diff options
Diffstat (limited to 'thirdparty/opus/silk/fixed')
29 files changed, 1383 insertions, 685 deletions
diff --git a/thirdparty/opus/silk/fixed/apply_sine_window_FIX.c b/thirdparty/opus/silk/fixed/apply_sine_window_FIX.c index 03e088a6de..4502b7130e 100644 --- a/thirdparty/opus/silk/fixed/apply_sine_window_FIX.c +++ b/thirdparty/opus/silk/fixed/apply_sine_window_FIX.c @@ -57,15 +57,15 @@ void silk_apply_sine_window( opus_int k, f_Q16, c_Q16; opus_int32 S0_Q16, S1_Q16; - celt_assert( win_type == 1 || win_type == 2 ); + silk_assert( win_type == 1 || win_type == 2 ); /* Length must be in a range from 16 to 120 and a multiple of 4 */ - celt_assert( length >= 16 && length <= 120 ); - celt_assert( ( length & 3 ) == 0 ); + silk_assert( length >= 16 && length <= 120 ); + silk_assert( ( length & 3 ) == 0 ); /* Frequency */ k = ( length >> 2 ) - 4; - celt_assert( k >= 0 && k <= 26 ); + silk_assert( k >= 0 && k <= 26 ); f_Q16 = (opus_int)freq_table_Q16[ k ]; /* Factor used for cosine approximation */ diff --git a/thirdparty/opus/silk/fixed/arm/warped_autocorrelation_FIX_arm.h b/thirdparty/opus/silk/fixed/arm/warped_autocorrelation_FIX_arm.h deleted file mode 100644 index 1992e43288..0000000000 --- a/thirdparty/opus/silk/fixed/arm/warped_autocorrelation_FIX_arm.h +++ /dev/null @@ -1,68 +0,0 @@ -/*********************************************************************** -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/thirdparty/opus/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c b/thirdparty/opus/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c deleted file mode 100644 index 00a70cb51f..0000000000 --- a/thirdparty/opus/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c +++ /dev/null @@ -1,260 +0,0 @@ -/*********************************************************************** -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/thirdparty/opus/silk/fixed/burg_modified_FIX.c b/thirdparty/opus/silk/fixed/burg_modified_FIX.c index 274d4b28e1..17d0e0993c 100644 --- a/thirdparty/opus/silk/fixed/burg_modified_FIX.c +++ b/thirdparty/opus/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 3 +#define N_BITS_HEAD_ROOM 2 #define MIN_RSHIFTS -16 #define MAX_RSHIFTS (32 - QA) @@ -65,7 +65,7 @@ void silk_burg_modified_c( opus_int32 xcorr[ SILK_MAX_ORDER_LPC ]; opus_int64 C0_64; - celt_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); + silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); /* Compute autocorrelations, added over subframes */ C0_64 = silk_inner_prod16_aligned_64( x, x, subfr_length*nb_subfr, arch ); diff --git a/thirdparty/opus/silk/fixed/corrMatrix_FIX.c b/thirdparty/opus/silk/fixed/corrMatrix_FIX.c index 1b4a29c232..c1d437c785 100644 --- a/thirdparty/opus/silk/fixed/corrMatrix_FIX.c +++ b/thirdparty/opus/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_ADD_RSHIFT32( inner_prod, silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts ); + inner_prod += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts ); } Xt[ lag ] = inner_prod; /* X[:,lag]'*t */ ptr1--; /* Go to next column of X */ @@ -77,54 +77,61 @@ 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_int32 *nrg, /* O Energy of x vector */ - opus_int *rshifts, /* O Right shifts of correlations and energy */ + opus_int *rshifts, /* I/O Right shifts of correlations */ int arch /* I Run-time architecture */ ) { - opus_int i, j, lag; + opus_int i, j, lag, rshifts_local, head_room_rshifts; opus_int32 energy; const opus_int16 *ptr1, *ptr2; /* Calculate energy to find shift used to fit in 32 bits */ - silk_sum_sqr_shift( nrg, rshifts, x, L + order - 1 ); - energy = *nrg; + 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; /* 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 ); + 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; } /* 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 ) ); - energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr1[ -j ] ), *rshifts ) ); + 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 ) ); 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 > 0 ) { + if( rshifts_local > 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 ); + energy += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts_local ); } /* 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 ) ); - energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr2[ -j ] ), *rshifts ) ); + 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 ) ); matrix_ptr( XX, lag + j, j, order ) = energy; matrix_ptr( XX, j, lag + j, order ) = energy; } @@ -146,5 +153,6 @@ void silk_corrMatrix_FIX( ptr2--;/* Update pointer to first sample of next column (lag) in X */ } } + *rshifts = rshifts_local; } diff --git a/thirdparty/opus/silk/fixed/encode_frame_FIX.c b/thirdparty/opus/silk/fixed/encode_frame_FIX.c index a02bf87dbb..5ef44b03fc 100644 --- a/thirdparty/opus/silk/fixed/encode_frame_FIX.c +++ b/thirdparty/opus/silk/fixed/encode_frame_FIX.c @@ -29,7 +29,6 @@ POSSIBILITY OF SUCH DAMAGE. #include "config.h" #endif -#include <stdlib.h> #include "main_FIX.h" #include "stack_alloc.h" #include "tuning_parameters.h" @@ -38,33 +37,26 @@ 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_int16 x16[], /* I Input signal */ + const opus_int32 xfw_Q3[], /* I Input signal */ opus_int condCoding /* I The type of conditional coding used so far for this frame */ ); void silk_encode_do_VAD_FIX( - silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ - opus_int activity /* I Decision of Opus voice activity detector */ + silk_encoder_state_FIX *psEnc /* I/O Pointer to Silk FIX encoder state */ ) { - const opus_int activity_threshold = SILK_FIX_CONST( SPEECH_ACTIVITY_DTX_THRES, 8 ); - /****************************/ /* Voice Activity Detection */ /****************************/ silk_VAD_GetSA_Q8( &psEnc->sCmn, psEnc->sCmn.inputBuf + 1, psEnc->sCmn.arch ); - /* If Opus VAD is inactive and Silk VAD is active: lower Silk VAD to just under the threshold */ - if( activity == VAD_NO_ACTIVITY && psEnc->sCmn.speech_activity_Q8 >= activity_threshold ) { - psEnc->sCmn.speech_activity_Q8 = activity_threshold - 1; - } /**************************************************/ /* Convert speech activity into VAD and DTX flags */ /**************************************************/ - if( psEnc->sCmn.speech_activity_Q8 < activity_threshold ) { + if( psEnc->sCmn.speech_activity_Q8 < SILK_FIX_CONST( SPEECH_ACTIVITY_DTX_THRES, 8 ) ) { psEnc->sCmn.indices.signalType = TYPE_NO_VOICE_ACTIVITY; psEnc->sCmn.noSpeechCounter++; - if( psEnc->sCmn.noSpeechCounter <= NB_SPEECH_FRAMES_BEFORE_DTX ) { + if( psEnc->sCmn.noSpeechCounter < NB_SPEECH_FRAMES_BEFORE_DTX ) { psEnc->sCmn.inDTX = 0; } else if( psEnc->sCmn.noSpeechCounter > MAX_CONSECUTIVE_DTX + NB_SPEECH_FRAMES_BEFORE_DTX ) { psEnc->sCmn.noSpeechCounter = NB_SPEECH_FRAMES_BEFORE_DTX; @@ -102,9 +94,6 @@ 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 */ @@ -129,6 +118,7 @@ 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; @@ -142,7 +132,7 @@ 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.ltp_mem_length, psEnc->sCmn.arch ); + silk_find_pitch_lags_FIX( psEnc, &sEncCtrl, res_pitch, x_frame, psEnc->sCmn.arch ); /************************/ /* Noise shape analysis */ @@ -152,17 +142,23 @@ opus_int silk_encode_frame_FIX( /***************************************************/ /* Find linear prediction coefficients (LPC + LTP) */ /***************************************************/ - silk_find_pred_coefs_FIX( psEnc, &sEncCtrl, res_pitch_frame, x_frame, condCoding ); + silk_find_pred_coefs_FIX( psEnc, &sEncCtrl, res_pitch, 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, x_frame, condCoding ); + silk_LBRR_encode_FIX( psEnc, &sEncCtrl, xfw_Q3, condCoding ); /* Loop over quantizer and entropy coding to control bitrate */ maxIter = 6; @@ -198,21 +194,17 @@ 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, x_frame, psEnc->sCmn.pulses, - sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR_Q13, sEncCtrl.HarmShapeGain_Q14, + 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, 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, x_frame, psEnc->sCmn.pulses, - sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR_Q13, sEncCtrl.HarmShapeGain_Q14, + 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, 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 */ /****************************************/ @@ -226,33 +218,6 @@ 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; } @@ -262,7 +227,7 @@ opus_int silk_encode_frame_FIX( if( found_lower && ( gainsID == gainsID_lower || nBits > maxBits ) ) { /* Restore output state from earlier iteration that did meet the bitrate budget */ silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) ); - celt_assert( sRangeEnc_copy2.offs <= 1275 ); + silk_assert( sRangeEnc_copy2.offs <= 1275 ); silk_memcpy( psRangeEnc->buf, ec_buf_copy, sRangeEnc_copy2.offs ); silk_memcpy( &psEnc->sCmn.sNSQ, &sNSQ_copy2, sizeof( silk_nsq_state ) ); psEnc->sShape.LastGainIndex = LastGainIndex_copy2; @@ -290,7 +255,7 @@ opus_int silk_encode_frame_FIX( gainsID_lower = gainsID; /* Copy part of the output state */ silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) ); - celt_assert( psRangeEnc->offs <= 1275 ); + silk_assert( psRangeEnc->offs <= 1275 ); silk_memcpy( ec_buf_copy, psRangeEnc->buf, psRangeEnc->offs ); silk_memcpy( &sNSQ_copy2, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) ); LastGainIndex_copy2 = psEnc->sShape.LastGainIndex; @@ -300,35 +265,15 @@ 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 ) { - 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 ); + gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) ); } - + 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 ); @@ -342,13 +287,7 @@ opus_int silk_encode_frame_FIX( } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { - 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 ); + sEncCtrl.Gains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q8 ), 8 ); } /* Quantize gains */ @@ -392,7 +331,7 @@ 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_int16 x16[], /* I Input signal */ + const opus_int32 xfw_Q3[], /* I Input signal */ opus_int condCoding /* I The type of conditional coding used so far for this frame */ ) { @@ -431,14 +370,14 @@ 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, x16, + silk_NSQ_del_dec( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q3, psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], psEncCtrl->PredCoef_Q12[ 0 ], psEncCtrl->LTPCoef_Q14, - psEncCtrl->AR_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, + psEncCtrl->AR2_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, x16, + silk_NSQ( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q3, psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], psEncCtrl->PredCoef_Q12[ 0 ], psEncCtrl->LTPCoef_Q14, - psEncCtrl->AR_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, + psEncCtrl->AR2_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/thirdparty/opus/silk/fixed/find_LPC_FIX.c b/thirdparty/opus/silk/fixed/find_LPC_FIX.c index c762a0f2a2..e11cdc86e6 100644 --- a/thirdparty/opus/silk/fixed/find_LPC_FIX.c +++ b/thirdparty/opus/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, psEncC->arch ); + silk_NLSF2A( a_tmp_Q12, NLSF0_Q15, psEncC->predictLPCOrder ); /* Calculate residual energy with NLSF interpolation */ silk_LPC_analysis_filter( LPC_res, x, a_tmp_Q12, 2 * subfr_length, psEncC->predictLPCOrder, psEncC->arch ); @@ -146,6 +146,6 @@ void silk_find_LPC_FIX( silk_A2NLSF( NLSF_Q15, a_Q16, psEncC->predictLPCOrder ); } - celt_assert( psEncC->indices.NLSFInterpCoef_Q2 == 4 || ( psEncC->useInterpolatedNLSFs && !psEncC->first_frame_after_reset && psEncC->nb_subfr == MAX_NB_SUBFR ) ); + silk_assert( psEncC->indices.NLSFInterpCoef_Q2 == 4 || ( psEncC->useInterpolatedNLSFs && !psEncC->first_frame_after_reset && psEncC->nb_subfr == MAX_NB_SUBFR ) ); RESTORE_STACK; } diff --git a/thirdparty/opus/silk/fixed/find_LTP_FIX.c b/thirdparty/opus/silk/fixed/find_LTP_FIX.c index 62d4afb250..1314a28137 100644 --- a/thirdparty/opus/silk/fixed/find_LTP_FIX.c +++ b/thirdparty/opus/silk/fixed/find_LTP_FIX.c @@ -32,68 +32,214 @@ 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_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 */ + 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 */ 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_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 */ int arch /* I Run-time architecture */ ) { - 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; + 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 ]; for( k = 0; k < nb_subfr; k++ ) { lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 ); - 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_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 ); } - silk_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, xXLTP_Q17_ptr, xX_shifts, arch ); /* xXLTP_Q17_ptr in Q( -xX_shifts ) */ + 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_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 */ - /* 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 + /* Limit temp such that the below scaling never wraps around */ + WLTP_max = 0; for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) { - XXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( XXLTP_Q17_ptr[ i ], temp, 17 ); + WLTP_max = silk_max( WLTP_ptr[ i ], WLTP_max ); } + 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++ ) { - xXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( xXLTP_Q17_ptr[ i ], temp, 17 ); + d_Q14[ k ] += b_Q14_ptr[ i ]; } -#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 ); + 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 ); } + + 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++ ) { - xXLTP_Q17_ptr[ i ] = (opus_int32)( silk_LSHIFT64( (opus_int64)xXLTP_Q17_ptr[ i ], 17 ) / temp ); + delta_b_Q14[ i ] = silk_max_16( b_Q14_ptr[ i ], 1638 ); /* 1638_Q14 = 0.1_Q0 */ + temp32 += delta_b_Q14[ i ]; /* Q14 */ } -#endif -TOC(div) - r_ptr += subfr_length; - XXLTP_Q17_ptr += LTP_ORDER * LTP_ORDER; - xXLTP_Q17_ptr += LTP_ORDER; + 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 ); + } + 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 ) ); } } diff --git a/thirdparty/opus/silk/fixed/find_pitch_lags_FIX.c b/thirdparty/opus/silk/fixed/find_pitch_lags_FIX.c index 6c3379f2bb..b8440a8247 100644 --- a/thirdparty/opus/silk/fixed/find_pitch_lags_FIX.c +++ b/thirdparty/opus/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_ptr; + const opus_int16 *x_buf, *x_buf_ptr; VARDECL( opus_int16, Wsig ); opus_int16 *Wsig_ptr; opus_int32 auto_corr[ MAX_FIND_PITCH_LPC_ORDER + 1 ]; @@ -59,7 +59,9 @@ void silk_find_pitch_lags_FIX( buf_len = psEnc->sCmn.la_pitch + psEnc->sCmn.frame_length + psEnc->sCmn.ltp_mem_length; /* Safety check */ - celt_assert( buf_len >= psEnc->sCmn.pitch_LPC_win_length ); + silk_assert( buf_len >= psEnc->sCmn.pitch_LPC_win_length ); + + x_buf = x - psEnc->sCmn.ltp_mem_length; /*************************************/ /* Estimate LPC AR coefficients */ @@ -70,19 +72,19 @@ void silk_find_pitch_lags_FIX( ALLOC( Wsig, psEnc->sCmn.pitch_LPC_win_length, opus_int16 ); /* First LA_LTP samples */ - x_ptr = x + buf_len - psEnc->sCmn.pitch_LPC_win_length; + x_buf_ptr = x_buf + buf_len - psEnc->sCmn.pitch_LPC_win_length; Wsig_ptr = Wsig; - silk_apply_sine_window( Wsig_ptr, x_ptr, 1, psEnc->sCmn.la_pitch ); + silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 1, psEnc->sCmn.la_pitch ); /* Middle un - windowed samples */ Wsig_ptr += psEnc->sCmn.la_pitch; - 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 ) ); + 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 ) ); /* Last LA_LTP samples */ Wsig_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ); - 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 ); + 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 ); /* Calculate autocorrelation sequence */ silk_autocorr( auto_corr, &scale, Wsig, psEnc->sCmn.pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1, arch ); @@ -110,7 +112,7 @@ void silk_find_pitch_lags_FIX( /*****************************************/ /* LPC analysis filtering */ /*****************************************/ - silk_LPC_analysis_filter( res, x, A_Q12, buf_len, psEnc->sCmn.pitchEstimationLPCOrder, psEnc->sCmn.arch ); + silk_LPC_analysis_filter( res, x_buf, 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/thirdparty/opus/silk/fixed/find_pred_coefs_FIX.c b/thirdparty/opus/silk/fixed/find_pred_coefs_FIX.c index 606d863347..d308e9cf5f 100644 --- a/thirdparty/opus/silk/fixed/find_pred_coefs_FIX.c +++ b/thirdparty/opus/silk/fixed/find_pred_coefs_FIX.c @@ -41,12 +41,13 @@ void silk_find_pred_coefs_FIX( ) { opus_int i; - opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ]; + opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ], Wght_Q15[ 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 min_gain_Q16, minInvGain_Q30; + opus_int32 tmp, min_gain_Q16, minInvGain_Q30; + opus_int LTP_corrs_rshift[ MAX_NB_SUBFR ]; SAVE_STACK; /* weighting for weighted least squares */ @@ -60,11 +61,13 @@ 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 ); - /* Limit inverse */ - invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 100 ); + /* Ensure Wght_Q15 a minimum value 1 */ + invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 363 ); /* 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 ] ); @@ -74,24 +77,24 @@ 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, xXLTP_Q17 ); - VARDECL( opus_int32, XXLTP_Q17 ); + VARDECL( opus_int32, WLTP ); /**********/ /* VOICED */ /**********/ - celt_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); + silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); - ALLOC( xXLTP_Q17, psEnc->sCmn.nb_subfr * LTP_ORDER, opus_int32 ); - ALLOC( XXLTP_Q17, psEnc->sCmn.nb_subfr * LTP_ORDER * LTP_ORDER, opus_int32 ); + ALLOC( WLTP, psEnc->sCmn.nb_subfr * LTP_ORDER * LTP_ORDER, opus_int32 ); /* LTP analysis */ - silk_find_LTP_FIX( XXLTP_Q17, xXLTP_Q17, res_pitch, - psEncCtrl->pitchL, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch ); + 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 ); /* 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, &psEncCtrl->LTPredCodGain_Q7, XXLTP_Q17, xXLTP_Q17, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch ); + &psEnc->sCmn.sum_log_gain_Q7, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr, + psEnc->sCmn.arch); /* Control LTP scaling */ silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding ); diff --git a/thirdparty/opus/silk/fixed/k2a_FIX.c b/thirdparty/opus/silk/fixed/k2a_FIX.c index 549f6eadaa..5fee599bcb 100644 --- a/thirdparty/opus/silk/fixed/k2a_FIX.c +++ b/thirdparty/opus/silk/fixed/k2a_FIX.c @@ -39,15 +39,14 @@ void silk_k2a( ) { opus_int k, n; - opus_int32 rc, tmp1, tmp2; + opus_int32 Atmp[ SILK_MAX_ORDER_LPC ]; for( k = 0; k < order; 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 ); + 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 ] ); } A_Q24[ k ] = -silk_LSHIFT( (opus_int32)rc_Q15[ k ], 9 ); } diff --git a/thirdparty/opus/silk/fixed/k2a_Q16_FIX.c b/thirdparty/opus/silk/fixed/k2a_Q16_FIX.c index 1595aa6212..3b03987544 100644 --- a/thirdparty/opus/silk/fixed/k2a_Q16_FIX.c +++ b/thirdparty/opus/silk/fixed/k2a_Q16_FIX.c @@ -39,16 +39,15 @@ void silk_k2a_Q16( ) { opus_int k, n; - opus_int32 rc, tmp1, tmp2; + opus_int32 Atmp[ SILK_MAX_ORDER_LPC ]; for( k = 0; k < order; k++ ) { - 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++ ) { + Atmp[ n ] = A_Q24[ n ]; } - A_Q24[ k ] = -silk_LSHIFT( rc, 8 ); + 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 ); } } diff --git a/thirdparty/opus/silk/fixed/main_FIX.h b/thirdparty/opus/silk/fixed/main_FIX.h index 6d2112e511..375b5eb32e 100644 --- a/thirdparty/opus/silk/fixed/main_FIX.h +++ b/thirdparty/opus/silk/fixed/main_FIX.h @@ -36,11 +36,6 @@ 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" @@ -52,9 +47,6 @@ 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 */ /*********************/ @@ -66,8 +58,7 @@ void silk_HP_variable_cutoff( /* Encoder main function */ void silk_encode_do_VAD_FIX( - silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ - opus_int activity /* I Decision of Opus voice activity detector */ + silk_encoder_state_FIX *psEnc /* I/O Pointer to Silk FIX encoder state */ ); /* Encoder main function */ @@ -90,11 +81,33 @@ 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 */ /**************************/ @@ -108,7 +121,7 @@ void silk_noise_shape_analysis_FIX( ); /* Autocorrelations for a warped frequency axis */ -void silk_warped_autocorrelation_FIX_c( +void silk_warped_autocorrelation_FIX( 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 */ @@ -117,11 +130,6 @@ void silk_warped_autocorrelation_FIX_c( 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 */ @@ -160,12 +168,16 @@ void silk_find_LPC_FIX( /* LTP analysis */ void silk_find_LTP_FIX( - 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 */ + 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 */ 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_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 */ int arch /* I Run-time architecture */ ); @@ -219,9 +231,9 @@ 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_int32 *nrg, /* O Energy of x vector */ - opus_int *rshifts, /* O Right shifts of correlations */ + opus_int *rshifts, /* I/O Right shifts of correlations */ int arch /* I Run-time architecture */ ); @@ -236,6 +248,22 @@ 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/thirdparty/opus/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h b/thirdparty/opus/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h index 3999b5bd09..c30481e437 100644 --- a/thirdparty/opus/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h +++ b/thirdparty/opus/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, arch ); + silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder ); } 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, arch ); - nrg = silk_LPC_inverse_pred_gain_Q24( AR1_Q24, psEnc->sCmn.shapingLPCOrder, arch ); + 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 ); diff --git a/thirdparty/opus/silk/fixed/mips/prefilter_FIX_mipsr1.h b/thirdparty/opus/silk/fixed/mips/prefilter_FIX_mipsr1.h new file mode 100644 index 0000000000..21b256885f --- /dev/null +++ b/thirdparty/opus/silk/fixed/mips/prefilter_FIX_mipsr1.h @@ -0,0 +1,184 @@ +/*********************************************************************** +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. +***********************************************************************/ +#ifndef __PREFILTER_FIX_MIPSR1_H__ +#define __PREFILTER_FIX_MIPSR1_H__ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "main_FIX.h" +#include "stack_alloc.h" +#include "tuning_parameters.h" + +#define OVERRIDE_silk_warped_LPC_analysis_filter_FIX +void silk_warped_LPC_analysis_filter_FIX( + 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) */ + int arch +) +{ + opus_int n, i; + opus_int32 acc_Q11, acc_Q22, tmp1, tmp2, tmp3, tmp4; + opus_int32 state_cur, state_next; + + (void)arch; + + /* Order must be even */ + /* Length must be even */ + + silk_assert( ( order & 1 ) == 0 ); + silk_assert( ( length & 1 ) == 0 ); + + for( n = 0; n < length; n+=2 ) { + /* Output of lowpass section */ + tmp2 = silk_SMLAWB( state[ 0 ], state[ 1 ], lambda_Q16 ); + state_cur = silk_LSHIFT( input[ n ], 14 ); + /* Output of allpass section */ + tmp1 = silk_SMLAWB( state[ 1 ], state[ 2 ] - tmp2, lambda_Q16 ); + state_next = tmp2; + acc_Q11 = silk_RSHIFT( order, 1 ); + acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ 0 ] ); + + + /* Output of lowpass section */ + tmp4 = silk_SMLAWB( state_cur, state_next, lambda_Q16 ); + state[ 0 ] = silk_LSHIFT( input[ n+1 ], 14 ); + /* Output of allpass section */ + tmp3 = silk_SMLAWB( state_next, tmp1 - tmp4, lambda_Q16 ); + state[ 1 ] = tmp4; + acc_Q22 = silk_RSHIFT( order, 1 ); + acc_Q22 = silk_SMLAWB( acc_Q22, tmp4, 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_cur = 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_next = tmp2; + acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ i ] ); + + + /* Output of allpass section */ + tmp4 = silk_SMLAWB( state_cur, state_next - tmp3, lambda_Q16 ); + state[ i ] = tmp3; + acc_Q22 = silk_SMLAWB( acc_Q22, tmp3, coef_Q13[ i - 1 ] ); + /* Output of allpass section */ + tmp3 = silk_SMLAWB( state_next, tmp1 - tmp4, lambda_Q16 ); + state[ i + 1 ] = tmp4; + acc_Q22 = silk_SMLAWB( acc_Q22, tmp4, coef_Q13[ i ] ); + } + 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 ); + + state[ order ] = tmp3; + acc_Q22 = silk_SMLAWB( acc_Q22, tmp3, coef_Q13[ order - 1 ] ); + res_Q2[ n+1 ] = silk_LSHIFT( (opus_int32)input[ n+1 ], 2 ) - silk_RSHIFT_ROUND( acc_Q22, 9 ); + } +} + + + +/* Prefilter for finding Quantizer input signal */ +#define OVERRIDE_silk_prefilt_FIX +static 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; + + if( lag > 0 ) { + for( i = 0; i < length; i++ ) { + /* 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 ); + + 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 ); + } + } + else + { + for( i = 0; i < length; i++ ) { + + 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( sLF_MA_shp_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 /* __PREFILTER_FIX_MIPSR1_H__ */ diff --git a/thirdparty/opus/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h b/thirdparty/opus/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h index 66eb2ed26d..e803ef0fce 100644 --- a/thirdparty/opus/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h +++ b/thirdparty/opus/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h @@ -41,8 +41,8 @@ POSSIBILITY OF SUCH DAMAGE. #define QS 14 /* Autocorrelations for a warped frequency axis */ -#define OVERRIDE_silk_warped_autocorrelation_FIX_c -void silk_warped_autocorrelation_FIX_c( +#define OVERRIDE_silk_warped_autocorrelation_FIX +void silk_warped_autocorrelation_FIX( 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 */ diff --git a/thirdparty/opus/silk/fixed/noise_shape_analysis_FIX.c b/thirdparty/opus/silk/fixed/noise_shape_analysis_FIX.c index 85fea0bf09..22a89f75ae 100644 --- a/thirdparty/opus/silk/fixed/noise_shape_analysis_FIX.c +++ b/thirdparty/opus/silk/fixed/noise_shape_analysis_FIX.c @@ -57,79 +57,88 @@ 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_Q24, + opus_int32 *coefs_syn_Q24, + opus_int32 *coefs_ana_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_Q16; + opus_int32 tmp, maxabs_Q24, chirp_Q16, gain_syn_Q16, gain_ana_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_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); + 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 ); } 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_Q24[ 0 ], lambda_Q16 ); - gain_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_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 ); for( i = 0; i < order; i++ ) { - coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ 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 ] ); } - 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_abs_int32( coefs_Q24[ i ] ); + tmp = silk_max( silk_abs_int32( coefs_syn_Q24[ i ] ), silk_abs_int32( coefs_ana_Q24[ i ] ) ); if( tmp > maxabs_Q24 ) { maxabs_Q24 = tmp; ind = i; } } - /* Use Q20 to avoid any overflow when multiplying by (ind + 1) later. */ - maxabs_Q20 = silk_RSHIFT(maxabs_Q24, 4); - if( maxabs_Q20 <= limit_Q20 ) { + if( maxabs_Q24 <= limit_Q24 ) { /* Coefficients are within range - done */ return; } /* Convert back to true warped coefficients */ for( i = 1; i < order; i++ ) { - coefs_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); + 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 ); } - gain_Q16 = silk_INVERSE32_varQ( gain_Q16, 32 ); + gain_syn_Q16 = silk_INVERSE32_varQ( gain_syn_Q16, 32 ); + gain_ana_Q16 = silk_INVERSE32_varQ( gain_ana_Q16, 32 ); for( i = 0; i < order; i++ ) { - coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ 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 ] ); } /* Apply bandwidth expansion */ chirp_Q16 = SILK_FIX_CONST( 0.99, 16 ) - silk_DIV32_varQ( - 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 ); + 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 ); /* Convert to monic warped coefficients */ lambda_Q16 = -lambda_Q16; for( i = order - 1; i > 0; i-- ) { - coefs_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); + 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 ); } 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_Q24[ 0 ], lambda_Q16 ); - gain_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); + 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 ); for( i = 0; i < order; i++ ) { - coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ 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 ] ); } } silk_assert( 0 ); } -/* Disable MIPS version until it's updated. */ -#if 0 && defined(MIPSr1_ASM) +#if defined(MIPSr1_ASM) #include "mips/noise_shape_analysis_FIX_mipsr1.h" #endif @@ -146,13 +155,14 @@ void silk_noise_shape_analysis_FIX( ) { silk_shape_state_FIX *psShapeSt = &psEnc->sShape; - 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_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_int32 auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ]; opus_int32 refl_coef_Q16[ MAX_SHAPE_LPC_ORDER ]; - opus_int32 AR_Q24[ MAX_SHAPE_LPC_ORDER ]; + opus_int32 AR1_Q24[ MAX_SHAPE_LPC_ORDER ]; + opus_int32 AR2_Q24[ MAX_SHAPE_LPC_ORDER ]; VARDECL( opus_int16, x_windowed ); const opus_int16 *x_ptr, *pitch_res_ptr; SAVE_STACK; @@ -199,14 +209,14 @@ 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; - nSegs = silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; - for( k = 0; k < nSegs; k++ ) { + for( k = 0; k < silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; k++ ) { silk_sum_sqr_shift( &nrg, &scale, pitch_res_ptr, nSamples ); nrg += silk_RSHIFT( nSamples, scale ); /* Q(-scale)*/ @@ -218,12 +228,18 @@ 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( energy_variation_Q7 > SILK_FIX_CONST( ENERGY_VARIATION_THRESHOLD_QNT_OFFSET, 7 ) * (nSegs-1) ) { + if( psEncCtrl->sparseness_Q8 > SILK_FIX_CONST( SPARSENESS_THRESHOLD_QNT_OFFSET, 8 ) ) { 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 ) ); } /*******************************/ @@ -231,8 +247,14 @@ 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 ) ); - BWExp_Q16 = silk_DIV32_varQ( SILK_FIX_CONST( BANDWIDTH_EXPANSION, 16 ), + BWExp1_Q16 = BWExp2_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 */ @@ -262,7 +284,7 @@ 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, arch ); + silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder ); } else { /* Calculate regular auto correlation */ silk_autocorr( auto_corr, &scale, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1, arch ); @@ -277,7 +299,7 @@ void silk_noise_shape_analysis_FIX( silk_assert( nrg >= 0 ); /* Convert reflection coefficients to prediction coefficients */ - silk_k2a_Q16( AR_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder ); + silk_k2a_Q16( AR2_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder ); Qnrg = -scale; /* range: -12...30*/ silk_assert( Qnrg >= -12 ); @@ -296,34 +318,40 @@ void silk_noise_shape_analysis_FIX( if( psEnc->sCmn.warping_Q16 > 0 ) { /* Adjust gain for warping */ - 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 ); + 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; } else { - 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 ); - } + psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( psEncCtrl->Gains_Q16[ k ], gain_mult_Q16 ); } - silk_assert( psEncCtrl->Gains_Q16[ k ] > 0 ); } - /* Bandwidth expansion */ - silk_bwexpander_32( AR_Q24, psEnc->sCmn.shapingLPCOrder, BWExp_Q16 ); + /* Bandwidth expansion for synthesis filter shaping */ + silk_bwexpander_32( AR2_Q24, psEnc->sCmn.shapingLPCOrder, BWExp2_Q16 ); - 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 ); + /* Compute noise shaping filter coefficients */ + silk_memcpy( AR1_Q24, AR2_Q24, psEnc->sCmn.shapingLPCOrder * sizeof( opus_int32 ) ); - /* 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 ); + /* 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 ); + + /* 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 ); + + /* 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 ) ); } } @@ -340,6 +368,11 @@ 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 */ @@ -377,6 +410,14 @@ 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 ), @@ -394,11 +435,14 @@ 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/thirdparty/opus/silk/fixed/pitch_analysis_core_FIX.c b/thirdparty/opus/silk/fixed/pitch_analysis_core_FIX.c index 14729046d2..01bb9fc0a8 100644 --- a/thirdparty/opus/silk/fixed/pitch_analysis_core_FIX.c +++ b/thirdparty/opus/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_unscaled, /* I Signal of length PE_FRAME_LENGTH_MS*Fs_kHz */ + const opus_int16 *frame, /* 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 */ @@ -94,17 +94,16 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 int arch /* I Run-time architecture */ ) { - VARDECL( opus_int16, frame_8kHz_buf ); + VARDECL( opus_int16, frame_8kHz ); VARDECL( opus_int16, frame_4kHz ); - VARDECL( opus_int16, frame_scaled ); opus_int32 filt_state[ 6 ]; - const opus_int16 *frame, *frame_8kHz; + const opus_int16 *input_frame_ptr; 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, energy_basis, energy_target; - opus_int d_srch[ PE_D_SRCH_LENGTH ], Cmax, length_d_srch, length_d_comp, shift; + 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; 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; @@ -120,13 +119,12 @@ 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 */ - celt_assert( Fs_kHz == 8 || Fs_kHz == 12 || Fs_kHz == 16 ); + silk_assert( Fs_kHz == 8 || Fs_kHz == 12 || Fs_kHz == 16 ); /* Check for valid complexity setting */ - celt_assert( complexity >= SILK_PE_MIN_COMPLEX ); - celt_assert( complexity <= SILK_PE_MAX_COMPLEX ); + silk_assert( complexity >= SILK_PE_MIN_COMPLEX ); + silk_assert( complexity <= SILK_PE_MAX_COMPLEX ); silk_assert( search_thres1_Q16 >= 0 && search_thres1_Q16 <= (1<<16) ); silk_assert( search_thres2_Q13 >= 0 && search_thres2_Q13 <= (1<<13) ); @@ -139,33 +137,17 @@ 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_buf, frame, frame_length ); - frame_8kHz = frame_8kHz_buf; + silk_resampler_down2( filt_state, frame_8kHz, frame, frame_length ); } else if( Fs_kHz == 12 ) { silk_memset( filt_state, 0, 6 * sizeof( opus_int32 ) ); - silk_resampler_down2_3( filt_state, frame_8kHz_buf, frame, frame_length ); - frame_8kHz = frame_8kHz_buf; + silk_resampler_down2_3( filt_state, frame_8kHz, frame, frame_length ); } else { - celt_assert( Fs_kHz == 8 ); - frame_8kHz = frame; + silk_assert( Fs_kHz == 8 ); + silk_memcpy( frame_8kHz, frame, frame_length_8kHz * sizeof(opus_int16) ); } /* Decimate again to 4 kHz */ @@ -178,6 +160,19 @@ 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 @@ -188,14 +183,14 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 target_ptr = &frame_4kHz[ silk_LSHIFT( SF_LENGTH_4KHZ, 2 ) ]; for( k = 0; k < nb_subfr >> 1; k++ ) { /* Check that we are within range of the array */ - celt_assert( target_ptr >= frame_4kHz ); - celt_assert( target_ptr + SF_LENGTH_8KHZ <= frame_4kHz + frame_length_4kHz ); + silk_assert( target_ptr >= frame_4kHz ); + silk_assert( target_ptr + SF_LENGTH_8KHZ <= frame_4kHz + frame_length_4kHz ); basis_ptr = target_ptr - MIN_LAG_4KHZ; /* Check that we are within range of the array */ - celt_assert( basis_ptr >= frame_4kHz ); - celt_assert( basis_ptr + SF_LENGTH_8KHZ <= frame_4kHz + frame_length_4kHz ); + silk_assert( basis_ptr >= frame_4kHz ); + silk_assert( basis_ptr + SF_LENGTH_8KHZ <= frame_4kHz + frame_length_4kHz ); celt_pitch_xcorr( target_ptr, target_ptr - MAX_LAG_4KHZ, xcorr32, SF_LENGTH_8KHZ, MAX_LAG_4KHZ - MIN_LAG_4KHZ + 1, arch ); @@ -249,7 +244,7 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 /* Sort */ length_d_srch = silk_ADD_LSHIFT32( 4, complexity, 1 ); - celt_assert( 3 * length_d_srch <= PE_D_SRCH_LENGTH ); + silk_assert( 3 * length_d_srch <= PE_D_SRCH_LENGTH ); silk_insertion_sort_decreasing_int16( C, d_srch, CSTRIDE_4KHZ, length_d_srch ); @@ -274,7 +269,7 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 break; } } - celt_assert( length_d_srch > 0 ); + silk_assert( length_d_srch > 0 ); ALLOC( d_comp, D_COMP_STRIDE, opus_int16 ); for( i = D_COMP_MIN; i < D_COMP_MAX; i++ ) { @@ -316,6 +311,18 @@ 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 *********************************************************************************/ @@ -325,8 +332,8 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 for( k = 0; k < nb_subfr; k++ ) { /* Check that we are within range of the array */ - celt_assert( target_ptr >= frame_8kHz ); - celt_assert( target_ptr + SF_LENGTH_8KHZ <= frame_8kHz + frame_length_8kHz ); + silk_assert( target_ptr >= frame_8kHz ); + silk_assert( target_ptr + SF_LENGTH_8KHZ <= frame_8kHz + frame_length_8kHz ); energy_target = silk_ADD32( silk_inner_prod_aligned( target_ptr, target_ptr, SF_LENGTH_8KHZ, arch ), 1 ); for( j = 0; j < length_d_comp; j++ ) { @@ -455,6 +462,24 @@ 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; @@ -494,14 +519,14 @@ 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, 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 ); + 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 ); 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 = &frame[ PE_LTP_MEM_LENGTH_MS * Fs_kHz ]; + target_ptr = &input_frame_ptr[ 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++ ) { @@ -550,7 +575,7 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 *lagIndex = (opus_int16)( lag - MIN_LAG_8KHZ ); *contourIndex = (opus_int8)CBimax; } - celt_assert( *lagIndex >= 0 ); + silk_assert( *lagIndex >= 0 ); /* return as voiced */ RESTORE_STACK; return 0; @@ -587,8 +612,8 @@ static void silk_P_Ana_calc_corr_st3( const opus_int8 *Lag_range_ptr, *Lag_CB_ptr; SAVE_STACK; - celt_assert( complexity >= SILK_PE_MIN_COMPLEX ); - celt_assert( complexity <= SILK_PE_MAX_COMPLEX ); + silk_assert( complexity >= SILK_PE_MIN_COMPLEX ); + silk_assert( complexity <= SILK_PE_MAX_COMPLEX ); if( nb_subfr == PE_MAX_NB_SUBFR ) { Lag_range_ptr = &silk_Lag_range_stage3[ complexity ][ 0 ][ 0 ]; @@ -596,7 +621,7 @@ static void silk_P_Ana_calc_corr_st3( nb_cbk_search = silk_nb_cbk_searchs_stage3[ complexity ]; cbk_size = PE_NB_CBKS_STAGE3_MAX; } else { - celt_assert( nb_subfr == PE_MAX_NB_SUBFR >> 1); + silk_assert( nb_subfr == PE_MAX_NB_SUBFR >> 1); Lag_range_ptr = &silk_Lag_range_stage3_10_ms[ 0 ][ 0 ]; Lag_CB_ptr = &silk_CB_lags_stage3_10_ms[ 0 ][ 0 ]; nb_cbk_search = PE_NB_CBKS_STAGE3_10MS; @@ -612,7 +637,7 @@ static void silk_P_Ana_calc_corr_st3( /* Calculate the correlations for each subframe */ lag_low = matrix_ptr( Lag_range_ptr, k, 0, 2 ); lag_high = matrix_ptr( Lag_range_ptr, k, 1, 2 ); - celt_assert(lag_high-lag_low+1 <= SCRATCH_SIZE); + silk_assert(lag_high-lag_low+1 <= SCRATCH_SIZE); celt_pitch_xcorr( target_ptr, target_ptr - start_lag - lag_high, xcorr32, sf_length, lag_high - lag_low + 1, arch ); for( j = lag_low; j <= lag_high; j++ ) { silk_assert( lag_counter < SCRATCH_SIZE ); @@ -659,8 +684,8 @@ static void silk_P_Ana_calc_energy_st3( const opus_int8 *Lag_range_ptr, *Lag_CB_ptr; SAVE_STACK; - celt_assert( complexity >= SILK_PE_MIN_COMPLEX ); - celt_assert( complexity <= SILK_PE_MAX_COMPLEX ); + silk_assert( complexity >= SILK_PE_MIN_COMPLEX ); + silk_assert( complexity <= SILK_PE_MAX_COMPLEX ); if( nb_subfr == PE_MAX_NB_SUBFR ) { Lag_range_ptr = &silk_Lag_range_stage3[ complexity ][ 0 ][ 0 ]; @@ -668,7 +693,7 @@ static void silk_P_Ana_calc_energy_st3( nb_cbk_search = silk_nb_cbk_searchs_stage3[ complexity ]; cbk_size = PE_NB_CBKS_STAGE3_MAX; } else { - celt_assert( nb_subfr == PE_MAX_NB_SUBFR >> 1); + silk_assert( nb_subfr == PE_MAX_NB_SUBFR >> 1); Lag_range_ptr = &silk_Lag_range_stage3_10_ms[ 0 ][ 0 ]; Lag_CB_ptr = &silk_CB_lags_stage3_10_ms[ 0 ][ 0 ]; nb_cbk_search = PE_NB_CBKS_STAGE3_10MS; diff --git a/thirdparty/opus/silk/fixed/prefilter_FIX.c b/thirdparty/opus/silk/fixed/prefilter_FIX.c new file mode 100644 index 0000000000..6a8e35152e --- /dev/null +++ b/thirdparty/opus/silk/fixed/prefilter_FIX.c @@ -0,0 +1,221 @@ +/*********************************************************************** +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/thirdparty/opus/silk/fixed/residual_energy16_FIX.c b/thirdparty/opus/silk/fixed/residual_energy16_FIX.c index 7f130f3d3d..ebffb2a66f 100644 --- a/thirdparty/opus/silk/fixed/residual_energy16_FIX.c +++ b/thirdparty/opus/silk/fixed/residual_energy16_FIX.c @@ -47,10 +47,10 @@ opus_int32 silk_residual_energy16_covar_FIX( const opus_int32 *pRow; /* Safety checks */ - celt_assert( D >= 0 ); - celt_assert( D <= 16 ); - celt_assert( cQ > 0 ); - celt_assert( cQ < 16 ); + silk_assert( D >= 0 ); + silk_assert( D <= 16 ); + silk_assert( cQ > 0 ); + silk_assert( cQ < 16 ); lshifts = 16 - cQ; Qxtra = lshifts; diff --git a/thirdparty/opus/silk/fixed/residual_energy_FIX.c b/thirdparty/opus/silk/fixed/residual_energy_FIX.c index 6c7cade9a0..41f74778e8 100644 --- a/thirdparty/opus/silk/fixed/residual_energy_FIX.c +++ b/thirdparty/opus/silk/fixed/residual_energy_FIX.c @@ -58,7 +58,7 @@ void silk_residual_energy_FIX( /* Filter input to create the LPC residual for each frame half, and measure subframe energies */ ALLOC( LPC_res, ( MAX_NB_SUBFR >> 1 ) * offset, opus_int16 ); - celt_assert( ( nb_subfr >> 1 ) * ( MAX_NB_SUBFR >> 1 ) == nb_subfr ); + silk_assert( ( nb_subfr >> 1 ) * ( MAX_NB_SUBFR >> 1 ) == nb_subfr ); for( i = 0; i < nb_subfr >> 1; i++ ) { /* Calculate half frame LPC residual signal including preceding samples */ silk_LPC_analysis_filter( LPC_res, x_ptr, a_Q12[ i ], ( MAX_NB_SUBFR >> 1 ) * offset, LPC_order, arch ); diff --git a/thirdparty/opus/silk/fixed/schur64_FIX.c b/thirdparty/opus/silk/fixed/schur64_FIX.c index 4b7e19ea59..764a10ef3e 100644 --- a/thirdparty/opus/silk/fixed/schur64_FIX.c +++ b/thirdparty/opus/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; - celt_assert( order >= 0 && order <= SILK_MAX_ORDER_LPC ); + silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 ); /* Check for invalid input */ if( c[ 0 ] <= 0 ) { @@ -51,10 +51,9 @@ opus_int32 silk_schur64( /* O returns residual ene return 0; } - k = 0; - do { + for( k = 0; k < order + 1; k++ ) { 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/thirdparty/opus/silk/fixed/schur_FIX.c b/thirdparty/opus/silk/fixed/schur_FIX.c index 2840f6b1aa..c4c0ef23b4 100644 --- a/thirdparty/opus/silk/fixed/schur_FIX.c +++ b/thirdparty/opus/silk/fixed/schur_FIX.c @@ -43,29 +43,28 @@ opus_int32 silk_schur( /* O Returns residual ene opus_int32 C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; opus_int32 Ctmp1, Ctmp2, rc_tmp_Q15; - celt_assert( order >= 0 && order <= SILK_MAX_ORDER_LPC ); + silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 ); /* 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 */ - do { + for( k = 0; k < order + 1; k++ ) { C[ k ][ 0 ] = C[ k ][ 1 ] = silk_RSHIFT( c[ k ], 1 ); - } while( ++k <= order ); + } } else if( lz > 2 ) { /* Shift to the left */ lz -= 2; - do { + for( k = 0; k < order + 1; k++ ) { C[ k ][ 0 ] = C[ k ][ 1 ] = silk_LSHIFT( c[ k ], lz ); - } while( ++k <= order ); + } } else { /* No need to shift */ - do { + for( k = 0; k < order + 1; k++ ) { C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ]; - } while( ++k <= order ); + } } for( k = 0; k < order; k++ ) { diff --git a/thirdparty/opus/silk/fixed/solve_LS_FIX.c b/thirdparty/opus/silk/fixed/solve_LS_FIX.c new file mode 100644 index 0000000000..51d7d49d02 --- /dev/null +++ b/thirdparty/opus/silk/fixed/solve_LS_FIX.c @@ -0,0 +1,249 @@ +/*********************************************************************** +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/thirdparty/opus/silk/fixed/structs_FIX.h b/thirdparty/opus/silk/fixed/structs_FIX.h index 2774a97b24..3294b25128 100644 --- a/thirdparty/opus/silk/fixed/structs_FIX.h +++ b/thirdparty/opus/silk/fixed/structs_FIX.h @@ -48,16 +48,30 @@ 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; /************************/ @@ -73,8 +87,11 @@ typedef struct { /* Noise shaping parameters */ /* Testing */ - silk_DWORD_ALIGN opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; + 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 ]; 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; @@ -82,6 +99,7 @@ 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/thirdparty/opus/silk/fixed/warped_autocorrelation_FIX.c b/thirdparty/opus/silk/fixed/warped_autocorrelation_FIX.c index 5c79553bc0..6ca6c1184d 100644 --- a/thirdparty/opus/silk/fixed/warped_autocorrelation_FIX.c +++ b/thirdparty/opus/silk/fixed/warped_autocorrelation_FIX.c @@ -31,14 +31,17 @@ 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 */ -#ifndef OVERRIDE_silk_warped_autocorrelation_FIX_c -void silk_warped_autocorrelation_FIX_c( +void silk_warped_autocorrelation_FIX( 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 */ @@ -53,7 +56,7 @@ void silk_warped_autocorrelation_FIX_c( opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; /* Order must be even */ - celt_assert( ( order & 1 ) == 0 ); + silk_assert( ( order & 1 ) == 0 ); silk_assert( 2 * QS - QC >= 0 ); /* Loop over samples */ @@ -89,4 +92,4 @@ void silk_warped_autocorrelation_FIX_c( } silk_assert( corr_QC[ 0 ] >= 0 ); /* If breaking, decrease QC*/ } -#endif /* OVERRIDE_silk_warped_autocorrelation_FIX_c */ +#endif /* OVERRIDE_silk_warped_autocorrelation_FIX */ diff --git a/thirdparty/opus/silk/fixed/x86/burg_modified_FIX_sse4_1.c b/thirdparty/opus/silk/fixed/x86/burg_modified_FIX_sse.c index bbb1ce0fcc..3c3583c5fc 100644 --- a/thirdparty/opus/silk/fixed/x86/burg_modified_FIX_sse4_1.c +++ b/thirdparty/opus/silk/fixed/x86/burg_modified_FIX_sse.c @@ -72,7 +72,7 @@ void silk_burg_modified_sse4_1( __m128i FIRST_3210, LAST_3210, ATMP_3210, TMP1_3210, TMP2_3210, T1_3210, T2_3210, PTR_3210, SUBFR_3210, X1_3210, X2_3210; __m128i CONST1 = _mm_set1_epi32(1); - celt_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); + silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); /* Compute autocorrelations, added over subframes */ silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length ); diff --git a/thirdparty/opus/silk/fixed/x86/prefilter_FIX_sse.c b/thirdparty/opus/silk/fixed/x86/prefilter_FIX_sse.c new file mode 100644 index 0000000000..488a603f5d --- /dev/null +++ b/thirdparty/opus/silk/fixed/x86/prefilter_FIX_sse.c @@ -0,0 +1,160 @@ +/* Copyright (c) 2014, Cisco Systems, INC + Written by XiangMingZhu WeiZhou MinPeng YanWang + + 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. + + 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 <xmmintrin.h> +#include <emmintrin.h> +#include <smmintrin.h> +#include "main.h" +#include "celt/x86/x86cpu.h" + +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) */ +) +{ + opus_int n, i; + opus_int32 acc_Q11, tmp1, tmp2; + + /* Order must be even */ + silk_assert( ( order & 1 ) == 0 ); + + if (order == 10) + { + if (0 == lambda_Q16) + { + __m128i coef_Q13_3210, coef_Q13_7654; + __m128i coef_Q13_0123, coef_Q13_4567; + __m128i state_0123, state_4567; + __m128i xmm_product1, xmm_product2; + __m128i xmm_tempa, xmm_tempb; + + register opus_int32 sum; + register opus_int32 state_8, state_9, state_a; + register opus_int64 coef_Q13_8, coef_Q13_9; + + silk_assert( length > 0 ); + + coef_Q13_3210 = OP_CVTEPI16_EPI32_M64( &coef_Q13[ 0 ] ); + coef_Q13_7654 = OP_CVTEPI16_EPI32_M64( &coef_Q13[ 4 ] ); + + coef_Q13_0123 = _mm_shuffle_epi32( coef_Q13_3210, _MM_SHUFFLE( 0, 1, 2, 3 ) ); + coef_Q13_4567 = _mm_shuffle_epi32( coef_Q13_7654, _MM_SHUFFLE( 0, 1, 2, 3 ) ); + + coef_Q13_8 = (opus_int64) coef_Q13[ 8 ]; + coef_Q13_9 = (opus_int64) coef_Q13[ 9 ]; + + state_0123 = _mm_loadu_si128( (__m128i *)(&state[ 0 ] ) ); + state_4567 = _mm_loadu_si128( (__m128i *)(&state[ 4 ] ) ); + + state_0123 = _mm_shuffle_epi32( state_0123, _MM_SHUFFLE( 0, 1, 2, 3 ) ); + state_4567 = _mm_shuffle_epi32( state_4567, _MM_SHUFFLE( 0, 1, 2, 3 ) ); + + state_8 = state[ 8 ]; + state_9 = state[ 9 ]; + state_a = 0; + + for( n = 0; n < length; n++ ) + { + xmm_product1 = _mm_mul_epi32( coef_Q13_0123, state_0123 ); /* 64-bit multiply, only 2 pairs */ + xmm_product2 = _mm_mul_epi32( coef_Q13_4567, state_4567 ); + + xmm_tempa = _mm_shuffle_epi32( state_0123, _MM_SHUFFLE( 0, 1, 2, 3 ) ); + xmm_tempb = _mm_shuffle_epi32( state_4567, _MM_SHUFFLE( 0, 1, 2, 3 ) ); + + xmm_product1 = _mm_srli_epi64( xmm_product1, 16 ); /* >> 16, zero extending works */ + xmm_product2 = _mm_srli_epi64( xmm_product2, 16 ); + + xmm_tempa = _mm_mul_epi32( coef_Q13_3210, xmm_tempa ); + xmm_tempb = _mm_mul_epi32( coef_Q13_7654, xmm_tempb ); + + xmm_tempa = _mm_srli_epi64( xmm_tempa, 16 ); + xmm_tempb = _mm_srli_epi64( xmm_tempb, 16 ); + + xmm_tempa = _mm_add_epi32( xmm_tempa, xmm_product1 ); + 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; + + xmm_tempa = _mm_add_epi32( xmm_tempa, _mm_shuffle_epi32( xmm_tempa, _MM_SHUFFLE( 0, 0, 0, 2 ) ) ); + sum += _mm_cvtsi128_si32( xmm_tempa); + res_Q2[ n ] = silk_LSHIFT( (opus_int32)input[ n ], 2 ) - silk_RSHIFT_ROUND( ( 5 + sum ), 9); + + /* move right */ + state_a = state_9; + state_9 = state_8; + state_8 = _mm_cvtsi128_si32( state_4567 ); + state_4567 = _mm_alignr_epi8( state_0123, state_4567, 4 ); + + state_0123 = _mm_alignr_epi8( _mm_cvtsi32_si128( silk_LSHIFT( input[ n ], 14 ) ), state_0123, 4 ); + } + + _mm_storeu_si128( (__m128i *)( &state[ 0 ] ), _mm_shuffle_epi32( state_0123, _MM_SHUFFLE( 0, 1, 2, 3 ) ) ); + _mm_storeu_si128( (__m128i *)( &state[ 4 ] ), _mm_shuffle_epi32( state_4567, _MM_SHUFFLE( 0, 1, 2, 3 ) ) ); + state[ 8 ] = state_8; + state[ 9 ] = state_9; + state[ 10 ] = state_a; + + return; + } + } + + 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 ); + } +} diff --git a/thirdparty/opus/silk/fixed/x86/vector_ops_FIX_sse4_1.c b/thirdparty/opus/silk/fixed/x86/vector_ops_FIX_sse.c index c1e90564d0..c1e90564d0 100644 --- a/thirdparty/opus/silk/fixed/x86/vector_ops_FIX_sse4_1.c +++ b/thirdparty/opus/silk/fixed/x86/vector_ops_FIX_sse.c |