rr[ k ], LTP_CORRS_HEAD_ROOM - LZs ); - rr_shifts += ( LTP_CORRS_HEAD_ROOM - LZs ); - } - corr_rshifts[ k ] = rr_shifts; - silk_corrMatrix_FIX( lag_ptr, subfr_length, LTP_ORDER, LTP_CORRS_HEAD_ROOM, WLTP_ptr, &corr_rshifts[ k ], arch ); /* WLTP_fix_ptr in Q( -corr_rshifts[ k ] ) */ - - /* The correlation vector always has lower max abs value than rr and/or RR so head room is assured */ - silk_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, Rr, corr_rshifts[ k ], arch ); /* Rr_fix_ptr in Q( -corr_rshifts[ k ] ) */ - if( corr_rshifts[ k ] > rr_shifts ) { - rr[ k ] = silk_RSHIFT( rr[ k ], corr_rshifts[ k ] - rr_shifts ); /* rr[ k ] in Q( -corr_rshifts[ k ] ) */ + silk_sum_sqr_shift( &xx, &xx_shifts, r_ptr, subfr_length + LTP_ORDER ); /* xx in Q( -xx_shifts ) */ + silk_corrMatrix_FIX( lag_ptr, subfr_length, LTP_ORDER, XXLTP_Q17_ptr, &nrg, &XX_shifts, arch ); /* XXLTP_Q17_ptr and nrg in Q( -XX_shifts ) */ + extra_shifts = xx_shifts - XX_shifts; + if( extra_shifts > 0 ) { + /* Shift XX */ + xX_shifts = xx_shifts; + for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) { + XXLTP_Q17_ptr[ i ] = silk_RSHIFT32( XXLTP_Q17_ptr[ i ], extra_shifts ); /* Q( -xX_shifts ) */ + } + nrg = silk_RSHIFT32( nrg, extra_shifts ); /* Q( -xX_shifts ) */ + } else if( extra_shifts < 0 ) { + /* Shift xx */ + xX_shifts = XX_shifts; + xx = silk_RSHIFT32( xx, -extra_shifts ); /* Q( -xX_shifts ) */ + } else { + xX_shifts = xx_shifts; } - silk_assert( rr[ k ] >= 0 ); - - regu = 1; - regu = silk_SMLAWB( regu, rr[ k ], SILK_FIX_CONST( LTP_DAMPING/3, 16 ) ); - regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, 0, 0, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) ); - regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, LTP_ORDER-1, LTP_ORDER-1, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) ); - silk_regularize_correlations_FIX( WLTP_ptr, &rr[k], regu, LTP_ORDER ); - - silk_solve_LDL_FIX( WLTP_ptr, LTP_ORDER, Rr, b_Q16 ); /* WLTP_fix_ptr and Rr_fix_ptr both in Q(-corr_rshifts[k]) */ - - /* Limit and store in Q14 */ - silk_fit_LTP( b_Q16, b_Q14_ptr ); - - /* Calculate residual energy */ - nrg[ k ] = silk_residual_energy16_covar_FIX( b_Q14_ptr, WLTP_ptr, Rr, rr[ k ], LTP_ORDER, 14 ); /* nrg_fix in Q( -corr_rshifts[ k ] ) */ - - /* temp = Wght[ k ] / ( nrg[ k ] * Wght[ k ] + 0.01f * subfr_length ); */ - extra_shifts = silk_min_int( corr_rshifts[ k ], LTP_CORRS_HEAD_ROOM ); - denom32 = silk_LSHIFT_SAT32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 + extra_shifts ) + /* Q( -corr_rshifts[ k ] + extra_shifts ) */ - silk_RSHIFT( silk_SMULWB( (opus_int32)subfr_length, 655 ), corr_rshifts[ k ] - extra_shifts ); /* Q( -corr_rshifts[ k ] + extra_shifts ) */ - denom32 = silk_max( denom32, 1 ); - silk_assert( ((opus_int64)Wght_Q15[ k ] << 16 ) < silk_int32_MAX ); /* Wght always < 0.5 in Q0 */ - temp32 = silk_DIV32( silk_LSHIFT( (opus_int32)Wght_Q15[ k ], 16 ), denom32 ); /* Q( 15 + 16 + corr_rshifts[k] - extra_shifts ) */ - temp32 = silk_RSHIFT( temp32, 31 + corr_rshifts[ k ] - extra_shifts - 26 ); /* Q26 */ + silk_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, xXLTP_Q17_ptr, xX_shifts, arch ); /* xXLTP_Q17_ptr in Q( -xX_shifts ) */ - /* Limit temp such that the below scaling never wraps around */ - WLTP_max = 0; + /* At this point all correlations are in Q(-xX_shifts) */ + temp = silk_SMLAWB( 1, nrg, SILK_FIX_CONST( LTP_CORR_INV_MAX, 16 ) ); + temp = silk_max( temp, xx ); +TIC(div) +#if 0 for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) { - WLTP_max = silk_max( WLTP_ptr[ i ], WLTP_max ); + XXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( XXLTP_Q17_ptr[ i ], temp, 17 ); } - lshift = silk_CLZ32( WLTP_max ) - 1 - 3; /* keep 3 bits free for vq_nearest_neighbor_fix */ - silk_assert( 26 - 18 + lshift >= 0 ); - if( 26 - 18 + lshift < 31 ) { - temp32 = silk_min_32( temp32, silk_LSHIFT( (opus_int32)1, 26 - 18 + lshift ) ); - } - - silk_scale_vector32_Q26_lshift_18( WLTP_ptr, temp32, LTP_ORDER * LTP_ORDER ); /* WLTP_ptr in Q( 18 - corr_rshifts[ k ] ) */ - - w[ k ] = matrix_ptr( WLTP_ptr, LTP_ORDER/2, LTP_ORDER/2, LTP_ORDER ); /* w in Q( 18 - corr_rshifts[ k ] ) */ - silk_assert( w[k] >= 0 ); - - r_ptr += subfr_length; - b_Q14_ptr += LTP_ORDER; - WLTP_ptr += LTP_ORDER * LTP_ORDER; - } - - maxRshifts = 0; - for( k = 0; k < nb_subfr; k++ ) { - maxRshifts = silk_max_int( corr_rshifts[ k ], maxRshifts ); - } - - /* Compute LTP coding gain */ - if( LTPredCodGain_Q7 != NULL ) { - LPC_LTP_res_nrg = 0; - LPC_res_nrg = 0; - silk_assert( LTP_CORRS_HEAD_ROOM >= 2 ); /* Check that no overflow will happen when adding */ - for( k = 0; k < nb_subfr; k++ ) { - LPC_res_nrg = silk_ADD32( LPC_res_nrg, silk_RSHIFT( silk_ADD32( silk_SMULWB( rr[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */ - LPC_LTP_res_nrg = silk_ADD32( LPC_LTP_res_nrg, silk_RSHIFT( silk_ADD32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */ - } - LPC_LTP_res_nrg = silk_max( LPC_LTP_res_nrg, 1 ); /* avoid division by zero */ - - div_Q16 = silk_DIV32_varQ( LPC_res_nrg, LPC_LTP_res_nrg, 16 ); - *LTPredCodGain_Q7 = ( opus_int )silk_SMULBB( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) ); - - silk_assert( *LTPredCodGain_Q7 == ( opus_int )silk_SAT16( silk_MUL( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) ) ) ); - } - - /* smoothing */ - /* d = sum( B, 1 ); */ - b_Q14_ptr = b_Q14; - for( k = 0; k < nb_subfr; k++ ) { - d_Q14[ k ] = 0; for( i = 0; i < LTP_ORDER; i++ ) { - d_Q14[ k ] += b_Q14_ptr[ i ]; - } - b_Q14_ptr += LTP_ORDER; - } - - /* m = ( w * d' ) / ( sum( w ) + 1e-3 ); */ - - /* Find maximum absolute value of d_Q14 and the bits used by w in Q0 */ - max_abs_d_Q14 = 0; - max_w_bits = 0; - for( k = 0; k < nb_subfr; k++ ) { - max_abs_d_Q14 = silk_max_32( max_abs_d_Q14, silk_abs( d_Q14[ k ] ) ); - /* w[ k ] is in Q( 18 - corr_rshifts[ k ] ) */ - /* Find bits needed in Q( 18 - maxRshifts ) */ - max_w_bits = silk_max_32( max_w_bits, 32 - silk_CLZ32( w[ k ] ) + corr_rshifts[ k ] - maxRshifts ); - } - - /* max_abs_d_Q14 = (5 << 15); worst case, i.e. LTP_ORDER * -silk_int16_MIN */ - silk_assert( max_abs_d_Q14 <= ( 5 << 15 ) ); - - /* How many bits is needed for w*d' in Q( 18 - maxRshifts ) in the worst case, of all d_Q14's being equal to max_abs_d_Q14 */ - extra_shifts = max_w_bits + 32 - silk_CLZ32( max_abs_d_Q14 ) - 14; - - /* Subtract what we got available; bits in output var plus maxRshifts */ - extra_shifts -= ( 32 - 1 - 2 + maxRshifts ); /* Keep sign bit free as well as 2 bits for accumulation */ - extra_shifts = silk_max_int( extra_shifts, 0 ); - - maxRshifts_wxtra = maxRshifts + extra_shifts; - - temp32 = silk_RSHIFT( 262, maxRshifts + extra_shifts ) + 1; /* 1e-3f in Q( 18 - (maxRshifts + extra_shifts) ) */ - wd = 0; - for( k = 0; k < nb_subfr; k++ ) { - /* w has at least 2 bits of headroom so no overflow should happen */ - temp32 = silk_ADD32( temp32, silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ) ); /* Q( 18 - maxRshifts_wxtra ) */ - wd = silk_ADD32( wd, silk_LSHIFT( silk_SMULWW( silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ), d_Q14[ k ] ), 2 ) ); /* Q( 18 - maxRshifts_wxtra ) */ - } - m_Q12 = silk_DIV32_varQ( wd, temp32, 12 ); - - b_Q14_ptr = b_Q14; - for( k = 0; k < nb_subfr; k++ ) { - /* w_fix[ k ] from Q( 18 - corr_rshifts[ k ] ) to Q( 16 ) */ - if( 2 - corr_rshifts[k] > 0 ) { - temp32 = silk_RSHIFT( w[ k ], 2 - corr_rshifts[ k ] ); - } else { - temp32 = silk_LSHIFT_SAT32( w[ k ], corr_rshifts[ k ] - 2 ); + xXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( xXLTP_Q17_ptr[ i ], temp, 17 ); } - - g_Q26 = silk_MUL( - silk_DIV32( - SILK_FIX_CONST( LTP_SMOOTHING, 26 ), - silk_RSHIFT( SILK_FIX_CONST( LTP_SMOOTHING, 26 ), 10 ) + temp32 ), /* Q10 */ - silk_LSHIFT_SAT32( silk_SUB_SAT32( (opus_int32)m_Q12, silk_RSHIFT( d_Q14[ k ], 2 ) ), 4 ) ); /* Q16 */ - - temp32 = 0; - for( i = 0; i < LTP_ORDER; i++ ) { - delta_b_Q14[ i ] = silk_max_16( b_Q14_ptr[ i ], 1638 ); /* 1638_Q14 = 0.1_Q0 */ - temp32 += delta_b_Q14[ i ]; /* Q14 */ +#else + for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) { + XXLTP_Q17_ptr[ i ] = (opus_int32)( silk_LSHIFT64( (opus_int64)XXLTP_Q17_ptr[ i ], 17 ) / temp ); } - temp32 = silk_DIV32( g_Q26, temp32 ); /* Q14 -> Q12 */ for( i = 0; i < LTP_ORDER; i++ ) { - b_Q14_ptr[ i ] = silk_LIMIT_32( (opus_int32)b_Q14_ptr[ i ] + silk_SMULWB( silk_LSHIFT_SAT32( temp32, 4 ), delta_b_Q14[ i ] ), -16000, 28000 ); + xXLTP_Q17_ptr[ i ] = (opus_int32)( silk_LSHIFT64( (opus_int64)xXLTP_Q17_ptr[ i ], 17 ) / temp ); } - b_Q14_ptr += LTP_ORDER; - } -} - -void silk_fit_LTP( - opus_int32 LTP_coefs_Q16[ LTP_ORDER ], - opus_int16 LTP_coefs_Q14[ LTP_ORDER ] -) -{ - opus_int i; - - for( i = 0; i < LTP_ORDER; i++ ) { - LTP_coefs_Q14[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( LTP_coefs_Q16[ i ], 2 ) ); +#endif +TOC(div) + r_ptr += subfr_length; + XXLTP_Q17_ptr += LTP_ORDER * LTP_ORDER; + xXLTP_Q17_ptr += LTP_ORDER; } } diff --git a/silk/fixed/find_pitch_lags_FIX.c b/silk/fixed/find_pitch_lags_FIX.c index b8440a82..9303e9db 100644 --- a/silk/fixed/find_pitch_lags_FIX.c +++ b/silk/fixed/find_pitch_lags_FIX.c @@ -44,7 +44,7 @@ void silk_find_pitch_lags_FIX( { opus_int buf_len, i, scale; opus_int32 thrhld_Q13, res_nrg; - const opus_int16 *x_buf, *x_buf_ptr; + const opus_int16 *x_ptr; VARDECL( opus_int16, Wsig ); opus_int16 *Wsig_ptr; opus_int32 auto_corr[ MAX_FIND_PITCH_LPC_ORDER + 1 ]; @@ -61,8 +61,6 @@ void silk_find_pitch_lags_FIX( /* Safety check */ silk_assert( buf_len >= psEnc->sCmn.pitch_LPC_win_length ); - x_buf = x - psEnc->sCmn.ltp_mem_length; - /*************************************/ /* Estimate LPC AR coefficients */ /*************************************/ @@ -72,19 +70,19 @@ void silk_find_pitch_lags_FIX( ALLOC( Wsig, psEnc->sCmn.pitch_LPC_win_length, opus_int16 ); /* First LA_LTP samples */ - x_buf_ptr = x_buf + buf_len - psEnc->sCmn.pitch_LPC_win_length; + x_ptr = x + buf_len - psEnc->sCmn.pitch_LPC_win_length; Wsig_ptr = Wsig; - silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 1, psEnc->sCmn.la_pitch ); + silk_apply_sine_window( Wsig_ptr, x_ptr, 1, psEnc->sCmn.la_pitch ); /* Middle un - windowed samples */ Wsig_ptr += psEnc->sCmn.la_pitch; - x_buf_ptr += psEnc->sCmn.la_pitch; - silk_memcpy( Wsig_ptr, x_buf_ptr, ( psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ) ) * sizeof( opus_int16 ) ); + x_ptr += psEnc->sCmn.la_pitch; + silk_memcpy( Wsig_ptr, x_ptr, ( psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ) ) * sizeof( opus_int16 ) ); /* Last LA_LTP samples */ Wsig_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ); - x_buf_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ); - silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 2, psEnc->sCmn.la_pitch ); + x_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ); + silk_apply_sine_window( Wsig_ptr, x_ptr, 2, psEnc->sCmn.la_pitch ); /* Calculate autocorrelation sequence */ silk_autocorr( auto_corr, &scale, Wsig, psEnc->sCmn.pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1, arch ); @@ -112,7 +110,7 @@ void silk_find_pitch_lags_FIX( /*****************************************/ /* LPC analysis filtering */ /*****************************************/ - silk_LPC_analysis_filter( res, x_buf, A_Q12, buf_len, psEnc->sCmn.pitchEstimationLPCOrder, psEnc->sCmn.arch ); + silk_LPC_analysis_filter( res, x, A_Q12, buf_len, psEnc->sCmn.pitchEstimationLPCOrder, psEnc->sCmn.arch ); if( psEnc->sCmn.indices.signalType != TYPE_NO_VOICE_ACTIVITY && psEnc->sCmn.first_frame_after_reset == 0 ) { /* Threshold for pitch estimator */ diff --git a/silk/fixed/find_pred_coefs_FIX.c b/silk/fixed/find_pred_coefs_FIX.c index d308e9cf..24c6aab3 100644 --- a/silk/fixed/find_pred_coefs_FIX.c +++ b/silk/fixed/find_pred_coefs_FIX.c @@ -41,13 +41,12 @@ void silk_find_pred_coefs_FIX( ) { opus_int i; - opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ], Wght_Q15[ MAX_NB_SUBFR ]; + opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ]; opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const opus_int16 *x_ptr; opus_int16 *x_pre_ptr; VARDECL( opus_int16, LPC_in_pre ); - opus_int32 tmp, min_gain_Q16, minInvGain_Q30; - opus_int LTP_corrs_rshift[ MAX_NB_SUBFR ]; + opus_int32 min_gain_Q16, minInvGain_Q30; SAVE_STACK; /* weighting for weighted least squares */ @@ -61,13 +60,11 @@ void silk_find_pred_coefs_FIX( /* Invert and normalize gains, and ensure that maximum invGains_Q16 is within range of a 16 bit int */ invGains_Q16[ i ] = silk_DIV32_varQ( min_gain_Q16, psEncCtrl->Gains_Q16[ i ], 16 - 2 ); - /* Ensure Wght_Q15 a minimum value 1 */ - invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 363 ); + /* Limit inverse */ + invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 100 ); /* Square the inverted gains */ silk_assert( invGains_Q16[ i ] == silk_SAT16( invGains_Q16[ i ] ) ); - tmp = silk_SMULWB( invGains_Q16[ i ], invGains_Q16[ i ] ); - Wght_Q15[ i ] = silk_RSHIFT( tmp, 1 ); /* Invert the inverted and normalized gains */ local_gains[ i ] = silk_DIV32( ( (opus_int32)1 << 16 ), invGains_Q16[ i ] ); @@ -77,24 +74,24 @@ void silk_find_pred_coefs_FIX( psEnc->sCmn.nb_subfr * psEnc->sCmn.predictLPCOrder + psEnc->sCmn.frame_length, opus_int16 ); if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { - VARDECL( opus_int32, WLTP ); + VARDECL( opus_int32, xXLTP_Q17 ); + VARDECL( opus_int32, XXLTP_Q17 ); /**********/ /* VOICED */ /**********/ silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); - ALLOC( WLTP, psEnc->sCmn.nb_subfr * LTP_ORDER * LTP_ORDER, opus_int32 ); + ALLOC( xXLTP_Q17, psEnc->sCmn.nb_subfr * LTP_ORDER, opus_int32 ); + ALLOC( XXLTP_Q17, psEnc->sCmn.nb_subfr * LTP_ORDER * LTP_ORDER, opus_int32 ); /* LTP analysis */ - silk_find_LTP_FIX( psEncCtrl->LTPCoef_Q14, WLTP, &psEncCtrl->LTPredCodGain_Q7, - res_pitch, psEncCtrl->pitchL, Wght_Q15, psEnc->sCmn.subfr_length, - psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length, LTP_corrs_rshift, psEnc->sCmn.arch ); + silk_find_LTP_FIX( XXLTP_Q17, xXLTP_Q17, res_pitch, + psEncCtrl->pitchL, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch ); /* Quantize LTP gain parameters */ silk_quant_LTP_gains( psEncCtrl->LTPCoef_Q14, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, - &psEnc->sCmn.sum_log_gain_Q7, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr, - psEnc->sCmn.arch); + &psEnc->sCmn.sum_log_gain_Q7, &psEncCtrl->LTPredCodGain_Q7, XXLTP_Q17, xXLTP_Q17, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch ); /* Control LTP scaling */ silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding ); diff --git a/silk/fixed/k2a_FIX.c b/silk/fixed/k2a_FIX.c index 5fee599b..549f6ead 100644 --- a/silk/fixed/k2a_FIX.c +++ b/silk/fixed/k2a_FIX.c @@ -39,14 +39,15 @@ void silk_k2a( ) { opus_int k, n; - opus_int32 Atmp[ SILK_MAX_ORDER_LPC ]; + opus_int32 rc, tmp1, tmp2; for( k = 0; k < order; k++ ) { - for( n = 0; n < k; n++ ) { - Atmp[ n ] = A_Q24[ n ]; - } - for( n = 0; n < k; n++ ) { - A_Q24[ n ] = silk_SMLAWB( A_Q24[ n ], silk_LSHIFT( Atmp[ k - n - 1 ], 1 ), rc_Q15[ k ] ); + rc = rc_Q15[ k ]; + for( n = 0; n < (k + 1) >> 1; n++ ) { + tmp1 = A_Q24[ n ]; + tmp2 = A_Q24[ k - n - 1 ]; + A_Q24[ n ] = silk_SMLAWB( tmp1, silk_LSHIFT( tmp2, 1 ), rc ); + A_Q24[ k - n - 1 ] = silk_SMLAWB( tmp2, silk_LSHIFT( tmp1, 1 ), rc ); } A_Q24[ k ] = -silk_LSHIFT( (opus_int32)rc_Q15[ k ], 9 ); } diff --git a/silk/fixed/k2a_Q16_FIX.c b/silk/fixed/k2a_Q16_FIX.c index 3b039875..1595aa62 100644 --- a/silk/fixed/k2a_Q16_FIX.c +++ b/silk/fixed/k2a_Q16_FIX.c @@ -39,15 +39,16 @@ void silk_k2a_Q16( ) { opus_int k, n; - opus_int32 Atmp[ SILK_MAX_ORDER_LPC ]; + opus_int32 rc, tmp1, tmp2; for( k = 0; k < order; k++ ) { - for( n = 0; n < k; n++ ) { - Atmp[ n ] = A_Q24[ n ]; + rc = rc_Q16[ k ]; + for( n = 0; n < (k + 1) >> 1; n++ ) { + tmp1 = A_Q24[ n ]; + tmp2 = A_Q24[ k - n - 1 ]; + A_Q24[ n ] = silk_SMLAWW( tmp1, tmp2, rc ); + A_Q24[ k - n - 1 ] = silk_SMLAWW( tmp2, tmp1, rc ); } - for( n = 0; n < k; n++ ) { - A_Q24[ n ] = silk_SMLAWW( A_Q24[ n ], Atmp[ k - n - 1 ], rc_Q16[ k ] ); - } - A_Q24[ k ] = -silk_LSHIFT( rc_Q16[ k ], 8 ); + A_Q24[ k ] = -silk_LSHIFT( rc, 8 ); } } diff --git a/silk/fixed/main_FIX.h b/silk/fixed/main_FIX.h index 375b5eb3..780afa39 100644 --- a/silk/fixed/main_FIX.h +++ b/silk/fixed/main_FIX.h @@ -36,6 +36,11 @@ POSSIBILITY OF SUCH DAMAGE. #include "debug.h" #include "entenc.h" +#if ((defined(OPUS_ARM_ASM) && defined(FIXED_POINT)) \ + || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) +#include "fixed/arm/warped_autocorrelation_FIX_arm.h" +#endif + #ifndef FORCE_CPP_BUILD #ifdef __cplusplus extern "C" @@ -47,6 +52,9 @@ extern "C" #define silk_encode_do_VAD_Fxx silk_encode_do_VAD_FIX #define silk_encode_frame_Fxx silk_encode_frame_FIX +#define QC 10 +#define QS 13 + /*********************/ /* Encoder Functions */ /*********************/ @@ -81,33 +89,11 @@ opus_int silk_init_encoder( opus_int silk_control_encoder( silk_encoder_state_Fxx *psEnc, /* I/O Pointer to Silk encoder state */ silk_EncControlStruct *encControl, /* I Control structure */ - const opus_int32 TargetRate_bps, /* I Target max bitrate (bps) */ const opus_int allow_bw_switch, /* I Flag to allow switching audio bandwidth */ const opus_int channelNb, /* I Channel number */ const opus_int force_fs_kHz ); -/****************/ -/* Prefiltering */ -/****************/ -void silk_prefilter_FIX( - silk_encoder_state_FIX *psEnc, /* I/O Encoder state */ - const silk_encoder_control_FIX *psEncCtrl, /* I Encoder control */ - opus_int32 xw_Q10[], /* O Weighted signal */ - const opus_int16 x[] /* I Speech signal */ -); - -void silk_warped_LPC_analysis_filter_FIX_c( - opus_int32 state[], /* I/O State [order + 1] */ - opus_int32 res_Q2[], /* O Residual signal [length] */ - const opus_int16 coef_Q13[], /* I Coefficients [order] */ - const opus_int16 input[], /* I Input signal [length] */ - const opus_int16 lambda_Q16, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -); - - /**************************/ /* Noise shaping analysis */ /**************************/ @@ -121,7 +107,7 @@ void silk_noise_shape_analysis_FIX( ); /* Autocorrelations for a warped frequency axis */ -void silk_warped_autocorrelation_FIX( +void silk_warped_autocorrelation_FIX_c( opus_int32 *corr, /* O Result [order + 1] */ opus_int *scale, /* O Scaling of the correlation vector */ const opus_int16 *input, /* I Input data to correlate */ @@ -130,6 +116,11 @@ void silk_warped_autocorrelation_FIX( const opus_int order /* I Correlation order (even) */ ); +#if !defined(OVERRIDE_silk_warped_autocorrelation_FIX) +#define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \ + ((void)(arch), silk_warped_autocorrelation_FIX_c(corr, scale, input, warping_Q16, length, order)) +#endif + /* Calculation of LTP state scaling */ void silk_LTP_scale_ctrl_FIX( silk_encoder_state_FIX *psEnc, /* I/O encoder state */ @@ -168,16 +159,12 @@ void silk_find_LPC_FIX( /* LTP analysis */ void silk_find_LTP_FIX( - opus_int16 b_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */ - opus_int32 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ - opus_int *LTPredCodGain_Q7, /* O LTP coding gain */ - const opus_int16 r_lpc[], /* I residual signal after LPC signal + state for first 10 ms */ + opus_int32 XXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Correlation matrix */ + opus_int32 xXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER ], /* O Correlation vector */ + const opus_int16 r_lpc[], /* I Residual signal after LPC */ const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ - const opus_int32 Wght_Q15[ MAX_NB_SUBFR ], /* I weights */ - const opus_int subfr_length, /* I subframe length */ - const opus_int nb_subfr, /* I number of subframes */ - const opus_int mem_offset, /* I number of samples in LTP memory */ - opus_int corr_rshifts[ MAX_NB_SUBFR ], /* O right shifts applied to correlations */ + const opus_int subfr_length, /* I Subframe length */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ); @@ -231,9 +218,9 @@ void silk_corrMatrix_FIX( const opus_int16 *x, /* I x vector [L + order - 1] used to form data matrix X */ const opus_int L, /* I Length of vectors */ const opus_int order, /* I Max lag for correlation */ - const opus_int head_room, /* I Desired headroom */ opus_int32 *XX, /* O Pointer to X'*X correlation matrix [ order x order ] */ - opus_int *rshifts, /* I/O Right shifts of correlations */ + opus_int32 *nrg, /* O Energy of x vector */ + opus_int *rshifts, /* O Right shifts of correlations */ int arch /* I Run-time architecture */ ); @@ -248,22 +235,6 @@ void silk_corrVector_FIX( int arch /* I Run-time architecture */ ); -/* Add noise to matrix diagonal */ -void silk_regularize_correlations_FIX( - opus_int32 *XX, /* I/O Correlation matrices */ - opus_int32 *xx, /* I/O Correlation values */ - opus_int32 noise, /* I Noise to add */ - opus_int D /* I Dimension of XX */ -); - -/* Solves Ax = b, assuming A is symmetric */ -void silk_solve_LDL_FIX( - opus_int32 *A, /* I Pointer to symetric square matrix A */ - opus_int M, /* I Size of matrix */ - const opus_int32 *b, /* I Pointer to b vector */ - opus_int32 *x_Q16 /* O Pointer to x solution vector */ -); - #ifndef FORCE_CPP_BUILD #ifdef __cplusplus } diff --git a/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h b/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h index c30481e4..3999b5bd 100644 --- a/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h +++ b/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h @@ -169,7 +169,7 @@ void silk_noise_shape_analysis_FIX( if( psEnc->sCmn.warping_Q16 > 0 ) { /* Calculate warped auto correlation */ - silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder ); + silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder, arch ); } else { /* Calculate regular auto correlation */ silk_autocorr( auto_corr, &scale, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1, arch ); @@ -224,8 +224,8 @@ void silk_noise_shape_analysis_FIX( silk_bwexpander_32( AR1_Q24, psEnc->sCmn.shapingLPCOrder, BWExp1_Q16 ); /* Ratio of prediction gains, in energy domain */ - pre_nrg_Q30 = silk_LPC_inverse_pred_gain_Q24( AR2_Q24, psEnc->sCmn.shapingLPCOrder ); - nrg = silk_LPC_inverse_pred_gain_Q24( AR1_Q24, psEnc->sCmn.shapingLPCOrder ); + pre_nrg_Q30 = silk_LPC_inverse_pred_gain_Q24( AR2_Q24, psEnc->sCmn.shapingLPCOrder, arch ); + nrg = silk_LPC_inverse_pred_gain_Q24( AR1_Q24, psEnc->sCmn.shapingLPCOrder, arch ); /*psEncCtrl->GainsPre[ k ] = 1.0f - 0.7f * ( 1.0f - pre_nrg / nrg ) = 0.3f + 0.7f * pre_nrg / nrg;*/ pre_nrg_Q30 = silk_LSHIFT32( silk_SMULWB( pre_nrg_Q30, SILK_FIX_CONST( 0.7, 15 ) ), 1 ); diff --git a/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h b/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h index e803ef0f..fcbd96c8 100644 --- a/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h +++ b/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h @@ -48,7 +48,8 @@ void silk_warped_autocorrelation_FIX( const opus_int16 *input, /* I Input data to correlate */ const opus_int warping_Q16, /* I Warping coefficient */ const opus_int length, /* I Length of input */ - const opus_int order /* I Correlation order (even) */ + const opus_int order, /* I Correlation order (even) */ + int arch /* I Run-time architecture */ ) { opus_int n, i, lsh; diff --git a/silk/fixed/noise_shape_analysis_FIX.c b/silk/fixed/noise_shape_analysis_FIX.c index 22a89f75..85fea0bf 100644 --- a/silk/fixed/noise_shape_analysis_FIX.c +++ b/silk/fixed/noise_shape_analysis_FIX.c @@ -57,88 +57,79 @@ static OPUS_INLINE opus_int32 warped_gain( /* gain in Q16*/ /* Convert warped filter coefficients to monic pseudo-warped coefficients and limit maximum */ /* amplitude of monic warped coefficients by using bandwidth expansion on the true coefficients */ static OPUS_INLINE void limit_warped_coefs( - opus_int32 *coefs_syn_Q24, - opus_int32 *coefs_ana_Q24, + opus_int32 *coefs_Q24, opus_int lambda_Q16, opus_int32 limit_Q24, opus_int order ) { opus_int i, iter, ind = 0; - opus_int32 tmp, maxabs_Q24, chirp_Q16, gain_syn_Q16, gain_ana_Q16; + opus_int32 tmp, maxabs_Q24, chirp_Q16, gain_Q16; opus_int32 nom_Q16, den_Q24; + opus_int32 limit_Q20, maxabs_Q20; /* Convert to monic coefficients */ lambda_Q16 = -lambda_Q16; for( i = order - 1; i > 0; i-- ) { - coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 ); - coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 ); + coefs_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); } lambda_Q16 = -lambda_Q16; - nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -(opus_int32)lambda_Q16, lambda_Q16 ); - den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_syn_Q24[ 0 ], lambda_Q16 ); - gain_syn_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); - den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_ana_Q24[ 0 ], lambda_Q16 ); - gain_ana_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); + nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -(opus_int32)lambda_Q16, lambda_Q16 ); + den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_Q24[ 0 ], lambda_Q16 ); + gain_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); for( i = 0; i < order; i++ ) { - coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] ); - coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] ); + coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ i ] ); } - + limit_Q20 = silk_RSHIFT(limit_Q24, 4); for( iter = 0; iter < 10; iter++ ) { /* Find maximum absolute value */ maxabs_Q24 = -1; for( i = 0; i < order; i++ ) { - tmp = silk_max( silk_abs_int32( coefs_syn_Q24[ i ] ), silk_abs_int32( coefs_ana_Q24[ i ] ) ); + tmp = silk_abs_int32( coefs_Q24[ i ] ); if( tmp > maxabs_Q24 ) { maxabs_Q24 = tmp; ind = i; } } - if( maxabs_Q24 <= limit_Q24 ) { + /* Use Q20 to avoid any overflow when multiplying by (ind + 1) later. */ + maxabs_Q20 = silk_RSHIFT(maxabs_Q24, 4); + if( maxabs_Q20 <= limit_Q20 ) { /* Coefficients are within range - done */ return; } /* Convert back to true warped coefficients */ for( i = 1; i < order; i++ ) { - coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 ); - coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 ); + coefs_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); } - gain_syn_Q16 = silk_INVERSE32_varQ( gain_syn_Q16, 32 ); - gain_ana_Q16 = silk_INVERSE32_varQ( gain_ana_Q16, 32 ); + gain_Q16 = silk_INVERSE32_varQ( gain_Q16, 32 ); for( i = 0; i < order; i++ ) { - coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] ); - coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] ); + coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ i ] ); } /* Apply bandwidth expansion */ chirp_Q16 = SILK_FIX_CONST( 0.99, 16 ) - silk_DIV32_varQ( - silk_SMULWB( maxabs_Q24 - limit_Q24, silk_SMLABB( SILK_FIX_CONST( 0.8, 10 ), SILK_FIX_CONST( 0.1, 10 ), iter ) ), - silk_MUL( maxabs_Q24, ind + 1 ), 22 ); - silk_bwexpander_32( coefs_syn_Q24, order, chirp_Q16 ); - silk_bwexpander_32( coefs_ana_Q24, order, chirp_Q16 ); + silk_SMULWB( maxabs_Q20 - limit_Q20, silk_SMLABB( SILK_FIX_CONST( 0.8, 10 ), SILK_FIX_CONST( 0.1, 10 ), iter ) ), + silk_MUL( maxabs_Q20, ind + 1 ), 22 ); + silk_bwexpander_32( coefs_Q24, order, chirp_Q16 ); /* Convert to monic warped coefficients */ lambda_Q16 = -lambda_Q16; for( i = order - 1; i > 0; i-- ) { - coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 ); - coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 ); + coefs_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); } lambda_Q16 = -lambda_Q16; nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -(opus_int32)lambda_Q16, lambda_Q16 ); - den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_syn_Q24[ 0 ], lambda_Q16 ); - gain_syn_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); - den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_ana_Q24[ 0 ], lambda_Q16 ); - gain_ana_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); + den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_Q24[ 0 ], lambda_Q16 ); + gain_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); for( i = 0; i < order; i++ ) { - coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] ); - coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] ); + coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ i ] ); } } silk_assert( 0 ); } -#if defined(MIPSr1_ASM) +/* Disable MIPS version until it's updated. */ +#if 0 && defined(MIPSr1_ASM) #include "mips/noise_shape_analysis_FIX_mipsr1.h" #endif @@ -155,14 +146,13 @@ void silk_noise_shape_analysis_FIX( ) { silk_shape_state_FIX *psShapeSt = &psEnc->sShape; - opus_int k, i, nSamples, Qnrg, b_Q14, warping_Q16, scale = 0; - opus_int32 SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32; - opus_int32 nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7; - opus_int32 delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8; + opus_int k, i, nSamples, nSegs, Qnrg, b_Q14, warping_Q16, scale = 0; + opus_int32 SNR_adj_dB_Q7, HarmShapeGain_Q16, Tilt_Q16, tmp32; + opus_int32 nrg, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7; + opus_int32 BWExp_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8; opus_int32 auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ]; opus_int32 refl_coef_Q16[ MAX_SHAPE_LPC_ORDER ]; - opus_int32 AR1_Q24[ MAX_SHAPE_LPC_ORDER ]; - opus_int32 AR2_Q24[ MAX_SHAPE_LPC_ORDER ]; + opus_int32 AR_Q24[ MAX_SHAPE_LPC_ORDER ]; VARDECL( opus_int16, x_windowed ); const opus_int16 *x_ptr, *pitch_res_ptr; SAVE_STACK; @@ -209,14 +199,14 @@ void silk_noise_shape_analysis_FIX( if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /* Initially set to 0; may be overruled in process_gains(..) */ psEnc->sCmn.indices.quantOffsetType = 0; - psEncCtrl->sparseness_Q8 = 0; } else { /* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */ nSamples = silk_LSHIFT( psEnc->sCmn.fs_kHz, 1 ); energy_variation_Q7 = 0; log_energy_prev_Q7 = 0; pitch_res_ptr = pitch_res; - for( k = 0; k < silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; k++ ) { + nSegs = silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; + for( k = 0; k < nSegs; k++ ) { silk_sum_sqr_shift( &nrg, &scale, pitch_res_ptr, nSamples ); nrg += silk_RSHIFT( nSamples, scale ); /* Q(-scale)*/ @@ -228,18 +218,12 @@ void silk_noise_shape_analysis_FIX( pitch_res_ptr += nSamples; } - psEncCtrl->sparseness_Q8 = silk_RSHIFT( silk_sigm_Q15( silk_SMULWB( energy_variation_Q7 - - SILK_FIX_CONST( 5.0, 7 ), SILK_FIX_CONST( 0.1, 16 ) ) ), 7 ); - /* Set quantization offset depending on sparseness measure */ - if( psEncCtrl->sparseness_Q8 > SILK_FIX_CONST( SPARSENESS_THRESHOLD_QNT_OFFSET, 8 ) ) { + if( energy_variation_Q7 > SILK_FIX_CONST( ENERGY_VARIATION_THRESHOLD_QNT_OFFSET, 7 ) * (nSegs-1) ) { psEnc->sCmn.indices.quantOffsetType = 0; } else { psEnc->sCmn.indices.quantOffsetType = 1; } - - /* Increase coding SNR for sparse signals */ - SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7, SILK_FIX_CONST( SPARSE_SNR_INCR_dB, 15 ), psEncCtrl->sparseness_Q8 - SILK_FIX_CONST( 0.5, 8 ) ); } /*******************************/ @@ -247,14 +231,8 @@ void silk_noise_shape_analysis_FIX( /*******************************/ /* More BWE for signals with high prediction gain */ strength_Q16 = silk_SMULWB( psEncCtrl->predGain_Q16, SILK_FIX_CONST( FIND_PITCH_WHITE_NOISE_FRACTION, 16 ) ); - BWExp1_Q16 = BWExp2_Q16 = silk_DIV32_varQ( SILK_FIX_CONST( BANDWIDTH_EXPANSION, 16 ), + BWExp_Q16 = silk_DIV32_varQ( SILK_FIX_CONST( BANDWIDTH_EXPANSION, 16 ), silk_SMLAWW( SILK_FIX_CONST( 1.0, 16 ), strength_Q16, strength_Q16 ), 16 ); - delta_Q16 = silk_SMULWB( SILK_FIX_CONST( 1.0, 16 ) - silk_SMULBB( 3, psEncCtrl->coding_quality_Q14 ), - SILK_FIX_CONST( LOW_RATE_BANDWIDTH_EXPANSION_DELTA, 16 ) ); - BWExp1_Q16 = silk_SUB32( BWExp1_Q16, delta_Q16 ); - BWExp2_Q16 = silk_ADD32( BWExp2_Q16, delta_Q16 ); - /* BWExp1 will be applied after BWExp2, so make it relative */ - BWExp1_Q16 = silk_DIV32_16( silk_LSHIFT( BWExp1_Q16, 14 ), silk_RSHIFT( BWExp2_Q16, 2 ) ); if( psEnc->sCmn.warping_Q16 > 0 ) { /* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */ @@ -284,7 +262,7 @@ void silk_noise_shape_analysis_FIX( if( psEnc->sCmn.warping_Q16 > 0 ) { /* Calculate warped auto correlation */ - silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder ); + silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder, arch ); } else { /* Calculate regular auto correlation */ silk_autocorr( auto_corr, &scale, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1, arch ); @@ -299,7 +277,7 @@ void silk_noise_shape_analysis_FIX( silk_assert( nrg >= 0 ); /* Convert reflection coefficients to prediction coefficients */ - silk_k2a_Q16( AR2_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder ); + silk_k2a_Q16( AR_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder ); Qnrg = -scale; /* range: -12...30*/ silk_assert( Qnrg >= -12 ); @@ -318,40 +296,34 @@ void silk_noise_shape_analysis_FIX( if( psEnc->sCmn.warping_Q16 > 0 ) { /* Adjust gain for warping */ - gain_mult_Q16 = warped_gain( AR2_Q24, warping_Q16, psEnc->sCmn.shapingLPCOrder ); - silk_assert( psEncCtrl->Gains_Q16[ k ] >= 0 ); - if ( silk_SMULWW( silk_RSHIFT_ROUND( psEncCtrl->Gains_Q16[ k ], 1 ), gain_mult_Q16 ) >= ( silk_int32_MAX >> 1 ) ) { - psEncCtrl->Gains_Q16[ k ] = silk_int32_MAX; + gain_mult_Q16 = warped_gain( AR_Q24, warping_Q16, psEnc->sCmn.shapingLPCOrder ); + silk_assert( psEncCtrl->Gains_Q16[ k ] > 0 ); + if( psEncCtrl->Gains_Q16[ k ] < SILK_FIX_CONST( 0.25, 16 ) ) { + psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( psEncCtrl->Gains_Q16[ k ], gain_mult_Q16 ); } else { - psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( psEncCtrl->Gains_Q16[ k ], gain_mult_Q16 ); + psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( silk_RSHIFT_ROUND( psEncCtrl->Gains_Q16[ k ], 1 ), gain_mult_Q16 ); + if ( psEncCtrl->Gains_Q16[ k ] >= ( silk_int32_MAX >> 1 ) ) { + psEncCtrl->Gains_Q16[ k ] = silk_int32_MAX; + } else { + psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT32( psEncCtrl->Gains_Q16[ k ], 1 ); + } } + silk_assert( psEncCtrl->Gains_Q16[ k ] > 0 ); } - /* Bandwidth expansion for synthesis filter shaping */ - silk_bwexpander_32( AR2_Q24, psEnc->sCmn.shapingLPCOrder, BWExp2_Q16 ); - - /* Compute noise shaping filter coefficients */ - silk_memcpy( AR1_Q24, AR2_Q24, psEnc->sCmn.shapingLPCOrder * sizeof( opus_int32 ) ); - - /* Bandwidth expansion for analysis filter shaping */ - silk_assert( BWExp1_Q16 <= SILK_FIX_CONST( 1.0, 16 ) ); - silk_bwexpander_32( AR1_Q24, psEnc->sCmn.shapingLPCOrder, BWExp1_Q16 ); - - /* Ratio of prediction gains, in energy domain */ - pre_nrg_Q30 = silk_LPC_inverse_pred_gain_Q24( AR2_Q24, psEnc->sCmn.shapingLPCOrder ); - nrg = silk_LPC_inverse_pred_gain_Q24( AR1_Q24, psEnc->sCmn.shapingLPCOrder ); - - /*psEncCtrl->GainsPre[ k ] = 1.0f - 0.7f * ( 1.0f - pre_nrg / nrg ) = 0.3f + 0.7f * pre_nrg / nrg;*/ - pre_nrg_Q30 = silk_LSHIFT32( silk_SMULWB( pre_nrg_Q30, SILK_FIX_CONST( 0.7, 15 ) ), 1 ); - psEncCtrl->GainsPre_Q14[ k ] = ( opus_int ) SILK_FIX_CONST( 0.3, 14 ) + silk_DIV32_varQ( pre_nrg_Q30, nrg, 14 ); + /* Bandwidth expansion */ + silk_bwexpander_32( AR_Q24, psEnc->sCmn.shapingLPCOrder, BWExp_Q16 ); - /* Convert to monic warped prediction coefficients and limit absolute values */ - limit_warped_coefs( AR2_Q24, AR1_Q24, warping_Q16, SILK_FIX_CONST( 3.999, 24 ), psEnc->sCmn.shapingLPCOrder ); + if( psEnc->sCmn.warping_Q16 > 0 ) { + /* Convert to monic warped prediction coefficients and limit absolute values */ + limit_warped_coefs( AR_Q24, warping_Q16, SILK_FIX_CONST( 3.999, 24 ), psEnc->sCmn.shapingLPCOrder ); - /* Convert from Q24 to Q13 and store in int16 */ - for( i = 0; i < psEnc->sCmn.shapingLPCOrder; i++ ) { - psEncCtrl->AR1_Q13[ k * MAX_SHAPE_LPC_ORDER + i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( AR1_Q24[ i ], 11 ) ); - psEncCtrl->AR2_Q13[ k * MAX_SHAPE_LPC_ORDER + i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( AR2_Q24[ i ], 11 ) ); + /* Convert from Q24 to Q13 and store in int16 */ + for( i = 0; i < psEnc->sCmn.shapingLPCOrder; i++ ) { + psEncCtrl->AR_Q13[ k * MAX_SHAPE_LPC_ORDER + i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( AR_Q24[ i ], 11 ) ); + } + } else { + silk_LPC_fit( &psEncCtrl->AR_Q13[ k * MAX_SHAPE_LPC_ORDER ], AR_Q24, 13, 24, psEnc->sCmn.shapingLPCOrder ); } } @@ -368,11 +340,6 @@ void silk_noise_shape_analysis_FIX( psEncCtrl->Gains_Q16[ k ] = silk_ADD_POS_SAT32( psEncCtrl->Gains_Q16[ k ], gain_add_Q16 ); } - gain_mult_Q16 = SILK_FIX_CONST( 1.0, 16 ) + silk_RSHIFT_ROUND( silk_MLA( SILK_FIX_CONST( INPUT_TILT, 26 ), - psEncCtrl->coding_quality_Q14, SILK_FIX_CONST( HIGH_RATE_INPUT_TILT, 12 ) ), 10 ); - for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - psEncCtrl->GainsPre_Q14[ k ] = silk_SMULWB( gain_mult_Q16, psEncCtrl->GainsPre_Q14[ k ] ); - } /************************************************/ /* Control low-frequency shaping and noise tilt */ @@ -410,14 +377,6 @@ void silk_noise_shape_analysis_FIX( /****************************/ /* HARMONIC SHAPING CONTROL */ /****************************/ - /* Control boosting of harmonic frequencies */ - HarmBoost_Q16 = silk_SMULWB( silk_SMULWB( SILK_FIX_CONST( 1.0, 17 ) - silk_LSHIFT( psEncCtrl->coding_quality_Q14, 3 ), - psEnc->LTPCorr_Q15 ), SILK_FIX_CONST( LOW_RATE_HARMONIC_BOOST, 16 ) ); - - /* More harmonic boost for noisy input signals */ - HarmBoost_Q16 = silk_SMLAWB( HarmBoost_Q16, - SILK_FIX_CONST( 1.0, 16 ) - silk_LSHIFT( psEncCtrl->input_quality_Q14, 2 ), SILK_FIX_CONST( LOW_INPUT_QUALITY_HARMONIC_BOOST, 16 ) ); - if( USE_HARM_SHAPING && psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /* More harmonic noise shaping for high bitrates or noisy input */ HarmShapeGain_Q16 = silk_SMLAWB( SILK_FIX_CONST( HARMONIC_SHAPING, 16 ), @@ -435,14 +394,11 @@ void silk_noise_shape_analysis_FIX( /* Smooth over subframes */ /*************************/ for( k = 0; k < MAX_NB_SUBFR; k++ ) { - psShapeSt->HarmBoost_smth_Q16 = - silk_SMLAWB( psShapeSt->HarmBoost_smth_Q16, HarmBoost_Q16 - psShapeSt->HarmBoost_smth_Q16, SILK_FIX_CONST( SUBFR_SMTH_COEF, 16 ) ); psShapeSt->HarmShapeGain_smth_Q16 = silk_SMLAWB( psShapeSt->HarmShapeGain_smth_Q16, HarmShapeGain_Q16 - psShapeSt->HarmShapeGain_smth_Q16, SILK_FIX_CONST( SUBFR_SMTH_COEF, 16 ) ); psShapeSt->Tilt_smth_Q16 = silk_SMLAWB( psShapeSt->Tilt_smth_Q16, Tilt_Q16 - psShapeSt->Tilt_smth_Q16, SILK_FIX_CONST( SUBFR_SMTH_COEF, 16 ) ); - psEncCtrl->HarmBoost_Q14[ k ] = ( opus_int )silk_RSHIFT_ROUND( psShapeSt->HarmBoost_smth_Q16, 2 ); psEncCtrl->HarmShapeGain_Q14[ k ] = ( opus_int )silk_RSHIFT_ROUND( psShapeSt->HarmShapeGain_smth_Q16, 2 ); psEncCtrl->Tilt_Q14[ k ] = ( opus_int )silk_RSHIFT_ROUND( psShapeSt->Tilt_smth_Q16, 2 ); } diff --git a/silk/fixed/pitch_analysis_core_FIX.c b/silk/fixed/pitch_analysis_core_FIX.c index 01bb9fc0..8df109e6 100644 --- a/silk/fixed/pitch_analysis_core_FIX.c +++ b/silk/fixed/pitch_analysis_core_FIX.c @@ -80,7 +80,7 @@ static void silk_P_Ana_calc_energy_st3( /* FIXED POINT CORE PITCH ANALYSIS FUNCTION */ /*************************************************************/ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 voiced, 1 unvoiced */ - const opus_int16 *frame, /* I Signal of length PE_FRAME_LENGTH_MS*Fs_kHz */ + const opus_int16 *frame_unscaled, /* I Signal of length PE_FRAME_LENGTH_MS*Fs_kHz */ opus_int *pitch_out, /* O 4 pitch lag values */ opus_int16 *lagIndex, /* O Lag Index */ opus_int8 *contourIndex, /* O Pitch contour Index */ @@ -94,16 +94,17 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 int arch /* I Run-time architecture */ ) { - VARDECL( opus_int16, frame_8kHz ); + VARDECL( opus_int16, frame_8kHz_buf ); VARDECL( opus_int16, frame_4kHz ); + VARDECL( opus_int16, frame_scaled ); opus_int32 filt_state[ 6 ]; - const opus_int16 *input_frame_ptr; + const opus_int16 *frame, *frame_8kHz; opus_int i, k, d, j; VARDECL( opus_int16, C ); VARDECL( opus_int32, xcorr32 ); const opus_int16 *target_ptr, *basis_ptr; - opus_int32 cross_corr, normalizer, energy, shift, energy_basis, energy_target; - opus_int d_srch[ PE_D_SRCH_LENGTH ], Cmax, length_d_srch, length_d_comp; + opus_int32 cross_corr, normalizer, energy, energy_basis, energy_target; + opus_int d_srch[ PE_D_SRCH_LENGTH ], Cmax, length_d_srch, length_d_comp, shift; VARDECL( opus_int16, d_comp ); opus_int32 sum, threshold, lag_counter; opus_int CBimax, CBimax_new, CBimax_old, lag, start_lag, end_lag, lag_new; @@ -119,6 +120,7 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 opus_int32 delta_lag_log2_sqr_Q7, lag_log2_Q7, prevLag_log2_Q7, prev_lag_bias_Q13; const opus_int8 *Lag_CB_ptr; SAVE_STACK; + /* Check for valid sampling frequency */ silk_assert( Fs_kHz == 8 || Fs_kHz == 12 || Fs_kHz == 16 ); @@ -137,17 +139,33 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 min_lag = PE_MIN_LAG_MS * Fs_kHz; max_lag = PE_MAX_LAG_MS * Fs_kHz - 1; + /* Downscale input if necessary */ + silk_sum_sqr_shift( &energy, &shift, frame_unscaled, frame_length ); + shift += 3 - silk_CLZ32( energy ); /* at least two bits headroom */ + ALLOC( frame_scaled, frame_length, opus_int16 ); + if( shift > 0 ) { + shift = silk_RSHIFT( shift + 1, 1 ); + for( i = 0; i < frame_length; i++ ) { + frame_scaled[ i ] = silk_RSHIFT( frame_unscaled[ i ], shift ); + } + frame = frame_scaled; + } else { + frame = frame_unscaled; + } + + ALLOC( frame_8kHz_buf, ( Fs_kHz == 8 ) ? 1 : frame_length_8kHz, opus_int16 ); /* Resample from input sampled at Fs_kHz to 8 kHz */ - ALLOC( frame_8kHz, frame_length_8kHz, opus_int16 ); if( Fs_kHz == 16 ) { silk_memset( filt_state, 0, 2 * sizeof( opus_int32 ) ); - silk_resampler_down2( filt_state, frame_8kHz, frame, frame_length ); + silk_resampler_down2( filt_state, frame_8kHz_buf, frame, frame_length ); + frame_8kHz = frame_8kHz_buf; } else if( Fs_kHz == 12 ) { silk_memset( filt_state, 0, 6 * sizeof( opus_int32 ) ); - silk_resampler_down2_3( filt_state, frame_8kHz, frame, frame_length ); + silk_resampler_down2_3( filt_state, frame_8kHz_buf, frame, frame_length ); + frame_8kHz = frame_8kHz_buf; } else { silk_assert( Fs_kHz == 8 ); - silk_memcpy( frame_8kHz, frame, frame_length_8kHz * sizeof(opus_int16) ); + frame_8kHz = frame; } /* Decimate again to 4 kHz */ @@ -160,19 +178,6 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 frame_4kHz[ i ] = silk_ADD_SAT16( frame_4kHz[ i ], frame_4kHz[ i - 1 ] ); } - /******************************************************************************* - ** Scale 4 kHz signal down to prevent correlations measures from overflowing - ** find scaling as max scaling for each 8kHz(?) subframe - *******************************************************************************/ - - /* Inner product is calculated with different lengths, so scale for the worst case */ - silk_sum_sqr_shift( &energy, &shift, frame_4kHz, frame_length_4kHz ); - if( shift > 0 ) { - shift = silk_RSHIFT( shift, 1 ); - for( i = 0; i < frame_length_4kHz; i++ ) { - frame_4kHz[ i ] = silk_RSHIFT( frame_4kHz[ i ], shift ); - } - } /****************************************************************************** * FIRST STAGE, operating in 4 khz @@ -311,18 +316,6 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 ** SECOND STAGE, operating at 8 kHz, on lag sections with high correlation *************************************************************************************/ - /****************************************************************************** - ** Scale signal down to avoid correlations measures from overflowing - *******************************************************************************/ - /* find scaling as max scaling for each subframe */ - silk_sum_sqr_shift( &energy, &shift, frame_8kHz, frame_length_8kHz ); - if( shift > 0 ) { - shift = silk_RSHIFT( shift, 1 ); - for( i = 0; i < frame_length_8kHz; i++ ) { - frame_8kHz[ i ] = silk_RSHIFT( frame_8kHz[ i ], shift ); - } - } - /********************************************************************************* * Find energy of each subframe projected onto its history, for a range of delays *********************************************************************************/ @@ -462,24 +455,6 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 silk_assert( *LTPCorr_Q15 >= 0 ); if( Fs_kHz > 8 ) { - VARDECL( opus_int16, scratch_mem ); - /***************************************************************************/ - /* Scale input signal down to avoid correlations measures from overflowing */ - /***************************************************************************/ - /* find scaling as max scaling for each subframe */ - silk_sum_sqr_shift( &energy, &shift, frame, frame_length ); - ALLOC( scratch_mem, shift > 0 ? frame_length : ALLOC_NONE, opus_int16 ); - if( shift > 0 ) { - /* Move signal to scratch mem because the input signal should be unchanged */ - shift = silk_RSHIFT( shift, 1 ); - for( i = 0; i < frame_length; i++ ) { - scratch_mem[ i ] = silk_RSHIFT( frame[ i ], shift ); - } - input_frame_ptr = scratch_mem; - } else { - input_frame_ptr = frame; - } - /* Search in original signal */ CBimax_old = CBimax; @@ -519,14 +494,14 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 /* Calculate the correlations and energies needed in stage 3 */ ALLOC( energies_st3, nb_subfr * nb_cbk_search, silk_pe_stage3_vals ); ALLOC( cross_corr_st3, nb_subfr * nb_cbk_search, silk_pe_stage3_vals ); - silk_P_Ana_calc_corr_st3( cross_corr_st3, input_frame_ptr, start_lag, sf_length, nb_subfr, complexity, arch ); - silk_P_Ana_calc_energy_st3( energies_st3, input_frame_ptr, start_lag, sf_length, nb_subfr, complexity, arch ); + silk_P_Ana_calc_corr_st3( cross_corr_st3, frame, start_lag, sf_length, nb_subfr, complexity, arch ); + silk_P_Ana_calc_energy_st3( energies_st3, frame, start_lag, sf_length, nb_subfr, complexity, arch ); lag_counter = 0; silk_assert( lag == silk_SAT16( lag ) ); contour_bias_Q15 = silk_DIV32_16( SILK_FIX_CONST( PE_FLATCONTOUR_BIAS, 15 ), lag ); - target_ptr = &input_frame_ptr[ PE_LTP_MEM_LENGTH_MS * Fs_kHz ]; + target_ptr = &frame[ PE_LTP_MEM_LENGTH_MS * Fs_kHz ]; energy_target = silk_ADD32( silk_inner_prod_aligned( target_ptr, target_ptr, nb_subfr * sf_length, arch ), 1 ); for( d = start_lag; d <= end_lag; d++ ) { for( j = 0; j < nb_cbk_search; j++ ) { diff --git a/silk/fixed/prefilter_FIX.c b/silk/fixed/prefilter_FIX.c deleted file mode 100644 index 6a8e3515..00000000 --- a/silk/fixed/prefilter_FIX.c +++ /dev/null @@ -1,221 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "main_FIX.h" -#include "stack_alloc.h" -#include "tuning_parameters.h" - -#if defined(MIPSr1_ASM) -#include "mips/prefilter_FIX_mipsr1.h" -#endif - - -#if !defined(OVERRIDE_silk_warped_LPC_analysis_filter_FIX) -#define silk_warped_LPC_analysis_filter_FIX(state, res_Q2, coef_Q13, input, lambda_Q16, length, order, arch) \ - ((void)(arch),silk_warped_LPC_analysis_filter_FIX_c(state, res_Q2, coef_Q13, input, lambda_Q16, length, order)) -#endif - -/* Prefilter for finding Quantizer input signal */ -static OPUS_INLINE void silk_prefilt_FIX( - silk_prefilter_state_FIX *P, /* I/O state */ - opus_int32 st_res_Q12[], /* I short term residual signal */ - opus_int32 xw_Q3[], /* O prefiltered signal */ - opus_int32 HarmShapeFIRPacked_Q12, /* I Harmonic shaping coeficients */ - opus_int Tilt_Q14, /* I Tilt shaping coeficient */ - opus_int32 LF_shp_Q14, /* I Low-frequancy shaping coeficients */ - opus_int lag, /* I Lag for harmonic shaping */ - opus_int length /* I Length of signals */ -); - -void silk_warped_LPC_analysis_filter_FIX_c( - opus_int32 state[], /* I/O State [order + 1] */ - opus_int32 res_Q2[], /* O Residual signal [length] */ - const opus_int16 coef_Q13[], /* I Coefficients [order] */ - const opus_int16 input[], /* I Input signal [length] */ - const opus_int16 lambda_Q16, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -) -{ - opus_int n, i; - opus_int32 acc_Q11, tmp1, tmp2; - - /* Order must be even */ - silk_assert( ( order & 1 ) == 0 ); - - for( n = 0; n < length; n++ ) { - /* Output of lowpass section */ - tmp2 = silk_SMLAWB( state[ 0 ], state[ 1 ], lambda_Q16 ); - state[ 0 ] = silk_LSHIFT( input[ n ], 14 ); - /* Output of allpass section */ - tmp1 = silk_SMLAWB( state[ 1 ], state[ 2 ] - tmp2, lambda_Q16 ); - state[ 1 ] = tmp2; - acc_Q11 = silk_RSHIFT( order, 1 ); - acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ 0 ] ); - /* Loop over allpass sections */ - for( i = 2; i < order; i += 2 ) { - /* Output of allpass section */ - tmp2 = silk_SMLAWB( state[ i ], state[ i + 1 ] - tmp1, lambda_Q16 ); - state[ i ] = tmp1; - acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ i - 1 ] ); - /* Output of allpass section */ - tmp1 = silk_SMLAWB( state[ i + 1 ], state[ i + 2 ] - tmp2, lambda_Q16 ); - state[ i + 1 ] = tmp2; - acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ i ] ); - } - state[ order ] = tmp1; - acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ order - 1 ] ); - res_Q2[ n ] = silk_LSHIFT( (opus_int32)input[ n ], 2 ) - silk_RSHIFT_ROUND( acc_Q11, 9 ); - } -} - -void silk_prefilter_FIX( - silk_encoder_state_FIX *psEnc, /* I/O Encoder state */ - const silk_encoder_control_FIX *psEncCtrl, /* I Encoder control */ - opus_int32 xw_Q3[], /* O Weighted signal */ - const opus_int16 x[] /* I Speech signal */ -) -{ - silk_prefilter_state_FIX *P = &psEnc->sPrefilt; - opus_int j, k, lag; - opus_int32 tmp_32; - const opus_int16 *AR1_shp_Q13; - const opus_int16 *px; - opus_int32 *pxw_Q3; - opus_int HarmShapeGain_Q12, Tilt_Q14; - opus_int32 HarmShapeFIRPacked_Q12, LF_shp_Q14; - VARDECL( opus_int32, x_filt_Q12 ); - VARDECL( opus_int32, st_res_Q2 ); - opus_int16 B_Q10[ 2 ]; - SAVE_STACK; - - /* Set up pointers */ - px = x; - pxw_Q3 = xw_Q3; - lag = P->lagPrev; - ALLOC( x_filt_Q12, psEnc->sCmn.subfr_length, opus_int32 ); - ALLOC( st_res_Q2, psEnc->sCmn.subfr_length, opus_int32 ); - for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - /* Update Variables that change per sub frame */ - if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { - lag = psEncCtrl->pitchL[ k ]; - } - - /* Noise shape parameters */ - HarmShapeGain_Q12 = silk_SMULWB( (opus_int32)psEncCtrl->HarmShapeGain_Q14[ k ], 16384 - psEncCtrl->HarmBoost_Q14[ k ] ); - silk_assert( HarmShapeGain_Q12 >= 0 ); - HarmShapeFIRPacked_Q12 = silk_RSHIFT( HarmShapeGain_Q12, 2 ); - HarmShapeFIRPacked_Q12 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q12, 1 ), 16 ); - Tilt_Q14 = psEncCtrl->Tilt_Q14[ k ]; - LF_shp_Q14 = psEncCtrl->LF_shp_Q14[ k ]; - AR1_shp_Q13 = &psEncCtrl->AR1_Q13[ k * MAX_SHAPE_LPC_ORDER ]; - - /* Short term FIR filtering*/ - silk_warped_LPC_analysis_filter_FIX( P->sAR_shp, st_res_Q2, AR1_shp_Q13, px, - psEnc->sCmn.warping_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.shapingLPCOrder, psEnc->sCmn.arch ); - - /* Reduce (mainly) low frequencies during harmonic emphasis */ - B_Q10[ 0 ] = silk_RSHIFT_ROUND( psEncCtrl->GainsPre_Q14[ k ], 4 ); - tmp_32 = silk_SMLABB( SILK_FIX_CONST( INPUT_TILT, 26 ), psEncCtrl->HarmBoost_Q14[ k ], HarmShapeGain_Q12 ); /* Q26 */ - tmp_32 = silk_SMLABB( tmp_32, psEncCtrl->coding_quality_Q14, SILK_FIX_CONST( HIGH_RATE_INPUT_TILT, 12 ) ); /* Q26 */ - tmp_32 = silk_SMULWB( tmp_32, -psEncCtrl->GainsPre_Q14[ k ] ); /* Q24 */ - tmp_32 = silk_RSHIFT_ROUND( tmp_32, 14 ); /* Q10 */ - B_Q10[ 1 ]= silk_SAT16( tmp_32 ); - x_filt_Q12[ 0 ] = silk_MLA( silk_MUL( st_res_Q2[ 0 ], B_Q10[ 0 ] ), P->sHarmHP_Q2, B_Q10[ 1 ] ); - for( j = 1; j < psEnc->sCmn.subfr_length; j++ ) { - x_filt_Q12[ j ] = silk_MLA( silk_MUL( st_res_Q2[ j ], B_Q10[ 0 ] ), st_res_Q2[ j - 1 ], B_Q10[ 1 ] ); - } - P->sHarmHP_Q2 = st_res_Q2[ psEnc->sCmn.subfr_length - 1 ]; - - silk_prefilt_FIX( P, x_filt_Q12, pxw_Q3, HarmShapeFIRPacked_Q12, Tilt_Q14, LF_shp_Q14, lag, psEnc->sCmn.subfr_length ); - - px += psEnc->sCmn.subfr_length; - pxw_Q3 += psEnc->sCmn.subfr_length; - } - - P->lagPrev = psEncCtrl->pitchL[ psEnc->sCmn.nb_subfr - 1 ]; - RESTORE_STACK; -} - -#ifndef OVERRIDE_silk_prefilt_FIX -/* Prefilter for finding Quantizer input signal */ -static OPUS_INLINE void silk_prefilt_FIX( - silk_prefilter_state_FIX *P, /* I/O state */ - opus_int32 st_res_Q12[], /* I short term residual signal */ - opus_int32 xw_Q3[], /* O prefiltered signal */ - opus_int32 HarmShapeFIRPacked_Q12, /* I Harmonic shaping coeficients */ - opus_int Tilt_Q14, /* I Tilt shaping coeficient */ - opus_int32 LF_shp_Q14, /* I Low-frequancy shaping coeficients */ - opus_int lag, /* I Lag for harmonic shaping */ - opus_int length /* I Length of signals */ -) -{ - opus_int i, idx, LTP_shp_buf_idx; - opus_int32 n_LTP_Q12, n_Tilt_Q10, n_LF_Q10; - opus_int32 sLF_MA_shp_Q12, sLF_AR_shp_Q12; - opus_int16 *LTP_shp_buf; - - /* To speed up use temp variables instead of using the struct */ - LTP_shp_buf = P->sLTP_shp; - LTP_shp_buf_idx = P->sLTP_shp_buf_idx; - sLF_AR_shp_Q12 = P->sLF_AR_shp_Q12; - sLF_MA_shp_Q12 = P->sLF_MA_shp_Q12; - - for( i = 0; i < length; i++ ) { - if( lag > 0 ) { - /* unrolled loop */ - silk_assert( HARM_SHAPE_FIR_TAPS == 3 ); - idx = lag + LTP_shp_buf_idx; - n_LTP_Q12 = silk_SMULBB( LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 - 1) & LTP_MASK ], HarmShapeFIRPacked_Q12 ); - n_LTP_Q12 = silk_SMLABT( n_LTP_Q12, LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 ) & LTP_MASK ], HarmShapeFIRPacked_Q12 ); - n_LTP_Q12 = silk_SMLABB( n_LTP_Q12, LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 + 1) & LTP_MASK ], HarmShapeFIRPacked_Q12 ); - } else { - n_LTP_Q12 = 0; - } - - n_Tilt_Q10 = silk_SMULWB( sLF_AR_shp_Q12, Tilt_Q14 ); - n_LF_Q10 = silk_SMLAWB( silk_SMULWT( sLF_AR_shp_Q12, LF_shp_Q14 ), sLF_MA_shp_Q12, LF_shp_Q14 ); - - sLF_AR_shp_Q12 = silk_SUB32( st_res_Q12[ i ], silk_LSHIFT( n_Tilt_Q10, 2 ) ); - sLF_MA_shp_Q12 = silk_SUB32( sLF_AR_shp_Q12, silk_LSHIFT( n_LF_Q10, 2 ) ); - - LTP_shp_buf_idx = ( LTP_shp_buf_idx - 1 ) & LTP_MASK; - LTP_shp_buf[ LTP_shp_buf_idx ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( sLF_MA_shp_Q12, 12 ) ); - - xw_Q3[i] = silk_RSHIFT_ROUND( silk_SUB32( sLF_MA_shp_Q12, n_LTP_Q12 ), 9 ); - } - - /* Copy temp variable back to state */ - P->sLF_AR_shp_Q12 = sLF_AR_shp_Q12; - P->sLF_MA_shp_Q12 = sLF_MA_shp_Q12; - P->sLTP_shp_buf_idx = LTP_shp_buf_idx; -} -#endif /* OVERRIDE_silk_prefilt_FIX */ diff --git a/silk/fixed/schur64_FIX.c b/silk/fixed/schur64_FIX.c index 764a10ef..b2cb12d9 100644 --- a/silk/fixed/schur64_FIX.c +++ b/silk/fixed/schur64_FIX.c @@ -43,7 +43,7 @@ opus_int32 silk_schur64( /* O returns residual ene opus_int32 C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; opus_int32 Ctmp1_Q30, Ctmp2_Q30, rc_tmp_Q31; - silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 ); + silk_assert( order >= 0 && order <= SILK_MAX_ORDER_LPC ); /* Check for invalid input */ if( c[ 0 ] <= 0 ) { @@ -51,9 +51,10 @@ opus_int32 silk_schur64( /* O returns residual ene return 0; } - for( k = 0; k < order + 1; k++ ) { + k = 0; + do { C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ]; - } + } while( ++k <= order ); for( k = 0; k < order; k++ ) { /* Check that we won't be getting an unstable rc, otherwise stop here. */ diff --git a/silk/fixed/schur_FIX.c b/silk/fixed/schur_FIX.c index c4c0ef23..59d44a6f 100644 --- a/silk/fixed/schur_FIX.c +++ b/silk/fixed/schur_FIX.c @@ -43,28 +43,29 @@ opus_int32 silk_schur( /* O Returns residual ene opus_int32 C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; opus_int32 Ctmp1, Ctmp2, rc_tmp_Q15; - silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 ); + silk_assert( order >= 0 && order <= SILK_MAX_ORDER_LPC ); /* Get number of leading zeros */ lz = silk_CLZ32( c[ 0 ] ); /* Copy correlations and adjust level to Q30 */ + k = 0; if( lz < 2 ) { /* lz must be 1, so shift one to the right */ - for( k = 0; k < order + 1; k++ ) { + do { C[ k ][ 0 ] = C[ k ][ 1 ] = silk_RSHIFT( c[ k ], 1 ); - } + } while( ++k <= order ); } else if( lz > 2 ) { /* Shift to the left */ lz -= 2; - for( k = 0; k < order + 1; k++ ) { + do { C[ k ][ 0 ] = C[ k ][ 1 ] = silk_LSHIFT( c[ k ], lz ); - } + } while( ++k <= order ); } else { /* No need to shift */ - for( k = 0; k < order + 1; k++ ) { + do { C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ]; - } + } while( ++k <= order ); } for( k = 0; k < order; k++ ) { diff --git a/silk/fixed/solve_LS_FIX.c b/silk/fixed/solve_LS_FIX.c deleted file mode 100644 index 51d7d49d..00000000 --- a/silk/fixed/solve_LS_FIX.c +++ /dev/null @@ -1,249 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "main_FIX.h" -#include "stack_alloc.h" -#include "tuning_parameters.h" - -/*****************************/ -/* Internal function headers */ -/*****************************/ - -typedef struct { - opus_int32 Q36_part; - opus_int32 Q48_part; -} inv_D_t; - -/* Factorize square matrix A into LDL form */ -static OPUS_INLINE void silk_LDL_factorize_FIX( - opus_int32 *A, /* I/O Pointer to Symetric Square Matrix */ - opus_int M, /* I Size of Matrix */ - opus_int32 *L_Q16, /* I/O Pointer to Square Upper triangular Matrix */ - inv_D_t *inv_D /* I/O Pointer to vector holding inverted diagonal elements of D */ -); - -/* Solve Lx = b, when L is lower triangular and has ones on the diagonal */ -static OPUS_INLINE void silk_LS_SolveFirst_FIX( - const opus_int32 *L_Q16, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const opus_int32 *b, /* I b Vector */ - opus_int32 *x_Q16 /* O x Vector */ -); - -/* Solve L^t*x = b, where L is lower triangular with ones on the diagonal */ -static OPUS_INLINE void silk_LS_SolveLast_FIX( - const opus_int32 *L_Q16, /* I Pointer to Lower Triangular Matrix */ - const opus_int M, /* I Dim of Matrix equation */ - const opus_int32 *b, /* I b Vector */ - opus_int32 *x_Q16 /* O x Vector */ -); - -static OPUS_INLINE void silk_LS_divide_Q16_FIX( - opus_int32 T[], /* I/O Numenator vector */ - inv_D_t *inv_D, /* I 1 / D vector */ - opus_int M /* I dimension */ -); - -/* Solves Ax = b, assuming A is symmetric */ -void silk_solve_LDL_FIX( - opus_int32 *A, /* I Pointer to symetric square matrix A */ - opus_int M, /* I Size of matrix */ - const opus_int32 *b, /* I Pointer to b vector */ - opus_int32 *x_Q16 /* O Pointer to x solution vector */ -) -{ - VARDECL( opus_int32, L_Q16 ); - opus_int32 Y[ MAX_MATRIX_SIZE ]; - inv_D_t inv_D[ MAX_MATRIX_SIZE ]; - SAVE_STACK; - - silk_assert( M <= MAX_MATRIX_SIZE ); - ALLOC( L_Q16, M * M, opus_int32 ); - - /*************************************************** - Factorize A by LDL such that A = L*D*L', - where L is lower triangular with ones on diagonal - ****************************************************/ - silk_LDL_factorize_FIX( A, M, L_Q16, inv_D ); - - /**************************************************** - * substitute D*L'*x = Y. ie: - L*D*L'*x = b => L*Y = b <=> Y = inv(L)*b - ******************************************************/ - silk_LS_SolveFirst_FIX( L_Q16, M, b, Y ); - - /**************************************************** - D*L'*x = Y <=> L'*x = inv(D)*Y, because D is - diagonal just multiply with 1/d_i - ****************************************************/ - silk_LS_divide_Q16_FIX( Y, inv_D, M ); - - /**************************************************** - x = inv(L') * inv(D) * Y - *****************************************************/ - silk_LS_SolveLast_FIX( L_Q16, M, Y, x_Q16 ); - RESTORE_STACK; -} - -static OPUS_INLINE void silk_LDL_factorize_FIX( - opus_int32 *A, /* I/O Pointer to Symetric Square Matrix */ - opus_int M, /* I Size of Matrix */ - opus_int32 *L_Q16, /* I/O Pointer to Square Upper triangular Matrix */ - inv_D_t *inv_D /* I/O Pointer to vector holding inverted diagonal elements of D */ -) -{ - opus_int i, j, k, status, loop_count; - const opus_int32 *ptr1, *ptr2; - opus_int32 diag_min_value, tmp_32, err; - opus_int32 v_Q0[ MAX_MATRIX_SIZE ], D_Q0[ MAX_MATRIX_SIZE ]; - opus_int32 one_div_diag_Q36, one_div_diag_Q40, one_div_diag_Q48; - - silk_assert( M <= MAX_MATRIX_SIZE ); - - status = 1; - diag_min_value = silk_max_32( silk_SMMUL( silk_ADD_SAT32( A[ 0 ], A[ silk_SMULBB( M, M ) - 1 ] ), SILK_FIX_CONST( FIND_LTP_COND_FAC, 31 ) ), 1 << 9 ); - for( loop_count = 0; loop_count < M && status == 1; loop_count++ ) { - status = 0; - for( j = 0; j < M; j++ ) { - ptr1 = matrix_adr( L_Q16, j, 0, M ); - tmp_32 = 0; - for( i = 0; i < j; i++ ) { - v_Q0[ i ] = silk_SMULWW( D_Q0[ i ], ptr1[ i ] ); /* Q0 */ - tmp_32 = silk_SMLAWW( tmp_32, v_Q0[ i ], ptr1[ i ] ); /* Q0 */ - } - tmp_32 = silk_SUB32( matrix_ptr( A, j, j, M ), tmp_32 ); - - if( tmp_32 < diag_min_value ) { - tmp_32 = silk_SUB32( silk_SMULBB( loop_count + 1, diag_min_value ), tmp_32 ); - /* Matrix not positive semi-definite, or ill conditioned */ - for( i = 0; i < M; i++ ) { - matrix_ptr( A, i, i, M ) = silk_ADD32( matrix_ptr( A, i, i, M ), tmp_32 ); - } - status = 1; - break; - } - D_Q0[ j ] = tmp_32; /* always < max(Correlation) */ - - /* two-step division */ - one_div_diag_Q36 = silk_INVERSE32_varQ( tmp_32, 36 ); /* Q36 */ - one_div_diag_Q40 = silk_LSHIFT( one_div_diag_Q36, 4 ); /* Q40 */ - err = silk_SUB32( (opus_int32)1 << 24, silk_SMULWW( tmp_32, one_div_diag_Q40 ) ); /* Q24 */ - one_div_diag_Q48 = silk_SMULWW( err, one_div_diag_Q40 ); /* Q48 */ - - /* Save 1/Ds */ - inv_D[ j ].Q36_part = one_div_diag_Q36; - inv_D[ j ].Q48_part = one_div_diag_Q48; - - matrix_ptr( L_Q16, j, j, M ) = 65536; /* 1.0 in Q16 */ - ptr1 = matrix_adr( A, j, 0, M ); - ptr2 = matrix_adr( L_Q16, j + 1, 0, M ); - for( i = j + 1; i < M; i++ ) { - tmp_32 = 0; - for( k = 0; k < j; k++ ) { - tmp_32 = silk_SMLAWW( tmp_32, v_Q0[ k ], ptr2[ k ] ); /* Q0 */ - } - tmp_32 = silk_SUB32( ptr1[ i ], tmp_32 ); /* always < max(Correlation) */ - - /* tmp_32 / D_Q0[j] : Divide to Q16 */ - matrix_ptr( L_Q16, i, j, M ) = silk_ADD32( silk_SMMUL( tmp_32, one_div_diag_Q48 ), - silk_RSHIFT( silk_SMULWW( tmp_32, one_div_diag_Q36 ), 4 ) ); - - /* go to next column */ - ptr2 += M; - } - } - } - - silk_assert( status == 0 ); -} - -static OPUS_INLINE void silk_LS_divide_Q16_FIX( - opus_int32 T[], /* I/O Numenator vector */ - inv_D_t *inv_D, /* I 1 / D vector */ - opus_int M /* I dimension */ -) -{ - opus_int i; - opus_int32 tmp_32; - opus_int32 one_div_diag_Q36, one_div_diag_Q48; - - for( i = 0; i < M; i++ ) { - one_div_diag_Q36 = inv_D[ i ].Q36_part; - one_div_diag_Q48 = inv_D[ i ].Q48_part; - - tmp_32 = T[ i ]; - T[ i ] = silk_ADD32( silk_SMMUL( tmp_32, one_div_diag_Q48 ), silk_RSHIFT( silk_SMULWW( tmp_32, one_div_diag_Q36 ), 4 ) ); - } -} - -/* Solve Lx = b, when L is lower triangular and has ones on the diagonal */ -static OPUS_INLINE void silk_LS_SolveFirst_FIX( - const opus_int32 *L_Q16, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const opus_int32 *b, /* I b Vector */ - opus_int32 *x_Q16 /* O x Vector */ -) -{ - opus_int i, j; - const opus_int32 *ptr32; - opus_int32 tmp_32; - - for( i = 0; i < M; i++ ) { - ptr32 = matrix_adr( L_Q16, i, 0, M ); - tmp_32 = 0; - for( j = 0; j < i; j++ ) { - tmp_32 = silk_SMLAWW( tmp_32, ptr32[ j ], x_Q16[ j ] ); - } - x_Q16[ i ] = silk_SUB32( b[ i ], tmp_32 ); - } -} - -/* Solve L^t*x = b, where L is lower triangular with ones on the diagonal */ -static OPUS_INLINE void silk_LS_SolveLast_FIX( - const opus_int32 *L_Q16, /* I Pointer to Lower Triangular Matrix */ - const opus_int M, /* I Dim of Matrix equation */ - const opus_int32 *b, /* I b Vector */ - opus_int32 *x_Q16 /* O x Vector */ -) -{ - opus_int i, j; - const opus_int32 *ptr32; - opus_int32 tmp_32; - - for( i = M - 1; i >= 0; i-- ) { - ptr32 = matrix_adr( L_Q16, 0, i, M ); - tmp_32 = 0; - for( j = M - 1; j > i; j-- ) { - tmp_32 = silk_SMLAWW( tmp_32, ptr32[ silk_SMULBB( j, M ) ], x_Q16[ j ] ); - } - x_Q16[ i ] = silk_SUB32( b[ i ], tmp_32 ); - } -} diff --git a/silk/fixed/structs_FIX.h b/silk/fixed/structs_FIX.h index 3294b251..2774a97b 100644 --- a/silk/fixed/structs_FIX.h +++ b/silk/fixed/structs_FIX.h @@ -48,30 +48,16 @@ typedef struct { } silk_shape_state_FIX; /********************************/ -/* Prefilter state */ -/********************************/ -typedef struct { - opus_int16 sLTP_shp[ LTP_BUF_LENGTH ]; - opus_int32 sAR_shp[ MAX_SHAPE_LPC_ORDER + 1 ]; - opus_int sLTP_shp_buf_idx; - opus_int32 sLF_AR_shp_Q12; - opus_int32 sLF_MA_shp_Q12; - opus_int32 sHarmHP_Q2; - opus_int32 rand_seed; - opus_int lagPrev; -} silk_prefilter_state_FIX; - -/********************************/ /* Encoder state FIX */ /********************************/ typedef struct { silk_encoder_state sCmn; /* Common struct, shared with floating-point code */ silk_shape_state_FIX sShape; /* Shape state */ - silk_prefilter_state_FIX sPrefilt; /* Prefilter State */ /* Buffer for find pitch and noise shape analysis */ silk_DWORD_ALIGN opus_int16 x_buf[ 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ];/* Buffer for find pitch and noise shape analysis */ opus_int LTPCorr_Q15; /* Normalized correlation from pitch lag estimator */ + opus_int32 resNrgSmth; } silk_encoder_state_FIX; /************************/ @@ -87,11 +73,8 @@ typedef struct { /* Noise shaping parameters */ /* Testing */ - silk_DWORD_ALIGN opus_int16 AR1_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; - silk_DWORD_ALIGN opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; + silk_DWORD_ALIGN opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ]; /* Packs two int16 coefficients per int32 value */ - opus_int GainsPre_Q14[ MAX_NB_SUBFR ]; - opus_int HarmBoost_Q14[ MAX_NB_SUBFR ]; opus_int Tilt_Q14[ MAX_NB_SUBFR ]; opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ]; opus_int Lambda_Q10; @@ -99,7 +82,6 @@ typedef struct { opus_int coding_quality_Q14; /* measures */ - opus_int sparseness_Q8; opus_int32 predGain_Q16; opus_int LTPredCodGain_Q7; opus_int32 ResNrg[ MAX_NB_SUBFR ]; /* Residual energy per subframe */ diff --git a/silk/fixed/warped_autocorrelation_FIX.c b/silk/fixed/warped_autocorrelation_FIX.c index 6ca6c118..994c299a 100644 --- a/silk/fixed/warped_autocorrelation_FIX.c +++ b/silk/fixed/warped_autocorrelation_FIX.c @@ -31,17 +31,13 @@ POSSIBILITY OF SUCH DAMAGE. #include "main_FIX.h" -#define QC 10 -#define QS 14 - #if defined(MIPSr1_ASM) #include "mips/warped_autocorrelation_FIX_mipsr1.h" #endif -#ifndef OVERRIDE_silk_warped_autocorrelation_FIX /* Autocorrelations for a warped frequency axis */ -void silk_warped_autocorrelation_FIX( +void silk_warped_autocorrelation_FIX_c( opus_int32 *corr, /* O Result [order + 1] */ opus_int *scale, /* O Scaling of the correlation vector */ const opus_int16 *input, /* I Input data to correlate */ @@ -92,4 +88,3 @@ void silk_warped_autocorrelation_FIX( } silk_assert( corr_QC[ 0 ] >= 0 ); /* If breaking, decrease QC*/ } -#endif /* OVERRIDE_silk_warped_autocorrelation_FIX */ diff --git a/silk/fixed/x86/prefilter_FIX_sse.c b/silk/fixed/x86/prefilter_FIX_sse.c index 488a603f..d8c9c2f5 100644 --- a/silk/fixed/x86/prefilter_FIX_sse.c +++ b/silk/fixed/x86/prefilter_FIX_sse.c @@ -107,8 +107,8 @@ void silk_warped_LPC_analysis_filter_FIX_sse4_1( xmm_tempb = _mm_add_epi32( xmm_tempb, xmm_product2 ); xmm_tempa = _mm_add_epi32( xmm_tempa, xmm_tempb ); - sum = (coef_Q13_8 * state_8) >> 16; - sum += (coef_Q13_9 * state_9) >> 16; + sum = (opus_int32)((coef_Q13_8 * state_8) >> 16); + sum += (opus_int32)((coef_Q13_9 * state_9) >> 16); xmm_tempa = _mm_add_epi32( xmm_tempa, _mm_shuffle_epi32( xmm_tempa, _MM_SHUFFLE( 0, 0, 0, 2 ) ) ); sum += _mm_cvtsi128_si32( xmm_tempa); diff --git a/silk/float/LPC_inv_pred_gain_FLP.c b/silk/float/LPC_inv_pred_gain_FLP.c index 25178bac..2be2122d 100644 --- a/silk/float/LPC_inv_pred_gain_FLP.c +++ b/silk/float/LPC_inv_pred_gain_FLP.c @@ -31,8 +31,7 @@ POSSIBILITY OF SUCH DAMAGE. #include "SigProc_FIX.h" #include "SigProc_FLP.h" - -#define RC_THRESHOLD 0.9999f +#include "define.h" /* compute inverse of LPC prediction gain, and */ /* test if LPC coefficients are stable (all poles within unit circle) */ @@ -43,34 +42,32 @@ silk_float silk_LPC_inverse_pred_gain_FLP( /* O return inverse prediction ga ) { opus_int k, n; - double invGain, rc, rc_mult1, rc_mult2; - silk_float Atmp[ 2 ][ SILK_MAX_ORDER_LPC ]; - silk_float *Aold, *Anew; + double invGain, rc, rc_mult1, rc_mult2, tmp1, tmp2; + silk_float Atmp[ SILK_MAX_ORDER_LPC ]; - Anew = Atmp[ order & 1 ]; - silk_memcpy( Anew, A, order * sizeof(silk_float) ); + silk_memcpy( Atmp, A, order * sizeof(silk_float) ); invGain = 1.0; for( k = order - 1; k > 0; k-- ) { - rc = -Anew[ k ]; - if( rc > RC_THRESHOLD || rc < -RC_THRESHOLD ) { + rc = -Atmp[ k ]; + rc_mult1 = 1.0f - rc * rc; + invGain *= rc_mult1; + if( invGain * MAX_PREDICTION_POWER_GAIN < 1.0f ) { return 0.0f; } - rc_mult1 = 1.0f - rc * rc; rc_mult2 = 1.0f / rc_mult1; - invGain *= rc_mult1; - /* swap pointers */ - Aold = Anew; - Anew = Atmp[ k & 1 ]; - for( n = 0; n < k; n++ ) { - Anew[ n ] = (silk_float)( ( Aold[ n ] - Aold[ k - n - 1 ] * rc ) * rc_mult2 ); + for( n = 0; n < (k + 1) >> 1; n++ ) { + tmp1 = Atmp[ n ]; + tmp2 = Atmp[ k - n - 1 ]; + Atmp[ n ] = (silk_float)( ( tmp1 - tmp2 * rc ) * rc_mult2 ); + Atmp[ k - n - 1 ] = (silk_float)( ( tmp2 - tmp1 * rc ) * rc_mult2 ); } } - rc = -Anew[ 0 ]; - if( rc > RC_THRESHOLD || rc < -RC_THRESHOLD ) { - return 0.0f; - } + rc = -Atmp[ 0 ]; rc_mult1 = 1.0f - rc * rc; invGain *= rc_mult1; + if( invGain * MAX_PREDICTION_POWER_GAIN < 1.0f ) { + return 0.0f; + } return (silk_float)invGain; } diff --git a/silk/float/SigProc_FLP.h b/silk/float/SigProc_FLP.h index f0cb3733..953de8b0 100644 --- a/silk/float/SigProc_FLP.h +++ b/silk/float/SigProc_FLP.h @@ -68,13 +68,6 @@ void silk_k2a_FLP( opus_int32 order /* I prediction order */ ); -/* Solve the normal equations using the Levinson-Durbin recursion */ -silk_float silk_levinsondurbin_FLP( /* O prediction error energy */ - silk_float A[], /* O prediction coefficients [order] */ - const silk_float corr[], /* I input auto-correlations [order + 1] */ - const opus_int order /* I prediction order */ -); - /* compute autocorrelation */ void silk_autocorrelation_FLP( silk_float *results, /* O result (length correlationCount) */ diff --git a/silk/float/encode_frame_FLP.c b/silk/float/encode_frame_FLP.c index 2092a4d9..c3ad50a9 100644 --- a/silk/float/encode_frame_FLP.c +++ b/silk/float/encode_frame_FLP.c @@ -29,6 +29,7 @@ POSSIBILITY OF SUCH DAMAGE. #include "config.h" #endif +#include <stdlib.h> #include "main_FLP.h" #include "tuning_parameters.h" @@ -85,7 +86,6 @@ opus_int silk_encode_frame_FLP( silk_encoder_control_FLP sEncCtrl; opus_int i, iter, maxIter, found_upper, found_lower, ret = 0; silk_float *x_frame, *res_pitch_frame; - silk_float xfw[ MAX_FRAME_LENGTH ]; silk_float res_pitch[ 2 * MAX_FRAME_LENGTH + LA_PITCH_MAX ]; ec_enc sRangeEnc_copy, sRangeEnc_copy2; silk_nsq_state sNSQ_copy, sNSQ_copy2; @@ -97,6 +97,9 @@ opus_int silk_encode_frame_FLP( opus_int8 LastGainIndex_copy2; opus_int32 pGains_Q16[ MAX_NB_SUBFR ]; opus_uint8 ec_buf_copy[ 1275 ]; + opus_int gain_lock[ MAX_NB_SUBFR ] = {0}; + opus_int16 best_gain_mult[ MAX_NB_SUBFR ]; + opus_int best_sum[ MAX_NB_SUBFR ]; /* This is totally unnecessary but many compilers (including gcc) are too dumb to realise it */ LastGainIndex_copy2 = nBits_lower = nBits_upper = gainMult_lower = gainMult_upper = 0; @@ -139,22 +142,17 @@ opus_int silk_encode_frame_FLP( /***************************************************/ /* Find linear prediction coefficients (LPC + LTP) */ /***************************************************/ - silk_find_pred_coefs_FLP( psEnc, &sEncCtrl, res_pitch, x_frame, condCoding ); + silk_find_pred_coefs_FLP( psEnc, &sEncCtrl, res_pitch_frame, x_frame, condCoding ); /****************************************/ /* Process gains */ /****************************************/ silk_process_gains_FLP( psEnc, &sEncCtrl, condCoding ); - /*****************************************/ - /* Prefiltering for noise shaper */ - /*****************************************/ - silk_prefilter_FLP( psEnc, &sEncCtrl, xfw, x_frame ); - /****************************************/ /* Low Bitrate Redundant Encoding */ /****************************************/ - silk_LBRR_encode_FLP( psEnc, &sEncCtrl, xfw, condCoding ); + silk_LBRR_encode_FLP( psEnc, &sEncCtrl, x_frame, condCoding ); /* Loop over quantizer and entroy coding to control bitrate */ maxIter = 6; @@ -188,7 +186,11 @@ opus_int silk_encode_frame_FLP( /*****************************************/ /* Noise shaping quantization */ /*****************************************/ - silk_NSQ_wrapper_FLP( psEnc, &sEncCtrl, &psEnc->sCmn.indices, &psEnc->sCmn.sNSQ, psEnc->sCmn.pulses, xfw ); + silk_NSQ_wrapper_FLP( psEnc, &sEncCtrl, &psEnc->sCmn.indices, &psEnc->sCmn.sNSQ, psEnc->sCmn.pulses, x_frame ); + + if ( iter == maxIter && !found_lower ) { + silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) ); + } /****************************************/ /* Encode Parameters */ @@ -203,6 +205,33 @@ opus_int silk_encode_frame_FLP( nBits = ec_tell( psRangeEnc ); + /* If we still bust after the last iteration, do some damage control. */ + if ( iter == maxIter && !found_lower && nBits > maxBits ) { + silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) ); + + /* Keep gains the same as the last frame. */ + psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev; + for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + psEnc->sCmn.indices.GainsIndices[ i ] = 4; + } + if (condCoding != CODE_CONDITIONALLY) { + psEnc->sCmn.indices.GainsIndices[ 0 ] = sEncCtrl.lastGainIndexPrev; + } + psEnc->sCmn.ec_prevLagIndex = ec_prevLagIndex_copy; + psEnc->sCmn.ec_prevSignalType = ec_prevSignalType_copy; + /* Clear all pulses. */ + for ( i = 0; i < psEnc->sCmn.frame_length; i++ ) { + psEnc->sCmn.pulses[ i ] = 0; + } + + silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding ); + + silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType, + psEnc->sCmn.pulses, psEnc->sCmn.frame_length ); + + nBits = ec_tell( psRangeEnc ); + } + if( useCBR == 0 && iter == 0 && nBits <= maxBits ) { break; } @@ -223,7 +252,9 @@ opus_int silk_encode_frame_FLP( if( nBits > maxBits ) { if( found_lower == 0 && iter >= 2 ) { /* Adjust the quantizer's rate/distortion tradeoff and discard previous "upper" results */ - sEncCtrl.Lambda *= 1.5f; + sEncCtrl.Lambda = silk_max_float(sEncCtrl.Lambda*1.5f, 1.5f); + /* Reducing dithering can help us hit the target. */ + psEnc->sCmn.indices.quantOffsetType = 0; found_upper = 0; gainsID_upper = -1; } else { @@ -250,15 +281,34 @@ opus_int silk_encode_frame_FLP( break; } + if ( !found_lower && nBits > maxBits ) { + int j; + for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + int sum=0; + for ( j = i*psEnc->sCmn.subfr_length; j < (i+1)*psEnc->sCmn.subfr_length; j++ ) { + sum += abs( psEnc->sCmn.pulses[j] ); + } + if ( iter == 0 || (sum < best_sum[i] && !gain_lock[i]) ) { + best_sum[i] = sum; + best_gain_mult[i] = gainMult_Q8; + } else { + gain_lock[i] = 1; + } + } + } if( ( found_lower & found_upper ) == 0 ) { /* Adjust gain according to high-rate rate/distortion curve */ - opus_int32 gain_factor_Q16; - gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) ); - gain_factor_Q16 = silk_min_32( gain_factor_Q16, SILK_FIX_CONST( 2, 16 ) ); if( nBits > maxBits ) { - gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) ); + if (gainMult_Q8 < 16384) { + gainMult_Q8 *= 2; + } else { + gainMult_Q8 = 32767; + } + } else { + opus_int32 gain_factor_Q16; + gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) ); + gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 ); } - gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 ); } else { /* Adjust gain by interpolating */ gainMult_Q8 = gainMult_lower + ( ( gainMult_upper - gainMult_lower ) * ( maxBits - nBits_lower ) ) / ( nBits_upper - nBits_lower ); @@ -272,7 +322,13 @@ opus_int silk_encode_frame_FLP( } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { - pGains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q8 ), 8 ); + opus_int16 tmp; + if ( gain_lock[i] ) { + tmp = best_gain_mult[i]; + } else { + tmp = gainMult_Q8; + } + pGains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], tmp ), 8 ); } /* Quantize gains */ diff --git a/silk/float/energy_FLP.c b/silk/float/energy_FLP.c index 24b8179f..7bc7173c 100644 --- a/silk/float/energy_FLP.c +++ b/silk/float/energy_FLP.c @@ -37,13 +37,12 @@ double silk_energy_FLP( opus_int dataSize ) { - opus_int i, dataSize4; + opus_int i; double result; /* 4x unrolled loop */ result = 0.0; - dataSize4 = dataSize & 0xFFFC; - for( i = 0; i < dataSize4; i += 4 ) { + for( i = 0; i < dataSize - 3; i += 4 ) { result += data[ i + 0 ] * (double)data[ i + 0 ] + data[ i + 1 ] * (double)data[ i + 1 ] + data[ i + 2 ] * (double)data[ i + 2 ] + diff --git a/silk/float/find_LPC_FLP.c b/silk/float/find_LPC_FLP.c index fcfe1c36..4d63964f 100644 --- a/silk/float/find_LPC_FLP.c +++ b/silk/float/find_LPC_FLP.c @@ -73,7 +73,7 @@ void silk_find_LPC_FLP( silk_interpolate( NLSF0_Q15, psEncC->prev_NLSFq_Q15, NLSF_Q15, k, psEncC->predictLPCOrder ); /* Convert to LPC for residual energy evaluation */ - silk_NLSF2A_FLP( a_tmp, NLSF0_Q15, psEncC->predictLPCOrder ); + silk_NLSF2A_FLP( a_tmp, NLSF0_Q15, psEncC->predictLPCOrder, psEncC->arch ); /* Calculate residual energy with LSF interpolation */ silk_LPC_analysis_filter_FLP( LPC_res, a_tmp, x, 2 * subfr_length, psEncC->predictLPCOrder ); diff --git a/silk/float/find_LTP_FLP.c b/silk/float/find_LTP_FLP.c index 72299960..f9706493 100644 --- a/silk/float/find_LTP_FLP.c +++ b/silk/float/find_LTP_FLP.c @@ -33,100 +33,32 @@ POSSIBILITY OF SUCH DAMAGE. #include "tuning_parameters.h" void silk_find_LTP_FLP( - silk_float b[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */ - silk_float WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ - silk_float *LTPredCodGain, /* O LTP coding gain */ - const silk_float r_lpc[], /* I LPC residual */ - const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ - const silk_float Wght[ MAX_NB_SUBFR ], /* I Weights */ + silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ + silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* O Weight for LTP quantization */ + const silk_float r_ptr[], /* I LPC residual */ + const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ const opus_int subfr_length, /* I Subframe length */ - const opus_int nb_subfr, /* I number of subframes */ - const opus_int mem_offset /* I Number of samples in LTP memory */ + const opus_int nb_subfr /* I number of subframes */ ) { - opus_int i, k; - silk_float *b_ptr, temp, *WLTP_ptr; - silk_float LPC_res_nrg, LPC_LTP_res_nrg; - silk_float d[ MAX_NB_SUBFR ], m, g, delta_b[ LTP_ORDER ]; - silk_float w[ MAX_NB_SUBFR ], nrg[ MAX_NB_SUBFR ], regu; - silk_float Rr[ LTP_ORDER ], rr[ MAX_NB_SUBFR ]; - const silk_float *r_ptr, *lag_ptr; + opus_int k; + silk_float *xX_ptr, *XX_ptr; + const silk_float *lag_ptr; + silk_float xx, temp; - b_ptr = b; - WLTP_ptr = WLTP; - r_ptr = &r_lpc[ mem_offset ]; + xX_ptr = xX; + XX_ptr = XX; for( k = 0; k < nb_subfr; k++ ) { lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 ); + silk_corrMatrix_FLP( lag_ptr, subfr_length, LTP_ORDER, XX_ptr ); + silk_corrVector_FLP( lag_ptr, r_ptr, subfr_length, LTP_ORDER, xX_ptr ); + xx = ( silk_float )silk_energy_FLP( r_ptr, subfr_length + LTP_ORDER ); + temp = 1.0f / silk_max( xx, LTP_CORR_INV_MAX * 0.5f * ( XX_ptr[ 0 ] + XX_ptr[ 24 ] ) + 1.0f ); + silk_scale_vector_FLP( XX_ptr, temp, LTP_ORDER * LTP_ORDER ); + silk_scale_vector_FLP( xX_ptr, temp, LTP_ORDER ); - silk_corrMatrix_FLP( lag_ptr, subfr_length, LTP_ORDER, WLTP_ptr ); - silk_corrVector_FLP( lag_ptr, r_ptr, subfr_length, LTP_ORDER, Rr ); - - rr[ k ] = ( silk_float )silk_energy_FLP( r_ptr, subfr_length ); - regu = 1.0f + rr[ k ] + - matrix_ptr( WLTP_ptr, 0, 0, LTP_ORDER ) + - matrix_ptr( WLTP_ptr, LTP_ORDER-1, LTP_ORDER-1, LTP_ORDER ); - regu *= LTP_DAMPING / 3; - silk_regularize_correlations_FLP( WLTP_ptr, &rr[ k ], regu, LTP_ORDER ); - silk_solve_LDL_FLP( WLTP_ptr, LTP_ORDER, Rr, b_ptr ); - - /* Calculate residual energy */ - nrg[ k ] = silk_residual_energy_covar_FLP( b_ptr, WLTP_ptr, Rr, rr[ k ], LTP_ORDER ); - - temp = Wght[ k ] / ( nrg[ k ] * Wght[ k ] + 0.01f * subfr_length ); - silk_scale_vector_FLP( WLTP_ptr, temp, LTP_ORDER * LTP_ORDER ); - w[ k ] = matrix_ptr( WLTP_ptr, LTP_ORDER / 2, LTP_ORDER / 2, LTP_ORDER ); - - r_ptr += subfr_length; - b_ptr += LTP_ORDER; - WLTP_ptr += LTP_ORDER * LTP_ORDER; - } - - /* Compute LTP coding gain */ - if( LTPredCodGain != NULL ) { - LPC_LTP_res_nrg = 1e-6f; - LPC_res_nrg = 0.0f; - for( k = 0; k < nb_subfr; k++ ) { - LPC_res_nrg += rr[ k ] * Wght[ k ]; - LPC_LTP_res_nrg += nrg[ k ] * Wght[ k ]; - } - - silk_assert( LPC_LTP_res_nrg > 0 ); - *LTPredCodGain = 3.0f * silk_log2( LPC_res_nrg / LPC_LTP_res_nrg ); - } - - /* Smoothing */ - /* d = sum( B, 1 ); */ - b_ptr = b; - for( k = 0; k < nb_subfr; k++ ) { - d[ k ] = 0; - for( i = 0; i < LTP_ORDER; i++ ) { - d[ k ] += b_ptr[ i ]; - } - b_ptr += LTP_ORDER; - } - /* m = ( w * d' ) / ( sum( w ) + 1e-3 ); */ - temp = 1e-3f; - for( k = 0; k < nb_subfr; k++ ) { - temp += w[ k ]; - } - m = 0; - for( k = 0; k < nb_subfr; k++ ) { - m += d[ k ] * w[ k ]; - } - m = m / temp; - - b_ptr = b; - for( k = 0; k < nb_subfr; k++ ) { - g = LTP_SMOOTHING / ( LTP_SMOOTHING + w[ k ] ) * ( m - d[ k ] ); - temp = 0; - for( i = 0; i < LTP_ORDER; i++ ) { - delta_b[ i ] = silk_max_float( b_ptr[ i ], 0.1f ); - temp += delta_b[ i ]; - } - temp = g / temp; - for( i = 0; i < LTP_ORDER; i++ ) { - b_ptr[ i ] = b_ptr[ i ] + delta_b[ i ] * temp; - } - b_ptr += LTP_ORDER; + r_ptr += subfr_length; + XX_ptr += LTP_ORDER * LTP_ORDER; + xX_ptr += LTP_ORDER; } } diff --git a/silk/float/find_pred_coefs_FLP.c b/silk/float/find_pred_coefs_FLP.c index 1af4fe5f..cb2e763b 100644 --- a/silk/float/find_pred_coefs_FLP.c +++ b/silk/float/find_pred_coefs_FLP.c @@ -41,8 +41,9 @@ void silk_find_pred_coefs_FLP( ) { opus_int i; - silk_float WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; - silk_float invGains[ MAX_NB_SUBFR ], Wght[ MAX_NB_SUBFR ]; + silk_float XXLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; + silk_float xXLTP[ MAX_NB_SUBFR * LTP_ORDER ]; + silk_float invGains[ MAX_NB_SUBFR ]; opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const silk_float *x_ptr; silk_float *x_pre_ptr, LPC_in_pre[ MAX_NB_SUBFR * MAX_LPC_ORDER + MAX_FRAME_LENGTH ]; @@ -52,7 +53,6 @@ void silk_find_pred_coefs_FLP( for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { silk_assert( psEncCtrl->Gains[ i ] > 0.0f ); invGains[ i ] = 1.0f / psEncCtrl->Gains[ i ]; - Wght[ i ] = invGains[ i ] * invGains[ i ]; } if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { @@ -62,13 +62,11 @@ void silk_find_pred_coefs_FLP( silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); /* LTP analysis */ - silk_find_LTP_FLP( psEncCtrl->LTPCoef, WLTP, &psEncCtrl->LTPredCodGain, res_pitch, - psEncCtrl->pitchL, Wght, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length ); + silk_find_LTP_FLP( XXLTP, xXLTP, res_pitch, psEncCtrl->pitchL, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr ); /* Quantize LTP gain parameters */ silk_quant_LTP_gains_FLP( psEncCtrl->LTPCoef, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, - &psEnc->sCmn.sum_log_gain_Q7, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr, - psEnc->sCmn.arch ); + &psEnc->sCmn.sum_log_gain_Q7, &psEncCtrl->LTPredCodGain, XXLTP, xXLTP, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch ); /* Control LTP scaling */ silk_LTP_scale_ctrl_FLP( psEnc, psEncCtrl, condCoding ); diff --git a/silk/float/inner_product_FLP.c b/silk/float/inner_product_FLP.c index 029c0129..cdd39d24 100644 --- a/silk/float/inner_product_FLP.c +++ b/silk/float/inner_product_FLP.c @@ -38,13 +38,12 @@ double silk_inner_product_FLP( opus_int dataSize ) { - opus_int i, dataSize4; + opus_int i; double result; /* 4x unrolled loop */ result = 0.0; - dataSize4 = dataSize & 0xFFFC; - for( i = 0; i < dataSize4; i += 4 ) { + for( i = 0; i < dataSize - 3; i += 4 ) { result += data1[ i + 0 ] * (double)data2[ i + 0 ] + data1[ i + 1 ] * (double)data2[ i + 1 ] + data1[ i + 2 ] * (double)data2[ i + 2 ] + diff --git a/silk/float/k2a_FLP.c b/silk/float/k2a_FLP.c index 12af4e76..1448008d 100644 --- a/silk/float/k2a_FLP.c +++ b/silk/float/k2a_FLP.c @@ -39,15 +39,16 @@ void silk_k2a_FLP( ) { opus_int k, n; - silk_float Atmp[ SILK_MAX_ORDER_LPC ]; + silk_float rck, tmp1, tmp2; for( k = 0; k < order; k++ ) { - for( n = 0; n < k; n++ ) { - Atmp[ n ] = A[ n ]; + rck = rc[ k ]; + for( n = 0; n < (k + 1) >> 1; n++ ) { + tmp1 = A[ n ]; + tmp2 = A[ k - n - 1 ]; + A[ n ] = tmp1 + tmp2 * rck; + A[ k - n - 1 ] = tmp2 + tmp1 * rck; } - for( n = 0; n < k; n++ ) { - A[ n ] += Atmp[ k - n - 1 ] * rc[ k ]; - } - A[ k ] = -rc[ k ]; + A[ k ] = -rck; } } diff --git a/silk/float/levinsondurbin_FLP.c b/silk/float/levinsondurbin_FLP.c deleted file mode 100644 index f0ba6069..00000000 --- a/silk/float/levinsondurbin_FLP.c +++ /dev/null @@ -1,81 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "SigProc_FLP.h" - -/* Solve the normal equations using the Levinson-Durbin recursion */ -silk_float silk_levinsondurbin_FLP( /* O prediction error energy */ - silk_float A[], /* O prediction coefficients [order] */ - const silk_float corr[], /* I input auto-correlations [order + 1] */ - const opus_int order /* I prediction order */ -) -{ - opus_int i, mHalf, m; - silk_float min_nrg, nrg, t, km, Atmp1, Atmp2; - - min_nrg = 1e-12f * corr[ 0 ] + 1e-9f; - nrg = corr[ 0 ]; - nrg = silk_max_float(min_nrg, nrg); - A[ 0 ] = corr[ 1 ] / nrg; - nrg -= A[ 0 ] * corr[ 1 ]; - nrg = silk_max_float(min_nrg, nrg); - - for( m = 1; m < order; m++ ) - { - t = corr[ m + 1 ]; - for( i = 0; i < m; i++ ) { - t -= A[ i ] * corr[ m - i ]; - } - - /* reflection coefficient */ - km = t / nrg; - - /* residual energy */ - nrg -= km * t; - nrg = silk_max_float(min_nrg, nrg); - - mHalf = m >> 1; - for( i = 0; i < mHalf; i++ ) { - Atmp1 = A[ i ]; - Atmp2 = A[ m - i - 1 ]; - A[ m - i - 1 ] -= km * Atmp1; - A[ i ] -= km * Atmp2; - } - if( m & 1 ) { - A[ mHalf ] -= km * A[ mHalf ]; - } - A[ m ] = km; - } - - /* return the residual energy */ - return nrg; -} - diff --git a/silk/float/main_FLP.h b/silk/float/main_FLP.h index e5a75972..f47fc93b 100644 --- a/silk/float/main_FLP.h +++ b/silk/float/main_FLP.h @@ -79,22 +79,11 @@ opus_int silk_init_encoder( opus_int silk_control_encoder( silk_encoder_state_FLP *psEnc, /* I/O Pointer to Silk encoder state FLP */ silk_EncControlStruct *encControl, /* I Control structure */ - const opus_int32 TargetRate_bps, /* I Target max bitrate (bps) */ const opus_int allow_bw_switch, /* I Flag to allow switching audio bandwidth */ const opus_int channelNb, /* I Channel number */ const opus_int force_fs_kHz ); -/****************/ -/* Prefiltering */ -/****************/ -void silk_prefilter_FLP( - silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ - const silk_encoder_control_FLP *psEncCtrl, /* I Encoder control FLP */ - silk_float xw[], /* O Weighted signal */ - const silk_float x[] /* I Speech signal */ -); - /**************************/ /* Noise shaping analysis */ /**************************/ @@ -153,15 +142,12 @@ void silk_find_LPC_FLP( /* LTP analysis */ void silk_find_LTP_FLP( - silk_float b[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */ - silk_float WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ - silk_float *LTPredCodGain, /* O LTP coding gain */ - const silk_float r_lpc[], /* I LPC residual */ + silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ + silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* O Weight for LTP quantization */ + const silk_float r_ptr[], /* I LPC residual */ const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ - const silk_float Wght[ MAX_NB_SUBFR ], /* I Weights */ const opus_int subfr_length, /* I Subframe length */ - const opus_int nb_subfr, /* I number of subframes */ - const opus_int mem_offset /* I Number of samples in LTP memory */ + const opus_int nb_subfr /* I number of subframes */ ); void silk_LTP_analysis_filter_FLP( @@ -198,14 +184,15 @@ void silk_LPC_analysis_filter_FLP( /* LTP tap quantizer */ void silk_quant_LTP_gains_FLP( - silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (Un-)quantized LTP gains */ + silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook index */ opus_int8 *periodicity_index, /* O Periodicity index */ opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */ - const silk_float W[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Error weights */ - const opus_int mu_Q10, /* I Mu value (R/D tradeoff) */ - const opus_int lowComplexity, /* I Flag for low complexity */ - const opus_int nb_subfr, /* I number of subframes */ + silk_float *pred_gain_dB, /* O LTP prediction gain */ + const silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Correlation matrix */ + const silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* I Correlation vector */ + const opus_int subfr_len, /* I Number of samples per subframe */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ); @@ -245,22 +232,6 @@ void silk_corrVector_FLP( silk_float *Xt /* O X'*t correlation vector [order] */ ); -/* Add noise to matrix diagonal */ -void silk_regularize_correlations_FLP( - silk_float *XX, /* I/O Correlation matrices */ - silk_float *xx, /* I/O Correlation values */ - const silk_float noise, /* I Noise energy to add */ - const opus_int D /* I Dimension of XX */ -); - -/* Function to solve linear equation Ax = b, where A is an MxM symmetric matrix */ -void silk_solve_LDL_FLP( - silk_float *A, /* I/O Symmetric square matrix, out: reg. */ - const opus_int M, /* I Size of matrix */ - const silk_float *b, /* I Pointer to b vector */ - silk_float *x /* O Pointer to x solution vector */ -); - /* Apply sine window to signal vector. */ /* Window types: */ /* 1 -> sine window from 0 to pi/2 */ @@ -285,7 +256,8 @@ void silk_A2NLSF_FLP( void silk_NLSF2A_FLP( silk_float *pAR, /* O LPC coefficients [ LPC_order ] */ const opus_int16 *NLSF_Q15, /* I NLSF vector [ LPC_order ] */ - const opus_int LPC_order /* I LPC order */ + const opus_int LPC_order, /* I LPC order */ + int arch /* I Run-time architecture */ ); /* Limit, stabilize, and quantize NLSFs */ diff --git a/silk/float/noise_shape_analysis_FLP.c b/silk/float/noise_shape_analysis_FLP.c index 65f6ea58..cb3d8a50 100644 --- a/silk/float/noise_shape_analysis_FLP.c +++ b/silk/float/noise_shape_analysis_FLP.c @@ -55,25 +55,21 @@ static OPUS_INLINE silk_float warped_gain( /* Convert warped filter coefficients to monic pseudo-warped coefficients and limit maximum */ /* amplitude of monic warped coefficients by using bandwidth expansion on the true coefficients */ static OPUS_INLINE void warped_true2monic_coefs( - silk_float *coefs_syn, - silk_float *coefs_ana, + silk_float *coefs, silk_float lambda, silk_float limit, opus_int order ) { opus_int i, iter, ind = 0; - silk_float tmp, maxabs, chirp, gain_syn, gain_ana; + silk_float tmp, maxabs, chirp, gain; /* Convert to monic coefficients */ for( i = order - 1; i > 0; i-- ) { - coefs_syn[ i - 1 ] -= lambda * coefs_syn[ i ]; - coefs_ana[ i - 1 ] -= lambda * coefs_ana[ i ]; + coefs[ i - 1 ] -= lambda * coefs[ i ]; } - gain_syn = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs_syn[ 0 ] ); - gain_ana = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs_ana[ 0 ] ); + gain = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs[ 0 ] ); for( i = 0; i < order; i++ ) { - coefs_syn[ i ] *= gain_syn; - coefs_ana[ i ] *= gain_ana; + coefs[ i ] *= gain; } /* Limit */ @@ -81,7 +77,7 @@ static OPUS_INLINE void warped_true2monic_coefs( /* Find maximum absolute value */ maxabs = -1.0f; for( i = 0; i < order; i++ ) { - tmp = silk_max( silk_abs_float( coefs_syn[ i ] ), silk_abs_float( coefs_ana[ i ] ) ); + tmp = silk_abs_float( coefs[ i ] ); if( tmp > maxabs ) { maxabs = tmp; ind = i; @@ -94,36 +90,59 @@ static OPUS_INLINE void warped_true2monic_coefs( /* Convert back to true warped coefficients */ for( i = 1; i < order; i++ ) { - coefs_syn[ i - 1 ] += lambda * coefs_syn[ i ]; - coefs_ana[ i - 1 ] += lambda * coefs_ana[ i ]; + coefs[ i - 1 ] += lambda * coefs[ i ]; } - gain_syn = 1.0f / gain_syn; - gain_ana = 1.0f / gain_ana; + gain = 1.0f / gain; for( i = 0; i < order; i++ ) { - coefs_syn[ i ] *= gain_syn; - coefs_ana[ i ] *= gain_ana; + coefs[ i ] *= gain; } /* Apply bandwidth expansion */ chirp = 0.99f - ( 0.8f + 0.1f * iter ) * ( maxabs - limit ) / ( maxabs * ( ind + 1 ) ); - silk_bwexpander_FLP( coefs_syn, order, chirp ); - silk_bwexpander_FLP( coefs_ana, order, chirp ); + silk_bwexpander_FLP( coefs, order, chirp ); /* Convert to monic warped coefficients */ for( i = order - 1; i > 0; i-- ) { - coefs_syn[ i - 1 ] -= lambda * coefs_syn[ i ]; - coefs_ana[ i - 1 ] -= lambda * coefs_ana[ i ]; + coefs[ i - 1 ] -= lambda * coefs[ i ]; } - gain_syn = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs_syn[ 0 ] ); - gain_ana = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs_ana[ 0 ] ); + gain = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs[ 0 ] ); for( i = 0; i < order; i++ ) { - coefs_syn[ i ] *= gain_syn; - coefs_ana[ i ] *= gain_ana; + coefs[ i ] *= gain; } } silk_assert( 0 ); } +static OPUS_INLINE void limit_coefs( + silk_float *coefs, + silk_float limit, + opus_int order +) { + opus_int i, iter, ind = 0; + silk_float tmp, maxabs, chirp; + + for( iter = 0; iter < 10; iter++ ) { + /* Find maximum absolute value */ + maxabs = -1.0f; + for( i = 0; i < order; i++ ) { + tmp = silk_abs_float( coefs[ i ] ); + if( tmp > maxabs ) { + maxabs = tmp; + ind = i; + } + } + if( maxabs <= limit ) { + /* Coefficients are within range - done */ + return; + } + + /* Apply bandwidth expansion */ + chirp = 0.99f - ( 0.8f + 0.1f * iter ) * ( maxabs - limit ) / ( maxabs * ( ind + 1 ) ); + silk_bwexpander_FLP( coefs, order, chirp ); + } + silk_assert( 0 ); +} + /* Compute noise shaping coefficients and initial gain values */ void silk_noise_shape_analysis_FLP( silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ @@ -133,12 +152,13 @@ void silk_noise_shape_analysis_FLP( ) { silk_shape_state_FLP *psShapeSt = &psEnc->sShape; - opus_int k, nSamples; - silk_float SNR_adj_dB, HarmBoost, HarmShapeGain, Tilt; - silk_float nrg, pre_nrg, log_energy, log_energy_prev, energy_variation; - silk_float delta, BWExp1, BWExp2, gain_mult, gain_add, strength, b, warping; + opus_int k, nSamples, nSegs; + silk_float SNR_adj_dB, HarmShapeGain, Tilt; + silk_float nrg, log_energy, log_energy_prev, energy_variation; + silk_float BWExp, gain_mult, gain_add, strength, b, warping; silk_float x_windowed[ SHAPE_LPC_WIN_MAX ]; silk_float auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ]; + silk_float rc[ MAX_SHAPE_LPC_ORDER + 1 ]; const silk_float *x_ptr, *pitch_res_ptr; /* Point to start of first LPC analysis block */ @@ -176,14 +196,14 @@ void silk_noise_shape_analysis_FLP( if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /* Initially set to 0; may be overruled in process_gains(..) */ psEnc->sCmn.indices.quantOffsetType = 0; - psEncCtrl->sparseness = 0.0f; } else { /* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */ nSamples = 2 * psEnc->sCmn.fs_kHz; energy_variation = 0.0f; log_energy_prev = 0.0f; pitch_res_ptr = pitch_res; - for( k = 0; k < silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; k++ ) { + nSegs = silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; + for( k = 0; k < nSegs; k++ ) { nrg = ( silk_float )nSamples + ( silk_float )silk_energy_FLP( pitch_res_ptr, nSamples ); log_energy = silk_log2( nrg ); if( k > 0 ) { @@ -192,17 +212,13 @@ void silk_noise_shape_analysis_FLP( log_energy_prev = log_energy; pitch_res_ptr += nSamples; } - psEncCtrl->sparseness = silk_sigmoid( 0.4f * ( energy_variation - 5.0f ) ); /* Set quantization offset depending on sparseness measure */ - if( psEncCtrl->sparseness > SPARSENESS_THRESHOLD_QNT_OFFSET ) { + if( energy_variation > ENERGY_VARIATION_THRESHOLD_QNT_OFFSET * (nSegs-1) ) { psEnc->sCmn.indices.quantOffsetType = 0; } else { psEnc->sCmn.indices.quantOffsetType = 1; } - - /* Increase coding SNR for sparse signals */ - SNR_adj_dB += SPARSE_SNR_INCR_dB * ( psEncCtrl->sparseness - 0.5f ); } /*******************************/ @@ -210,19 +226,10 @@ void silk_noise_shape_analysis_FLP( /*******************************/ /* More BWE for signals with high prediction gain */ strength = FIND_PITCH_WHITE_NOISE_FRACTION * psEncCtrl->predGain; /* between 0.0 and 1.0 */ - BWExp1 = BWExp2 = BANDWIDTH_EXPANSION / ( 1.0f + strength * strength ); - delta = LOW_RATE_BANDWIDTH_EXPANSION_DELTA * ( 1.0f - 0.75f * psEncCtrl->coding_quality ); - BWExp1 -= delta; - BWExp2 += delta; - /* BWExp1 will be applied after BWExp2, so make it relative */ - BWExp1 /= BWExp2; - - if( psEnc->sCmn.warping_Q16 > 0 ) { - /* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */ - warping = (silk_float)psEnc->sCmn.warping_Q16 / 65536.0f + 0.01f * psEncCtrl->coding_quality; - } else { - warping = 0.0f; - } + BWExp = BANDWIDTH_EXPANSION / ( 1.0f + strength * strength ); + + /* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */ + warping = (silk_float)psEnc->sCmn.warping_Q16 / 65536.0f + 0.01f * psEncCtrl->coding_quality; /********************************************/ /* Compute noise shaping AR coefs and gains */ @@ -252,37 +259,28 @@ void silk_noise_shape_analysis_FLP( } /* Add white noise, as a fraction of energy */ - auto_corr[ 0 ] += auto_corr[ 0 ] * SHAPE_WHITE_NOISE_FRACTION; + auto_corr[ 0 ] += auto_corr[ 0 ] * SHAPE_WHITE_NOISE_FRACTION + 1.0f; /* Convert correlations to prediction coefficients, and compute residual energy */ - nrg = silk_levinsondurbin_FLP( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], auto_corr, psEnc->sCmn.shapingLPCOrder ); + nrg = silk_schur_FLP( rc, auto_corr, psEnc->sCmn.shapingLPCOrder ); + silk_k2a_FLP( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], rc, psEnc->sCmn.shapingLPCOrder ); psEncCtrl->Gains[ k ] = ( silk_float )sqrt( nrg ); if( psEnc->sCmn.warping_Q16 > 0 ) { /* Adjust gain for warping */ - psEncCtrl->Gains[ k ] *= warped_gain( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], warping, psEnc->sCmn.shapingLPCOrder ); + psEncCtrl->Gains[ k ] *= warped_gain( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], warping, psEnc->sCmn.shapingLPCOrder ); } /* Bandwidth expansion for synthesis filter shaping */ - silk_bwexpander_FLP( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp2 ); - - /* Compute noise shaping filter coefficients */ - silk_memcpy( - &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], - &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], - psEnc->sCmn.shapingLPCOrder * sizeof( silk_float ) ); + silk_bwexpander_FLP( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp ); - /* Bandwidth expansion for analysis filter shaping */ - silk_bwexpander_FLP( &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp1 ); - - /* Ratio of prediction gains, in energy domain */ - pre_nrg = silk_LPC_inverse_pred_gain_FLP( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder ); - nrg = silk_LPC_inverse_pred_gain_FLP( &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder ); - psEncCtrl->GainsPre[ k ] = 1.0f - 0.7f * ( 1.0f - pre_nrg / nrg ); - - /* Convert to monic warped prediction coefficients and limit absolute values */ - warped_true2monic_coefs( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], - warping, 3.999f, psEnc->sCmn.shapingLPCOrder ); + if( psEnc->sCmn.warping_Q16 > 0 ) { + /* Convert to monic warped prediction coefficients and limit absolute values */ + warped_true2monic_coefs( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], warping, 3.999f, psEnc->sCmn.shapingLPCOrder ); + } else { + /* Limit absolute values */ + limit_coefs( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], 3.999f, psEnc->sCmn.shapingLPCOrder ); + } } /*****************/ @@ -296,11 +294,6 @@ void silk_noise_shape_analysis_FLP( psEncCtrl->Gains[ k ] += gain_add; } - gain_mult = 1.0f + INPUT_TILT + psEncCtrl->coding_quality * HIGH_RATE_INPUT_TILT; - for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - psEncCtrl->GainsPre[ k ] *= gain_mult; - } - /************************************************/ /* Control low-frequency shaping and noise tilt */ /************************************************/ @@ -331,12 +324,6 @@ void silk_noise_shape_analysis_FLP( /****************************/ /* HARMONIC SHAPING CONTROL */ /****************************/ - /* Control boosting of harmonic frequencies */ - HarmBoost = LOW_RATE_HARMONIC_BOOST * ( 1.0f - psEncCtrl->coding_quality ) * psEnc->LTPCorr; - - /* More harmonic boost for noisy input signals */ - HarmBoost += LOW_INPUT_QUALITY_HARMONIC_BOOST * ( 1.0f - psEncCtrl->input_quality ); - if( USE_HARM_SHAPING && psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /* Harmonic noise shaping */ HarmShapeGain = HARMONIC_SHAPING; @@ -355,8 +342,6 @@ void silk_noise_shape_analysis_FLP( /* Smooth over subframes */ /*************************/ for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - psShapeSt->HarmBoost_smth += SUBFR_SMTH_COEF * ( HarmBoost - psShapeSt->HarmBoost_smth ); - psEncCtrl->HarmBoost[ k ] = psShapeSt->HarmBoost_smth; psShapeSt->HarmShapeGain_smth += SUBFR_SMTH_COEF * ( HarmShapeGain - psShapeSt->HarmShapeGain_smth ); psEncCtrl->HarmShapeGain[ k ] = psShapeSt->HarmShapeGain_smth; psShapeSt->Tilt_smth += SUBFR_SMTH_COEF * ( Tilt - psShapeSt->Tilt_smth ); diff --git a/silk/float/pitch_analysis_core_FLP.c b/silk/float/pitch_analysis_core_FLP.c index d0e637a2..b3716937 100644 --- a/silk/float/pitch_analysis_core_FLP.c +++ b/silk/float/pitch_analysis_core_FLP.c @@ -159,7 +159,7 @@ opus_int silk_pitch_analysis_core_FLP( /* O Voicing estimate: 0 voiced, /* Low-pass filter */ for( i = frame_length_4kHz - 1; i > 0; i-- ) { - frame_4kHz[ i ] += frame_4kHz[ i - 1 ]; + frame_4kHz[ i ] = silk_ADD_SAT16( frame_4kHz[ i ], frame_4kHz[ i - 1 ] ); } /****************************************************************************** diff --git a/silk/float/prefilter_FLP.c b/silk/float/prefilter_FLP.c deleted file mode 100644 index 8bc32fb4..00000000 --- a/silk/float/prefilter_FLP.c +++ /dev/null @@ -1,206 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "main_FLP.h" -#include "tuning_parameters.h" - -/* -* Prefilter for finding Quantizer input signal -*/ -static OPUS_INLINE void silk_prefilt_FLP( - silk_prefilter_state_FLP *P, /* I/O state */ - silk_float st_res[], /* I */ - silk_float xw[], /* O */ - silk_float *HarmShapeFIR, /* I */ - silk_float Tilt, /* I */ - silk_float LF_MA_shp, /* I */ - silk_float LF_AR_shp, /* I */ - opus_int lag, /* I */ - opus_int length /* I */ -); - -static void silk_warped_LPC_analysis_filter_FLP( - silk_float state[], /* I/O State [order + 1] */ - silk_float res[], /* O Residual signal [length] */ - const silk_float coef[], /* I Coefficients [order] */ - const silk_float input[], /* I Input signal [length] */ - const silk_float lambda, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -) -{ - opus_int n, i; - silk_float acc, tmp1, tmp2; - - /* Order must be even */ - silk_assert( ( order & 1 ) == 0 ); - - for( n = 0; n < length; n++ ) { - /* Output of lowpass section */ - tmp2 = state[ 0 ] + lambda * state[ 1 ]; - state[ 0 ] = input[ n ]; - /* Output of allpass section */ - tmp1 = state[ 1 ] + lambda * ( state[ 2 ] - tmp2 ); - state[ 1 ] = tmp2; - acc = coef[ 0 ] * tmp2; - /* Loop over allpass sections */ - for( i = 2; i < order; i += 2 ) { - /* Output of allpass section */ - tmp2 = state[ i ] + lambda * ( state[ i + 1 ] - tmp1 ); - state[ i ] = tmp1; - acc += coef[ i - 1 ] * tmp1; - /* Output of allpass section */ - tmp1 = state[ i + 1 ] + lambda * ( state[ i + 2 ] - tmp2 ); - state[ i + 1 ] = tmp2; - acc += coef[ i ] * tmp2; - } - state[ order ] = tmp1; - acc += coef[ order - 1 ] * tmp1; - res[ n ] = input[ n ] - acc; - } -} - -/* -* silk_prefilter. Main prefilter function -*/ -void silk_prefilter_FLP( - silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ - const silk_encoder_control_FLP *psEncCtrl, /* I Encoder control FLP */ - silk_float xw[], /* O Weighted signal */ - const silk_float x[] /* I Speech signal */ -) -{ - silk_prefilter_state_FLP *P = &psEnc->sPrefilt; - opus_int j, k, lag; - silk_float HarmShapeGain, Tilt, LF_MA_shp, LF_AR_shp; - silk_float B[ 2 ]; - const silk_float *AR1_shp; - const silk_float *px; - silk_float *pxw; - silk_float HarmShapeFIR[ 3 ]; - silk_float st_res[ MAX_SUB_FRAME_LENGTH + MAX_LPC_ORDER ]; - - /* Set up pointers */ - px = x; - pxw = xw; - lag = P->lagPrev; - for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - /* Update Variables that change per sub frame */ - if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { - lag = psEncCtrl->pitchL[ k ]; - } - - /* Noise shape parameters */ - HarmShapeGain = psEncCtrl->HarmShapeGain[ k ] * ( 1.0f - psEncCtrl->HarmBoost[ k ] ); - HarmShapeFIR[ 0 ] = 0.25f * HarmShapeGain; - HarmShapeFIR[ 1 ] = 32767.0f / 65536.0f * HarmShapeGain; - HarmShapeFIR[ 2 ] = 0.25f * HarmShapeGain; - Tilt = psEncCtrl->Tilt[ k ]; - LF_MA_shp = psEncCtrl->LF_MA_shp[ k ]; - LF_AR_shp = psEncCtrl->LF_AR_shp[ k ]; - AR1_shp = &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ]; - - /* Short term FIR filtering */ - silk_warped_LPC_analysis_filter_FLP( P->sAR_shp, st_res, AR1_shp, px, - (silk_float)psEnc->sCmn.warping_Q16 / 65536.0f, psEnc->sCmn.subfr_length, psEnc->sCmn.shapingLPCOrder ); - - /* Reduce (mainly) low frequencies during harmonic emphasis */ - B[ 0 ] = psEncCtrl->GainsPre[ k ]; - B[ 1 ] = -psEncCtrl->GainsPre[ k ] * - ( psEncCtrl->HarmBoost[ k ] * HarmShapeGain + INPUT_TILT + psEncCtrl->coding_quality * HIGH_RATE_INPUT_TILT ); - pxw[ 0 ] = B[ 0 ] * st_res[ 0 ] + B[ 1 ] * P->sHarmHP; - for( j = 1; j < psEnc->sCmn.subfr_length; j++ ) { - pxw[ j ] = B[ 0 ] * st_res[ j ] + B[ 1 ] * st_res[ j - 1 ]; - } - P->sHarmHP = st_res[ psEnc->sCmn.subfr_length - 1 ]; - - silk_prefilt_FLP( P, pxw, pxw, HarmShapeFIR, Tilt, LF_MA_shp, LF_AR_shp, lag, psEnc->sCmn.subfr_length ); - - px += psEnc->sCmn.subfr_length; - pxw += psEnc->sCmn.subfr_length; - } - P->lagPrev = psEncCtrl->pitchL[ psEnc->sCmn.nb_subfr - 1 ]; -} - -/* -* Prefilter for finding Quantizer input signal -*/ -static OPUS_INLINE void silk_prefilt_FLP( - silk_prefilter_state_FLP *P, /* I/O state */ - silk_float st_res[], /* I */ - silk_float xw[], /* O */ - silk_float *HarmShapeFIR, /* I */ - silk_float Tilt, /* I */ - silk_float LF_MA_shp, /* I */ - silk_float LF_AR_shp, /* I */ - opus_int lag, /* I */ - opus_int length /* I */ -) -{ - opus_int i; - opus_int idx, LTP_shp_buf_idx; - silk_float n_Tilt, n_LF, n_LTP; - silk_float sLF_AR_shp, sLF_MA_shp; - silk_float *LTP_shp_buf; - - /* To speed up use temp variables instead of using the struct */ - LTP_shp_buf = P->sLTP_shp; - LTP_shp_buf_idx = P->sLTP_shp_buf_idx; - sLF_AR_shp = P->sLF_AR_shp; - sLF_MA_shp = P->sLF_MA_shp; - - for( i = 0; i < length; i++ ) { - if( lag > 0 ) { - silk_assert( HARM_SHAPE_FIR_TAPS == 3 ); - idx = lag + LTP_shp_buf_idx; - n_LTP = LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 - 1) & LTP_MASK ] * HarmShapeFIR[ 0 ]; - n_LTP += LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 ) & LTP_MASK ] * HarmShapeFIR[ 1 ]; - n_LTP += LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 + 1) & LTP_MASK ] * HarmShapeFIR[ 2 ]; - } else { - n_LTP = 0; - } - - n_Tilt = sLF_AR_shp * Tilt; - n_LF = sLF_AR_shp * LF_AR_shp + sLF_MA_shp * LF_MA_shp; - - sLF_AR_shp = st_res[ i ] - n_Tilt; - sLF_MA_shp = sLF_AR_shp - n_LF; - - LTP_shp_buf_idx = ( LTP_shp_buf_idx - 1 ) & LTP_MASK; - LTP_shp_buf[ LTP_shp_buf_idx ] = sLF_MA_shp; - - xw[ i ] = sLF_MA_shp - n_LTP; - } - /* Copy temp variable back to state */ - P->sLF_AR_shp = sLF_AR_shp; - P->sLF_MA_shp = sLF_MA_shp; - P->sLTP_shp_buf_idx = LTP_shp_buf_idx; -} diff --git a/silk/float/schur_FLP.c b/silk/float/schur_FLP.c index ee436f83..c1e0bbb5 100644 --- a/silk/float/schur_FLP.c +++ b/silk/float/schur_FLP.c @@ -38,22 +38,23 @@ silk_float silk_schur_FLP( /* O returns residual energy ) { opus_int k, n; - silk_float C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; - silk_float Ctmp1, Ctmp2, rc_tmp; + double C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; + double Ctmp1, Ctmp2, rc_tmp; - silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 ); + silk_assert( order >= 0 && order <= SILK_MAX_ORDER_LPC ); /* Copy correlations */ - for( k = 0; k < order+1; k++ ) { + k = 0; + do { C[ k ][ 0 ] = C[ k ][ 1 ] = auto_corr[ k ]; - } + } while( ++k <= order ); for( k = 0; k < order; k++ ) { /* Get reflection coefficient */ rc_tmp = -C[ k + 1 ][ 0 ] / silk_max_float( C[ 0 ][ 1 ], 1e-9f ); /* Save the output */ - refl_coef[ k ] = rc_tmp; + refl_coef[ k ] = (silk_float)rc_tmp; /* Update correlations */ for( n = 0; n < order - k; n++ ) { @@ -65,6 +66,5 @@ silk_float silk_schur_FLP( /* O returns residual energy } /* Return residual energy */ - return C[ 0 ][ 1 ]; + return (silk_float)C[ 0 ][ 1 ]; } - diff --git a/silk/float/solve_LS_FLP.c b/silk/float/solve_LS_FLP.c deleted file mode 100644 index 7c90d665..00000000 --- a/silk/float/solve_LS_FLP.c +++ /dev/null @@ -1,207 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "main_FLP.h" -#include "tuning_parameters.h" - -/********************************************************************** - * LDL Factorisation. Finds the upper triangular matrix L and the diagonal - * Matrix D (only the diagonal elements returned in a vector)such that - * the symmetric matric A is given by A = L*D*L'. - **********************************************************************/ -static OPUS_INLINE void silk_LDL_FLP( - silk_float *A, /* I/O Pointer to Symetric Square Matrix */ - opus_int M, /* I Size of Matrix */ - silk_float *L, /* I/O Pointer to Square Upper triangular Matrix */ - silk_float *Dinv /* I/O Pointer to vector holding the inverse diagonal elements of D */ -); - -/********************************************************************** - * Function to solve linear equation Ax = b, when A is a MxM lower - * triangular matrix, with ones on the diagonal. - **********************************************************************/ -static OPUS_INLINE void silk_SolveWithLowerTriangularWdiagOnes_FLP( - const silk_float *L, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const silk_float *b, /* I b Vector */ - silk_float *x /* O x Vector */ -); - -/********************************************************************** - * Function to solve linear equation (A^T)x = b, when A is a MxM lower - * triangular, with ones on the diagonal. (ie then A^T is upper triangular) - **********************************************************************/ -static OPUS_INLINE void silk_SolveWithUpperTriangularFromLowerWdiagOnes_FLP( - const silk_float *L, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const silk_float *b, /* I b Vector */ - silk_float *x /* O x Vector */ -); - -/********************************************************************** - * Function to solve linear equation Ax = b, when A is a MxM - * symmetric square matrix - using LDL factorisation - **********************************************************************/ -void silk_solve_LDL_FLP( - silk_float *A, /* I/O Symmetric square matrix, out: reg. */ - const opus_int M, /* I Size of matrix */ - const silk_float *b, /* I Pointer to b vector */ - silk_float *x /* O Pointer to x solution vector */ -) -{ - opus_int i; - silk_float L[ MAX_MATRIX_SIZE ][ MAX_MATRIX_SIZE ]; - silk_float T[ MAX_MATRIX_SIZE ]; - silk_float Dinv[ MAX_MATRIX_SIZE ]; /* inverse diagonal elements of D*/ - - silk_assert( M <= MAX_MATRIX_SIZE ); - - /*************************************************** - Factorize A by LDL such that A = L*D*(L^T), - where L is lower triangular with ones on diagonal - ****************************************************/ - silk_LDL_FLP( A, M, &L[ 0 ][ 0 ], Dinv ); - - /**************************************************** - * substitute D*(L^T) = T. ie: - L*D*(L^T)*x = b => L*T = b <=> T = inv(L)*b - ******************************************************/ - silk_SolveWithLowerTriangularWdiagOnes_FLP( &L[ 0 ][ 0 ], M, b, T ); - - /**************************************************** - D*(L^T)*x = T <=> (L^T)*x = inv(D)*T, because D is - diagonal just multiply with 1/d_i - ****************************************************/ - for( i = 0; i < M; i++ ) { - T[ i ] = T[ i ] * Dinv[ i ]; - } - /**************************************************** - x = inv(L') * inv(D) * T - *****************************************************/ - silk_SolveWithUpperTriangularFromLowerWdiagOnes_FLP( &L[ 0 ][ 0 ], M, T, x ); -} - -static OPUS_INLINE void silk_SolveWithUpperTriangularFromLowerWdiagOnes_FLP( - const silk_float *L, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const silk_float *b, /* I b Vector */ - silk_float *x /* O x Vector */ -) -{ - opus_int i, j; - silk_float temp; - const silk_float *ptr1; - - for( i = M - 1; i >= 0; i-- ) { - ptr1 = matrix_adr( L, 0, i, M ); - temp = 0; - for( j = M - 1; j > i ; j-- ) { - temp += ptr1[ j * M ] * x[ j ]; - } - temp = b[ i ] - temp; - x[ i ] = temp; - } -} - -static OPUS_INLINE void silk_SolveWithLowerTriangularWdiagOnes_FLP( - const silk_float *L, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const silk_float *b, /* I b Vector */ - silk_float *x /* O x Vector */ -) -{ - opus_int i, j; - silk_float temp; - const silk_float *ptr1; - - for( i = 0; i < M; i++ ) { - ptr1 = matrix_adr( L, i, 0, M ); - temp = 0; - for( j = 0; j < i; j++ ) { - temp += ptr1[ j ] * x[ j ]; - } - temp = b[ i ] - temp; - x[ i ] = temp; - } -} - -static OPUS_INLINE void silk_LDL_FLP( - silk_float *A, /* I/O Pointer to Symetric Square Matrix */ - opus_int M, /* I Size of Matrix */ - silk_float *L, /* I/O Pointer to Square Upper triangular Matrix */ - silk_float *Dinv /* I/O Pointer to vector holding the inverse diagonal elements of D */ -) -{ - opus_int i, j, k, loop_count, err = 1; - silk_float *ptr1, *ptr2; - double temp, diag_min_value; - silk_float v[ MAX_MATRIX_SIZE ], D[ MAX_MATRIX_SIZE ]; /* temp arrays*/ - - silk_assert( M <= MAX_MATRIX_SIZE ); - - diag_min_value = FIND_LTP_COND_FAC * 0.5f * ( A[ 0 ] + A[ M * M - 1 ] ); - for( loop_count = 0; loop_count < M && err == 1; loop_count++ ) { - err = 0; - for( j = 0; j < M; j++ ) { - ptr1 = matrix_adr( L, j, 0, M ); - temp = matrix_ptr( A, j, j, M ); /* element in row j column j*/ - for( i = 0; i < j; i++ ) { - v[ i ] = ptr1[ i ] * D[ i ]; - temp -= ptr1[ i ] * v[ i ]; - } - if( temp < diag_min_value ) { - /* Badly conditioned matrix: add white noise and run again */ - temp = ( loop_count + 1 ) * diag_min_value - temp; - for( i = 0; i < M; i++ ) { - matrix_ptr( A, i, i, M ) += ( silk_float )temp; - } - err = 1; - break; - } - D[ j ] = ( silk_float )temp; - Dinv[ j ] = ( silk_float )( 1.0f / temp ); - matrix_ptr( L, j, j, M ) = 1.0f; - - ptr1 = matrix_adr( A, j, 0, M ); - ptr2 = matrix_adr( L, j + 1, 0, M); - for( i = j + 1; i < M; i++ ) { - temp = 0.0; - for( k = 0; k < j; k++ ) { - temp += ptr2[ k ] * v[ k ]; - } - matrix_ptr( L, i, j, M ) = ( silk_float )( ( ptr1[ i ] - temp ) * Dinv[ j ] ); - ptr2 += M; /* go to next column*/ - } - } - } - silk_assert( err == 0 ); -} - diff --git a/silk/float/structs_FLP.h b/silk/float/structs_FLP.h index 14d647ce..3150b386 100644 --- a/silk/float/structs_FLP.h +++ b/silk/float/structs_FLP.h @@ -42,32 +42,16 @@ extern "C" /********************************/ typedef struct { opus_int8 LastGainIndex; - silk_float HarmBoost_smth; silk_float HarmShapeGain_smth; silk_float Tilt_smth; } silk_shape_state_FLP; /********************************/ -/* Prefilter state */ -/********************************/ -typedef struct { - silk_float sLTP_shp[ LTP_BUF_LENGTH ]; - silk_float sAR_shp[ MAX_SHAPE_LPC_ORDER + 1 ]; - opus_int sLTP_shp_buf_idx; - silk_float sLF_AR_shp; - silk_float sLF_MA_shp; - silk_float sHarmHP; - opus_int32 rand_seed; - opus_int lagPrev; -} silk_prefilter_state_FLP; - -/********************************/ /* Encoder state FLP */ /********************************/ typedef struct { silk_encoder_state sCmn; /* Common struct, shared with fixed-point code */ silk_shape_state_FLP sShape; /* Noise shaping state */ - silk_prefilter_state_FLP sPrefilt; /* Prefilter State */ /* Buffer for find pitch and noise shape analysis */ silk_float x_buf[ 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ];/* Buffer for find pitch and noise shape analysis */ @@ -86,12 +70,9 @@ typedef struct { opus_int pitchL[ MAX_NB_SUBFR ]; /* Noise shaping parameters */ - silk_float AR1[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; - silk_float AR2[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; + silk_float AR[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; silk_float LF_MA_shp[ MAX_NB_SUBFR ]; silk_float LF_AR_shp[ MAX_NB_SUBFR ]; - silk_float GainsPre[ MAX_NB_SUBFR ]; - silk_float HarmBoost[ MAX_NB_SUBFR ]; silk_float Tilt[ MAX_NB_SUBFR ]; silk_float HarmShapeGain[ MAX_NB_SUBFR ]; silk_float Lambda; @@ -99,7 +80,6 @@ typedef struct { silk_float coding_quality; /* Measures */ - silk_float sparseness; silk_float predGain; silk_float LTPredCodGain; silk_float ResNrg[ MAX_NB_SUBFR ]; /* Residual energy per subframe */ diff --git a/silk/float/wrappers_FLP.c b/silk/float/wrappers_FLP.c index 6666b8ef..ad90b874 100644 --- a/silk/float/wrappers_FLP.c +++ b/silk/float/wrappers_FLP.c @@ -54,13 +54,14 @@ void silk_A2NLSF_FLP( void silk_NLSF2A_FLP( silk_float *pAR, /* O LPC coefficients [ LPC_order ] */ const opus_int16 *NLSF_Q15, /* I NLSF vector [ LPC_order ] */ - const opus_int LPC_order /* I LPC order */ + const opus_int LPC_order, /* I LPC order */ + int arch /* I Run-time architecture */ ) { opus_int i; opus_int16 a_fix_Q12[ MAX_LPC_ORDER ]; - silk_NLSF2A( a_fix_Q12, NLSF_Q15, LPC_order ); + silk_NLSF2A( a_fix_Q12, NLSF_Q15, LPC_order, arch ); for( i = 0; i < LPC_order; i++ ) { pAR[ i ] = ( silk_float )a_fix_Q12[ i ] * ( 1.0f / 4096.0f ); @@ -102,14 +103,14 @@ void silk_NSQ_wrapper_FLP( ) { opus_int i, j; - opus_int32 x_Q3[ MAX_FRAME_LENGTH ]; + opus_int16 x16[ MAX_FRAME_LENGTH ]; opus_int32 Gains_Q16[ MAX_NB_SUBFR ]; silk_DWORD_ALIGN opus_int16 PredCoef_Q12[ 2 ][ MAX_LPC_ORDER ]; opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ]; opus_int LTP_scale_Q14; /* Noise shaping parameters */ - opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; + opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ]; /* Packs two int16 coefficients per int32 value */ opus_int Lambda_Q10; opus_int Tilt_Q14[ MAX_NB_SUBFR ]; @@ -119,7 +120,7 @@ void silk_NSQ_wrapper_FLP( /* Noise shape parameters */ for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { for( j = 0; j < psEnc->sCmn.shapingLPCOrder; j++ ) { - AR2_Q13[ i * MAX_SHAPE_LPC_ORDER + j ] = silk_float2int( psEncCtrl->AR2[ i * MAX_SHAPE_LPC_ORDER + j ] * 8192.0f ); + AR_Q13[ i * MAX_SHAPE_LPC_ORDER + j ] = silk_float2int( psEncCtrl->AR[ i * MAX_SHAPE_LPC_ORDER + j ] * 8192.0f ); } } @@ -155,16 +156,16 @@ void silk_NSQ_wrapper_FLP( /* Convert input to fix */ for( i = 0; i < psEnc->sCmn.frame_length; i++ ) { - x_Q3[ i ] = silk_float2int( 8.0f * x[ i ] ); + x16[ i ] = silk_float2int( x[ i ] ); } /* Call NSQ */ if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { - silk_NSQ_del_dec( &psEnc->sCmn, psNSQ, psIndices, x_Q3, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, - AR2_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14, psEnc->sCmn.arch ); + silk_NSQ_del_dec( &psEnc->sCmn, psNSQ, psIndices, x16, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, + AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14, psEnc->sCmn.arch ); } else { - silk_NSQ( &psEnc->sCmn, psNSQ, psIndices, x_Q3, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, - AR2_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14, psEnc->sCmn.arch ); + silk_NSQ( &psEnc->sCmn, psNSQ, psIndices, x16, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, + AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14, psEnc->sCmn.arch ); } } @@ -172,31 +173,35 @@ void silk_NSQ_wrapper_FLP( /* Floating-point Silk LTP quantiation wrapper */ /***********************************************/ void silk_quant_LTP_gains_FLP( - silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (Un-)quantized LTP gains */ + silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook index */ opus_int8 *periodicity_index, /* O Periodicity index */ opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */ - const silk_float W[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Error weights */ - const opus_int mu_Q10, /* I Mu value (R/D tradeoff) */ - const opus_int lowComplexity, /* I Flag for low complexity */ - const opus_int nb_subfr, /* I number of subframes */ + silk_float *pred_gain_dB, /* O LTP prediction gain */ + const silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Correlation matrix */ + const silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* I Correlation vector */ + const opus_int subfr_len, /* I Number of samples per subframe */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ) { - opus_int i; + opus_int i, pred_gain_dB_Q7; opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ]; - opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ]; + opus_int32 XX_Q17[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; + opus_int32 xX_Q17[ MAX_NB_SUBFR * LTP_ORDER ]; - for( i = 0; i < nb_subfr * LTP_ORDER; i++ ) { - B_Q14[ i ] = (opus_int16)silk_float2int( B[ i ] * 16384.0f ); - } for( i = 0; i < nb_subfr * LTP_ORDER * LTP_ORDER; i++ ) { - W_Q18[ i ] = (opus_int32)silk_float2int( W[ i ] * 262144.0f ); + XX_Q17[ i ] = (opus_int32)silk_float2int( XX[ i ] * 131072.0f ); + } + for( i = 0; i < nb_subfr * LTP_ORDER; i++ ) { + xX_Q17[ i ] = (opus_int32)silk_float2int( xX[ i ] * 131072.0f ); } - silk_quant_LTP_gains( B_Q14, cbk_index, periodicity_index, sum_log_gain_Q7, W_Q18, mu_Q10, lowComplexity, nb_subfr, arch ); + silk_quant_LTP_gains( B_Q14, cbk_index, periodicity_index, sum_log_gain_Q7, &pred_gain_dB_Q7, XX_Q17, xX_Q17, subfr_len, nb_subfr, arch ); for( i = 0; i < nb_subfr * LTP_ORDER; i++ ) { B[ i ] = (silk_float)B_Q14[ i ] * ( 1.0f / 16384.0f ); } + + *pred_gain_dB = (silk_float)pred_gain_dB_Q7 * ( 1.0f / 128.0f ); } diff --git a/silk/gain_quant.c b/silk/gain_quant.c index 64ccd061..ee65245a 100644 --- a/silk/gain_quant.c +++ b/silk/gain_quant.c @@ -76,6 +76,7 @@ void silk_gains_quant( /* Accumulate deltas */ if( ind[ k ] > double_step_size_threshold ) { *prev_ind += silk_LSHIFT( ind[ k ], 1 ) - double_step_size_threshold; + *prev_ind = silk_min_int( *prev_ind, N_LEVELS_QGAIN - 1 ); } else { *prev_ind += ind[ k ]; } diff --git a/silk/init_decoder.c b/silk/init_decoder.c index f887c678..16c03dcd 100644 --- a/silk/init_decoder.c +++ b/silk/init_decoder.c @@ -44,6 +44,7 @@ opus_int silk_init_decoder( /* Used to deactivate LSF interpolation */ psDec->first_frame_after_reset = 1; psDec->prev_gain_Q16 = 65536; + psDec->arch = opus_select_arch(); /* Reset CNG state */ silk_CNG_Reset( psDec ); diff --git a/silk/lin2log.c b/silk/lin2log.c index d4fe5153..0d5155aa 100644 --- a/silk/lin2log.c +++ b/silk/lin2log.c @@ -41,6 +41,6 @@ opus_int32 silk_lin2log( silk_CLZ_FRAC( inLin, &lz, &frac_Q7 ); /* Piece-wise parabolic approximation */ - return silk_LSHIFT( 31 - lz, 7 ) + silk_SMLAWB( frac_Q7, silk_MUL( frac_Q7, 128 - frac_Q7 ), 179 ); + return silk_ADD_LSHIFT32( silk_SMLAWB( frac_Q7, silk_MUL( frac_Q7, 128 - frac_Q7 ), 179 ), 31 - lz, 7 ); } diff --git a/silk/macros.h b/silk/macros.h index d3ca3475..3c67b6e5 100644 --- a/silk/macros.h +++ b/silk/macros.h @@ -36,14 +36,6 @@ POSSIBILITY OF SUCH DAMAGE. #include "opus_defines.h" #include "arch.h" -#if OPUS_GNUC_PREREQ(3, 0) -#define opus_likely(x) (__builtin_expect(!!(x), 1)) -#define opus_unlikely(x) (__builtin_expect(!!(x), 0)) -#else -#define opus_likely(x) (!!(x)) -#define opus_unlikely(x) (!!(x)) -#endif - /* This is an OPUS_INLINE header file for general platform. */ /* (a32 * (opus_int32)((opus_int16)(b32))) >> 16 output have to be 32bit int */ diff --git a/silk/main.h b/silk/main.h index 2f90d68f..1a33eed5 100644 --- a/silk/main.h +++ b/silk/main.h @@ -42,6 +42,10 @@ POSSIBILITY OF SUCH DAMAGE. #include "x86/main_sse.h" #endif +#if (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) +#include "arm/NSQ_del_dec_arm.h" +#endif + /* Convert Left/Right stereo signal to adaptive Mid/Side representation */ void silk_stereo_LR_to_MS( stereo_enc_state *state, /* I/O State */ @@ -109,22 +113,22 @@ void silk_stereo_decode_mid_only( /* Encodes signs of excitation */ void silk_encode_signs( - ec_enc *psRangeEnc, /* I/O Compressor data structure */ - const opus_int8 pulses[], /* I pulse signal */ - opus_int length, /* I length of input */ - const opus_int signalType, /* I Signal type */ - const opus_int quantOffsetType, /* I Quantization offset type */ - const opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ + ec_enc *psRangeEnc, /* I/O Compressor data structure */ + const opus_int8 pulses[], /* I pulse signal */ + opus_int length, /* I length of input */ + const opus_int signalType, /* I Signal type */ + const opus_int quantOffsetType, /* I Quantization offset type */ + const opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ ); /* Decodes signs of excitation */ void silk_decode_signs( - ec_dec *psRangeDec, /* I/O Compressor data structure */ - opus_int16 pulses[], /* I/O pulse signal */ - opus_int length, /* I length of input */ - const opus_int signalType, /* I Signal type */ - const opus_int quantOffsetType, /* I Quantization offset type */ - const opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ + ec_dec *psRangeDec, /* I/O Compressor data structure */ + opus_int16 pulses[], /* I/O pulse signal */ + opus_int length, /* I length of input */ + const opus_int signalType, /* I Signal type */ + const opus_int quantOffsetType, /* I Quantization offset type */ + const opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ ); /* Check encoder control struct */ @@ -205,37 +209,37 @@ void silk_interpolate( /* LTP tap quantizer */ void silk_quant_LTP_gains( - opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (un)quantized LTP gains */ + opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */ opus_int8 *periodicity_index, /* O Periodicity Index */ opus_int32 *sum_gain_dB_Q7, /* I/O Cumulative max prediction gain */ - const opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Error Weights in Q18 */ - opus_int mu_Q9, /* I Mu value (R/D tradeoff) */ - opus_int lowComplexity, /* I Flag for low complexity */ - const opus_int nb_subfr, /* I number of subframes */ + opus_int *pred_gain_dB_Q7, /* O LTP prediction gain */ + const opus_int32 XX_Q17[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Correlation matrix in Q18 */ + const opus_int32 xX_Q17[ MAX_NB_SUBFR*LTP_ORDER ], /* I Correlation vector in Q18 */ + const opus_int subfr_len, /* I Number of samples per subframe */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ); /* Entropy constrained matrix-weighted VQ, for a single input data vector */ void silk_VQ_WMat_EC_c( opus_int8 *ind, /* O index of best codebook vector */ - opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */ + opus_int32 *res_nrg_Q15, /* O best residual energy */ + opus_int32 *rate_dist_Q8, /* O best total bitrate */ opus_int *gain_Q7, /* O sum of absolute LTP coefficients */ - const opus_int16 *in_Q14, /* I input vector to be quantized */ - const opus_int32 *W_Q18, /* I weighting matrix */ + const opus_int32 *XX_Q17, /* I correlation matrix */ + const opus_int32 *xX_Q17, /* I correlation vector */ const opus_int8 *cb_Q7, /* I codebook */ const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */ const opus_uint8 *cl_Q5, /* I code length for each codebook vector */ - const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */ + const opus_int subfr_len, /* I number of samples per subframe */ const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ - opus_int L /* I number of vectors in codebook */ + const opus_int L /* I number of vectors in codebook */ ); #if !defined(OVERRIDE_silk_VQ_WMat_EC) -#define silk_VQ_WMat_EC(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \ - mu_Q9, max_gain_Q7, L, arch) \ - ((void)(arch),silk_VQ_WMat_EC_c(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \ - mu_Q9, max_gain_Q7, L)) +#define silk_VQ_WMat_EC(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, subfr_len, max_gain_Q7, L, arch) \ + ((void)(arch),silk_VQ_WMat_EC_c(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, subfr_len, max_gain_Q7, L)) #endif /************************************/ @@ -243,14 +247,14 @@ void silk_VQ_WMat_EC_c( /************************************/ void silk_NSQ_c( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q3[], /* I Prefiltered input signal */ + const opus_int16 x16[], /* I Input */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ - const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ @@ -261,22 +265,22 @@ void silk_NSQ_c( ); #if !defined(OVERRIDE_silk_NSQ) -#define silk_NSQ(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \ +#define silk_NSQ(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \ HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \ - ((void)(arch),silk_NSQ_c(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \ + ((void)(arch),silk_NSQ_c(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \ HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14)) #endif /* Noise shaping using delayed decision */ void silk_NSQ_del_dec_c( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q3[], /* I Prefiltered input signal */ + const opus_int16 x16[], /* I Input */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ - const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ @@ -287,9 +291,9 @@ void silk_NSQ_del_dec_c( ); #if !defined(OVERRIDE_silk_NSQ_del_dec) -#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \ +#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \ HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \ - ((void)(arch),silk_NSQ_del_dec_c(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \ + ((void)(arch),silk_NSQ_del_dec_c(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \ HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14)) #endif @@ -346,6 +350,7 @@ void silk_NLSF_VQ( opus_int32 err_Q26[], /* O Quantization errors [K] */ const opus_int16 in_Q15[], /* I Input vectors to be quantized [LPC_order] */ const opus_uint8 pCB_Q8[], /* I Codebook vectors [K*LPC_order] */ + const opus_int16 pWght_Q9[], /* I Codebook weights [K*LPC_order] */ const opus_int K, /* I Number of codebook vectors */ const opus_int LPC_order /* I Number of LPCs */ ); diff --git a/silk/mips/NSQ_del_dec_mipsr1.h b/silk/mips/NSQ_del_dec_mipsr1.h index ad1cfe2a..cd70713a 100644 --- a/silk/mips/NSQ_del_dec_mipsr1.h +++ b/silk/mips/NSQ_del_dec_mipsr1.h @@ -61,7 +61,7 @@ static inline void silk_noise_shape_quantizer_del_dec( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay, /* I */ int arch /* I */ ) @@ -323,8 +323,9 @@ static inline void silk_noise_shape_quantizer_del_dec( psSS[ 1 ].xq_Q14 = xq_Q14; } - *smpl_buf_idx = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK; /* Index to newest samples */ - last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK; /* Index to decisionDelay old samples */ + *smpl_buf_idx = ( *smpl_buf_idx - 1 ) % DECISION_DELAY; + if( *smpl_buf_idx < 0 ) *smpl_buf_idx += DECISION_DELAY; + last_smple_idx = ( *smpl_buf_idx + decisionDelay ) % DECISION_DELAY; /* Find winner */ RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10; diff --git a/silk/mips/sigproc_fix_mipsr1.h b/silk/mips/sigproc_fix_mipsr1.h index 3b0a6953..51520c0a 100644 --- a/silk/mips/sigproc_fix_mipsr1.h +++ b/silk/mips/sigproc_fix_mipsr1.h @@ -28,11 +28,6 @@ POSSIBILITY OF SUCH DAMAGE. #ifndef SILK_SIGPROC_FIX_MIPSR1_H #define SILK_SIGPROC_FIX_MIPSR1_H -#ifdef __cplusplus -extern "C" -{ -#endif - #undef silk_SAT16 static inline short int silk_SAT16(int a) { diff --git a/silk/process_NLSFs.c b/silk/process_NLSFs.c index 0ab71f01..2f10f8df 100644 --- a/silk/process_NLSFs.c +++ b/silk/process_NLSFs.c @@ -89,7 +89,7 @@ void silk_process_NLSFs( NLSF_mu_Q20, psEncC->NLSF_MSVQ_Survivors, psEncC->indices.signalType ); /* Convert quantized NLSFs back to LPC coefficients */ - silk_NLSF2A( PredCoef_Q12[ 1 ], pNLSF_Q15, psEncC->predictLPCOrder ); + silk_NLSF2A( PredCoef_Q12[ 1 ], pNLSF_Q15, psEncC->predictLPCOrder, psEncC->arch ); if( doInterpolate ) { /* Calculate the interpolated, quantized LSF vector for the first half */ @@ -97,7 +97,7 @@ void silk_process_NLSFs( psEncC->indices.NLSFInterpCoef_Q2, psEncC->predictLPCOrder ); /* Convert back to LPC coefficients */ - silk_NLSF2A( PredCoef_Q12[ 0 ], pNLSF0_temp_Q15, psEncC->predictLPCOrder ); + silk_NLSF2A( PredCoef_Q12[ 0 ], pNLSF0_temp_Q15, psEncC->predictLPCOrder, psEncC->arch ); } else { /* Copy LPC coefficients for first half from second half */ diff --git a/silk/quant_LTP_gains.c b/silk/quant_LTP_gains.c index 513a8c44..d6b8eff8 100644 --- a/silk/quant_LTP_gains.c +++ b/silk/quant_LTP_gains.c @@ -33,14 +33,15 @@ POSSIBILITY OF SUCH DAMAGE. #include "tuning_parameters.h" void silk_quant_LTP_gains( - opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (un)quantized LTP gains */ + opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */ opus_int8 *periodicity_index, /* O Periodicity Index */ opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */ - const opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Error Weights in Q18 */ - opus_int mu_Q9, /* I Mu value (R/D tradeoff) */ - opus_int lowComplexity, /* I Flag for low complexity */ - const opus_int nb_subfr, /* I number of subframes */ + opus_int *pred_gain_dB_Q7, /* O LTP prediction gain */ + const opus_int32 XX_Q17[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Correlation matrix in Q18 */ + const opus_int32 xX_Q17[ MAX_NB_SUBFR*LTP_ORDER ], /* I Correlation vector in Q18 */ + const opus_int subfr_len, /* I Number of samples per subframe */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ) { @@ -49,16 +50,16 @@ void silk_quant_LTP_gains( const opus_uint8 *cl_ptr_Q5; const opus_int8 *cbk_ptr_Q7; const opus_uint8 *cbk_gain_ptr_Q7; - const opus_int16 *b_Q14_ptr; - const opus_int32 *W_Q18_ptr; - opus_int32 rate_dist_Q14_subfr, rate_dist_Q14, min_rate_dist_Q14; - opus_int32 sum_log_gain_tmp_Q7, best_sum_log_gain_Q7, max_gain_Q7, gain_Q7; + const opus_int32 *XX_Q17_ptr, *xX_Q17_ptr; + opus_int32 res_nrg_Q15_subfr, res_nrg_Q15, rate_dist_Q7_subfr, rate_dist_Q7, min_rate_dist_Q7; + opus_int32 sum_log_gain_tmp_Q7, best_sum_log_gain_Q7, max_gain_Q7; + opus_int gain_Q7; /***************************************************/ /* iterate over different codebooks with different */ /* rates/distortions, and choose best */ /***************************************************/ - min_rate_dist_Q14 = silk_int32_MAX; + min_rate_dist_Q7 = silk_int32_MAX; best_sum_log_gain_Q7 = 0; for( k = 0; k < 3; k++ ) { /* Safety margin for pitch gain control, to take into account factors @@ -70,53 +71,47 @@ void silk_quant_LTP_gains( cbk_gain_ptr_Q7 = silk_LTP_vq_gain_ptrs_Q7[ k ]; cbk_size = silk_LTP_vq_sizes[ k ]; - /* Set up pointer to first subframe */ - W_Q18_ptr = W_Q18; - b_Q14_ptr = B_Q14; + /* Set up pointers to first subframe */ + XX_Q17_ptr = XX_Q17; + xX_Q17_ptr = xX_Q17; - rate_dist_Q14 = 0; + res_nrg_Q15 = 0; + rate_dist_Q7 = 0; sum_log_gain_tmp_Q7 = *sum_log_gain_Q7; for( j = 0; j < nb_subfr; j++ ) { max_gain_Q7 = silk_log2lin( ( SILK_FIX_CONST( MAX_SUM_LOG_GAIN_DB / 6.0, 7 ) - sum_log_gain_tmp_Q7 ) + SILK_FIX_CONST( 7, 7 ) ) - gain_safety; - silk_VQ_WMat_EC( &temp_idx[ j ], /* O index of best codebook vector */ - &rate_dist_Q14_subfr, /* O best weighted quantization error + mu * rate */ + &res_nrg_Q15_subfr, /* O residual energy */ + &rate_dist_Q7_subfr, /* O best weighted quantization error + mu * rate */ &gain_Q7, /* O sum of absolute LTP coefficients */ - b_Q14_ptr, /* I input vector to be quantized */ - W_Q18_ptr, /* I weighting matrix */ + XX_Q17_ptr, /* I correlation matrix */ + xX_Q17_ptr, /* I correlation vector */ cbk_ptr_Q7, /* I codebook */ cbk_gain_ptr_Q7, /* I codebook effective gains */ cl_ptr_Q5, /* I code length for each codebook vector */ - mu_Q9, /* I tradeoff between weighted error and rate */ + subfr_len, /* I number of samples per subframe */ max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ cbk_size, /* I number of vectors in codebook */ arch /* I Run-time architecture */ ); - rate_dist_Q14 = silk_ADD_POS_SAT32( rate_dist_Q14, rate_dist_Q14_subfr ); + res_nrg_Q15 = silk_ADD_POS_SAT32( res_nrg_Q15, res_nrg_Q15_subfr ); + rate_dist_Q7 = silk_ADD_POS_SAT32( rate_dist_Q7, rate_dist_Q7_subfr ); sum_log_gain_tmp_Q7 = silk_max(0, sum_log_gain_tmp_Q7 + silk_lin2log( gain_safety + gain_Q7 ) - SILK_FIX_CONST( 7, 7 )); - b_Q14_ptr += LTP_ORDER; - W_Q18_ptr += LTP_ORDER * LTP_ORDER; + XX_Q17_ptr += LTP_ORDER * LTP_ORDER; + xX_Q17_ptr += LTP_ORDER; } - /* Avoid never finding a codebook */ - rate_dist_Q14 = silk_min( silk_int32_MAX - 1, rate_dist_Q14 ); - - if( rate_dist_Q14 < min_rate_dist_Q14 ) { - min_rate_dist_Q14 = rate_dist_Q14; + if( rate_dist_Q7 <= min_rate_dist_Q7 ) { + min_rate_dist_Q7 = rate_dist_Q7; *periodicity_index = (opus_int8)k; silk_memcpy( cbk_index, temp_idx, nb_subfr * sizeof( opus_int8 ) ); best_sum_log_gain_Q7 = sum_log_gain_tmp_Q7; } - - /* Break early in low-complexity mode if rate distortion is below threshold */ - if( lowComplexity && ( rate_dist_Q14 < silk_LTP_gain_middle_avg_RD_Q14 ) ) { - break; - } } cbk_ptr_Q7 = silk_LTP_vq_ptrs_Q7[ *periodicity_index ]; @@ -125,5 +120,13 @@ void silk_quant_LTP_gains( B_Q14[ j * LTP_ORDER + k ] = silk_LSHIFT( cbk_ptr_Q7[ cbk_index[ j ] * LTP_ORDER + k ], 7 ); } } + + if( nb_subfr == 2 ) { + res_nrg_Q15 = silk_RSHIFT32( res_nrg_Q15, 1 ); + } else { + res_nrg_Q15 = silk_RSHIFT32( res_nrg_Q15, 2 ); + } + *sum_log_gain_Q7 = best_sum_log_gain_Q7; + *pred_gain_dB_Q7 = (opus_int)silk_SMULBB( -3, silk_lin2log( res_nrg_Q15 ) - ( 15 << 7 ) ); } diff --git a/silk/structs.h b/silk/structs.h index 827829dc..4ff590b0 100644 --- a/silk/structs.h +++ b/silk/structs.h @@ -48,6 +48,7 @@ typedef struct { opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ]; opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ]; opus_int32 sLF_AR_shp_Q14; + opus_int32 sDiff_shp_Q14; opus_int lagPrev; opus_int sLTP_buf_idx; opus_int sLTP_shp_buf_idx; @@ -86,6 +87,7 @@ typedef struct { const opus_int16 quantStepSize_Q16; const opus_int16 invQuantStepSize_Q6; const opus_uint8 *CB1_NLSF_Q8; + const opus_int16 *CB1_Wght_Q9; const opus_uint8 *CB1_iCDF; const opus_uint8 *pred_Q8; const opus_uint8 *ec_sel; @@ -169,8 +171,6 @@ typedef struct { opus_int pitchEstimationComplexity; /* Complexity level for pitch estimator */ opus_int pitchEstimationLPCOrder; /* Whitening filter order for pitch estimator */ opus_int32 pitchEstimationThreshold_Q16; /* Threshold for pitch estimator */ - opus_int LTPQuantLowComplexity; /* Flag for low complexity LTP quantization */ - opus_int mu_LTP_Q9; /* Rate-distortion tradeoff in LTP quantization */ opus_int32 sum_log_gain_Q7; /* Cumulative max prediction gain */ opus_int NLSF_MSVQ_Survivors; /* Number of survivors in NLSF MSVQ */ opus_int first_frame_after_reset; /* Flag for deactivating NLSF interpolation, pitch prediction */ @@ -301,6 +301,7 @@ typedef struct { /* Stuff used for PLC */ opus_int lossCnt; opus_int prevSignalType; + int arch; silk_PLC_struct sPLC; diff --git a/silk/sum_sqr_shift.c b/silk/sum_sqr_shift.c index 129df191..4fd0c3d7 100644 --- a/silk/sum_sqr_shift.c +++ b/silk/sum_sqr_shift.c @@ -41,43 +41,40 @@ void silk_sum_sqr_shift( ) { opus_int i, shft; - opus_int32 nrg_tmp, nrg; + opus_uint32 nrg_tmp; + opus_int32 nrg; - nrg = 0; - shft = 0; - len--; - for( i = 0; i < len; i += 2 ) { - nrg = silk_SMLABB_ovflw( nrg, x[ i ], x[ i ] ); - nrg = silk_SMLABB_ovflw( nrg, x[ i + 1 ], x[ i + 1 ] ); - if( nrg < 0 ) { - /* Scale down */ - nrg = (opus_int32)silk_RSHIFT_uint( (opus_uint32)nrg, 2 ); - shft = 2; - i+=2; - break; - } + /* Do a first run with the maximum shift we could have. */ + shft = 31-silk_CLZ32(len); + /* Let's be conservative with rounding and start with nrg=len. */ + nrg = len; + for( i = 0; i < len - 1; i += 2 ) { + nrg_tmp = silk_SMULBB( x[ i ], x[ i ] ); + nrg_tmp = silk_SMLABB_ovflw( nrg_tmp, x[ i + 1 ], x[ i + 1 ] ); + nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, nrg_tmp, shft ); } - for( ; i < len; i += 2 ) { + if( i < len ) { + /* One sample left to process */ + nrg_tmp = silk_SMULBB( x[ i ], x[ i ] ); + nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, nrg_tmp, shft ); + } + silk_assert( nrg >= 0 ); + /* Make sure the result will fit in a 32-bit signed integer with two bits + of headroom. */ + shft = silk_max_32(0, shft+3 - silk_CLZ32(nrg)); + nrg = 0; + for( i = 0 ; i < len - 1; i += 2 ) { nrg_tmp = silk_SMULBB( x[ i ], x[ i ] ); nrg_tmp = silk_SMLABB_ovflw( nrg_tmp, x[ i + 1 ], x[ i + 1 ] ); - nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, (opus_uint32)nrg_tmp, shft ); - if( nrg < 0 ) { - /* Scale down */ - nrg = (opus_int32)silk_RSHIFT_uint( (opus_uint32)nrg, 2 ); - shft += 2; - } + nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, nrg_tmp, shft ); } - if( i == len ) { + if( i < len ) { /* One sample left to process */ nrg_tmp = silk_SMULBB( x[ i ], x[ i ] ); nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, nrg_tmp, shft ); } - /* Make sure to have at least one extra leading zero (two leading zeros in total) */ - if( nrg & 0xC0000000 ) { - nrg = silk_RSHIFT_uint( (opus_uint32)nrg, 2 ); - shft += 2; - } + silk_assert( nrg >= 0 ); /* Output arguments */ *shift = shft; diff --git a/silk/tables.h b/silk/tables.h index 7fea6fda..8b0380ee 100644 --- a/silk/tables.h +++ b/silk/tables.h @@ -76,10 +76,8 @@ extern const opus_uint8 silk_NLSF_EXT_iCDF[ 7 ]; extern const opus_uint8 silk_LTP_per_index_iCDF[ 3 ]; /* 3 */ extern const opus_uint8 * const silk_LTP_gain_iCDF_ptrs[ NB_LTP_CBKS ]; /* 3 */ extern const opus_uint8 * const silk_LTP_gain_BITS_Q5_ptrs[ NB_LTP_CBKS ]; /* 3 */ -extern const opus_int16 silk_LTP_gain_middle_avg_RD_Q14; extern const opus_int8 * const silk_LTP_vq_ptrs_Q7[ NB_LTP_CBKS ]; /* 168 */ extern const opus_uint8 * const silk_LTP_vq_gain_ptrs_Q7[NB_LTP_CBKS]; - extern const opus_int8 silk_LTP_vq_sizes[ NB_LTP_CBKS ]; /* 3 */ extern const opus_uint8 silk_LTPscale_iCDF[ 3 ]; /* 4 */ diff --git a/silk/tables_LTP.c b/silk/tables_LTP.c index 0e6a0254..5e12c864 100644 --- a/silk/tables_LTP.c +++ b/silk/tables_LTP.c @@ -51,8 +51,6 @@ static const opus_uint8 silk_LTP_gain_iCDF_2[32] = { 24, 20, 16, 12, 9, 5, 2, 0 }; -const opus_int16 silk_LTP_gain_middle_avg_RD_Q14 = 12304; - static const opus_uint8 silk_LTP_gain_BITS_Q5_0[8] = { 15, 131, 138, 138, 155, 155, 173, 173 }; diff --git a/silk/tables_NLSF_CB_NB_MB.c b/silk/tables_NLSF_CB_NB_MB.c index 8c59d207..195d5b95 100644 --- a/silk/tables_NLSF_CB_NB_MB.c +++ b/silk/tables_NLSF_CB_NB_MB.c @@ -74,6 +74,41 @@ static const opus_uint8 silk_NLSF_CB1_NB_MB_Q8[ 320 ] = { 64, 84, 104, 118, 156, 177, 201, 230 }; +static const opus_int16 silk_NLSF_CB1_Wght_Q9[ 320 ] = { + 2897, 2314, 2314, 2314, 2287, 2287, 2314, 2300, 2327, 2287, + 2888, 2580, 2394, 2367, 2314, 2274, 2274, 2274, 2274, 2194, + 2487, 2340, 2340, 2314, 2314, 2314, 2340, 2340, 2367, 2354, + 3216, 2766, 2340, 2340, 2314, 2274, 2221, 2207, 2261, 2194, + 2460, 2474, 2367, 2394, 2394, 2394, 2394, 2367, 2407, 2314, + 3479, 3056, 2127, 2207, 2274, 2274, 2274, 2287, 2314, 2261, + 3282, 3141, 2580, 2394, 2247, 2221, 2207, 2194, 2194, 2114, + 4096, 3845, 2221, 2620, 2620, 2407, 2314, 2394, 2367, 2074, + 3178, 3244, 2367, 2221, 2553, 2434, 2340, 2314, 2167, 2221, + 3338, 3488, 2726, 2194, 2261, 2460, 2354, 2367, 2207, 2101, + 2354, 2420, 2327, 2367, 2394, 2420, 2420, 2420, 2460, 2367, + 3779, 3629, 2434, 2527, 2367, 2274, 2274, 2300, 2207, 2048, + 3254, 3225, 2713, 2846, 2447, 2327, 2300, 2300, 2274, 2127, + 3263, 3300, 2753, 2806, 2447, 2261, 2261, 2247, 2127, 2101, + 2873, 2981, 2633, 2367, 2407, 2354, 2194, 2247, 2247, 2114, + 3225, 3197, 2633, 2580, 2274, 2181, 2247, 2221, 2221, 2141, + 3178, 3310, 2740, 2407, 2274, 2274, 2274, 2287, 2194, 2114, + 3141, 3272, 2460, 2061, 2287, 2500, 2367, 2487, 2434, 2181, + 3507, 3282, 2314, 2700, 2647, 2474, 2367, 2394, 2340, 2127, + 3423, 3535, 3038, 3056, 2300, 1950, 2221, 2274, 2274, 2274, + 3404, 3366, 2087, 2687, 2873, 2354, 2420, 2274, 2474, 2540, + 3760, 3488, 1950, 2660, 2897, 2527, 2394, 2367, 2460, 2261, + 3028, 3272, 2740, 2888, 2740, 2154, 2127, 2287, 2234, 2247, + 3695, 3657, 2025, 1969, 2660, 2700, 2580, 2500, 2327, 2367, + 3207, 3413, 2354, 2074, 2888, 2888, 2340, 2487, 2247, 2167, + 3338, 3366, 2846, 2780, 2327, 2154, 2274, 2287, 2114, 2061, + 2327, 2300, 2181, 2167, 2181, 2367, 2633, 2700, 2700, 2553, + 2407, 2434, 2221, 2261, 2221, 2221, 2340, 2420, 2607, 2700, + 3038, 3244, 2806, 2888, 2474, 2074, 2300, 2314, 2354, 2380, + 2221, 2154, 2127, 2287, 2500, 2793, 2793, 2620, 2580, 2367, + 3676, 3713, 2234, 1838, 2181, 2753, 2726, 2673, 2513, 2207, + 2793, 3160, 2726, 2553, 2846, 2513, 2181, 2394, 2221, 2181 +}; + static const opus_uint8 silk_NLSF_CB1_iCDF_NB_MB[ 64 ] = { 212, 178, 148, 129, 108, 96, 85, 82, 79, 77, 61, 59, 57, 56, 51, 49, @@ -150,6 +185,7 @@ const silk_NLSF_CB_struct silk_NLSF_CB_NB_MB = SILK_FIX_CONST( 0.18, 16 ), SILK_FIX_CONST( 1.0 / 0.18, 6 ), silk_NLSF_CB1_NB_MB_Q8, + silk_NLSF_CB1_Wght_Q9, silk_NLSF_CB1_iCDF_NB_MB, silk_NLSF_PRED_NB_MB_Q8, silk_NLSF_CB2_SELECT_NB_MB, diff --git a/silk/tables_NLSF_CB_WB.c b/silk/tables_NLSF_CB_WB.c index 50af87eb..5cc9f57b 100644 --- a/silk/tables_NLSF_CB_WB.c +++ b/silk/tables_NLSF_CB_WB.c @@ -98,6 +98,41 @@ static const opus_uint8 silk_NLSF_CB1_WB_Q8[ 512 ] = { 110, 119, 129, 141, 175, 198, 218, 237 }; +static const opus_int16 silk_NLSF_CB1_WB_Wght_Q9[ 512 ] = { + 3657, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2963, 2963, 2925, 2846, + 3216, 3085, 2972, 3056, 3056, 3010, 3010, 3010, 2963, 2963, 3010, 2972, 2888, 2846, 2846, 2726, + 3920, 4014, 2981, 3207, 3207, 2934, 3056, 2846, 3122, 3244, 2925, 2846, 2620, 2553, 2780, 2925, + 3516, 3197, 3010, 3103, 3019, 2888, 2925, 2925, 2925, 2925, 2888, 2888, 2888, 2888, 2888, 2753, + 5054, 5054, 2934, 3573, 3385, 3056, 3085, 2793, 3160, 3160, 2972, 2846, 2513, 2540, 2753, 2888, + 4428, 4149, 2700, 2753, 2972, 3010, 2925, 2846, 2981, 3019, 2925, 2925, 2925, 2925, 2888, 2726, + 3620, 3019, 2972, 3056, 3056, 2873, 2806, 3056, 3216, 3047, 2981, 3291, 3291, 2981, 3310, 2991, + 5227, 5014, 2540, 3338, 3526, 3385, 3197, 3094, 3376, 2981, 2700, 2647, 2687, 2793, 2846, 2673, + 5081, 5174, 4615, 4428, 2460, 2897, 3047, 3207, 3169, 2687, 2740, 2888, 2846, 2793, 2846, 2700, + 3122, 2888, 2963, 2925, 2925, 2925, 2925, 2963, 2963, 2963, 2963, 2925, 2925, 2963, 2963, 2963, + 4202, 3207, 2981, 3103, 3010, 2888, 2888, 2925, 2972, 2873, 2916, 3019, 2972, 3010, 3197, 2873, + 3760, 3760, 3244, 3103, 2981, 2888, 2925, 2888, 2972, 2934, 2793, 2793, 2846, 2888, 2888, 2660, + 3854, 4014, 3207, 3122, 3244, 2934, 3047, 2963, 2963, 3085, 2846, 2793, 2793, 2793, 2793, 2580, + 3845, 4080, 3357, 3516, 3094, 2740, 3010, 2934, 3122, 3085, 2846, 2846, 2647, 2647, 2846, 2806, + 5147, 4894, 3225, 3845, 3441, 3169, 2897, 3413, 3451, 2700, 2580, 2673, 2740, 2846, 2806, 2753, + 4109, 3789, 3291, 3160, 2925, 2888, 2888, 2925, 2793, 2740, 2793, 2740, 2793, 2846, 2888, 2806, + 5081, 5054, 3047, 3545, 3244, 3056, 3085, 2944, 3103, 2897, 2740, 2740, 2740, 2846, 2793, 2620, + 4309, 4309, 2860, 2527, 3207, 3376, 3376, 3075, 3075, 3376, 3056, 2846, 2647, 2580, 2726, 2753, + 3056, 2916, 2806, 2888, 2740, 2687, 2897, 3103, 3150, 3150, 3216, 3169, 3056, 3010, 2963, 2846, + 4375, 3882, 2925, 2888, 2846, 2888, 2846, 2846, 2888, 2888, 2888, 2846, 2888, 2925, 2888, 2846, + 2981, 2916, 2916, 2981, 2981, 3056, 3122, 3216, 3150, 3056, 3010, 2972, 2972, 2972, 2925, 2740, + 4229, 4149, 3310, 3347, 2925, 2963, 2888, 2981, 2981, 2846, 2793, 2740, 2846, 2846, 2846, 2793, + 4080, 4014, 3103, 3010, 2925, 2925, 2925, 2888, 2925, 2925, 2846, 2846, 2846, 2793, 2888, 2780, + 4615, 4575, 3169, 3441, 3207, 2981, 2897, 3038, 3122, 2740, 2687, 2687, 2687, 2740, 2793, 2700, + 4149, 4269, 3789, 3657, 2726, 2780, 2888, 2888, 3010, 2972, 2925, 2846, 2687, 2687, 2793, 2888, + 4215, 3554, 2753, 2846, 2846, 2888, 2888, 2888, 2925, 2925, 2888, 2925, 2925, 2925, 2963, 2888, + 5174, 4921, 2261, 3432, 3789, 3479, 3347, 2846, 3310, 3479, 3150, 2897, 2460, 2487, 2753, 2925, + 3451, 3685, 3122, 3197, 3357, 3047, 3207, 3207, 2981, 3216, 3085, 2925, 2925, 2687, 2540, 2434, + 2981, 3010, 2793, 2793, 2740, 2793, 2846, 2972, 3056, 3103, 3150, 3150, 3150, 3103, 3010, 3010, + 2944, 2873, 2687, 2726, 2780, 3010, 3432, 3545, 3357, 3244, 3056, 3010, 2963, 2925, 2888, 2846, + 3019, 2944, 2897, 3010, 3010, 2972, 3019, 3103, 3056, 3056, 3010, 2888, 2846, 2925, 2925, 2888, + 3920, 3967, 3010, 3197, 3357, 3216, 3291, 3291, 3479, 3704, 3441, 2726, 2181, 2460, 2580, 2607 +}; + static const opus_uint8 silk_NLSF_CB1_iCDF_WB[ 64 ] = { 225, 204, 201, 184, 183, 175, 158, 154, 153, 135, 119, 115, 113, 110, 109, 99, @@ -188,6 +223,7 @@ const silk_NLSF_CB_struct silk_NLSF_CB_WB = SILK_FIX_CONST( 0.15, 16 ), SILK_FIX_CONST( 1.0 / 0.15, 6 ), silk_NLSF_CB1_WB_Q8, + silk_NLSF_CB1_WB_Wght_Q9, silk_NLSF_CB1_iCDF_WB, silk_NLSF_PRED_WB_Q8, silk_NLSF_CB2_SELECT_WB, diff --git a/silk/tests/test_unit_LPC_inv_pred_gain.c b/silk/tests/test_unit_LPC_inv_pred_gain.c new file mode 100644 index 00000000..67067cea --- /dev/null +++ b/silk/tests/test_unit_LPC_inv_pred_gain.c @@ -0,0 +1,129 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc., Jean-Marc Valin +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include <stdio.h> +#include <stdlib.h> +#include "celt/stack_alloc.h" +#include "cpu_support.h" +#include "SigProc_FIX.h" + +/* Computes the impulse response of the filter so we + can catch filters that are definitely unstable. Some + unstable filters may be classified as stable, but not + the other way around. */ +int check_stability(opus_int16 *A_Q12, int order) { + int i; + int j; + int sum_a, sum_abs_a; + sum_a = sum_abs_a = 0; + for( j = 0; j < order; j++ ) { + sum_a += A_Q12[ j ]; + sum_abs_a += silk_abs( A_Q12[ j ] ); + } + /* Check DC stability. */ + if( sum_a >= 4096 ) { + return 0; + } + /* If the sum of absolute values is less than 1, the filter + has to be stable. */ + if( sum_abs_a < 4096 ) { + return 1; + } + double y[SILK_MAX_ORDER_LPC] = {0}; + y[0] = 1; + for( i = 0; i < 10000; i++ ) { + double sum = 0; + for( j = 0; j < order; j++ ) { + sum += y[ j ]*A_Q12[ j ]; + } + for( j = order - 1; j > 0; j-- ) { + y[ j ] = y[ j - 1 ]; + } + y[ 0 ] = sum*(1./4096); + /* If impulse response reaches +/- 10000, the filter + is definitely unstable. */ + if( !(y[ 0 ] < 10000 && y[ 0 ] > -10000) ) { + return 0; + } + /* Test every 8 sample for low amplitude. */ + if( ( i & 0x7 ) == 0 ) { + double amp = 0; + for( j = 0; j < order; j++ ) { + amp += fabs(y[j]); + } + if( amp < 0.00001 ) { + return 1; + } + } + } + return 1; +} + +int main(void) { + const int arch = opus_select_arch(); + /* Set to 10000 so all branches in C function are triggered */ + const int loop_num = 10000; + int count = 0; + ALLOC_STACK; + + /* FIXME: Make the seed random (with option to set it explicitly) + so we get wider coverage. */ + srand(0); + + printf("Testing silk_LPC_inverse_pred_gain() optimization ...\n"); + for( count = 0; count < loop_num; count++ ) { + unsigned int i; + opus_int order; + unsigned int shift; + opus_int16 A_Q12[ SILK_MAX_ORDER_LPC ]; + opus_int32 gain; + + for( order = 2; order <= SILK_MAX_ORDER_LPC; order += 2 ) { /* order must be even. */ + for( shift = 0; shift < 16; shift++ ) { /* Different dynamic range. */ + for( i = 0; i < SILK_MAX_ORDER_LPC; i++ ) { + A_Q12[i] = ((opus_int16)rand()) >> shift; + } + gain = silk_LPC_inverse_pred_gain(A_Q12, order, arch); + /* Look for filters that silk_LPC_inverse_pred_gain() thinks are + stable but definitely aren't. */ + if( gain != 0 && !check_stability(A_Q12, order) ) { + fprintf(stderr, "**Loop %4d failed!**\n", count); + return 1; + } + } + } + if( !(count % 500) ) { + printf("Loop %4d passed\n", count); + } + } + printf("silk_LPC_inverse_pred_gain() optimization passed\n"); + return 0; +} diff --git a/silk/tuning_parameters.h b/silk/tuning_parameters.h index 5b8f4042..d70275fd 100644 --- a/silk/tuning_parameters.h +++ b/silk/tuning_parameters.h @@ -53,19 +53,12 @@ extern "C" /* LPC analysis regularization */ #define FIND_LPC_COND_FAC 1e-5f -/* LTP analysis defines */ -#define FIND_LTP_COND_FAC 1e-5f -#define LTP_DAMPING 0.05f -#define LTP_SMOOTHING 0.1f - -/* LTP quantization settings */ -#define MU_LTP_QUANT_NB 0.03f -#define MU_LTP_QUANT_MB 0.025f -#define MU_LTP_QUANT_WB 0.02f - /* Max cumulative LTP gain */ #define MAX_SUM_LOG_GAIN_DB 250.0f +/* LTP analysis defines */ +#define LTP_CORR_INV_MAX 0.03f + /***********************/ /* High pass filtering */ /***********************/ @@ -103,25 +96,16 @@ extern "C" #define SPARSE_SNR_INCR_dB 2.0f /* threshold for sparseness measure above which to use lower quantization offset during unvoiced */ -#define SPARSENESS_THRESHOLD_QNT_OFFSET 0.75f +#define ENERGY_VARIATION_THRESHOLD_QNT_OFFSET 0.6f /* warping control */ #define WARPING_MULTIPLIER 0.015f /* fraction added to first autocorrelation value */ -#define SHAPE_WHITE_NOISE_FRACTION 5e-5f +#define SHAPE_WHITE_NOISE_FRACTION 3e-5f /* noise shaping filter chirp factor */ -#define BANDWIDTH_EXPANSION 0.95f - -/* difference between chirp factors for analysis and synthesis noise shaping filters at low bitrates */ -#define LOW_RATE_BANDWIDTH_EXPANSION_DELTA 0.01f - -/* extra harmonic boosting (signal shaping) at low bitrates */ -#define LOW_RATE_HARMONIC_BOOST 0.1f - -/* extra harmonic boosting (signal shaping) for noisy input signals */ -#define LOW_INPUT_QUALITY_HARMONIC_BOOST 0.1f +#define BANDWIDTH_EXPANSION 0.94f /* harmonic noise shaping */ #define HARMONIC_SHAPING 0.3f diff --git a/silk/x86/NSQ_del_dec_sse.c b/silk/x86/NSQ_del_dec_sse.c index 21d4a8bc..c5212bee 100644 --- a/silk/x86/NSQ_del_dec_sse.c +++ b/silk/x86/NSQ_del_dec_sse.c @@ -107,12 +107,12 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay /* I */ ); void silk_NSQ_del_dec_sse4_1( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -234,7 +234,8 @@ void silk_NSQ_del_dec_sse4_1( psDD = &psDelDec[ Winner_ind ]; last_smple_idx = smpl_buf_idx + decisionDelay; for( i = 0; i < decisionDelay; i++ ) { - last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; + last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY; + if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY; pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) ); @@ -285,7 +286,8 @@ void silk_NSQ_del_dec_sse4_1( last_smple_idx = smpl_buf_idx + decisionDelay; Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 ); for( i = 0; i < decisionDelay; i++ ) { - last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; + last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY; + if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY; pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) ); @@ -299,7 +301,6 @@ void silk_NSQ_del_dec_sse4_1( NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; /* Save quantized speech signal */ - /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[psEncC->ltp_mem_length], psEncC->frame_length * sizeof( opus_int16 ) ) */ silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); RESTORE_STACK; @@ -333,7 +334,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay /* I */ ) { @@ -638,8 +639,9 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1( psSS[ 1 ].xq_Q14 = xq_Q14; } } - *smpl_buf_idx = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK; /* Index to newest samples */ - last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK; /* Index to decisionDelay old samples */ + *smpl_buf_idx = ( *smpl_buf_idx - 1 ) % DECISION_DELAY; + if( *smpl_buf_idx < 0 ) *smpl_buf_idx += DECISION_DELAY; + last_smple_idx = ( *smpl_buf_idx + decisionDelay ) % DECISION_DELAY; /* Find winner */ RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10; diff --git a/silk/x86/NSQ_sse.c b/silk/x86/NSQ_sse.c index bb3c5f19..6a9e6e96 100644 --- a/silk/x86/NSQ_sse.c +++ b/silk/x86/NSQ_sse.c @@ -71,7 +71,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_10_16_sse4_1( ); void silk_NSQ_sse4_1( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -233,7 +233,6 @@ void silk_NSQ_sse4_1( NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; /* Save quantized speech and noise shaping signals */ - /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[ psEncC->ltp_mem_length ], psEncC->frame_length * sizeof( opus_int16 ) ) */ silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); RESTORE_STACK; diff --git a/silk/x86/main_sse.h b/silk/x86/main_sse.h index d8d61310..2f15d448 100644 --- a/silk/x86/main_sse.h +++ b/silk/x86/main_sse.h @@ -34,6 +34,7 @@ # if defined(OPUS_X86_MAY_HAVE_SSE4_1) +#if 0 /* FIXME: SSE disabled until silk_VQ_WMat_EC_sse4_1() gets updated. */ # define OVERRIDE_silk_VQ_WMat_EC void silk_VQ_WMat_EC_sse4_1( @@ -79,11 +80,13 @@ extern void (*const SILK_VQ_WMAT_EC_IMPL[OPUS_ARCHMASK + 1])( mu_Q9, max_gain_Q7, L)) #endif +#endif +#if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */ # define OVERRIDE_silk_NSQ void silk_NSQ_sse4_1( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -110,7 +113,7 @@ void silk_NSQ_sse4_1( #else extern void (*const SILK_NSQ_IMPL[OPUS_ARCHMASK + 1])( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -137,7 +140,7 @@ extern void (*const SILK_NSQ_IMPL[OPUS_ARCHMASK + 1])( # define OVERRIDE_silk_NSQ_del_dec void silk_NSQ_del_dec_sse4_1( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -164,7 +167,7 @@ void silk_NSQ_del_dec_sse4_1( #else extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -187,6 +190,7 @@ extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])( HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14)) #endif +#endif void silk_noise_shape_quantizer( silk_nsq_state *NSQ, /* I/O NSQ state */ @@ -238,39 +242,6 @@ extern opus_int (*const SILK_VAD_GETSA_Q8_IMPL[OPUS_ARCHMASK + 1])( silk_encoder_state *psEnC, const opus_int16 pIn[]); -# define OVERRIDE_silk_warped_LPC_analysis_filter_FIX - -#endif - -void silk_warped_LPC_analysis_filter_FIX_sse4_1( - opus_int32 state[], /* I/O State [order + 1] */ - opus_int32 res_Q2[], /* O Residual signal [length] */ - const opus_int16 coef_Q13[], /* I Coefficients [order] */ - const opus_int16 input[], /* I Input signal [length] */ - const opus_int16 lambda_Q16, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -); - -#if defined(OPUS_X86_PRESUME_SSE4_1) -#define silk_warped_LPC_analysis_filter_FIX(state, res_Q2, coef_Q13, input, lambda_Q16, length, order, arch) \ - ((void)(arch),silk_warped_LPC_analysis_filter_FIX_c(state, res_Q2, coef_Q13, input, lambda_Q16, length, order)) - -#else - -extern void (*const SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[OPUS_ARCHMASK + 1])( - opus_int32 state[], /* I/O State [order + 1] */ - opus_int32 res_Q2[], /* O Residual signal [length] */ - const opus_int16 coef_Q13[], /* I Coefficients [order] */ - const opus_int16 input[], /* I Input signal [length] */ - const opus_int16 lambda_Q16, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -); - -# define silk_warped_LPC_analysis_filter_FIX(state, res_Q2, coef_Q13, input, lambda_Q16, length, order, arch) \ - ((*SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[(arch) & OPUS_ARCHMASK])(state, res_Q2, coef_Q13, input, lambda_Q16, length, order)) - #endif # endif diff --git a/silk/x86/x86_silk_map.c b/silk/x86/x86_silk_map.c index 818841f2..32dcc3ca 100644 --- a/silk/x86/x86_silk_map.c +++ b/silk/x86/x86_silk_map.c @@ -66,8 +66,9 @@ opus_int (*const SILK_VAD_GETSA_Q8_IMPL[ OPUS_ARCHMASK + 1 ] )( MAY_HAVE_SSE4_1( silk_VAD_GetSA_Q8 ) /* avx */ }; +#if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */ void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -89,7 +90,9 @@ void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )( MAY_HAVE_SSE4_1( silk_NSQ ), /* sse4.1 */ MAY_HAVE_SSE4_1( silk_NSQ ) /* avx */ }; +#endif +#if 0 /* FIXME: SSE disabled until silk_VQ_WMat_EC_sse4_1() gets updated. */ void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )( opus_int8 *ind, /* O index of best codebook vector */ opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */ @@ -109,9 +112,11 @@ void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )( MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ), /* sse4.1 */ MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ) /* avx */ }; +#endif +#if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */ void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -133,25 +138,10 @@ void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )( MAY_HAVE_SSE4_1( silk_NSQ_del_dec ), /* sse4.1 */ MAY_HAVE_SSE4_1( silk_NSQ_del_dec ) /* avx */ }; +#endif #if defined(FIXED_POINT) -void (*const SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[ OPUS_ARCHMASK + 1 ] )( - opus_int32 state[], /* I/O State [order + 1] */ - opus_int32 res_Q2[], /* O Residual signal [length] */ - const opus_int16 coef_Q13[], /* I Coefficients [order] */ - const opus_int16 input[], /* I Input signal [length] */ - const opus_int16 lambda_Q16, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -) = { - silk_warped_LPC_analysis_filter_FIX_c, /* non-sse */ - silk_warped_LPC_analysis_filter_FIX_c, - silk_warped_LPC_analysis_filter_FIX_c, - MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ), /* sse4.1 */ - MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ) /* avx */ -}; - void (*const SILK_BURG_MODIFIED_IMPL[ OPUS_ARCHMASK + 1 ] )( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ |