diff options
Diffstat (limited to 'media/libopus/silk/fixed')
31 files changed, 5396 insertions, 0 deletions
diff --git a/media/libopus/silk/fixed/LTP_analysis_filter_FIX.c b/media/libopus/silk/fixed/LTP_analysis_filter_FIX.c new file mode 100644 index 0000000000..5574e7069f --- /dev/null +++ b/media/libopus/silk/fixed/LTP_analysis_filter_FIX.c @@ -0,0 +1,90 @@ +/*********************************************************************** +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" + +void silk_LTP_analysis_filter_FIX( + opus_int16 *LTP_res, /* O LTP residual signal of length MAX_NB_SUBFR * ( pre_length + subfr_length ) */ + const opus_int16 *x, /* I Pointer to input signal with at least max( pitchL ) preceding samples */ + const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],/* I LTP_ORDER LTP coefficients for each MAX_NB_SUBFR subframe */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag, one for each subframe */ + const opus_int32 invGains_Q16[ MAX_NB_SUBFR ], /* I Inverse quantization gains, one for each subframe */ + const opus_int subfr_length, /* I Length of each subframe */ + const opus_int nb_subfr, /* I Number of subframes */ + const opus_int pre_length /* I Length of the preceding samples starting at &x[0] for each subframe */ +) +{ + const opus_int16 *x_ptr, *x_lag_ptr; + opus_int16 Btmp_Q14[ LTP_ORDER ]; + opus_int16 *LTP_res_ptr; + opus_int k, i; + opus_int32 LTP_est; + + x_ptr = x; + LTP_res_ptr = LTP_res; + for( k = 0; k < nb_subfr; k++ ) { + + x_lag_ptr = x_ptr - pitchL[ k ]; + + Btmp_Q14[ 0 ] = LTPCoef_Q14[ k * LTP_ORDER ]; + Btmp_Q14[ 1 ] = LTPCoef_Q14[ k * LTP_ORDER + 1 ]; + Btmp_Q14[ 2 ] = LTPCoef_Q14[ k * LTP_ORDER + 2 ]; + Btmp_Q14[ 3 ] = LTPCoef_Q14[ k * LTP_ORDER + 3 ]; + Btmp_Q14[ 4 ] = LTPCoef_Q14[ k * LTP_ORDER + 4 ]; + + /* LTP analysis FIR filter */ + for( i = 0; i < subfr_length + pre_length; i++ ) { + LTP_res_ptr[ i ] = x_ptr[ i ]; + + /* Long-term prediction */ + LTP_est = silk_SMULBB( x_lag_ptr[ LTP_ORDER / 2 ], Btmp_Q14[ 0 ] ); + LTP_est = silk_SMLABB_ovflw( LTP_est, x_lag_ptr[ 1 ], Btmp_Q14[ 1 ] ); + LTP_est = silk_SMLABB_ovflw( LTP_est, x_lag_ptr[ 0 ], Btmp_Q14[ 2 ] ); + LTP_est = silk_SMLABB_ovflw( LTP_est, x_lag_ptr[ -1 ], Btmp_Q14[ 3 ] ); + LTP_est = silk_SMLABB_ovflw( LTP_est, x_lag_ptr[ -2 ], Btmp_Q14[ 4 ] ); + + LTP_est = silk_RSHIFT_ROUND( LTP_est, 14 ); /* round and -> Q0*/ + + /* Subtract long-term prediction */ + LTP_res_ptr[ i ] = (opus_int16)silk_SAT16( (opus_int32)x_ptr[ i ] - LTP_est ); + + /* Scale residual */ + LTP_res_ptr[ i ] = silk_SMULWB( invGains_Q16[ k ], LTP_res_ptr[ i ] ); + + x_lag_ptr++; + } + + /* Update pointers */ + LTP_res_ptr += subfr_length + pre_length; + x_ptr += subfr_length; + } +} + diff --git a/media/libopus/silk/fixed/LTP_scale_ctrl_FIX.c b/media/libopus/silk/fixed/LTP_scale_ctrl_FIX.c new file mode 100644 index 0000000000..db1016e0b2 --- /dev/null +++ b/media/libopus/silk/fixed/LTP_scale_ctrl_FIX.c @@ -0,0 +1,58 @@ +/*********************************************************************** +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" + +/* Calculation of LTP state scaling */ +void silk_LTP_scale_ctrl_FIX( + silk_encoder_state_FIX *psEnc, /* I/O encoder state */ + silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ + opus_int condCoding /* I The type of conditional coding to use */ +) +{ + opus_int round_loss; + + if( condCoding == CODE_INDEPENDENTLY ) { + /* Only scale if first frame in packet */ + round_loss = psEnc->sCmn.PacketLoss_perc * psEnc->sCmn.nFramesPerPacket; + if ( psEnc->sCmn.LBRR_flag ) { + /* LBRR reduces the effective loss. In practice, it does not square the loss because + losses aren't independent, but that still seems to work best. We also never go below 2%. */ + round_loss = 2 + silk_SMULBB( round_loss, round_loss ) / 100; + } + psEnc->sCmn.indices.LTP_scaleIndex = silk_SMULBB( psEncCtrl->LTPredCodGain_Q7, round_loss ) > silk_log2lin( 128*7 + 2900-psEnc->sCmn.SNR_dB_Q7 ); + psEnc->sCmn.indices.LTP_scaleIndex += silk_SMULBB( psEncCtrl->LTPredCodGain_Q7, round_loss ) > silk_log2lin( 128*7 + 3900-psEnc->sCmn.SNR_dB_Q7 ); + } else { + /* Default is minimum scaling */ + psEnc->sCmn.indices.LTP_scaleIndex = 0; + } + psEncCtrl->LTP_scale_Q14 = silk_LTPScales_table_Q14[ psEnc->sCmn.indices.LTP_scaleIndex ]; +} diff --git a/media/libopus/silk/fixed/apply_sine_window_FIX.c b/media/libopus/silk/fixed/apply_sine_window_FIX.c new file mode 100644 index 0000000000..03e088a6de --- /dev/null +++ b/media/libopus/silk/fixed/apply_sine_window_FIX.c @@ -0,0 +1,101 @@ +/*********************************************************************** +Copyright (c) 2006-2011, Skype Limited. All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "SigProc_FIX.h" + +/* Apply sine window to signal vector. */ +/* Window types: */ +/* 1 -> sine window from 0 to pi/2 */ +/* 2 -> sine window from pi/2 to pi */ +/* Every other sample is linearly interpolated, for speed. */ +/* Window length must be between 16 and 120 (incl) and a multiple of 4. */ + +/* Matlab code for table: + for k=16:9*4:16+2*9*4, fprintf(' %7.d,', -round(65536*pi ./ (k:4:k+8*4))); fprintf('\n'); end +*/ +static const opus_int16 freq_table_Q16[ 27 ] = { + 12111, 9804, 8235, 7100, 6239, 5565, 5022, 4575, 4202, + 3885, 3612, 3375, 3167, 2984, 2820, 2674, 2542, 2422, + 2313, 2214, 2123, 2038, 1961, 1889, 1822, 1760, 1702, +}; + +void silk_apply_sine_window( + opus_int16 px_win[], /* O Pointer to windowed signal */ + const opus_int16 px[], /* I Pointer to input signal */ + const opus_int win_type, /* I Selects a window type */ + const opus_int length /* I Window length, multiple of 4 */ +) +{ + opus_int k, f_Q16, c_Q16; + opus_int32 S0_Q16, S1_Q16; + + celt_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 ); + + /* Frequency */ + k = ( length >> 2 ) - 4; + celt_assert( k >= 0 && k <= 26 ); + f_Q16 = (opus_int)freq_table_Q16[ k ]; + + /* Factor used for cosine approximation */ + c_Q16 = silk_SMULWB( (opus_int32)f_Q16, -f_Q16 ); + silk_assert( c_Q16 >= -32768 ); + + /* initialize state */ + if( win_type == 1 ) { + /* start from 0 */ + S0_Q16 = 0; + /* approximation of sin(f) */ + S1_Q16 = f_Q16 + silk_RSHIFT( length, 3 ); + } else { + /* start from 1 */ + S0_Q16 = ( (opus_int32)1 << 16 ); + /* approximation of cos(f) */ + S1_Q16 = ( (opus_int32)1 << 16 ) + silk_RSHIFT( c_Q16, 1 ) + silk_RSHIFT( length, 4 ); + } + + /* Uses the recursive equation: sin(n*f) = 2 * cos(f) * sin((n-1)*f) - sin((n-2)*f) */ + /* 4 samples at a time */ + for( k = 0; k < length; k += 4 ) { + px_win[ k ] = (opus_int16)silk_SMULWB( silk_RSHIFT( S0_Q16 + S1_Q16, 1 ), px[ k ] ); + px_win[ k + 1 ] = (opus_int16)silk_SMULWB( S1_Q16, px[ k + 1] ); + S0_Q16 = silk_SMULWB( S1_Q16, c_Q16 ) + silk_LSHIFT( S1_Q16, 1 ) - S0_Q16 + 1; + S0_Q16 = silk_min( S0_Q16, ( (opus_int32)1 << 16 ) ); + + px_win[ k + 2 ] = (opus_int16)silk_SMULWB( silk_RSHIFT( S0_Q16 + S1_Q16, 1 ), px[ k + 2] ); + px_win[ k + 3 ] = (opus_int16)silk_SMULWB( S0_Q16, px[ k + 3 ] ); + S1_Q16 = silk_SMULWB( S0_Q16, c_Q16 ) + silk_LSHIFT( S0_Q16, 1 ) - S1_Q16; + S1_Q16 = silk_min( S1_Q16, ( (opus_int32)1 << 16 ) ); + } +} diff --git a/media/libopus/silk/fixed/arm/warped_autocorrelation_FIX_arm.h b/media/libopus/silk/fixed/arm/warped_autocorrelation_FIX_arm.h new file mode 100644 index 0000000000..1992e43288 --- /dev/null +++ b/media/libopus/silk/fixed/arm/warped_autocorrelation_FIX_arm.h @@ -0,0 +1,68 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifndef SILK_WARPED_AUTOCORRELATION_FIX_ARM_H +# define SILK_WARPED_AUTOCORRELATION_FIX_ARM_H + +# include "celt/arm/armcpu.h" + +# if defined(FIXED_POINT) + +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +void silk_warped_autocorrelation_FIX_neon( + opus_int32 *corr, /* O Result [order + 1] */ + opus_int *scale, /* O Scaling of the correlation vector */ + const opus_int16 *input, /* I Input data to correlate */ + const opus_int warping_Q16, /* I Warping coefficient */ + const opus_int length, /* I Length of input */ + const opus_int order /* I Correlation order (even) */ +); + +# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON) +# define OVERRIDE_silk_warped_autocorrelation_FIX (1) +# define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \ + ((void)(arch), PRESUME_NEON(silk_warped_autocorrelation_FIX)(corr, scale, input, warping_Q16, length, order)) +# endif +# endif + +# if !defined(OVERRIDE_silk_warped_autocorrelation_FIX) +/*Is run-time CPU detection enabled on this platform?*/ +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern void (*const SILK_WARPED_AUTOCORRELATION_FIX_IMPL[OPUS_ARCHMASK+1])(opus_int32*, opus_int*, const opus_int16*, const opus_int, const opus_int, const opus_int); +# define OVERRIDE_silk_warped_autocorrelation_FIX (1) +# define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \ + ((*SILK_WARPED_AUTOCORRELATION_FIX_IMPL[(arch)&OPUS_ARCHMASK])(corr, scale, input, warping_Q16, length, order)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_silk_warped_autocorrelation_FIX (1) +# define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \ + ((void)(arch), silk_warped_autocorrelation_FIX_neon(corr, scale, input, warping_Q16, length, order)) +# endif +# endif + +# endif /* end FIXED_POINT */ + +#endif /* end SILK_WARPED_AUTOCORRELATION_FIX_ARM_H */ diff --git a/media/libopus/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c b/media/libopus/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c new file mode 100644 index 0000000000..6f3be025cc --- /dev/null +++ b/media/libopus/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c @@ -0,0 +1,265 @@ +/*********************************************************************** +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 ); + + /* The additional +4 is to ensure a later vld1q_s32 call does not overflow. */ + /* Strictly, only +3 is needed but +4 simplifies initialization using the 4x32 neon load. */ + ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER + 4, 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 += 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 ]; + + /* The additional +4 is to ensure a later vld1q_s32 call does not overflow. */ + ALLOC( state, length + order + 4, 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/media/libopus/silk/fixed/autocorr_FIX.c b/media/libopus/silk/fixed/autocorr_FIX.c new file mode 100644 index 0000000000..de95c98693 --- /dev/null +++ b/media/libopus/silk/fixed/autocorr_FIX.c @@ -0,0 +1,48 @@ +/*********************************************************************** +Copyright (c) 2006-2011, Skype Limited. All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "SigProc_FIX.h" +#include "celt_lpc.h" + +/* Compute autocorrelation */ +void silk_autocorr( + opus_int32 *results, /* O Result (length correlationCount) */ + opus_int *scale, /* O Scaling of the correlation vector */ + const opus_int16 *inputData, /* I Input data to correlate */ + const opus_int inputDataSize, /* I Length of input */ + const opus_int correlationCount, /* I Number of correlation taps to compute */ + int arch /* I Run-time architecture */ +) +{ + opus_int corrCount; + corrCount = silk_min_int( inputDataSize, correlationCount ); + *scale = _celt_autocorr(inputData, results, NULL, 0, corrCount-1, inputDataSize, arch); +} diff --git a/media/libopus/silk/fixed/burg_modified_FIX.c b/media/libopus/silk/fixed/burg_modified_FIX.c new file mode 100644 index 0000000000..185a12b178 --- /dev/null +++ b/media/libopus/silk/fixed/burg_modified_FIX.c @@ -0,0 +1,280 @@ +/*********************************************************************** +Copyright (c) 2006-2011, Skype Limited. All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "SigProc_FIX.h" +#include "define.h" +#include "tuning_parameters.h" +#include "pitch.h" + +#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 MIN_RSHIFTS -16 +#define MAX_RSHIFTS (32 - QA) + +/* Compute reflection coefficients from input signal */ +void silk_burg_modified_c( + opus_int32 *res_nrg, /* O Residual energy */ + opus_int *res_nrg_Q, /* O Residual energy Q value */ + opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ + const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */ + const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ + const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ + const opus_int nb_subfr, /* I Number of subframes stacked in x */ + const opus_int D, /* I Order */ + int arch /* I Run-time architecture */ +) +{ + opus_int k, n, s, lz, rshifts, reached_max_gain; + opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; + const opus_int16 *x_ptr; + opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; + opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; + opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; + opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; + opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; + opus_int32 xcorr[ SILK_MAX_ORDER_LPC ]; + opus_int64 C0_64; + + celt_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); + + /* Compute autocorrelations, added over subframes */ + C0_64 = silk_inner_prod16( x, x, subfr_length*nb_subfr, arch ); + lz = silk_CLZ64(C0_64); + rshifts = 32 + 1 + N_BITS_HEAD_ROOM - lz; + if (rshifts > MAX_RSHIFTS) rshifts = MAX_RSHIFTS; + if (rshifts < MIN_RSHIFTS) rshifts = MIN_RSHIFTS; + + if (rshifts > 0) { + C0 = (opus_int32)silk_RSHIFT64(C0_64, rshifts ); + } else { + C0 = silk_LSHIFT32((opus_int32)C0_64, -rshifts ); + } + + CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ + silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); + if( rshifts > 0 ) { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + for( n = 1; n < D + 1; n++ ) { + C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( + silk_inner_prod16( x_ptr, x_ptr + n, subfr_length - n, arch ), rshifts ); + } + } + } else { + for( s = 0; s < nb_subfr; s++ ) { + int i; + opus_int32 d; + x_ptr = x + s * subfr_length; + celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch ); + for( n = 1; n < D + 1; n++ ) { + for ( i = n + subfr_length - D, d = 0; i < subfr_length; i++ ) + d = MAC16_16( d, x_ptr[ i ], x_ptr[ i - n ] ); + xcorr[ n - 1 ] += d; + } + for( n = 1; n < D + 1; n++ ) { + C_first_row[ n - 1 ] += silk_LSHIFT32( xcorr[ n - 1 ], -rshifts ); + } + } + } + silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); + + /* Initialize */ + CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ + + invGain_Q30 = (opus_int32)1 << 30; + reached_max_gain = 0; + for( n = 0; n < D; n++ ) { + /* Update first row of correlation matrix (without first element) */ + /* Update last row of correlation matrix (without last element, stored in reversed order) */ + /* Update C * Af */ + /* Update C * flipud(Af) (stored in reversed order) */ + if( rshifts > -2 ) { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts) */ + x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts) */ + tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16) */ + tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16) */ + for( k = 0; k < n; k++ ) { + C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ + C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ + Atmp_QA = Af_QA[ k ]; + tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ + tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16) */ + } + tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts) */ + tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts) */ + for( k = 0; k <= n; k++ ) { + CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift ) */ + CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift ) */ + } + } + } else { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts ) */ + x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts ) */ + tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17 */ + tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17 */ + for( k = 0; k < n; k++ ) { + C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ + C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ + Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */ + /* We sometimes get overflows in the multiplications (even beyond +/- 2^32), + but they cancel each other and the real result seems to always fit in a 32-bit + signed integer. This was determined experimentally, not theoretically (unfortunately). */ + tmp1 = silk_MLA_ovflw( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */ + tmp2 = silk_MLA_ovflw( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */ + } + tmp1 = -tmp1; /* Q17 */ + tmp2 = -tmp2; /* Q17 */ + for( k = 0; k <= n; k++ ) { + CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1, + silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift ) */ + CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2, + silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */ + } + } + } + + /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ + tmp1 = C_first_row[ n ]; /* Q( -rshifts ) */ + tmp2 = C_last_row[ n ]; /* Q( -rshifts ) */ + num = 0; /* Q( -rshifts ) */ + nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts ) */ + for( k = 0; k < n; k++ ) { + Atmp_QA = Af_QA[ k ]; + lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1; + lz = silk_min( 32 - QA, lz ); + Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz ) */ + + tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ + tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ + num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ + nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ), + Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts ) */ + } + CAf[ n + 1 ] = tmp1; /* Q( -rshifts ) */ + CAb[ n + 1 ] = tmp2; /* Q( -rshifts ) */ + num = silk_ADD32( num, tmp2 ); /* Q( -rshifts ) */ + num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts ) */ + + /* Calculate the next order reflection (parcor) coefficient */ + if( silk_abs( num ) < nrg ) { + rc_Q31 = silk_DIV32_varQ( num, nrg, 31 ); + } else { + rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN; + } + + /* Update inverse prediction gain */ + tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); + tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 ); + if( tmp1 <= minInvGain_Q30 ) { + /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ + tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */ + rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */ + if( rc_Q31 > 0 ) { + /* Newton-Raphson iteration */ + rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */ + rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */ + if( num < 0 ) { + /* Ensure adjusted reflection coefficients has the original sign */ + rc_Q31 = -rc_Q31; + } + } + invGain_Q30 = minInvGain_Q30; + reached_max_gain = 1; + } else { + invGain_Q30 = tmp1; + } + + /* Update the AR coefficients */ + for( k = 0; k < (n + 1) >> 1; k++ ) { + tmp1 = Af_QA[ k ]; /* QA */ + tmp2 = Af_QA[ n - k - 1 ]; /* QA */ + Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA */ + Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA */ + } + Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA */ + + if( reached_max_gain ) { + /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ + for( k = n + 1; k < D; k++ ) { + Af_QA[ k ] = 0; + } + break; + } + + /* Update C * Af and C * Ab */ + for( k = 0; k <= n + 1; k++ ) { + tmp1 = CAf[ k ]; /* Q( -rshifts ) */ + tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts ) */ + CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts ) */ + CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts ) */ + } + } + + if( reached_max_gain ) { + for( k = 0; k < D; k++ ) { + /* Scale coefficients */ + A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); + } + /* Subtract energy of preceding samples from C0 */ + if( rshifts > 0 ) { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16( x_ptr, x_ptr, D, arch ), rshifts ); + } + } else { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D, arch), -rshifts); + } + } + /* Approximate residual energy */ + *res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 ); + *res_nrg_Q = -rshifts; + } else { + /* Return residual energy */ + nrg = CAf[ 0 ]; /* Q( -rshifts ) */ + tmp1 = (opus_int32)1 << 16; /* Q16 */ + for( k = 0; k < D; k++ ) { + Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16 */ + nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts ) */ + tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16 */ + A_Q16[ k ] = -Atmp1; + } + *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ), -tmp1 );/* Q( -rshifts ) */ + *res_nrg_Q = -rshifts; + } +} diff --git a/media/libopus/silk/fixed/corrMatrix_FIX.c b/media/libopus/silk/fixed/corrMatrix_FIX.c new file mode 100644 index 0000000000..1b4a29c232 --- /dev/null +++ b/media/libopus/silk/fixed/corrMatrix_FIX.c @@ -0,0 +1,150 @@ +/*********************************************************************** +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 + +/********************************************************************** + * Correlation Matrix Computations for LS estimate. + **********************************************************************/ + +#include "main_FIX.h" + +/* Calculates correlation vector X'*t */ +void silk_corrVector_FIX( + const opus_int16 *x, /* I x vector [L + order - 1] used to form data matrix X */ + const opus_int16 *t, /* I Target vector [L] */ + const opus_int L, /* I Length of vectors */ + const opus_int order, /* I Max lag for correlation */ + opus_int32 *Xt, /* O Pointer to X'*t correlation vector [order] */ + const opus_int rshifts, /* I Right shifts of correlations */ + int arch /* I Run-time architecture */ +) +{ + opus_int lag, i; + const opus_int16 *ptr1, *ptr2; + opus_int32 inner_prod; + + ptr1 = &x[ order - 1 ]; /* Points to first sample of column 0 of X: X[:,0] */ + ptr2 = t; + /* Calculate X'*t */ + if( rshifts > 0 ) { + /* Right shifting used */ + 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 ); + } + Xt[ lag ] = inner_prod; /* X[:,lag]'*t */ + ptr1--; /* Go to next column of X */ + } + } else { + silk_assert( rshifts == 0 ); + for( lag = 0; lag < order; lag++ ) { + Xt[ lag ] = silk_inner_prod_aligned( ptr1, ptr2, L, arch ); /* X[:,lag]'*t */ + ptr1--; /* Go to next column of X */ + } + } +} + +/* Calculates correlation matrix X'*X */ +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 */ + 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 */ + int arch /* I Run-time architecture */ +) +{ + opus_int i, j, lag; + opus_int32 energy; + const opus_int16 *ptr1, *ptr2; + + /* Calculate energy to find shift used to fit in 32 bits */ + silk_sum_sqr_shift( nrg, rshifts, x, L + order - 1 ); + energy = *nrg; + + /* Calculate energy of first column (0) of X: X[:,0]'*X[:,0] */ + /* Remove contribution of first order - 1 samples */ + for( i = 0; i < order - 1; i++ ) { + energy -= silk_RSHIFT32( silk_SMULBB( x[ i ], x[ i ] ), *rshifts ); + } + + /* 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 ) ); + 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 ) { + /* 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 ); + } + /* 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 ) ); + matrix_ptr( XX, lag + j, j, order ) = energy; + matrix_ptr( XX, j, lag + j, order ) = energy; + } + ptr2--; /* Update pointer to first sample of next column (lag) in X */ + } + } else { + for( lag = 1; lag < order; lag++ ) { + /* Inner product of column 0 and column lag: X[:,0]'*X[:,lag] */ + energy = silk_inner_prod_aligned( ptr1, ptr2, L, arch ); + matrix_ptr( XX, lag, 0, order ) = energy; + matrix_ptr( XX, 0, lag, order ) = energy; + /* Calculate remaining off diagonal: X[:,j]'*X[:,j + lag] */ + for( j = 1; j < ( order - lag ); j++ ) { + energy = silk_SUB32( energy, silk_SMULBB( ptr1[ L - j ], ptr2[ L - j ] ) ); + energy = silk_SMLABB( energy, ptr1[ -j ], ptr2[ -j ] ); + matrix_ptr( XX, lag + j, j, order ) = energy; + matrix_ptr( XX, j, lag + j, order ) = energy; + } + ptr2--;/* Update pointer to first sample of next column (lag) in X */ + } + } +} + diff --git a/media/libopus/silk/fixed/encode_frame_FIX.c b/media/libopus/silk/fixed/encode_frame_FIX.c new file mode 100644 index 0000000000..a02bf87dbb --- /dev/null +++ b/media/libopus/silk/fixed/encode_frame_FIX.c @@ -0,0 +1,448 @@ +/*********************************************************************** +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 <stdlib.h> +#include "main_FIX.h" +#include "stack_alloc.h" +#include "tuning_parameters.h" + +/* Low Bitrate Redundancy (LBRR) encoding. Reuse all parameters but encode with lower bitrate */ +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 */ + 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 */ +) +{ + 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 ) { + psEnc->sCmn.indices.signalType = TYPE_NO_VOICE_ACTIVITY; + psEnc->sCmn.noSpeechCounter++; + 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; + psEnc->sCmn.inDTX = 0; + } + psEnc->sCmn.VAD_flags[ psEnc->sCmn.nFramesEncoded ] = 0; + } else { + psEnc->sCmn.noSpeechCounter = 0; + psEnc->sCmn.inDTX = 0; + psEnc->sCmn.indices.signalType = TYPE_UNVOICED; + psEnc->sCmn.VAD_flags[ psEnc->sCmn.nFramesEncoded ] = 1; + } +} + +/****************/ +/* Encode frame */ +/****************/ +opus_int silk_encode_frame_FIX( + silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ + opus_int32 *pnBytesOut, /* O Pointer to number of payload bytes; */ + ec_enc *psRangeEnc, /* I/O compressor data structure */ + opus_int condCoding, /* I The type of conditional coding to use */ + opus_int maxBits, /* I If > 0: maximum number of output bits */ + opus_int useCBR /* I Flag to force constant-bitrate operation */ +) +{ + silk_encoder_control_FIX sEncCtrl; + opus_int i, iter, maxIter, found_upper, found_lower, ret = 0; + opus_int16 *x_frame; + ec_enc sRangeEnc_copy, sRangeEnc_copy2; + silk_nsq_state sNSQ_copy, sNSQ_copy2; + opus_int32 seed_copy, nBits, nBits_lower, nBits_upper, gainMult_lower, gainMult_upper; + opus_int32 gainsID, gainsID_lower, gainsID_upper; + opus_int16 gainMult_Q8; + 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 */ + LastGainIndex_copy2 = nBits_lower = nBits_upper = gainMult_lower = gainMult_upper = 0; + + psEnc->sCmn.indices.Seed = psEnc->sCmn.frameCounter++ & 3; + + /**************************************************************/ + /* Set up Input Pointers, and insert frame in input buffer */ + /*************************************************************/ + /* start of frame to encode */ + x_frame = psEnc->x_buf + psEnc->sCmn.ltp_mem_length; + + /***************************************/ + /* Ensure smooth bandwidth transitions */ + /***************************************/ + silk_LP_variable_cutoff( &psEnc->sCmn.sLP, psEnc->sCmn.inputBuf + 1, psEnc->sCmn.frame_length ); + + /*******************************************/ + /* Copy new frame to front of input buffer */ + /*******************************************/ + 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_int16, res_pitch ); + VARDECL( opus_uint8, ec_buf_copy ); + opus_int16 *res_pitch_frame; + + ALLOC( res_pitch, + psEnc->sCmn.la_pitch + psEnc->sCmn.frame_length + + psEnc->sCmn.ltp_mem_length, opus_int16 ); + /* start of pitch LPC residual frame */ + res_pitch_frame = res_pitch + psEnc->sCmn.ltp_mem_length; + + /*****************************************/ + /* 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 ); + + /************************/ + /* Noise shape analysis */ + /************************/ + silk_noise_shape_analysis_FIX( psEnc, &sEncCtrl, res_pitch_frame, x_frame, psEnc->sCmn.arch ); + + /***************************************************/ + /* Find linear prediction coefficients (LPC + LTP) */ + /***************************************************/ + silk_find_pred_coefs_FIX( psEnc, &sEncCtrl, res_pitch_frame, x_frame, condCoding ); + + /****************************************/ + /* Process gains */ + /****************************************/ + silk_process_gains_FIX( psEnc, &sEncCtrl, condCoding ); + + /****************************************/ + /* Low Bitrate Redundant Encoding */ + /****************************************/ + silk_LBRR_encode_FIX( psEnc, &sEncCtrl, x_frame, condCoding ); + + /* Loop over quantizer and entropy coding to control bitrate */ + maxIter = 6; + gainMult_Q8 = SILK_FIX_CONST( 1, 8 ); + found_lower = 0; + found_upper = 0; + gainsID = silk_gains_ID( psEnc->sCmn.indices.GainsIndices, psEnc->sCmn.nb_subfr ); + gainsID_lower = -1; + gainsID_upper = -1; + /* Copy part of the input state */ + silk_memcpy( &sRangeEnc_copy, psRangeEnc, sizeof( ec_enc ) ); + silk_memcpy( &sNSQ_copy, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) ); + seed_copy = psEnc->sCmn.indices.Seed; + ec_prevLagIndex_copy = psEnc->sCmn.ec_prevLagIndex; + ec_prevSignalType_copy = psEnc->sCmn.ec_prevSignalType; + ALLOC( ec_buf_copy, 1275, opus_uint8 ); + for( iter = 0; ; iter++ ) { + if( gainsID == gainsID_lower ) { + nBits = nBits_lower; + } else if( gainsID == gainsID_upper ) { + nBits = nBits_upper; + } else { + /* Restore part of the input state */ + if( iter > 0 ) { + silk_memcpy( psRangeEnc, &sRangeEnc_copy, sizeof( ec_enc ) ); + silk_memcpy( &psEnc->sCmn.sNSQ, &sNSQ_copy, sizeof( silk_nsq_state ) ); + psEnc->sCmn.indices.Seed = seed_copy; + psEnc->sCmn.ec_prevLagIndex = ec_prevLagIndex_copy; + psEnc->sCmn.ec_prevSignalType = ec_prevSignalType_copy; + } + + /*****************************************/ + /* 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, + 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, + 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 */ + /****************************************/ + silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding ); + + /****************************************/ + /* Encode Excitation Signal */ + /****************************************/ + silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType, + psEnc->sCmn.pulses, psEnc->sCmn.frame_length ); + + 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; + } + } + + if( iter == maxIter ) { + 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_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; + } + break; + } + + if( nBits > maxBits ) { + if( found_lower == 0 && iter >= 2 ) { + /* Adjust the quantizer's rate/distortion tradeoff and discard previous "upper" results */ + sEncCtrl.Lambda_Q10 = silk_ADD_RSHIFT32( sEncCtrl.Lambda_Q10, sEncCtrl.Lambda_Q10, 1 ); + found_upper = 0; + gainsID_upper = -1; + } else { + found_upper = 1; + nBits_upper = nBits; + gainMult_upper = gainMult_Q8; + gainsID_upper = gainsID; + } + } else if( nBits < maxBits - 5 ) { + found_lower = 1; + nBits_lower = nBits; + gainMult_lower = gainMult_Q8; + if( gainsID != gainsID_lower ) { + gainsID_lower = gainsID; + /* Copy part of the output state */ + silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) ); + celt_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; + } + } else { + /* Within 5 bits of budget: close enough */ + 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 */ + 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 ); + } + + } 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 ); + /* New gain multplier must be between 25% and 75% of old range (note that gainMult_upper < gainMult_lower) */ + if( gainMult_Q8 > silk_ADD_RSHIFT32( gainMult_lower, gainMult_upper - gainMult_lower, 2 ) ) { + gainMult_Q8 = silk_ADD_RSHIFT32( gainMult_lower, gainMult_upper - gainMult_lower, 2 ); + } else + if( gainMult_Q8 < silk_SUB_RSHIFT32( gainMult_upper, gainMult_upper - gainMult_lower, 2 ) ) { + gainMult_Q8 = silk_SUB_RSHIFT32( gainMult_upper, gainMult_upper - gainMult_lower, 2 ); + } + } + + 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 ); + } + + /* Quantize gains */ + psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev; + silk_gains_quant( psEnc->sCmn.indices.GainsIndices, sEncCtrl.Gains_Q16, + &psEnc->sShape.LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr ); + + /* Unique identifier of gains vector */ + gainsID = silk_gains_ID( psEnc->sCmn.indices.GainsIndices, psEnc->sCmn.nb_subfr ); + } + } + + /* Update input buffer */ + silk_memmove( psEnc->x_buf, &psEnc->x_buf[ psEnc->sCmn.frame_length ], + ( psEnc->sCmn.ltp_mem_length + LA_SHAPE_MS * psEnc->sCmn.fs_kHz ) * sizeof( opus_int16 ) ); + + /* Exit without entropy coding */ + if( psEnc->sCmn.prefillFlag ) { + /* No payload */ + *pnBytesOut = 0; + RESTORE_STACK; + return ret; + } + + /* Parameters needed for next frame */ + psEnc->sCmn.prevLag = sEncCtrl.pitchL[ psEnc->sCmn.nb_subfr - 1 ]; + psEnc->sCmn.prevSignalType = psEnc->sCmn.indices.signalType; + + /****************************************/ + /* Finalize payload */ + /****************************************/ + psEnc->sCmn.first_frame_after_reset = 0; + /* Payload size */ + *pnBytesOut = silk_RSHIFT( ec_tell( psRangeEnc ) + 7, 3 ); + + RESTORE_STACK; + return ret; +} + +/* Low-Bitrate Redundancy (LBRR) encoding. Reuse all parameters but encode excitation at lower bitrate */ +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 */ + opus_int condCoding /* I The type of conditional coding used so far for this frame */ +) +{ + opus_int32 TempGains_Q16[ MAX_NB_SUBFR ]; + SideInfoIndices *psIndices_LBRR = &psEnc->sCmn.indices_LBRR[ psEnc->sCmn.nFramesEncoded ]; + silk_nsq_state sNSQ_LBRR; + + /*******************************************/ + /* Control use of inband LBRR */ + /*******************************************/ + if( psEnc->sCmn.LBRR_enabled && psEnc->sCmn.speech_activity_Q8 > SILK_FIX_CONST( LBRR_SPEECH_ACTIVITY_THRES, 8 ) ) { + psEnc->sCmn.LBRR_flags[ psEnc->sCmn.nFramesEncoded ] = 1; + + /* Copy noise shaping quantizer state and quantization indices from regular encoding */ + silk_memcpy( &sNSQ_LBRR, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) ); + silk_memcpy( psIndices_LBRR, &psEnc->sCmn.indices, sizeof( SideInfoIndices ) ); + + /* Save original gains */ + silk_memcpy( TempGains_Q16, psEncCtrl->Gains_Q16, psEnc->sCmn.nb_subfr * sizeof( opus_int32 ) ); + + if( psEnc->sCmn.nFramesEncoded == 0 || psEnc->sCmn.LBRR_flags[ psEnc->sCmn.nFramesEncoded - 1 ] == 0 ) { + /* First frame in packet or previous frame not LBRR coded */ + psEnc->sCmn.LBRRprevLastGainIndex = psEnc->sShape.LastGainIndex; + + /* Increase Gains to get target LBRR rate */ + psIndices_LBRR->GainsIndices[ 0 ] = psIndices_LBRR->GainsIndices[ 0 ] + psEnc->sCmn.LBRR_GainIncreases; + psIndices_LBRR->GainsIndices[ 0 ] = silk_min_int( psIndices_LBRR->GainsIndices[ 0 ], N_LEVELS_QGAIN - 1 ); + } + + /* Decode to get gains in sync with decoder */ + /* Overwrite unquantized gains with quantized gains */ + silk_gains_dequant( psEncCtrl->Gains_Q16, psIndices_LBRR->GainsIndices, + &psEnc->sCmn.LBRRprevLastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr ); + + /*****************************************/ + /* Noise shaping quantization */ + /*****************************************/ + if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { + silk_NSQ_del_dec( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, x16, + psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], psEncCtrl->PredCoef_Q12[ 0 ], psEncCtrl->LTPCoef_Q14, + psEncCtrl->AR_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, + psEncCtrl->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14, psEnc->sCmn.arch ); + } else { + silk_NSQ( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, x16, + 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->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14, psEnc->sCmn.arch ); + } + + /* Restore original gains */ + silk_memcpy( psEncCtrl->Gains_Q16, TempGains_Q16, psEnc->sCmn.nb_subfr * sizeof( opus_int32 ) ); + } +} diff --git a/media/libopus/silk/fixed/find_LPC_FIX.c b/media/libopus/silk/fixed/find_LPC_FIX.c new file mode 100644 index 0000000000..c762a0f2a2 --- /dev/null +++ b/media/libopus/silk/fixed/find_LPC_FIX.c @@ -0,0 +1,151 @@ +/*********************************************************************** +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" + +/* Finds LPC vector from correlations, and converts to NLSF */ +void silk_find_LPC_FIX( + silk_encoder_state *psEncC, /* I/O Encoder state */ + opus_int16 NLSF_Q15[], /* O NLSFs */ + const opus_int16 x[], /* I Input signal */ + const opus_int32 minInvGain_Q30 /* I Inverse of max prediction gain */ +) +{ + opus_int k, subfr_length; + opus_int32 a_Q16[ MAX_LPC_ORDER ]; + opus_int isInterpLower, shift; + opus_int32 res_nrg0, res_nrg1; + opus_int rshift0, rshift1; + + /* Used only for LSF interpolation */ + opus_int32 a_tmp_Q16[ MAX_LPC_ORDER ], res_nrg_interp, res_nrg, res_tmp_nrg; + opus_int res_nrg_interp_Q, res_nrg_Q, res_tmp_nrg_Q; + opus_int16 a_tmp_Q12[ MAX_LPC_ORDER ]; + opus_int16 NLSF0_Q15[ MAX_LPC_ORDER ]; + SAVE_STACK; + + subfr_length = psEncC->subfr_length + psEncC->predictLPCOrder; + + /* Default: no interpolation */ + psEncC->indices.NLSFInterpCoef_Q2 = 4; + + /* Burg AR analysis for the full frame */ + silk_burg_modified( &res_nrg, &res_nrg_Q, a_Q16, x, minInvGain_Q30, subfr_length, psEncC->nb_subfr, psEncC->predictLPCOrder, psEncC->arch ); + + if( psEncC->useInterpolatedNLSFs && !psEncC->first_frame_after_reset && psEncC->nb_subfr == MAX_NB_SUBFR ) { + VARDECL( opus_int16, LPC_res ); + + /* Optimal solution for last 10 ms */ + silk_burg_modified( &res_tmp_nrg, &res_tmp_nrg_Q, a_tmp_Q16, x + 2 * subfr_length, minInvGain_Q30, subfr_length, 2, psEncC->predictLPCOrder, psEncC->arch ); + + /* subtract residual energy here, as that's easier than adding it to the */ + /* residual energy of the first 10 ms in each iteration of the search below */ + shift = res_tmp_nrg_Q - res_nrg_Q; + if( shift >= 0 ) { + if( shift < 32 ) { + res_nrg = res_nrg - silk_RSHIFT( res_tmp_nrg, shift ); + } + } else { + silk_assert( shift > -32 ); + res_nrg = silk_RSHIFT( res_nrg, -shift ) - res_tmp_nrg; + res_nrg_Q = res_tmp_nrg_Q; + } + + /* Convert to NLSFs */ + silk_A2NLSF( NLSF_Q15, a_tmp_Q16, psEncC->predictLPCOrder ); + + ALLOC( LPC_res, 2 * subfr_length, opus_int16 ); + + /* Search over interpolation indices to find the one with lowest residual energy */ + for( k = 3; k >= 0; k-- ) { + /* Interpolate NLSFs for first half */ + 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 ); + + /* Calculate residual energy with NLSF interpolation */ + silk_LPC_analysis_filter( LPC_res, x, a_tmp_Q12, 2 * subfr_length, psEncC->predictLPCOrder, psEncC->arch ); + + silk_sum_sqr_shift( &res_nrg0, &rshift0, LPC_res + psEncC->predictLPCOrder, subfr_length - psEncC->predictLPCOrder ); + silk_sum_sqr_shift( &res_nrg1, &rshift1, LPC_res + psEncC->predictLPCOrder + subfr_length, subfr_length - psEncC->predictLPCOrder ); + + /* Add subframe energies from first half frame */ + shift = rshift0 - rshift1; + if( shift >= 0 ) { + res_nrg1 = silk_RSHIFT( res_nrg1, shift ); + res_nrg_interp_Q = -rshift0; + } else { + res_nrg0 = silk_RSHIFT( res_nrg0, -shift ); + res_nrg_interp_Q = -rshift1; + } + res_nrg_interp = silk_ADD32( res_nrg0, res_nrg1 ); + + /* Compare with first half energy without NLSF interpolation, or best interpolated value so far */ + shift = res_nrg_interp_Q - res_nrg_Q; + if( shift >= 0 ) { + if( silk_RSHIFT( res_nrg_interp, shift ) < res_nrg ) { + isInterpLower = silk_TRUE; + } else { + isInterpLower = silk_FALSE; + } + } else { + if( -shift < 32 ) { + if( res_nrg_interp < silk_RSHIFT( res_nrg, -shift ) ) { + isInterpLower = silk_TRUE; + } else { + isInterpLower = silk_FALSE; + } + } else { + isInterpLower = silk_FALSE; + } + } + + /* Determine whether current interpolated NLSFs are best so far */ + if( isInterpLower == silk_TRUE ) { + /* Interpolation has lower residual energy */ + res_nrg = res_nrg_interp; + res_nrg_Q = res_nrg_interp_Q; + psEncC->indices.NLSFInterpCoef_Q2 = (opus_int8)k; + } + } + } + + if( psEncC->indices.NLSFInterpCoef_Q2 == 4 ) { + /* NLSF interpolation is currently inactive, calculate NLSFs from full frame AR coefficients */ + 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 ) ); + RESTORE_STACK; +} diff --git a/media/libopus/silk/fixed/find_LTP_FIX.c b/media/libopus/silk/fixed/find_LTP_FIX.c new file mode 100644 index 0000000000..62d4afb250 --- /dev/null +++ b/media/libopus/silk/fixed/find_LTP_FIX.c @@ -0,0 +1,99 @@ +/*********************************************************************** +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 "tuning_parameters.h" + +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 */ + 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 */ + 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; + 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_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, xXLTP_Q17_ptr, xX_shifts, arch ); /* xXLTP_Q17_ptr in Q( -xX_shifts ) */ + + /* At this point all correlations are in Q(-xX_shifts) */ + temp = silk_SMLAWB( 1, nrg, SILK_FIX_CONST( LTP_CORR_INV_MAX, 16 ) ); + temp = silk_max( temp, xx ); +TIC(div) +#if 0 + for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) { + XXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( XXLTP_Q17_ptr[ i ], temp, 17 ); + } + for( i = 0; i < LTP_ORDER; i++ ) { + xXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( xXLTP_Q17_ptr[ i ], temp, 17 ); + } +#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 ); + } + for( i = 0; i < LTP_ORDER; i++ ) { + xXLTP_Q17_ptr[ i ] = (opus_int32)( silk_LSHIFT64( (opus_int64)xXLTP_Q17_ptr[ i ], 17 ) / temp ); + } +#endif +TOC(div) + r_ptr += subfr_length; + XXLTP_Q17_ptr += LTP_ORDER * LTP_ORDER; + xXLTP_Q17_ptr += LTP_ORDER; + } +} diff --git a/media/libopus/silk/fixed/find_pitch_lags_FIX.c b/media/libopus/silk/fixed/find_pitch_lags_FIX.c new file mode 100644 index 0000000000..6c3379f2bb --- /dev/null +++ b/media/libopus/silk/fixed/find_pitch_lags_FIX.c @@ -0,0 +1,143 @@ +/*********************************************************************** +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" + +/* Find pitch lags */ +void silk_find_pitch_lags_FIX( + silk_encoder_state_FIX *psEnc, /* I/O encoder state */ + silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ + opus_int16 res[], /* O residual */ + const opus_int16 x[], /* I Speech signal */ + int arch /* I Run-time architecture */ +) +{ + opus_int buf_len, i, scale; + opus_int32 thrhld_Q13, res_nrg; + const opus_int16 *x_ptr; + VARDECL( opus_int16, Wsig ); + opus_int16 *Wsig_ptr; + opus_int32 auto_corr[ MAX_FIND_PITCH_LPC_ORDER + 1 ]; + opus_int16 rc_Q15[ MAX_FIND_PITCH_LPC_ORDER ]; + opus_int32 A_Q24[ MAX_FIND_PITCH_LPC_ORDER ]; + opus_int16 A_Q12[ MAX_FIND_PITCH_LPC_ORDER ]; + SAVE_STACK; + + /******************************************/ + /* Set up buffer lengths etc based on Fs */ + /******************************************/ + 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 ); + + /*************************************/ + /* Estimate LPC AR coefficients */ + /*************************************/ + + /* Calculate windowed signal */ + + 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; + Wsig_ptr = Wsig; + silk_apply_sine_window( Wsig_ptr, x_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 ) ); + + /* 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 ); + + /* Calculate autocorrelation sequence */ + silk_autocorr( auto_corr, &scale, Wsig, psEnc->sCmn.pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1, arch ); + + /* Add white noise, as fraction of energy */ + auto_corr[ 0 ] = silk_SMLAWB( auto_corr[ 0 ], auto_corr[ 0 ], SILK_FIX_CONST( FIND_PITCH_WHITE_NOISE_FRACTION, 16 ) ) + 1; + + /* Calculate the reflection coefficients using schur */ + res_nrg = silk_schur( rc_Q15, auto_corr, psEnc->sCmn.pitchEstimationLPCOrder ); + + /* Prediction gain */ + psEncCtrl->predGain_Q16 = silk_DIV32_varQ( auto_corr[ 0 ], silk_max_int( res_nrg, 1 ), 16 ); + + /* Convert reflection coefficients to prediction coefficients */ + silk_k2a( A_Q24, rc_Q15, psEnc->sCmn.pitchEstimationLPCOrder ); + + /* Convert From 32 bit Q24 to 16 bit Q12 coefs */ + for( i = 0; i < psEnc->sCmn.pitchEstimationLPCOrder; i++ ) { + A_Q12[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT( A_Q24[ i ], 12 ) ); + } + + /* Do BWE */ + silk_bwexpander( A_Q12, psEnc->sCmn.pitchEstimationLPCOrder, SILK_FIX_CONST( FIND_PITCH_BANDWIDTH_EXPANSION, 16 ) ); + + /*****************************************/ + /* LPC analysis filtering */ + /*****************************************/ + silk_LPC_analysis_filter( res, x, A_Q12, buf_len, psEnc->sCmn.pitchEstimationLPCOrder, psEnc->sCmn.arch ); + + if( psEnc->sCmn.indices.signalType != TYPE_NO_VOICE_ACTIVITY && psEnc->sCmn.first_frame_after_reset == 0 ) { + /* Threshold for pitch estimator */ + thrhld_Q13 = SILK_FIX_CONST( 0.6, 13 ); + thrhld_Q13 = silk_SMLABB( thrhld_Q13, SILK_FIX_CONST( -0.004, 13 ), psEnc->sCmn.pitchEstimationLPCOrder ); + thrhld_Q13 = silk_SMLAWB( thrhld_Q13, SILK_FIX_CONST( -0.1, 21 ), psEnc->sCmn.speech_activity_Q8 ); + thrhld_Q13 = silk_SMLABB( thrhld_Q13, SILK_FIX_CONST( -0.15, 13 ), silk_RSHIFT( psEnc->sCmn.prevSignalType, 1 ) ); + thrhld_Q13 = silk_SMLAWB( thrhld_Q13, SILK_FIX_CONST( -0.1, 14 ), psEnc->sCmn.input_tilt_Q15 ); + thrhld_Q13 = silk_SAT16( thrhld_Q13 ); + + /*****************************************/ + /* Call pitch estimator */ + /*****************************************/ + if( silk_pitch_analysis_core( res, psEncCtrl->pitchL, &psEnc->sCmn.indices.lagIndex, &psEnc->sCmn.indices.contourIndex, + &psEnc->LTPCorr_Q15, psEnc->sCmn.prevLag, psEnc->sCmn.pitchEstimationThreshold_Q16, + (opus_int)thrhld_Q13, psEnc->sCmn.fs_kHz, psEnc->sCmn.pitchEstimationComplexity, psEnc->sCmn.nb_subfr, + psEnc->sCmn.arch) == 0 ) + { + psEnc->sCmn.indices.signalType = TYPE_VOICED; + } else { + psEnc->sCmn.indices.signalType = TYPE_UNVOICED; + } + } else { + silk_memset( psEncCtrl->pitchL, 0, sizeof( psEncCtrl->pitchL ) ); + psEnc->sCmn.indices.lagIndex = 0; + psEnc->sCmn.indices.contourIndex = 0; + psEnc->LTPCorr_Q15 = 0; + } + RESTORE_STACK; +} diff --git a/media/libopus/silk/fixed/find_pred_coefs_FIX.c b/media/libopus/silk/fixed/find_pred_coefs_FIX.c new file mode 100644 index 0000000000..ad363fb7bc --- /dev/null +++ b/media/libopus/silk/fixed/find_pred_coefs_FIX.c @@ -0,0 +1,146 @@ +/*********************************************************************** +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" + +void silk_find_pred_coefs_FIX( + silk_encoder_state_FIX *psEnc, /* I/O encoder state */ + silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ + const opus_int16 res_pitch[], /* I Residual from pitch analysis */ + const opus_int16 x[], /* I Speech signal */ + opus_int condCoding /* I The type of conditional coding to use */ +) +{ + opus_int i; + opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ]; + /* Set to NLSF_Q15 to zero so we don't copy junk to the state. */ + opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]={0}; + const opus_int16 *x_ptr; + opus_int16 *x_pre_ptr; + VARDECL( opus_int16, LPC_in_pre ); + opus_int32 min_gain_Q16, minInvGain_Q30; + SAVE_STACK; + + /* weighting for weighted least squares */ + min_gain_Q16 = silk_int32_MAX >> 6; + for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + min_gain_Q16 = silk_min( min_gain_Q16, psEncCtrl->Gains_Q16[ i ] ); + } + for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + /* Divide to Q16 */ + silk_assert( psEncCtrl->Gains_Q16[ i ] > 0 ); + /* 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 ); + + /* Square the inverted gains */ + silk_assert( invGains_Q16[ i ] == silk_SAT16( invGains_Q16[ i ] ) ); + + /* Invert the inverted and normalized gains */ + local_gains[ i ] = silk_DIV32( ( (opus_int32)1 << 16 ), invGains_Q16[ i ] ); + } + + ALLOC( LPC_in_pre, + 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 ); + + /**********/ + /* VOICED */ + /**********/ + celt_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 ); + + /* 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 ); + + /* 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 ); + + /* Control LTP scaling */ + silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding ); + + /* Create LTP residual */ + silk_LTP_analysis_filter_FIX( LPC_in_pre, x - psEnc->sCmn.predictLPCOrder, psEncCtrl->LTPCoef_Q14, + psEncCtrl->pitchL, invGains_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); + + } else { + /************/ + /* UNVOICED */ + /************/ + /* Create signal with prepended subframes, scaled by inverse gains */ + x_ptr = x - psEnc->sCmn.predictLPCOrder; + x_pre_ptr = LPC_in_pre; + for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + silk_scale_copy_vector16( x_pre_ptr, x_ptr, invGains_Q16[ i ], + psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder ); + x_pre_ptr += psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder; + x_ptr += psEnc->sCmn.subfr_length; + } + + silk_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( opus_int16 ) ); + psEncCtrl->LTPredCodGain_Q7 = 0; + psEnc->sCmn.sum_log_gain_Q7 = 0; + } + + /* Limit on total predictive coding gain */ + if( psEnc->sCmn.first_frame_after_reset ) { + minInvGain_Q30 = SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN_AFTER_RESET, 30 ); + } else { + minInvGain_Q30 = silk_log2lin( silk_SMLAWB( 16 << 7, (opus_int32)psEncCtrl->LTPredCodGain_Q7, SILK_FIX_CONST( 1.0 / 3, 16 ) ) ); /* Q16 */ + minInvGain_Q30 = silk_DIV32_varQ( minInvGain_Q30, + silk_SMULWW( SILK_FIX_CONST( MAX_PREDICTION_POWER_GAIN, 0 ), + silk_SMLAWB( SILK_FIX_CONST( 0.25, 18 ), SILK_FIX_CONST( 0.75, 18 ), psEncCtrl->coding_quality_Q14 ) ), 14 ); + } + + /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ + silk_find_LPC_FIX( &psEnc->sCmn, NLSF_Q15, LPC_in_pre, minInvGain_Q30 ); + + /* Quantize LSFs */ + silk_process_NLSFs( &psEnc->sCmn, psEncCtrl->PredCoef_Q12, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 ); + + /* Calculate residual energy using quantized LPC coefficients */ + silk_residual_energy_FIX( psEncCtrl->ResNrg, psEncCtrl->ResNrgQ, LPC_in_pre, psEncCtrl->PredCoef_Q12, local_gains, + psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder, psEnc->sCmn.arch ); + + /* Copy to prediction struct for use in next frame for interpolation */ + silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); + RESTORE_STACK; +} diff --git a/media/libopus/silk/fixed/k2a_FIX.c b/media/libopus/silk/fixed/k2a_FIX.c new file mode 100644 index 0000000000..549f6eadaa --- /dev/null +++ b/media/libopus/silk/fixed/k2a_FIX.c @@ -0,0 +1,54 @@ +/*********************************************************************** +Copyright (c) 2006-2011, Skype Limited. All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "SigProc_FIX.h" + +/* Step up function, converts reflection coefficients to prediction coefficients */ +void silk_k2a( + opus_int32 *A_Q24, /* O Prediction coefficients [order] Q24 */ + const opus_int16 *rc_Q15, /* I Reflection coefficients [order] Q15 */ + const opus_int32 order /* I Prediction order */ +) +{ + opus_int k, n; + opus_int32 rc, tmp1, tmp2; + + 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 ); + } + A_Q24[ k ] = -silk_LSHIFT( (opus_int32)rc_Q15[ k ], 9 ); + } +} diff --git a/media/libopus/silk/fixed/k2a_Q16_FIX.c b/media/libopus/silk/fixed/k2a_Q16_FIX.c new file mode 100644 index 0000000000..1595aa6212 --- /dev/null +++ b/media/libopus/silk/fixed/k2a_Q16_FIX.c @@ -0,0 +1,54 @@ +/*********************************************************************** +Copyright (c) 2006-2011, Skype Limited. All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "SigProc_FIX.h" + +/* Step up function, converts reflection coefficients to prediction coefficients */ +void silk_k2a_Q16( + opus_int32 *A_Q24, /* O Prediction coefficients [order] Q24 */ + const opus_int32 *rc_Q16, /* I Reflection coefficients [order] Q16 */ + const opus_int32 order /* I Prediction order */ +) +{ + opus_int k, n; + opus_int32 rc, tmp1, tmp2; + + 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 ); + } + A_Q24[ k ] = -silk_LSHIFT( rc, 8 ); + } +} diff --git a/media/libopus/silk/fixed/main_FIX.h b/media/libopus/silk/fixed/main_FIX.h new file mode 100644 index 0000000000..6d2112e511 --- /dev/null +++ b/media/libopus/silk/fixed/main_FIX.h @@ -0,0 +1,244 @@ +/*********************************************************************** +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 SILK_MAIN_FIX_H +#define SILK_MAIN_FIX_H + +#include "SigProc_FIX.h" +#include "structs_FIX.h" +#include "control.h" +#include "main.h" +#include "PLC.h" +#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" +{ +#endif +#endif + +#define silk_encoder_state_Fxx silk_encoder_state_FIX +#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 */ +/*********************/ + +/* High-pass filter with cutoff frequency adaptation based on pitch lag statistics */ +void silk_HP_variable_cutoff( + silk_encoder_state_Fxx state_Fxx[] /* I/O Encoder states */ +); + +/* 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 */ +); + +/* Encoder main function */ +opus_int silk_encode_frame_FIX( + silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ + opus_int32 *pnBytesOut, /* O Pointer to number of payload bytes; */ + ec_enc *psRangeEnc, /* I/O compressor data structure */ + opus_int condCoding, /* I The type of conditional coding to use */ + opus_int maxBits, /* I If > 0: maximum number of output bits */ + opus_int useCBR /* I Flag to force constant-bitrate operation */ +); + +/* Initializes the Silk encoder state */ +opus_int silk_init_encoder( + silk_encoder_state_Fxx *psEnc, /* I/O Pointer to Silk FIX encoder state */ + int arch /* I Run-time architecture */ +); + +/* Control the Silk 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_int allow_bw_switch, /* I Flag to allow switching audio bandwidth */ + const opus_int channelNb, /* I Channel number */ + const opus_int force_fs_kHz +); + +/**************************/ +/* Noise shaping analysis */ +/**************************/ +/* Compute noise shaping coefficients and initial gain values */ +void silk_noise_shape_analysis_FIX( + silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */ + silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control FIX */ + const opus_int16 *pitch_res, /* I LPC residual from pitch analysis */ + const opus_int16 *x, /* I Input signal [ frame_length + la_shape ] */ + int arch /* I Run-time architecture */ +); + +/* Autocorrelations for a warped frequency axis */ +void silk_warped_autocorrelation_FIX_c( + opus_int32 *corr, /* O Result [order + 1] */ + opus_int *scale, /* O Scaling of the correlation vector */ + const opus_int16 *input, /* I Input data to correlate */ + 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(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 */ + silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ + opus_int condCoding /* I The type of conditional coding to use */ +); + +/**********************************************/ +/* Prediction Analysis */ +/**********************************************/ +/* Find pitch lags */ +void silk_find_pitch_lags_FIX( + silk_encoder_state_FIX *psEnc, /* I/O encoder state */ + silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ + opus_int16 res[], /* O residual */ + const opus_int16 x[], /* I Speech signal */ + int arch /* I Run-time architecture */ +); + +/* Find LPC and LTP coefficients */ +void silk_find_pred_coefs_FIX( + silk_encoder_state_FIX *psEnc, /* I/O encoder state */ + silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ + const opus_int16 res_pitch[], /* I Residual from pitch analysis */ + const opus_int16 x[], /* I Speech signal */ + opus_int condCoding /* I The type of conditional coding to use */ +); + +/* LPC analysis */ +void silk_find_LPC_FIX( + silk_encoder_state *psEncC, /* I/O Encoder state */ + opus_int16 NLSF_Q15[], /* O NLSFs */ + const opus_int16 x[], /* I Input signal */ + const opus_int32 minInvGain_Q30 /* I Inverse of max prediction gain */ +); + +/* 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 */ + 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 */ + int arch /* I Run-time architecture */ +); + +void silk_LTP_analysis_filter_FIX( + opus_int16 *LTP_res, /* O LTP residual signal of length MAX_NB_SUBFR * ( pre_length + subfr_length ) */ + const opus_int16 *x, /* I Pointer to input signal with at least max( pitchL ) preceding samples */ + const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],/* I LTP_ORDER LTP coefficients for each MAX_NB_SUBFR subframe */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag, one for each subframe */ + const opus_int32 invGains_Q16[ MAX_NB_SUBFR ], /* I Inverse quantization gains, one for each subframe */ + const opus_int subfr_length, /* I Length of each subframe */ + const opus_int nb_subfr, /* I Number of subframes */ + const opus_int pre_length /* I Length of the preceding samples starting at &x[0] for each subframe */ +); + +/* Calculates residual energies of input subframes where all subframes have LPC_order */ +/* of preceding samples */ +void silk_residual_energy_FIX( + opus_int32 nrgs[ MAX_NB_SUBFR ], /* O Residual energy per subframe */ + opus_int nrgsQ[ MAX_NB_SUBFR ], /* O Q value per subframe */ + const opus_int16 x[], /* I Input signal */ + opus_int16 a_Q12[ 2 ][ MAX_LPC_ORDER ], /* I AR coefs for each frame half */ + const opus_int32 gains[ MAX_NB_SUBFR ], /* I Quantization gains */ + const opus_int subfr_length, /* I Subframe length */ + const opus_int nb_subfr, /* I Number of subframes */ + const opus_int LPC_order, /* I LPC order */ + int arch /* I Run-time architecture */ +); + +/* Residual energy: nrg = wxx - 2 * wXx * c + c' * wXX * c */ +opus_int32 silk_residual_energy16_covar_FIX( + const opus_int16 *c, /* I Prediction vector */ + const opus_int32 *wXX, /* I Correlation matrix */ + const opus_int32 *wXx, /* I Correlation vector */ + opus_int32 wxx, /* I Signal energy */ + opus_int D, /* I Dimension */ + opus_int cQ /* I Q value for c vector 0 - 15 */ +); + +/* Processing of gains */ +void silk_process_gains_FIX( + silk_encoder_state_FIX *psEnc, /* I/O Encoder state */ + silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control */ + opus_int condCoding /* I The type of conditional coding to use */ +); + +/******************/ +/* Linear Algebra */ +/******************/ +/* Calculates correlation matrix X'*X */ +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 */ + 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 */ + int arch /* I Run-time architecture */ +); + +/* Calculates correlation vector X'*t */ +void silk_corrVector_FIX( + const opus_int16 *x, /* I x vector [L + order - 1] used to form data matrix X */ + const opus_int16 *t, /* I Target vector [L] */ + const opus_int L, /* I Length of vectors */ + const opus_int order, /* I Max lag for correlation */ + opus_int32 *Xt, /* O Pointer to X'*t correlation vector [order] */ + const opus_int rshifts, /* I Right shifts of correlations */ + int arch /* I Run-time architecture */ +); + +#ifndef FORCE_CPP_BUILD +#ifdef __cplusplus +} +#endif /* __cplusplus */ +#endif /* FORCE_CPP_BUILD */ +#endif /* SILK_MAIN_FIX_H */ diff --git a/media/libopus/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h b/media/libopus/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h new file mode 100644 index 0000000000..3999b5bd09 --- /dev/null +++ b/media/libopus/silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h @@ -0,0 +1,336 @@ +/*********************************************************************** +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. +***********************************************************************/ + + +/**************************************************************/ +/* Compute noise shaping coefficients and initial gain values */ +/**************************************************************/ +#define OVERRIDE_silk_noise_shape_analysis_FIX + +void silk_noise_shape_analysis_FIX( + silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */ + silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control FIX */ + const opus_int16 *pitch_res, /* I LPC residual from pitch analysis */ + const opus_int16 *x, /* I Input signal [ frame_length + la_shape ] */ + int arch /* I Run-time architecture */ +) +{ + silk_shape_state_FIX *psShapeSt = &psEnc->sShape; + opus_int k, i, nSamples, Qnrg, b_Q14, warping_Q16, scale = 0; + opus_int32 SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32; + opus_int32 nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7; + opus_int32 delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8; + opus_int32 auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ]; + opus_int32 refl_coef_Q16[ MAX_SHAPE_LPC_ORDER ]; + opus_int32 AR1_Q24[ MAX_SHAPE_LPC_ORDER ]; + opus_int32 AR2_Q24[ MAX_SHAPE_LPC_ORDER ]; + VARDECL( opus_int16, x_windowed ); + const opus_int16 *x_ptr, *pitch_res_ptr; + SAVE_STACK; + + /* Point to start of first LPC analysis block */ + x_ptr = x - psEnc->sCmn.la_shape; + + /****************/ + /* GAIN CONTROL */ + /****************/ + SNR_adj_dB_Q7 = psEnc->sCmn.SNR_dB_Q7; + + /* Input quality is the average of the quality in the lowest two VAD bands */ + psEncCtrl->input_quality_Q14 = ( opus_int )silk_RSHIFT( (opus_int32)psEnc->sCmn.input_quality_bands_Q15[ 0 ] + + psEnc->sCmn.input_quality_bands_Q15[ 1 ], 2 ); + + /* Coding quality level, between 0.0_Q0 and 1.0_Q0, but in Q14 */ + psEncCtrl->coding_quality_Q14 = silk_RSHIFT( silk_sigm_Q15( silk_RSHIFT_ROUND( SNR_adj_dB_Q7 - + SILK_FIX_CONST( 20.0, 7 ), 4 ) ), 1 ); + + /* Reduce coding SNR during low speech activity */ + if( psEnc->sCmn.useCBR == 0 ) { + b_Q8 = SILK_FIX_CONST( 1.0, 8 ) - psEnc->sCmn.speech_activity_Q8; + b_Q8 = silk_SMULWB( silk_LSHIFT( b_Q8, 8 ), b_Q8 ); + SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7, + silk_SMULBB( SILK_FIX_CONST( -BG_SNR_DECR_dB, 7 ) >> ( 4 + 1 ), b_Q8 ), /* Q11*/ + silk_SMULWB( SILK_FIX_CONST( 1.0, 14 ) + psEncCtrl->input_quality_Q14, psEncCtrl->coding_quality_Q14 ) ); /* Q12*/ + } + + if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { + /* Reduce gains for periodic signals */ + SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7, SILK_FIX_CONST( HARM_SNR_INCR_dB, 8 ), psEnc->LTPCorr_Q15 ); + } else { + /* For unvoiced signals and low-quality input, adjust the quality slower than SNR_dB setting */ + SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7, + silk_SMLAWB( SILK_FIX_CONST( 6.0, 9 ), -SILK_FIX_CONST( 0.4, 18 ), psEnc->sCmn.SNR_dB_Q7 ), + SILK_FIX_CONST( 1.0, 14 ) - psEncCtrl->input_quality_Q14 ); + } + + /*************************/ + /* SPARSENESS PROCESSING */ + /*************************/ + /* Set quantizer offset */ + if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { + /* Initially set to 0; may be overruled in process_gains(..) */ + psEnc->sCmn.indices.quantOffsetType = 0; + psEncCtrl->sparseness_Q8 = 0; + } else { + /* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */ + nSamples = silk_LSHIFT( psEnc->sCmn.fs_kHz, 1 ); + energy_variation_Q7 = 0; + log_energy_prev_Q7 = 0; + pitch_res_ptr = pitch_res; + for( k = 0; k < silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; k++ ) { + silk_sum_sqr_shift( &nrg, &scale, pitch_res_ptr, nSamples ); + nrg += silk_RSHIFT( nSamples, scale ); /* Q(-scale)*/ + + log_energy_Q7 = silk_lin2log( nrg ); + if( k > 0 ) { + energy_variation_Q7 += silk_abs( log_energy_Q7 - log_energy_prev_Q7 ); + } + log_energy_prev_Q7 = log_energy_Q7; + pitch_res_ptr += nSamples; + } + + psEncCtrl->sparseness_Q8 = silk_RSHIFT( silk_sigm_Q15( silk_SMULWB( energy_variation_Q7 - + SILK_FIX_CONST( 5.0, 7 ), SILK_FIX_CONST( 0.1, 16 ) ) ), 7 ); + + /* Set quantization offset depending on sparseness measure */ + if( psEncCtrl->sparseness_Q8 > SILK_FIX_CONST( SPARSENESS_THRESHOLD_QNT_OFFSET, 8 ) ) { + 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 ) ); + } + + /*******************************/ + /* Control bandwidth expansion */ + /*******************************/ + /* More BWE for signals with high prediction gain */ + strength_Q16 = silk_SMULWB( psEncCtrl->predGain_Q16, SILK_FIX_CONST( FIND_PITCH_WHITE_NOISE_FRACTION, 16 ) ); + BWExp1_Q16 = BWExp2_Q16 = silk_DIV32_varQ( SILK_FIX_CONST( BANDWIDTH_EXPANSION, 16 ), + 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 */ + warping_Q16 = silk_SMLAWB( psEnc->sCmn.warping_Q16, (opus_int32)psEncCtrl->coding_quality_Q14, SILK_FIX_CONST( 0.01, 18 ) ); + } else { + warping_Q16 = 0; + } + + /********************************************/ + /* Compute noise shaping AR coefs and gains */ + /********************************************/ + ALLOC( x_windowed, psEnc->sCmn.shapeWinLength, opus_int16 ); + for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { + /* Apply window: sine slope followed by flat part followed by cosine slope */ + opus_int shift, slope_part, flat_part; + flat_part = psEnc->sCmn.fs_kHz * 3; + slope_part = silk_RSHIFT( psEnc->sCmn.shapeWinLength - flat_part, 1 ); + + silk_apply_sine_window( x_windowed, x_ptr, 1, slope_part ); + shift = slope_part; + silk_memcpy( x_windowed + shift, x_ptr + shift, flat_part * sizeof(opus_int16) ); + shift += flat_part; + silk_apply_sine_window( x_windowed + shift, x_ptr + shift, 2, slope_part ); + + /* Update pointer: next LPC analysis block */ + x_ptr += psEnc->sCmn.subfr_length; + + 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 ); + } else { + /* Calculate regular auto correlation */ + silk_autocorr( auto_corr, &scale, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1, arch ); + } + + /* Add white noise, as a fraction of energy */ + auto_corr[0] = silk_ADD32( auto_corr[0], silk_max_32( silk_SMULWB( silk_RSHIFT( auto_corr[ 0 ], 4 ), + SILK_FIX_CONST( SHAPE_WHITE_NOISE_FRACTION, 20 ) ), 1 ) ); + + /* Calculate the reflection coefficients using schur */ + nrg = silk_schur64( refl_coef_Q16, auto_corr, psEnc->sCmn.shapingLPCOrder ); + silk_assert( nrg >= 0 ); + + /* Convert reflection coefficients to prediction coefficients */ + silk_k2a_Q16( AR2_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder ); + + Qnrg = -scale; /* range: -12...30*/ + silk_assert( Qnrg >= -12 ); + silk_assert( Qnrg <= 30 ); + + /* Make sure that Qnrg is an even number */ + if( Qnrg & 1 ) { + Qnrg -= 1; + nrg >>= 1; + } + + tmp32 = silk_SQRT_APPROX( nrg ); + Qnrg >>= 1; /* range: -6...15*/ + + psEncCtrl->Gains_Q16[ k ] = (silk_LSHIFT32( silk_LIMIT( (tmp32), silk_RSHIFT32( silk_int32_MIN, (16 - Qnrg) ), \ + silk_RSHIFT32( silk_int32_MAX, (16 - Qnrg) ) ), (16 - Qnrg) )); + + if( psEnc->sCmn.warping_Q16 > 0 ) { + /* Adjust gain for warping */ + gain_mult_Q16 = warped_gain( AR2_Q24, warping_Q16, psEnc->sCmn.shapingLPCOrder ); + silk_assert( psEncCtrl->Gains_Q16[ k ] >= 0 ); + if ( silk_SMULWW( silk_RSHIFT_ROUND( psEncCtrl->Gains_Q16[ k ], 1 ), gain_mult_Q16 ) >= ( silk_int32_MAX >> 1 ) ) { + psEncCtrl->Gains_Q16[ k ] = silk_int32_MAX; + } else { + psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( psEncCtrl->Gains_Q16[ k ], gain_mult_Q16 ); + } + } + + /* Bandwidth expansion for synthesis filter shaping */ + silk_bwexpander_32( AR2_Q24, psEnc->sCmn.shapingLPCOrder, BWExp2_Q16 ); + + /* Compute noise shaping filter coefficients */ + silk_memcpy( AR1_Q24, AR2_Q24, psEnc->sCmn.shapingLPCOrder * sizeof( opus_int32 ) ); + + /* Bandwidth expansion for analysis filter shaping */ + silk_assert( BWExp1_Q16 <= SILK_FIX_CONST( 1.0, 16 ) ); + silk_bwexpander_32( AR1_Q24, psEnc->sCmn.shapingLPCOrder, BWExp1_Q16 ); + + /* Ratio of prediction gains, in energy domain */ + pre_nrg_Q30 = silk_LPC_inverse_pred_gain_Q24( AR2_Q24, psEnc->sCmn.shapingLPCOrder, arch ); + nrg = silk_LPC_inverse_pred_gain_Q24( AR1_Q24, psEnc->sCmn.shapingLPCOrder, arch ); + + /*psEncCtrl->GainsPre[ k ] = 1.0f - 0.7f * ( 1.0f - pre_nrg / nrg ) = 0.3f + 0.7f * pre_nrg / nrg;*/ + pre_nrg_Q30 = silk_LSHIFT32( silk_SMULWB( pre_nrg_Q30, SILK_FIX_CONST( 0.7, 15 ) ), 1 ); + 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 ) ); + } + } + + /*****************/ + /* Gain tweaking */ + /*****************/ + /* Increase gains during low speech activity and put lower limit on gains */ + gain_mult_Q16 = silk_log2lin( -silk_SMLAWB( -SILK_FIX_CONST( 16.0, 7 ), SNR_adj_dB_Q7, SILK_FIX_CONST( 0.16, 16 ) ) ); + gain_add_Q16 = silk_log2lin( silk_SMLAWB( SILK_FIX_CONST( 16.0, 7 ), SILK_FIX_CONST( MIN_QGAIN_DB, 7 ), SILK_FIX_CONST( 0.16, 16 ) ) ); + silk_assert( gain_mult_Q16 > 0 ); + for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { + psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( psEncCtrl->Gains_Q16[ k ], gain_mult_Q16 ); + silk_assert( psEncCtrl->Gains_Q16[ k ] >= 0 ); + 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 */ + /************************************************/ + /* Less low frequency shaping for noisy inputs */ + strength_Q16 = silk_MUL( SILK_FIX_CONST( LOW_FREQ_SHAPING, 4 ), silk_SMLAWB( SILK_FIX_CONST( 1.0, 12 ), + SILK_FIX_CONST( LOW_QUALITY_LOW_FREQ_SHAPING_DECR, 13 ), psEnc->sCmn.input_quality_bands_Q15[ 0 ] - SILK_FIX_CONST( 1.0, 15 ) ) ); + strength_Q16 = silk_RSHIFT( silk_MUL( strength_Q16, psEnc->sCmn.speech_activity_Q8 ), 8 ); + if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { + /* Reduce low frequencies quantization noise for periodic signals, depending on pitch lag */ + /*f = 400; freqz([1, -0.98 + 2e-4 * f], [1, -0.97 + 7e-4 * f], 2^12, Fs); axis([0, 1000, -10, 1])*/ + opus_int fs_kHz_inv = silk_DIV32_16( SILK_FIX_CONST( 0.2, 14 ), psEnc->sCmn.fs_kHz ); + for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { + b_Q14 = fs_kHz_inv + silk_DIV32_16( SILK_FIX_CONST( 3.0, 14 ), psEncCtrl->pitchL[ k ] ); + /* Pack two coefficients in one int32 */ + psEncCtrl->LF_shp_Q14[ k ] = silk_LSHIFT( SILK_FIX_CONST( 1.0, 14 ) - b_Q14 - silk_SMULWB( strength_Q16, b_Q14 ), 16 ); + psEncCtrl->LF_shp_Q14[ k ] |= (opus_uint16)( b_Q14 - SILK_FIX_CONST( 1.0, 14 ) ); + } + silk_assert( SILK_FIX_CONST( HARM_HP_NOISE_COEF, 24 ) < SILK_FIX_CONST( 0.5, 24 ) ); /* Guarantees that second argument to SMULWB() is within range of an opus_int16*/ + Tilt_Q16 = - SILK_FIX_CONST( HP_NOISE_COEF, 16 ) - + silk_SMULWB( SILK_FIX_CONST( 1.0, 16 ) - SILK_FIX_CONST( HP_NOISE_COEF, 16 ), + silk_SMULWB( SILK_FIX_CONST( HARM_HP_NOISE_COEF, 24 ), psEnc->sCmn.speech_activity_Q8 ) ); + } else { + b_Q14 = silk_DIV32_16( 21299, psEnc->sCmn.fs_kHz ); /* 1.3_Q0 = 21299_Q14*/ + /* Pack two coefficients in one int32 */ + psEncCtrl->LF_shp_Q14[ 0 ] = silk_LSHIFT( SILK_FIX_CONST( 1.0, 14 ) - b_Q14 - + silk_SMULWB( strength_Q16, silk_SMULWB( SILK_FIX_CONST( 0.6, 16 ), b_Q14 ) ), 16 ); + psEncCtrl->LF_shp_Q14[ 0 ] |= (opus_uint16)( b_Q14 - SILK_FIX_CONST( 1.0, 14 ) ); + for( k = 1; k < psEnc->sCmn.nb_subfr; k++ ) { + psEncCtrl->LF_shp_Q14[ k ] = psEncCtrl->LF_shp_Q14[ 0 ]; + } + Tilt_Q16 = -SILK_FIX_CONST( HP_NOISE_COEF, 16 ); + } + + /****************************/ + /* 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 ), + SILK_FIX_CONST( 1.0, 16 ) - silk_SMULWB( SILK_FIX_CONST( 1.0, 18 ) - silk_LSHIFT( psEncCtrl->coding_quality_Q14, 4 ), + psEncCtrl->input_quality_Q14 ), SILK_FIX_CONST( HIGH_RATE_OR_LOW_QUALITY_HARMONIC_SHAPING, 16 ) ); + + /* Less harmonic noise shaping for less periodic signals */ + HarmShapeGain_Q16 = silk_SMULWB( silk_LSHIFT( HarmShapeGain_Q16, 1 ), + silk_SQRT_APPROX( silk_LSHIFT( psEnc->LTPCorr_Q15, 15 ) ) ); + } else { + HarmShapeGain_Q16 = 0; + } + + /*************************/ + /* 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 ); + } + RESTORE_STACK; +} diff --git a/media/libopus/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h b/media/libopus/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h new file mode 100644 index 0000000000..66eb2ed26d --- /dev/null +++ b/media/libopus/silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h @@ -0,0 +1,165 @@ +/*********************************************************************** +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 __WARPED_AUTOCORRELATION_FIX_MIPSR1_H__ +#define __WARPED_AUTOCORRELATION_FIX_MIPSR1_H__ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "main_FIX.h" + +#undef QC +#define QC 10 + +#undef QS +#define QS 14 + +/* Autocorrelations for a warped frequency axis */ +#define OVERRIDE_silk_warped_autocorrelation_FIX_c +void silk_warped_autocorrelation_FIX_c( + opus_int32 *corr, /* O Result [order + 1] */ + opus_int *scale, /* O Scaling of the correlation vector */ + const opus_int16 *input, /* I Input data to correlate */ + const opus_int warping_Q16, /* I Warping coefficient */ + const opus_int length, /* I Length of input */ + const opus_int order /* I Correlation order (even) */ +) +{ + opus_int n, i, lsh; + opus_int32 tmp1_QS=0, tmp2_QS=0, tmp3_QS=0, tmp4_QS=0, tmp5_QS=0, tmp6_QS=0, tmp7_QS=0, tmp8_QS=0, start_1=0, start_2=0, start_3=0; + opus_int32 state_QS[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; + opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; + opus_int64 temp64; + + opus_int32 val; + val = 2 * QS - QC; + + /* Order must be even */ + silk_assert( ( order & 1 ) == 0 ); + silk_assert( 2 * QS - QC >= 0 ); + + /* Loop over samples */ + for( n = 0; n < length; n=n+4 ) { + + tmp1_QS = silk_LSHIFT32( (opus_int32)input[ n ], QS ); + start_1 = tmp1_QS; + tmp3_QS = silk_LSHIFT32( (opus_int32)input[ n+1], QS ); + start_2 = tmp3_QS; + tmp5_QS = silk_LSHIFT32( (opus_int32)input[ n+2], QS ); + start_3 = tmp5_QS; + tmp7_QS = silk_LSHIFT32( (opus_int32)input[ n+3], QS ); + + /* Loop over allpass sections */ + for( i = 0; i < order; i += 2 ) { + /* Output of allpass section */ + tmp2_QS = silk_SMLAWB( state_QS[ i ], state_QS[ i + 1 ] - tmp1_QS, warping_Q16 ); + corr_QC[ i ] = __builtin_mips_madd( corr_QC[ i ], tmp1_QS, start_1); + + tmp4_QS = silk_SMLAWB( tmp1_QS, tmp2_QS - tmp3_QS, warping_Q16 ); + corr_QC[ i ] = __builtin_mips_madd( corr_QC[ i ], tmp3_QS, start_2); + + tmp6_QS = silk_SMLAWB( tmp3_QS, tmp4_QS - tmp5_QS, warping_Q16 ); + corr_QC[ i ] = __builtin_mips_madd( corr_QC[ i ], tmp5_QS, start_3); + + tmp8_QS = silk_SMLAWB( tmp5_QS, tmp6_QS - tmp7_QS, warping_Q16 ); + state_QS[ i ] = tmp7_QS; + corr_QC[ i ] = __builtin_mips_madd( corr_QC[ i ], tmp7_QS, state_QS[0]); + + /* Output of allpass section */ + tmp1_QS = silk_SMLAWB( state_QS[ i + 1 ], state_QS[ i + 2 ] - tmp2_QS, warping_Q16 ); + corr_QC[ i+1 ] = __builtin_mips_madd( corr_QC[ i+1 ], tmp2_QS, start_1); + + tmp3_QS = silk_SMLAWB( tmp2_QS, tmp1_QS - tmp4_QS, warping_Q16 ); + corr_QC[ i+1 ] = __builtin_mips_madd( corr_QC[ i+1 ], tmp4_QS, start_2); + + tmp5_QS = silk_SMLAWB( tmp4_QS, tmp3_QS - tmp6_QS, warping_Q16 ); + corr_QC[ i+1 ] = __builtin_mips_madd( corr_QC[ i+1 ], tmp6_QS, start_3); + + tmp7_QS = silk_SMLAWB( tmp6_QS, tmp5_QS - tmp8_QS, warping_Q16 ); + state_QS[ i + 1 ] = tmp8_QS; + corr_QC[ i+1 ] = __builtin_mips_madd( corr_QC[ i+1 ], tmp8_QS, state_QS[ 0 ]); + + } + state_QS[ order ] = tmp7_QS; + + corr_QC[ order ] = __builtin_mips_madd( corr_QC[ order ], tmp1_QS, start_1); + corr_QC[ order ] = __builtin_mips_madd( corr_QC[ order ], tmp3_QS, start_2); + corr_QC[ order ] = __builtin_mips_madd( corr_QC[ order ], tmp5_QS, start_3); + corr_QC[ order ] = __builtin_mips_madd( corr_QC[ order ], tmp7_QS, state_QS[ 0 ]); + } + + for(;n< length; n++ ) { + + tmp1_QS = silk_LSHIFT32( (opus_int32)input[ n ], QS ); + + /* Loop over allpass sections */ + for( i = 0; i < order; i += 2 ) { + + /* Output of allpass section */ + tmp2_QS = silk_SMLAWB( state_QS[ i ], state_QS[ i + 1 ] - tmp1_QS, warping_Q16 ); + state_QS[ i ] = tmp1_QS; + corr_QC[ i ] = __builtin_mips_madd( corr_QC[ i ], tmp1_QS, state_QS[ 0 ]); + + /* Output of allpass section */ + tmp1_QS = silk_SMLAWB( state_QS[ i + 1 ], state_QS[ i + 2 ] - tmp2_QS, warping_Q16 ); + state_QS[ i + 1 ] = tmp2_QS; + corr_QC[ i+1 ] = __builtin_mips_madd( corr_QC[ i+1 ], tmp2_QS, state_QS[ 0 ]); + } + state_QS[ order ] = tmp1_QS; + corr_QC[ order ] = __builtin_mips_madd( corr_QC[ order ], tmp1_QS, state_QS[ 0 ]); + } + + temp64 = corr_QC[ 0 ]; + temp64 = __builtin_mips_shilo(temp64, val); + + lsh = silk_CLZ64( temp64 ) - 35; + lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC ); + *scale = -( QC + lsh ); + silk_assert( *scale >= -30 && *scale <= 12 ); + if( lsh >= 0 ) { + for( i = 0; i < order + 1; i++ ) { + temp64 = corr_QC[ i ]; + //temp64 = __builtin_mips_shilo(temp64, val); + temp64 = (val >= 0) ? (temp64 >> val) : (temp64 << -val); + corr[ i ] = (opus_int32)silk_CHECK_FIT32( __builtin_mips_shilo( temp64, -lsh ) ); + } + } else { + for( i = 0; i < order + 1; i++ ) { + temp64 = corr_QC[ i ]; + //temp64 = __builtin_mips_shilo(temp64, val); + temp64 = (val >= 0) ? (temp64 >> val) : (temp64 << -val); + corr[ i ] = (opus_int32)silk_CHECK_FIT32( __builtin_mips_shilo( temp64, -lsh ) ); + } + } + + corr_QC[ 0 ] = __builtin_mips_shilo(corr_QC[ 0 ], val); + + silk_assert( corr_QC[ 0 ] >= 0 ); /* If breaking, decrease QC*/ +} +#endif /* __WARPED_AUTOCORRELATION_FIX_MIPSR1_H__ */ diff --git a/media/libopus/silk/fixed/noise_shape_analysis_FIX.c b/media/libopus/silk/fixed/noise_shape_analysis_FIX.c new file mode 100644 index 0000000000..85fea0bf09 --- /dev/null +++ b/media/libopus/silk/fixed/noise_shape_analysis_FIX.c @@ -0,0 +1,407 @@ +/*********************************************************************** +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" + +/* Compute gain to make warped filter coefficients have a zero mean log frequency response on a */ +/* non-warped frequency scale. (So that it can be implemented with a minimum-phase monic filter.) */ +/* Note: A monic filter is one with the first coefficient equal to 1.0. In Silk we omit the first */ +/* coefficient in an array of coefficients, for monic filters. */ +static OPUS_INLINE opus_int32 warped_gain( /* gain in Q16*/ + const opus_int32 *coefs_Q24, + opus_int lambda_Q16, + opus_int order +) { + opus_int i; + opus_int32 gain_Q24; + + lambda_Q16 = -lambda_Q16; + gain_Q24 = coefs_Q24[ order - 1 ]; + for( i = order - 2; i >= 0; i-- ) { + gain_Q24 = silk_SMLAWB( coefs_Q24[ i ], gain_Q24, lambda_Q16 ); + } + gain_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), gain_Q24, -lambda_Q16 ); + return silk_INVERSE32_varQ( gain_Q24, 40 ); +} + +/* 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_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 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 ); + } + 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 ); + for( i = 0; i < order; i++ ) { + coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ i ] ); + } + limit_Q20 = silk_RSHIFT(limit_Q24, 4); + for( iter = 0; iter < 10; iter++ ) { + /* Find maximum absolute value */ + maxabs_Q24 = -1; + for( i = 0; i < order; i++ ) { + tmp = silk_abs_int32( coefs_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 ) { + /* 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 ); + } + gain_Q16 = silk_INVERSE32_varQ( gain_Q16, 32 ); + for( i = 0; i < order; i++ ) { + coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ i ] ); + } + + /* Apply bandwidth expansion */ + chirp_Q16 = SILK_FIX_CONST( 0.99, 16 ) - silk_DIV32_varQ( + silk_SMULWB( maxabs_Q20 - limit_Q20, silk_SMLABB( SILK_FIX_CONST( 0.8, 10 ), SILK_FIX_CONST( 0.1, 10 ), iter ) ), + silk_MUL( maxabs_Q20, ind + 1 ), 22 ); + silk_bwexpander_32( coefs_Q24, order, chirp_Q16 ); + + /* Convert to monic warped coefficients */ + lambda_Q16 = -lambda_Q16; + for( i = order - 1; i > 0; i-- ) { + coefs_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); + } + lambda_Q16 = -lambda_Q16; + nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -(opus_int32)lambda_Q16, lambda_Q16 ); + den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_Q24[ 0 ], lambda_Q16 ); + gain_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 ] ); + } + } + silk_assert( 0 ); +} + +/* Disable MIPS version until it's updated. */ +#if 0 && defined(MIPSr1_ASM) +#include "mips/noise_shape_analysis_FIX_mipsr1.h" +#endif + +/**************************************************************/ +/* Compute noise shaping coefficients and initial gain values */ +/**************************************************************/ +#ifndef OVERRIDE_silk_noise_shape_analysis_FIX +void silk_noise_shape_analysis_FIX( + silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */ + silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control FIX */ + const opus_int16 *pitch_res, /* I LPC residual from pitch analysis */ + const opus_int16 *x, /* I Input signal [ frame_length + la_shape ] */ + int arch /* I Run-time architecture */ +) +{ + 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_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 ]; + VARDECL( opus_int16, x_windowed ); + const opus_int16 *x_ptr, *pitch_res_ptr; + SAVE_STACK; + + /* Point to start of first LPC analysis block */ + x_ptr = x - psEnc->sCmn.la_shape; + + /****************/ + /* GAIN CONTROL */ + /****************/ + SNR_adj_dB_Q7 = psEnc->sCmn.SNR_dB_Q7; + + /* Input quality is the average of the quality in the lowest two VAD bands */ + psEncCtrl->input_quality_Q14 = ( opus_int )silk_RSHIFT( (opus_int32)psEnc->sCmn.input_quality_bands_Q15[ 0 ] + + psEnc->sCmn.input_quality_bands_Q15[ 1 ], 2 ); + + /* Coding quality level, between 0.0_Q0 and 1.0_Q0, but in Q14 */ + psEncCtrl->coding_quality_Q14 = silk_RSHIFT( silk_sigm_Q15( silk_RSHIFT_ROUND( SNR_adj_dB_Q7 - + SILK_FIX_CONST( 20.0, 7 ), 4 ) ), 1 ); + + /* Reduce coding SNR during low speech activity */ + if( psEnc->sCmn.useCBR == 0 ) { + b_Q8 = SILK_FIX_CONST( 1.0, 8 ) - psEnc->sCmn.speech_activity_Q8; + b_Q8 = silk_SMULWB( silk_LSHIFT( b_Q8, 8 ), b_Q8 ); + SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7, + silk_SMULBB( SILK_FIX_CONST( -BG_SNR_DECR_dB, 7 ) >> ( 4 + 1 ), b_Q8 ), /* Q11*/ + silk_SMULWB( SILK_FIX_CONST( 1.0, 14 ) + psEncCtrl->input_quality_Q14, psEncCtrl->coding_quality_Q14 ) ); /* Q12*/ + } + + if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { + /* Reduce gains for periodic signals */ + SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7, SILK_FIX_CONST( HARM_SNR_INCR_dB, 8 ), psEnc->LTPCorr_Q15 ); + } else { + /* For unvoiced signals and low-quality input, adjust the quality slower than SNR_dB setting */ + SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7, + silk_SMLAWB( SILK_FIX_CONST( 6.0, 9 ), -SILK_FIX_CONST( 0.4, 18 ), psEnc->sCmn.SNR_dB_Q7 ), + SILK_FIX_CONST( 1.0, 14 ) - psEncCtrl->input_quality_Q14 ); + } + + /*************************/ + /* SPARSENESS PROCESSING */ + /*************************/ + /* Set quantizer offset */ + if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { + /* Initially set to 0; may be overruled in process_gains(..) */ + psEnc->sCmn.indices.quantOffsetType = 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++ ) { + silk_sum_sqr_shift( &nrg, &scale, pitch_res_ptr, nSamples ); + nrg += silk_RSHIFT( nSamples, scale ); /* Q(-scale)*/ + + log_energy_Q7 = silk_lin2log( nrg ); + if( k > 0 ) { + energy_variation_Q7 += silk_abs( log_energy_Q7 - log_energy_prev_Q7 ); + } + log_energy_prev_Q7 = log_energy_Q7; + pitch_res_ptr += nSamples; + } + + /* Set quantization offset depending on sparseness measure */ + if( energy_variation_Q7 > SILK_FIX_CONST( ENERGY_VARIATION_THRESHOLD_QNT_OFFSET, 7 ) * (nSegs-1) ) { + psEnc->sCmn.indices.quantOffsetType = 0; + } else { + psEnc->sCmn.indices.quantOffsetType = 1; + } + } + + /*******************************/ + /* Control bandwidth expansion */ + /*******************************/ + /* 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 ), + silk_SMLAWW( SILK_FIX_CONST( 1.0, 16 ), strength_Q16, strength_Q16 ), 16 ); + + if( psEnc->sCmn.warping_Q16 > 0 ) { + /* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */ + warping_Q16 = silk_SMLAWB( psEnc->sCmn.warping_Q16, (opus_int32)psEncCtrl->coding_quality_Q14, SILK_FIX_CONST( 0.01, 18 ) ); + } else { + warping_Q16 = 0; + } + + /********************************************/ + /* Compute noise shaping AR coefs and gains */ + /********************************************/ + ALLOC( x_windowed, psEnc->sCmn.shapeWinLength, opus_int16 ); + for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { + /* Apply window: sine slope followed by flat part followed by cosine slope */ + opus_int shift, slope_part, flat_part; + flat_part = psEnc->sCmn.fs_kHz * 3; + slope_part = silk_RSHIFT( psEnc->sCmn.shapeWinLength - flat_part, 1 ); + + silk_apply_sine_window( x_windowed, x_ptr, 1, slope_part ); + shift = slope_part; + silk_memcpy( x_windowed + shift, x_ptr + shift, flat_part * sizeof(opus_int16) ); + shift += flat_part; + silk_apply_sine_window( x_windowed + shift, x_ptr + shift, 2, slope_part ); + + /* Update pointer: next LPC analysis block */ + x_ptr += psEnc->sCmn.subfr_length; + + 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 ); + } else { + /* Calculate regular auto correlation */ + silk_autocorr( auto_corr, &scale, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1, arch ); + } + + /* Add white noise, as a fraction of energy */ + auto_corr[0] = silk_ADD32( auto_corr[0], silk_max_32( silk_SMULWB( silk_RSHIFT( auto_corr[ 0 ], 4 ), + SILK_FIX_CONST( SHAPE_WHITE_NOISE_FRACTION, 20 ) ), 1 ) ); + + /* Calculate the reflection coefficients using schur */ + nrg = silk_schur64( refl_coef_Q16, auto_corr, psEnc->sCmn.shapingLPCOrder ); + silk_assert( nrg >= 0 ); + + /* Convert reflection coefficients to prediction coefficients */ + silk_k2a_Q16( AR_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder ); + + Qnrg = -scale; /* range: -12...30*/ + silk_assert( Qnrg >= -12 ); + silk_assert( Qnrg <= 30 ); + + /* Make sure that Qnrg is an even number */ + if( Qnrg & 1 ) { + Qnrg -= 1; + nrg >>= 1; + } + + tmp32 = silk_SQRT_APPROX( nrg ); + Qnrg >>= 1; /* range: -6...15*/ + + psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT_SAT32( tmp32, 16 - Qnrg ); + + 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 ); + } 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 ); + } + } + silk_assert( psEncCtrl->Gains_Q16[ k ] > 0 ); + } + + /* Bandwidth expansion */ + silk_bwexpander_32( AR_Q24, psEnc->sCmn.shapingLPCOrder, BWExp_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 ); + + /* 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 ); + } + } + + /*****************/ + /* Gain tweaking */ + /*****************/ + /* Increase gains during low speech activity and put lower limit on gains */ + gain_mult_Q16 = silk_log2lin( -silk_SMLAWB( -SILK_FIX_CONST( 16.0, 7 ), SNR_adj_dB_Q7, SILK_FIX_CONST( 0.16, 16 ) ) ); + gain_add_Q16 = silk_log2lin( silk_SMLAWB( SILK_FIX_CONST( 16.0, 7 ), SILK_FIX_CONST( MIN_QGAIN_DB, 7 ), SILK_FIX_CONST( 0.16, 16 ) ) ); + silk_assert( gain_mult_Q16 > 0 ); + for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { + psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( psEncCtrl->Gains_Q16[ k ], gain_mult_Q16 ); + silk_assert( psEncCtrl->Gains_Q16[ k ] >= 0 ); + psEncCtrl->Gains_Q16[ k ] = silk_ADD_POS_SAT32( psEncCtrl->Gains_Q16[ k ], gain_add_Q16 ); + } + + + /************************************************/ + /* Control low-frequency shaping and noise tilt */ + /************************************************/ + /* Less low frequency shaping for noisy inputs */ + strength_Q16 = silk_MUL( SILK_FIX_CONST( LOW_FREQ_SHAPING, 4 ), silk_SMLAWB( SILK_FIX_CONST( 1.0, 12 ), + SILK_FIX_CONST( LOW_QUALITY_LOW_FREQ_SHAPING_DECR, 13 ), psEnc->sCmn.input_quality_bands_Q15[ 0 ] - SILK_FIX_CONST( 1.0, 15 ) ) ); + strength_Q16 = silk_RSHIFT( silk_MUL( strength_Q16, psEnc->sCmn.speech_activity_Q8 ), 8 ); + if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { + /* Reduce low frequencies quantization noise for periodic signals, depending on pitch lag */ + /*f = 400; freqz([1, -0.98 + 2e-4 * f], [1, -0.97 + 7e-4 * f], 2^12, Fs); axis([0, 1000, -10, 1])*/ + opus_int fs_kHz_inv = silk_DIV32_16( SILK_FIX_CONST( 0.2, 14 ), psEnc->sCmn.fs_kHz ); + for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { + b_Q14 = fs_kHz_inv + silk_DIV32_16( SILK_FIX_CONST( 3.0, 14 ), psEncCtrl->pitchL[ k ] ); + /* Pack two coefficients in one int32 */ + psEncCtrl->LF_shp_Q14[ k ] = silk_LSHIFT( SILK_FIX_CONST( 1.0, 14 ) - b_Q14 - silk_SMULWB( strength_Q16, b_Q14 ), 16 ); + psEncCtrl->LF_shp_Q14[ k ] |= (opus_uint16)( b_Q14 - SILK_FIX_CONST( 1.0, 14 ) ); + } + silk_assert( SILK_FIX_CONST( HARM_HP_NOISE_COEF, 24 ) < SILK_FIX_CONST( 0.5, 24 ) ); /* Guarantees that second argument to SMULWB() is within range of an opus_int16*/ + Tilt_Q16 = - SILK_FIX_CONST( HP_NOISE_COEF, 16 ) - + silk_SMULWB( SILK_FIX_CONST( 1.0, 16 ) - SILK_FIX_CONST( HP_NOISE_COEF, 16 ), + silk_SMULWB( SILK_FIX_CONST( HARM_HP_NOISE_COEF, 24 ), psEnc->sCmn.speech_activity_Q8 ) ); + } else { + b_Q14 = silk_DIV32_16( 21299, psEnc->sCmn.fs_kHz ); /* 1.3_Q0 = 21299_Q14*/ + /* Pack two coefficients in one int32 */ + psEncCtrl->LF_shp_Q14[ 0 ] = silk_LSHIFT( SILK_FIX_CONST( 1.0, 14 ) - b_Q14 - + silk_SMULWB( strength_Q16, silk_SMULWB( SILK_FIX_CONST( 0.6, 16 ), b_Q14 ) ), 16 ); + psEncCtrl->LF_shp_Q14[ 0 ] |= (opus_uint16)( b_Q14 - SILK_FIX_CONST( 1.0, 14 ) ); + for( k = 1; k < psEnc->sCmn.nb_subfr; k++ ) { + psEncCtrl->LF_shp_Q14[ k ] = psEncCtrl->LF_shp_Q14[ 0 ]; + } + Tilt_Q16 = -SILK_FIX_CONST( HP_NOISE_COEF, 16 ); + } + + /****************************/ + /* HARMONIC SHAPING CONTROL */ + /****************************/ + 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 ), + SILK_FIX_CONST( 1.0, 16 ) - silk_SMULWB( SILK_FIX_CONST( 1.0, 18 ) - silk_LSHIFT( psEncCtrl->coding_quality_Q14, 4 ), + psEncCtrl->input_quality_Q14 ), SILK_FIX_CONST( HIGH_RATE_OR_LOW_QUALITY_HARMONIC_SHAPING, 16 ) ); + + /* Less harmonic noise shaping for less periodic signals */ + HarmShapeGain_Q16 = silk_SMULWB( silk_LSHIFT( HarmShapeGain_Q16, 1 ), + silk_SQRT_APPROX( silk_LSHIFT( psEnc->LTPCorr_Q15, 15 ) ) ); + } else { + HarmShapeGain_Q16 = 0; + } + + /*************************/ + /* Smooth over subframes */ + /*************************/ + for( k = 0; k < MAX_NB_SUBFR; k++ ) { + 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->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 ); + } + RESTORE_STACK; +} +#endif /* OVERRIDE_silk_noise_shape_analysis_FIX */ diff --git a/media/libopus/silk/fixed/pitch_analysis_core_FIX.c b/media/libopus/silk/fixed/pitch_analysis_core_FIX.c new file mode 100644 index 0000000000..14729046d2 --- /dev/null +++ b/media/libopus/silk/fixed/pitch_analysis_core_FIX.c @@ -0,0 +1,721 @@ +/*********************************************************************** +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 + +/*********************************************************** +* Pitch analyser function +********************************************************** */ +#include "SigProc_FIX.h" +#include "pitch_est_defines.h" +#include "stack_alloc.h" +#include "debug.h" +#include "pitch.h" + +#define SCRATCH_SIZE 22 +#define SF_LENGTH_4KHZ ( PE_SUBFR_LENGTH_MS * 4 ) +#define SF_LENGTH_8KHZ ( PE_SUBFR_LENGTH_MS * 8 ) +#define MIN_LAG_4KHZ ( PE_MIN_LAG_MS * 4 ) +#define MIN_LAG_8KHZ ( PE_MIN_LAG_MS * 8 ) +#define MAX_LAG_4KHZ ( PE_MAX_LAG_MS * 4 ) +#define MAX_LAG_8KHZ ( PE_MAX_LAG_MS * 8 - 1 ) +#define CSTRIDE_4KHZ ( MAX_LAG_4KHZ + 1 - MIN_LAG_4KHZ ) +#define CSTRIDE_8KHZ ( MAX_LAG_8KHZ + 3 - ( MIN_LAG_8KHZ - 2 ) ) +#define D_COMP_MIN ( MIN_LAG_8KHZ - 3 ) +#define D_COMP_MAX ( MAX_LAG_8KHZ + 4 ) +#define D_COMP_STRIDE ( D_COMP_MAX - D_COMP_MIN ) + +typedef opus_int32 silk_pe_stage3_vals[ PE_NB_STAGE3_LAGS ]; + +/************************************************************/ +/* Internally used functions */ +/************************************************************/ +static void silk_P_Ana_calc_corr_st3( + silk_pe_stage3_vals cross_corr_st3[], /* O 3 DIM correlation array */ + const opus_int16 frame[], /* I vector to correlate */ + opus_int start_lag, /* I lag offset to search around */ + opus_int sf_length, /* I length of a 5 ms subframe */ + opus_int nb_subfr, /* I number of subframes */ + opus_int complexity, /* I Complexity setting */ + int arch /* I Run-time architecture */ +); + +static void silk_P_Ana_calc_energy_st3( + silk_pe_stage3_vals energies_st3[], /* O 3 DIM energy array */ + const opus_int16 frame[], /* I vector to calc energy in */ + opus_int start_lag, /* I lag offset to search around */ + opus_int sf_length, /* I length of one 5 ms subframe */ + opus_int nb_subfr, /* I number of subframes */ + opus_int complexity, /* I Complexity setting */ + int arch /* I Run-time architecture */ +); + +/*************************************************************/ +/* 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 */ + opus_int *pitch_out, /* O 4 pitch lag values */ + opus_int16 *lagIndex, /* O Lag Index */ + opus_int8 *contourIndex, /* O Pitch contour Index */ + opus_int *LTPCorr_Q15, /* I/O Normalized correlation; input: value from previous frame */ + opus_int prevLag, /* I Last lag of previous frame; set to zero is unvoiced */ + const opus_int32 search_thres1_Q16, /* I First stage threshold for lag candidates 0 - 1 */ + const opus_int search_thres2_Q13, /* I Final threshold for lag candidates 0 - 1 */ + const opus_int Fs_kHz, /* I Sample frequency (kHz) */ + const opus_int complexity, /* I Complexity setting, 0-2, where 2 is highest */ + const opus_int nb_subfr, /* I number of 5 ms subframes */ + int arch /* I Run-time architecture */ +) +{ + VARDECL( opus_int16, frame_8kHz_buf ); + VARDECL( opus_int16, frame_4kHz ); + VARDECL( opus_int16, frame_scaled ); + opus_int32 filt_state[ 6 ]; + const opus_int16 *frame, *frame_8kHz; + opus_int i, k, d, j; + VARDECL( opus_int16, C ); + VARDECL( opus_int32, xcorr32 ); + const opus_int16 *target_ptr, *basis_ptr; + opus_int32 cross_corr, normalizer, energy, energy_basis, energy_target; + opus_int d_srch[ PE_D_SRCH_LENGTH ], Cmax, length_d_srch, length_d_comp, shift; + VARDECL( opus_int16, d_comp ); + opus_int32 sum, threshold, lag_counter; + opus_int CBimax, CBimax_new, CBimax_old, lag, start_lag, end_lag, lag_new; + opus_int32 CC[ PE_NB_CBKS_STAGE2_EXT ], CCmax, CCmax_b, CCmax_new_b, CCmax_new; + VARDECL( silk_pe_stage3_vals, energies_st3 ); + VARDECL( silk_pe_stage3_vals, cross_corr_st3 ); + opus_int frame_length, frame_length_8kHz, frame_length_4kHz; + opus_int sf_length; + opus_int min_lag; + opus_int max_lag; + opus_int32 contour_bias_Q15, diff; + opus_int nb_cbk_search, cbk_size; + 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 ); + + /* Check for valid complexity setting */ + celt_assert( complexity >= SILK_PE_MIN_COMPLEX ); + celt_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) ); + + /* Set up frame lengths max / min lag for the sampling frequency */ + frame_length = ( PE_LTP_MEM_LENGTH_MS + nb_subfr * PE_SUBFR_LENGTH_MS ) * Fs_kHz; + frame_length_4kHz = ( PE_LTP_MEM_LENGTH_MS + nb_subfr * PE_SUBFR_LENGTH_MS ) * 4; + frame_length_8kHz = ( PE_LTP_MEM_LENGTH_MS + nb_subfr * PE_SUBFR_LENGTH_MS ) * 8; + sf_length = PE_SUBFR_LENGTH_MS * Fs_kHz; + 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 */ + 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; + } 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; + } else { + celt_assert( Fs_kHz == 8 ); + frame_8kHz = frame; + } + + /* Decimate again to 4 kHz */ + silk_memset( filt_state, 0, 2 * sizeof( opus_int32 ) );/* Set state to zero */ + ALLOC( frame_4kHz, frame_length_4kHz, opus_int16 ); + silk_resampler_down2( filt_state, frame_4kHz, frame_8kHz, frame_length_8kHz ); + + /* Low-pass filter */ + for( i = frame_length_4kHz - 1; i > 0; i-- ) { + frame_4kHz[ i ] = silk_ADD_SAT16( frame_4kHz[ i ], frame_4kHz[ i - 1 ] ); + } + + + /****************************************************************************** + * FIRST STAGE, operating in 4 khz + ******************************************************************************/ + ALLOC( C, nb_subfr * CSTRIDE_8KHZ, opus_int16 ); + ALLOC( xcorr32, MAX_LAG_4KHZ-MIN_LAG_4KHZ+1, opus_int32 ); + silk_memset( C, 0, (nb_subfr >> 1) * CSTRIDE_4KHZ * sizeof( opus_int16 ) ); + 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 ); + + 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 ); + + celt_pitch_xcorr( target_ptr, target_ptr - MAX_LAG_4KHZ, xcorr32, SF_LENGTH_8KHZ, MAX_LAG_4KHZ - MIN_LAG_4KHZ + 1, arch ); + + /* Calculate first vector products before loop */ + cross_corr = xcorr32[ MAX_LAG_4KHZ - MIN_LAG_4KHZ ]; + normalizer = silk_inner_prod_aligned( target_ptr, target_ptr, SF_LENGTH_8KHZ, arch ); + normalizer = silk_ADD32( normalizer, silk_inner_prod_aligned( basis_ptr, basis_ptr, SF_LENGTH_8KHZ, arch ) ); + normalizer = silk_ADD32( normalizer, silk_SMULBB( SF_LENGTH_8KHZ, 4000 ) ); + + matrix_ptr( C, k, 0, CSTRIDE_4KHZ ) = + (opus_int16)silk_DIV32_varQ( cross_corr, normalizer, 13 + 1 ); /* Q13 */ + + /* From now on normalizer is computed recursively */ + for( d = MIN_LAG_4KHZ + 1; d <= MAX_LAG_4KHZ; d++ ) { + basis_ptr--; + + /* Check that we are within range of the array */ + silk_assert( basis_ptr >= frame_4kHz ); + silk_assert( basis_ptr + SF_LENGTH_8KHZ <= frame_4kHz + frame_length_4kHz ); + + cross_corr = xcorr32[ MAX_LAG_4KHZ - d ]; + + /* Add contribution of new sample and remove contribution from oldest sample */ + normalizer = silk_ADD32( normalizer, + silk_SMULBB( basis_ptr[ 0 ], basis_ptr[ 0 ] ) - + silk_SMULBB( basis_ptr[ SF_LENGTH_8KHZ ], basis_ptr[ SF_LENGTH_8KHZ ] ) ); + + matrix_ptr( C, k, d - MIN_LAG_4KHZ, CSTRIDE_4KHZ) = + (opus_int16)silk_DIV32_varQ( cross_corr, normalizer, 13 + 1 ); /* Q13 */ + } + /* Update target pointer */ + target_ptr += SF_LENGTH_8KHZ; + } + + /* Combine two subframes into single correlation measure and apply short-lag bias */ + if( nb_subfr == PE_MAX_NB_SUBFR ) { + for( i = MAX_LAG_4KHZ; i >= MIN_LAG_4KHZ; i-- ) { + sum = (opus_int32)matrix_ptr( C, 0, i - MIN_LAG_4KHZ, CSTRIDE_4KHZ ) + + (opus_int32)matrix_ptr( C, 1, i - MIN_LAG_4KHZ, CSTRIDE_4KHZ ); /* Q14 */ + sum = silk_SMLAWB( sum, sum, silk_LSHIFT( -i, 4 ) ); /* Q14 */ + C[ i - MIN_LAG_4KHZ ] = (opus_int16)sum; /* Q14 */ + } + } else { + /* Only short-lag bias */ + for( i = MAX_LAG_4KHZ; i >= MIN_LAG_4KHZ; i-- ) { + sum = silk_LSHIFT( (opus_int32)C[ i - MIN_LAG_4KHZ ], 1 ); /* Q14 */ + sum = silk_SMLAWB( sum, sum, silk_LSHIFT( -i, 4 ) ); /* Q14 */ + C[ i - MIN_LAG_4KHZ ] = (opus_int16)sum; /* Q14 */ + } + } + + /* Sort */ + length_d_srch = silk_ADD_LSHIFT32( 4, complexity, 1 ); + celt_assert( 3 * length_d_srch <= PE_D_SRCH_LENGTH ); + silk_insertion_sort_decreasing_int16( C, d_srch, CSTRIDE_4KHZ, + length_d_srch ); + + /* Escape if correlation is very low already here */ + Cmax = (opus_int)C[ 0 ]; /* Q14 */ + if( Cmax < SILK_FIX_CONST( 0.2, 14 ) ) { + silk_memset( pitch_out, 0, nb_subfr * sizeof( opus_int ) ); + *LTPCorr_Q15 = 0; + *lagIndex = 0; + *contourIndex = 0; + RESTORE_STACK; + return 1; + } + + threshold = silk_SMULWB( search_thres1_Q16, Cmax ); + for( i = 0; i < length_d_srch; i++ ) { + /* Convert to 8 kHz indices for the sorted correlation that exceeds the threshold */ + if( C[ i ] > threshold ) { + d_srch[ i ] = silk_LSHIFT( d_srch[ i ] + MIN_LAG_4KHZ, 1 ); + } else { + length_d_srch = i; + break; + } + } + celt_assert( length_d_srch > 0 ); + + ALLOC( d_comp, D_COMP_STRIDE, opus_int16 ); + for( i = D_COMP_MIN; i < D_COMP_MAX; i++ ) { + d_comp[ i - D_COMP_MIN ] = 0; + } + for( i = 0; i < length_d_srch; i++ ) { + d_comp[ d_srch[ i ] - D_COMP_MIN ] = 1; + } + + /* Convolution */ + for( i = D_COMP_MAX - 1; i >= MIN_LAG_8KHZ; i-- ) { + d_comp[ i - D_COMP_MIN ] += + d_comp[ i - 1 - D_COMP_MIN ] + d_comp[ i - 2 - D_COMP_MIN ]; + } + + length_d_srch = 0; + for( i = MIN_LAG_8KHZ; i < MAX_LAG_8KHZ + 1; i++ ) { + if( d_comp[ i + 1 - D_COMP_MIN ] > 0 ) { + d_srch[ length_d_srch ] = i; + length_d_srch++; + } + } + + /* Convolution */ + for( i = D_COMP_MAX - 1; i >= MIN_LAG_8KHZ; i-- ) { + d_comp[ i - D_COMP_MIN ] += d_comp[ i - 1 - D_COMP_MIN ] + + d_comp[ i - 2 - D_COMP_MIN ] + d_comp[ i - 3 - D_COMP_MIN ]; + } + + length_d_comp = 0; + for( i = MIN_LAG_8KHZ; i < D_COMP_MAX; i++ ) { + if( d_comp[ i - D_COMP_MIN ] > 0 ) { + d_comp[ length_d_comp ] = i - 2; + length_d_comp++; + } + } + + /********************************************************************************** + ** SECOND STAGE, operating at 8 kHz, on lag sections with high correlation + *************************************************************************************/ + + /********************************************************************************* + * Find energy of each subframe projected onto its history, for a range of delays + *********************************************************************************/ + silk_memset( C, 0, nb_subfr * CSTRIDE_8KHZ * sizeof( opus_int16 ) ); + + target_ptr = &frame_8kHz[ PE_LTP_MEM_LENGTH_MS * 8 ]; + 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 ); + + 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++ ) { + d = d_comp[ j ]; + basis_ptr = target_ptr - d; + + /* Check that we are within range of the array */ + silk_assert( basis_ptr >= frame_8kHz ); + silk_assert( basis_ptr + SF_LENGTH_8KHZ <= frame_8kHz + frame_length_8kHz ); + + cross_corr = silk_inner_prod_aligned( target_ptr, basis_ptr, SF_LENGTH_8KHZ, arch ); + if( cross_corr > 0 ) { + energy_basis = silk_inner_prod_aligned( basis_ptr, basis_ptr, SF_LENGTH_8KHZ, arch ); + matrix_ptr( C, k, d - ( MIN_LAG_8KHZ - 2 ), CSTRIDE_8KHZ ) = + (opus_int16)silk_DIV32_varQ( cross_corr, + silk_ADD32( energy_target, + energy_basis ), + 13 + 1 ); /* Q13 */ + } else { + matrix_ptr( C, k, d - ( MIN_LAG_8KHZ - 2 ), CSTRIDE_8KHZ ) = 0; + } + } + target_ptr += SF_LENGTH_8KHZ; + } + + /* search over lag range and lags codebook */ + /* scale factor for lag codebook, as a function of center lag */ + + CCmax = silk_int32_MIN; + CCmax_b = silk_int32_MIN; + + CBimax = 0; /* To avoid returning undefined lag values */ + lag = -1; /* To check if lag with strong enough correlation has been found */ + + if( prevLag > 0 ) { + if( Fs_kHz == 12 ) { + prevLag = silk_DIV32_16( silk_LSHIFT( prevLag, 1 ), 3 ); + } else if( Fs_kHz == 16 ) { + prevLag = silk_RSHIFT( prevLag, 1 ); + } + prevLag_log2_Q7 = silk_lin2log( (opus_int32)prevLag ); + } else { + prevLag_log2_Q7 = 0; + } + silk_assert( search_thres2_Q13 == silk_SAT16( search_thres2_Q13 ) ); + /* Set up stage 2 codebook based on number of subframes */ + if( nb_subfr == PE_MAX_NB_SUBFR ) { + cbk_size = PE_NB_CBKS_STAGE2_EXT; + Lag_CB_ptr = &silk_CB_lags_stage2[ 0 ][ 0 ]; + if( Fs_kHz == 8 && complexity > SILK_PE_MIN_COMPLEX ) { + /* If input is 8 khz use a larger codebook here because it is last stage */ + nb_cbk_search = PE_NB_CBKS_STAGE2_EXT; + } else { + nb_cbk_search = PE_NB_CBKS_STAGE2; + } + } else { + cbk_size = PE_NB_CBKS_STAGE2_10MS; + Lag_CB_ptr = &silk_CB_lags_stage2_10_ms[ 0 ][ 0 ]; + nb_cbk_search = PE_NB_CBKS_STAGE2_10MS; + } + + for( k = 0; k < length_d_srch; k++ ) { + d = d_srch[ k ]; + for( j = 0; j < nb_cbk_search; j++ ) { + CC[ j ] = 0; + for( i = 0; i < nb_subfr; i++ ) { + opus_int d_subfr; + /* Try all codebooks */ + d_subfr = d + matrix_ptr( Lag_CB_ptr, i, j, cbk_size ); + CC[ j ] = CC[ j ] + + (opus_int32)matrix_ptr( C, i, + d_subfr - ( MIN_LAG_8KHZ - 2 ), + CSTRIDE_8KHZ ); + } + } + /* Find best codebook */ + CCmax_new = silk_int32_MIN; + CBimax_new = 0; + for( i = 0; i < nb_cbk_search; i++ ) { + if( CC[ i ] > CCmax_new ) { + CCmax_new = CC[ i ]; + CBimax_new = i; + } + } + + /* Bias towards shorter lags */ + lag_log2_Q7 = silk_lin2log( d ); /* Q7 */ + silk_assert( lag_log2_Q7 == silk_SAT16( lag_log2_Q7 ) ); + silk_assert( nb_subfr * SILK_FIX_CONST( PE_SHORTLAG_BIAS, 13 ) == silk_SAT16( nb_subfr * SILK_FIX_CONST( PE_SHORTLAG_BIAS, 13 ) ) ); + CCmax_new_b = CCmax_new - silk_RSHIFT( silk_SMULBB( nb_subfr * SILK_FIX_CONST( PE_SHORTLAG_BIAS, 13 ), lag_log2_Q7 ), 7 ); /* Q13 */ + + /* Bias towards previous lag */ + silk_assert( nb_subfr * SILK_FIX_CONST( PE_PREVLAG_BIAS, 13 ) == silk_SAT16( nb_subfr * SILK_FIX_CONST( PE_PREVLAG_BIAS, 13 ) ) ); + if( prevLag > 0 ) { + delta_lag_log2_sqr_Q7 = lag_log2_Q7 - prevLag_log2_Q7; + silk_assert( delta_lag_log2_sqr_Q7 == silk_SAT16( delta_lag_log2_sqr_Q7 ) ); + delta_lag_log2_sqr_Q7 = silk_RSHIFT( silk_SMULBB( delta_lag_log2_sqr_Q7, delta_lag_log2_sqr_Q7 ), 7 ); + prev_lag_bias_Q13 = silk_RSHIFT( silk_SMULBB( nb_subfr * SILK_FIX_CONST( PE_PREVLAG_BIAS, 13 ), *LTPCorr_Q15 ), 15 ); /* Q13 */ + prev_lag_bias_Q13 = silk_DIV32( silk_MUL( prev_lag_bias_Q13, delta_lag_log2_sqr_Q7 ), delta_lag_log2_sqr_Q7 + SILK_FIX_CONST( 0.5, 7 ) ); + CCmax_new_b -= prev_lag_bias_Q13; /* Q13 */ + } + + if( CCmax_new_b > CCmax_b && /* Find maximum biased correlation */ + CCmax_new > silk_SMULBB( nb_subfr, search_thres2_Q13 ) && /* Correlation needs to be high enough to be voiced */ + silk_CB_lags_stage2[ 0 ][ CBimax_new ] <= MIN_LAG_8KHZ /* Lag must be in range */ + ) { + CCmax_b = CCmax_new_b; + CCmax = CCmax_new; + lag = d; + CBimax = CBimax_new; + } + } + + if( lag == -1 ) { + /* No suitable candidate found */ + silk_memset( pitch_out, 0, nb_subfr * sizeof( opus_int ) ); + *LTPCorr_Q15 = 0; + *lagIndex = 0; + *contourIndex = 0; + RESTORE_STACK; + return 1; + } + + /* Output normalized correlation */ + *LTPCorr_Q15 = (opus_int)silk_LSHIFT( silk_DIV32_16( CCmax, nb_subfr ), 2 ); + silk_assert( *LTPCorr_Q15 >= 0 ); + + if( Fs_kHz > 8 ) { + /* Search in original signal */ + + CBimax_old = CBimax; + /* Compensate for decimation */ + silk_assert( lag == silk_SAT16( lag ) ); + if( Fs_kHz == 12 ) { + lag = silk_RSHIFT( silk_SMULBB( lag, 3 ), 1 ); + } else if( Fs_kHz == 16 ) { + lag = silk_LSHIFT( lag, 1 ); + } else { + lag = silk_SMULBB( lag, 3 ); + } + + lag = silk_LIMIT_int( lag, min_lag, max_lag ); + start_lag = silk_max_int( lag - 2, min_lag ); + end_lag = silk_min_int( lag + 2, max_lag ); + lag_new = lag; /* to avoid undefined lag */ + CBimax = 0; /* to avoid undefined lag */ + + CCmax = silk_int32_MIN; + /* pitch lags according to second stage */ + for( k = 0; k < nb_subfr; k++ ) { + pitch_out[ k ] = lag + 2 * silk_CB_lags_stage2[ k ][ CBimax_old ]; + } + + /* Set up codebook parameters according to complexity setting and frame length */ + if( nb_subfr == PE_MAX_NB_SUBFR ) { + nb_cbk_search = (opus_int)silk_nb_cbk_searchs_stage3[ complexity ]; + cbk_size = PE_NB_CBKS_STAGE3_MAX; + Lag_CB_ptr = &silk_CB_lags_stage3[ 0 ][ 0 ]; + } else { + nb_cbk_search = PE_NB_CBKS_STAGE3_10MS; + cbk_size = PE_NB_CBKS_STAGE3_10MS; + Lag_CB_ptr = &silk_CB_lags_stage3_10_ms[ 0 ][ 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 ); + + 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 ]; + 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++ ) { + cross_corr = 0; + energy = energy_target; + for( k = 0; k < nb_subfr; k++ ) { + cross_corr = silk_ADD32( cross_corr, + matrix_ptr( cross_corr_st3, k, j, + nb_cbk_search )[ lag_counter ] ); + energy = silk_ADD32( energy, + matrix_ptr( energies_st3, k, j, + nb_cbk_search )[ lag_counter ] ); + silk_assert( energy >= 0 ); + } + if( cross_corr > 0 ) { + CCmax_new = silk_DIV32_varQ( cross_corr, energy, 13 + 1 ); /* Q13 */ + /* Reduce depending on flatness of contour */ + diff = silk_int16_MAX - silk_MUL( contour_bias_Q15, j ); /* Q15 */ + silk_assert( diff == silk_SAT16( diff ) ); + CCmax_new = silk_SMULWB( CCmax_new, diff ); /* Q14 */ + } else { + CCmax_new = 0; + } + + if( CCmax_new > CCmax && ( d + silk_CB_lags_stage3[ 0 ][ j ] ) <= max_lag ) { + CCmax = CCmax_new; + lag_new = d; + CBimax = j; + } + } + lag_counter++; + } + + for( k = 0; k < nb_subfr; k++ ) { + pitch_out[ k ] = lag_new + matrix_ptr( Lag_CB_ptr, k, CBimax, cbk_size ); + pitch_out[ k ] = silk_LIMIT( pitch_out[ k ], min_lag, PE_MAX_LAG_MS * Fs_kHz ); + } + *lagIndex = (opus_int16)( lag_new - min_lag); + *contourIndex = (opus_int8)CBimax; + } else { /* Fs_kHz == 8 */ + /* Save Lags */ + for( k = 0; k < nb_subfr; k++ ) { + pitch_out[ k ] = lag + matrix_ptr( Lag_CB_ptr, k, CBimax, cbk_size ); + pitch_out[ k ] = silk_LIMIT( pitch_out[ k ], MIN_LAG_8KHZ, PE_MAX_LAG_MS * 8 ); + } + *lagIndex = (opus_int16)( lag - MIN_LAG_8KHZ ); + *contourIndex = (opus_int8)CBimax; + } + celt_assert( *lagIndex >= 0 ); + /* return as voiced */ + RESTORE_STACK; + return 0; +} + +/*********************************************************************** + * Calculates the correlations used in stage 3 search. In order to cover + * the whole lag codebook for all the searched offset lags (lag +- 2), + * the following correlations are needed in each sub frame: + * + * sf1: lag range [-8,...,7] total 16 correlations + * sf2: lag range [-4,...,4] total 9 correlations + * sf3: lag range [-3,....4] total 8 correltions + * sf4: lag range [-6,....8] total 15 correlations + * + * In total 48 correlations. The direct implementation computed in worst + * case 4*12*5 = 240 correlations, but more likely around 120. + ***********************************************************************/ +static void silk_P_Ana_calc_corr_st3( + silk_pe_stage3_vals cross_corr_st3[], /* O 3 DIM correlation array */ + const opus_int16 frame[], /* I vector to correlate */ + opus_int start_lag, /* I lag offset to search around */ + opus_int sf_length, /* I length of a 5 ms subframe */ + opus_int nb_subfr, /* I number of subframes */ + opus_int complexity, /* I Complexity setting */ + int arch /* I Run-time architecture */ +) +{ + const opus_int16 *target_ptr; + opus_int i, j, k, lag_counter, lag_low, lag_high; + opus_int nb_cbk_search, delta, idx, cbk_size; + VARDECL( opus_int32, scratch_mem ); + VARDECL( opus_int32, xcorr32 ); + 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 ); + + if( nb_subfr == PE_MAX_NB_SUBFR ) { + Lag_range_ptr = &silk_Lag_range_stage3[ complexity ][ 0 ][ 0 ]; + Lag_CB_ptr = &silk_CB_lags_stage3[ 0 ][ 0 ]; + 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); + 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; + cbk_size = PE_NB_CBKS_STAGE3_10MS; + } + ALLOC( scratch_mem, SCRATCH_SIZE, opus_int32 ); + ALLOC( xcorr32, SCRATCH_SIZE, opus_int32 ); + + target_ptr = &frame[ silk_LSHIFT( sf_length, 2 ) ]; /* Pointer to middle of frame */ + for( k = 0; k < nb_subfr; k++ ) { + lag_counter = 0; + + /* 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); + 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 ); + scratch_mem[ lag_counter ] = xcorr32[ lag_high - j ]; + lag_counter++; + } + + delta = matrix_ptr( Lag_range_ptr, k, 0, 2 ); + for( i = 0; i < nb_cbk_search; i++ ) { + /* Fill out the 3 dim array that stores the correlations for */ + /* each code_book vector for each start lag */ + idx = matrix_ptr( Lag_CB_ptr, k, i, cbk_size ) - delta; + for( j = 0; j < PE_NB_STAGE3_LAGS; j++ ) { + silk_assert( idx + j < SCRATCH_SIZE ); + silk_assert( idx + j < lag_counter ); + matrix_ptr( cross_corr_st3, k, i, nb_cbk_search )[ j ] = + scratch_mem[ idx + j ]; + } + } + target_ptr += sf_length; + } + RESTORE_STACK; +} + +/********************************************************************/ +/* Calculate the energies for first two subframes. The energies are */ +/* calculated recursively. */ +/********************************************************************/ +static void silk_P_Ana_calc_energy_st3( + silk_pe_stage3_vals energies_st3[], /* O 3 DIM energy array */ + const opus_int16 frame[], /* I vector to calc energy in */ + opus_int start_lag, /* I lag offset to search around */ + opus_int sf_length, /* I length of one 5 ms subframe */ + opus_int nb_subfr, /* I number of subframes */ + opus_int complexity, /* I Complexity setting */ + int arch /* I Run-time architecture */ +) +{ + const opus_int16 *target_ptr, *basis_ptr; + opus_int32 energy; + opus_int k, i, j, lag_counter; + opus_int nb_cbk_search, delta, idx, cbk_size, lag_diff; + VARDECL( opus_int32, scratch_mem ); + 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 ); + + if( nb_subfr == PE_MAX_NB_SUBFR ) { + Lag_range_ptr = &silk_Lag_range_stage3[ complexity ][ 0 ][ 0 ]; + Lag_CB_ptr = &silk_CB_lags_stage3[ 0 ][ 0 ]; + 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); + 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; + cbk_size = PE_NB_CBKS_STAGE3_10MS; + } + ALLOC( scratch_mem, SCRATCH_SIZE, opus_int32 ); + + target_ptr = &frame[ silk_LSHIFT( sf_length, 2 ) ]; + for( k = 0; k < nb_subfr; k++ ) { + lag_counter = 0; + + /* Calculate the energy for first lag */ + basis_ptr = target_ptr - ( start_lag + matrix_ptr( Lag_range_ptr, k, 0, 2 ) ); + energy = silk_inner_prod_aligned( basis_ptr, basis_ptr, sf_length, arch ); + silk_assert( energy >= 0 ); + scratch_mem[ lag_counter ] = energy; + lag_counter++; + + lag_diff = ( matrix_ptr( Lag_range_ptr, k, 1, 2 ) - matrix_ptr( Lag_range_ptr, k, 0, 2 ) + 1 ); + for( i = 1; i < lag_diff; i++ ) { + /* remove part outside new window */ + energy -= silk_SMULBB( basis_ptr[ sf_length - i ], basis_ptr[ sf_length - i ] ); + silk_assert( energy >= 0 ); + + /* add part that comes into window */ + energy = silk_ADD_SAT32( energy, silk_SMULBB( basis_ptr[ -i ], basis_ptr[ -i ] ) ); + silk_assert( energy >= 0 ); + silk_assert( lag_counter < SCRATCH_SIZE ); + scratch_mem[ lag_counter ] = energy; + lag_counter++; + } + + delta = matrix_ptr( Lag_range_ptr, k, 0, 2 ); + for( i = 0; i < nb_cbk_search; i++ ) { + /* Fill out the 3 dim array that stores the correlations for */ + /* each code_book vector for each start lag */ + idx = matrix_ptr( Lag_CB_ptr, k, i, cbk_size ) - delta; + for( j = 0; j < PE_NB_STAGE3_LAGS; j++ ) { + silk_assert( idx + j < SCRATCH_SIZE ); + silk_assert( idx + j < lag_counter ); + matrix_ptr( energies_st3, k, i, nb_cbk_search )[ j ] = + scratch_mem[ idx + j ]; + silk_assert( + matrix_ptr( energies_st3, k, i, nb_cbk_search )[ j ] >= 0 ); + } + } + target_ptr += sf_length; + } + RESTORE_STACK; +} diff --git a/media/libopus/silk/fixed/process_gains_FIX.c b/media/libopus/silk/fixed/process_gains_FIX.c new file mode 100644 index 0000000000..05aba31788 --- /dev/null +++ b/media/libopus/silk/fixed/process_gains_FIX.c @@ -0,0 +1,117 @@ +/*********************************************************************** +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 "tuning_parameters.h" + +/* Processing of gains */ +void silk_process_gains_FIX( + silk_encoder_state_FIX *psEnc, /* I/O Encoder state */ + silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control */ + opus_int condCoding /* I The type of conditional coding to use */ +) +{ + silk_shape_state_FIX *psShapeSt = &psEnc->sShape; + opus_int k; + opus_int32 s_Q16, InvMaxSqrVal_Q16, gain, gain_squared, ResNrg, ResNrgPart, quant_offset_Q10; + + /* Gain reduction when LTP coding gain is high */ + if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { + /*s = -0.5f * silk_sigmoid( 0.25f * ( psEncCtrl->LTPredCodGain - 12.0f ) ); */ + s_Q16 = -silk_sigm_Q15( silk_RSHIFT_ROUND( psEncCtrl->LTPredCodGain_Q7 - SILK_FIX_CONST( 12.0, 7 ), 4 ) ); + for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { + psEncCtrl->Gains_Q16[ k ] = silk_SMLAWB( psEncCtrl->Gains_Q16[ k ], psEncCtrl->Gains_Q16[ k ], s_Q16 ); + } + } + + /* Limit the quantized signal */ + /* InvMaxSqrVal = pow( 2.0f, 0.33f * ( 21.0f - SNR_dB ) ) / subfr_length; */ + InvMaxSqrVal_Q16 = silk_DIV32_16( silk_log2lin( + silk_SMULWB( SILK_FIX_CONST( 21 + 16 / 0.33, 7 ) - psEnc->sCmn.SNR_dB_Q7, SILK_FIX_CONST( 0.33, 16 ) ) ), psEnc->sCmn.subfr_length ); + + for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { + /* Soft limit on ratio residual energy and squared gains */ + ResNrg = psEncCtrl->ResNrg[ k ]; + ResNrgPart = silk_SMULWW( ResNrg, InvMaxSqrVal_Q16 ); + if( psEncCtrl->ResNrgQ[ k ] > 0 ) { + ResNrgPart = silk_RSHIFT_ROUND( ResNrgPart, psEncCtrl->ResNrgQ[ k ] ); + } else { + if( ResNrgPart >= silk_RSHIFT( silk_int32_MAX, -psEncCtrl->ResNrgQ[ k ] ) ) { + ResNrgPart = silk_int32_MAX; + } else { + ResNrgPart = silk_LSHIFT( ResNrgPart, -psEncCtrl->ResNrgQ[ k ] ); + } + } + gain = psEncCtrl->Gains_Q16[ k ]; + gain_squared = silk_ADD_SAT32( ResNrgPart, silk_SMMUL( gain, gain ) ); + if( gain_squared < silk_int16_MAX ) { + /* recalculate with higher precision */ + gain_squared = silk_SMLAWW( silk_LSHIFT( ResNrgPart, 16 ), gain, gain ); + silk_assert( gain_squared > 0 ); + gain = silk_SQRT_APPROX( gain_squared ); /* Q8 */ + gain = silk_min( gain, silk_int32_MAX >> 8 ); + psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT_SAT32( gain, 8 ); /* Q16 */ + } else { + gain = silk_SQRT_APPROX( gain_squared ); /* Q0 */ + gain = silk_min( gain, silk_int32_MAX >> 16 ); + psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT_SAT32( gain, 16 ); /* Q16 */ + } + } + + /* Save unquantized gains and gain Index */ + silk_memcpy( psEncCtrl->GainsUnq_Q16, psEncCtrl->Gains_Q16, psEnc->sCmn.nb_subfr * sizeof( opus_int32 ) ); + psEncCtrl->lastGainIndexPrev = psShapeSt->LastGainIndex; + + /* Quantize gains */ + silk_gains_quant( psEnc->sCmn.indices.GainsIndices, psEncCtrl->Gains_Q16, + &psShapeSt->LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr ); + + /* Set quantizer offset for voiced signals. Larger offset when LTP coding gain is low or tilt is high (ie low-pass) */ + if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { + if( psEncCtrl->LTPredCodGain_Q7 + silk_RSHIFT( psEnc->sCmn.input_tilt_Q15, 8 ) > SILK_FIX_CONST( 1.0, 7 ) ) { + psEnc->sCmn.indices.quantOffsetType = 0; + } else { + psEnc->sCmn.indices.quantOffsetType = 1; + } + } + + /* Quantizer boundary adjustment */ + quant_offset_Q10 = silk_Quantization_Offsets_Q10[ psEnc->sCmn.indices.signalType >> 1 ][ psEnc->sCmn.indices.quantOffsetType ]; + psEncCtrl->Lambda_Q10 = SILK_FIX_CONST( LAMBDA_OFFSET, 10 ) + + silk_SMULBB( SILK_FIX_CONST( LAMBDA_DELAYED_DECISIONS, 10 ), psEnc->sCmn.nStatesDelayedDecision ) + + silk_SMULWB( SILK_FIX_CONST( LAMBDA_SPEECH_ACT, 18 ), psEnc->sCmn.speech_activity_Q8 ) + + silk_SMULWB( SILK_FIX_CONST( LAMBDA_INPUT_QUALITY, 12 ), psEncCtrl->input_quality_Q14 ) + + silk_SMULWB( SILK_FIX_CONST( LAMBDA_CODING_QUALITY, 12 ), psEncCtrl->coding_quality_Q14 ) + + silk_SMULWB( SILK_FIX_CONST( LAMBDA_QUANT_OFFSET, 16 ), quant_offset_Q10 ); + + silk_assert( psEncCtrl->Lambda_Q10 > 0 ); + silk_assert( psEncCtrl->Lambda_Q10 < SILK_FIX_CONST( 2, 10 ) ); +} diff --git a/media/libopus/silk/fixed/regularize_correlations_FIX.c b/media/libopus/silk/fixed/regularize_correlations_FIX.c new file mode 100644 index 0000000000..a2836b05f4 --- /dev/null +++ b/media/libopus/silk/fixed/regularize_correlations_FIX.c @@ -0,0 +1,47 @@ +/*********************************************************************** +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" + +/* 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 */ +) +{ + opus_int i; + for( i = 0; i < D; i++ ) { + matrix_ptr( &XX[ 0 ], i, i, D ) = silk_ADD32( matrix_ptr( &XX[ 0 ], i, i, D ), noise ); + } + xx[ 0 ] += noise; +} diff --git a/media/libopus/silk/fixed/residual_energy16_FIX.c b/media/libopus/silk/fixed/residual_energy16_FIX.c new file mode 100644 index 0000000000..7f130f3d3d --- /dev/null +++ b/media/libopus/silk/fixed/residual_energy16_FIX.c @@ -0,0 +1,103 @@ +/*********************************************************************** +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" + +/* Residual energy: nrg = wxx - 2 * wXx * c + c' * wXX * c */ +opus_int32 silk_residual_energy16_covar_FIX( + const opus_int16 *c, /* I Prediction vector */ + const opus_int32 *wXX, /* I Correlation matrix */ + const opus_int32 *wXx, /* I Correlation vector */ + opus_int32 wxx, /* I Signal energy */ + opus_int D, /* I Dimension */ + opus_int cQ /* I Q value for c vector 0 - 15 */ +) +{ + opus_int i, j, lshifts, Qxtra; + opus_int32 c_max, w_max, tmp, tmp2, nrg; + opus_int cn[ MAX_MATRIX_SIZE ]; + const opus_int32 *pRow; + + /* Safety checks */ + celt_assert( D >= 0 ); + celt_assert( D <= 16 ); + celt_assert( cQ > 0 ); + celt_assert( cQ < 16 ); + + lshifts = 16 - cQ; + Qxtra = lshifts; + + c_max = 0; + for( i = 0; i < D; i++ ) { + c_max = silk_max_32( c_max, silk_abs( (opus_int32)c[ i ] ) ); + } + Qxtra = silk_min_int( Qxtra, silk_CLZ32( c_max ) - 17 ); + + w_max = silk_max_32( wXX[ 0 ], wXX[ D * D - 1 ] ); + Qxtra = silk_min_int( Qxtra, silk_CLZ32( silk_MUL( D, silk_RSHIFT( silk_SMULWB( w_max, c_max ), 4 ) ) ) - 5 ); + Qxtra = silk_max_int( Qxtra, 0 ); + for( i = 0; i < D; i++ ) { + cn[ i ] = silk_LSHIFT( ( opus_int )c[ i ], Qxtra ); + silk_assert( silk_abs(cn[i]) <= ( silk_int16_MAX + 1 ) ); /* Check that silk_SMLAWB can be used */ + } + lshifts -= Qxtra; + + /* Compute wxx - 2 * wXx * c */ + tmp = 0; + for( i = 0; i < D; i++ ) { + tmp = silk_SMLAWB( tmp, wXx[ i ], cn[ i ] ); + } + nrg = silk_RSHIFT( wxx, 1 + lshifts ) - tmp; /* Q: -lshifts - 1 */ + + /* Add c' * wXX * c, assuming wXX is symmetric */ + tmp2 = 0; + for( i = 0; i < D; i++ ) { + tmp = 0; + pRow = &wXX[ i * D ]; + for( j = i + 1; j < D; j++ ) { + tmp = silk_SMLAWB( tmp, pRow[ j ], cn[ j ] ); + } + tmp = silk_SMLAWB( tmp, silk_RSHIFT( pRow[ i ], 1 ), cn[ i ] ); + tmp2 = silk_SMLAWB( tmp2, tmp, cn[ i ] ); + } + nrg = silk_ADD_LSHIFT32( nrg, tmp2, lshifts ); /* Q: -lshifts - 1 */ + + /* Keep one bit free always, because we add them for LSF interpolation */ + if( nrg < 1 ) { + nrg = 1; + } else if( nrg > silk_RSHIFT( silk_int32_MAX, lshifts + 2 ) ) { + nrg = silk_int32_MAX >> 1; + } else { + nrg = silk_LSHIFT( nrg, lshifts + 1 ); /* Q0 */ + } + return nrg; + +} diff --git a/media/libopus/silk/fixed/residual_energy_FIX.c b/media/libopus/silk/fixed/residual_energy_FIX.c new file mode 100644 index 0000000000..6c7cade9a0 --- /dev/null +++ b/media/libopus/silk/fixed/residual_energy_FIX.c @@ -0,0 +1,98 @@ +/*********************************************************************** +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" + +/* Calculates residual energies of input subframes where all subframes have LPC_order */ +/* of preceding samples */ +void silk_residual_energy_FIX( + opus_int32 nrgs[ MAX_NB_SUBFR ], /* O Residual energy per subframe */ + opus_int nrgsQ[ MAX_NB_SUBFR ], /* O Q value per subframe */ + const opus_int16 x[], /* I Input signal */ + opus_int16 a_Q12[ 2 ][ MAX_LPC_ORDER ], /* I AR coefs for each frame half */ + const opus_int32 gains[ MAX_NB_SUBFR ], /* I Quantization gains */ + const opus_int subfr_length, /* I Subframe length */ + const opus_int nb_subfr, /* I Number of subframes */ + const opus_int LPC_order, /* I LPC order */ + int arch /* I Run-time architecture */ +) +{ + opus_int offset, i, j, rshift, lz1, lz2; + opus_int16 *LPC_res_ptr; + VARDECL( opus_int16, LPC_res ); + const opus_int16 *x_ptr; + opus_int32 tmp32; + SAVE_STACK; + + x_ptr = x; + offset = LPC_order + subfr_length; + + /* 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 ); + 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 ); + + /* Point to first subframe of the just calculated LPC residual signal */ + LPC_res_ptr = LPC_res + LPC_order; + for( j = 0; j < ( MAX_NB_SUBFR >> 1 ); j++ ) { + /* Measure subframe energy */ + silk_sum_sqr_shift( &nrgs[ i * ( MAX_NB_SUBFR >> 1 ) + j ], &rshift, LPC_res_ptr, subfr_length ); + + /* Set Q values for the measured energy */ + nrgsQ[ i * ( MAX_NB_SUBFR >> 1 ) + j ] = -rshift; + + /* Move to next subframe */ + LPC_res_ptr += offset; + } + /* Move to next frame half */ + x_ptr += ( MAX_NB_SUBFR >> 1 ) * offset; + } + + /* Apply the squared subframe gains */ + for( i = 0; i < nb_subfr; i++ ) { + /* Fully upscale gains and energies */ + lz1 = silk_CLZ32( nrgs[ i ] ) - 1; + lz2 = silk_CLZ32( gains[ i ] ) - 1; + + tmp32 = silk_LSHIFT32( gains[ i ], lz2 ); + + /* Find squared gains */ + tmp32 = silk_SMMUL( tmp32, tmp32 ); /* Q( 2 * lz2 - 32 )*/ + + /* Scale energies */ + nrgs[ i ] = silk_SMMUL( tmp32, silk_LSHIFT32( nrgs[ i ], lz1 ) ); /* Q( nrgsQ[ i ] + lz1 + 2 * lz2 - 32 - 32 )*/ + nrgsQ[ i ] += lz1 + 2 * lz2 - 32 - 32; + } + RESTORE_STACK; +} diff --git a/media/libopus/silk/fixed/schur64_FIX.c b/media/libopus/silk/fixed/schur64_FIX.c new file mode 100644 index 0000000000..4b7e19ea59 --- /dev/null +++ b/media/libopus/silk/fixed/schur64_FIX.c @@ -0,0 +1,93 @@ +/*********************************************************************** +Copyright (c) 2006-2011, Skype Limited. All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "SigProc_FIX.h" + +/* Slower than schur(), but more accurate. */ +/* Uses SMULL(), available on armv4 */ +opus_int32 silk_schur64( /* O returns residual energy */ + opus_int32 rc_Q16[], /* O Reflection coefficients [order] Q16 */ + const opus_int32 c[], /* I Correlations [order+1] */ + opus_int32 order /* I Prediction order */ +) +{ + opus_int k, n; + 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 ); + + /* Check for invalid input */ + if( c[ 0 ] <= 0 ) { + silk_memset( rc_Q16, 0, order * sizeof( opus_int32 ) ); + return 0; + } + + k = 0; + do { + C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ]; + } while( ++k <= order ); + + for( k = 0; k < order; k++ ) { + /* Check that we won't be getting an unstable rc, otherwise stop here. */ + if (silk_abs_int32(C[ k + 1 ][ 0 ]) >= C[ 0 ][ 1 ]) { + if ( C[ k + 1 ][ 0 ] > 0 ) { + rc_Q16[ k ] = -SILK_FIX_CONST( .99f, 16 ); + } else { + rc_Q16[ k ] = SILK_FIX_CONST( .99f, 16 ); + } + k++; + break; + } + + /* Get reflection coefficient: divide two Q30 values and get result in Q31 */ + rc_tmp_Q31 = silk_DIV32_varQ( -C[ k + 1 ][ 0 ], C[ 0 ][ 1 ], 31 ); + + /* Save the output */ + rc_Q16[ k ] = silk_RSHIFT_ROUND( rc_tmp_Q31, 15 ); + + /* Update correlations */ + for( n = 0; n < order - k; n++ ) { + Ctmp1_Q30 = C[ n + k + 1 ][ 0 ]; + Ctmp2_Q30 = C[ n ][ 1 ]; + + /* Multiply and add the highest int32 */ + C[ n + k + 1 ][ 0 ] = Ctmp1_Q30 + silk_SMMUL( silk_LSHIFT( Ctmp2_Q30, 1 ), rc_tmp_Q31 ); + C[ n ][ 1 ] = Ctmp2_Q30 + silk_SMMUL( silk_LSHIFT( Ctmp1_Q30, 1 ), rc_tmp_Q31 ); + } + } + + for(; k < order; k++ ) { + rc_Q16[ k ] = 0; + } + + return silk_max_32( 1, C[ 0 ][ 1 ] ); +} diff --git a/media/libopus/silk/fixed/schur_FIX.c b/media/libopus/silk/fixed/schur_FIX.c new file mode 100644 index 0000000000..2840f6b1aa --- /dev/null +++ b/media/libopus/silk/fixed/schur_FIX.c @@ -0,0 +1,107 @@ +/*********************************************************************** +Copyright (c) 2006-2011, Skype Limited. All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "SigProc_FIX.h" + +/* Faster than schur64(), but much less accurate. */ +/* uses SMLAWB(), requiring armv5E and higher. */ +opus_int32 silk_schur( /* O Returns residual energy */ + opus_int16 *rc_Q15, /* O reflection coefficients [order] Q15 */ + const opus_int32 *c, /* I correlations [order+1] */ + const opus_int32 order /* I prediction order */ +) +{ + opus_int k, n, lz; + 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 ); + + /* 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 { + 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 { + C[ k ][ 0 ] = C[ k ][ 1 ] = silk_LSHIFT( c[ k ], lz ); + } while( ++k <= order ); + } else { + /* No need to shift */ + do { + C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ]; + } while( ++k <= order ); + } + + for( k = 0; k < order; k++ ) { + /* Check that we won't be getting an unstable rc, otherwise stop here. */ + if (silk_abs_int32(C[ k + 1 ][ 0 ]) >= C[ 0 ][ 1 ]) { + if ( C[ k + 1 ][ 0 ] > 0 ) { + rc_Q15[ k ] = -SILK_FIX_CONST( .99f, 15 ); + } else { + rc_Q15[ k ] = SILK_FIX_CONST( .99f, 15 ); + } + k++; + break; + } + + /* Get reflection coefficient */ + rc_tmp_Q15 = -silk_DIV32_16( C[ k + 1 ][ 0 ], silk_max_32( silk_RSHIFT( C[ 0 ][ 1 ], 15 ), 1 ) ); + + /* Clip (shouldn't happen for properly conditioned inputs) */ + rc_tmp_Q15 = silk_SAT16( rc_tmp_Q15 ); + + /* Store */ + rc_Q15[ k ] = (opus_int16)rc_tmp_Q15; + + /* Update correlations */ + for( n = 0; n < order - k; n++ ) { + Ctmp1 = C[ n + k + 1 ][ 0 ]; + Ctmp2 = C[ n ][ 1 ]; + C[ n + k + 1 ][ 0 ] = silk_SMLAWB( Ctmp1, silk_LSHIFT( Ctmp2, 1 ), rc_tmp_Q15 ); + C[ n ][ 1 ] = silk_SMLAWB( Ctmp2, silk_LSHIFT( Ctmp1, 1 ), rc_tmp_Q15 ); + } + } + + for(; k < order; k++ ) { + rc_Q15[ k ] = 0; + } + + /* return residual energy */ + return silk_max_32( 1, C[ 0 ][ 1 ] ); +} diff --git a/media/libopus/silk/fixed/structs_FIX.h b/media/libopus/silk/fixed/structs_FIX.h new file mode 100644 index 0000000000..2774a97b24 --- /dev/null +++ b/media/libopus/silk/fixed/structs_FIX.h @@ -0,0 +1,116 @@ +/*********************************************************************** +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 SILK_STRUCTS_FIX_H +#define SILK_STRUCTS_FIX_H + +#include "typedef.h" +#include "main.h" +#include "structs.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + +/********************************/ +/* Noise shaping analysis state */ +/********************************/ +typedef struct { + opus_int8 LastGainIndex; + opus_int32 HarmBoost_smth_Q16; + opus_int32 HarmShapeGain_smth_Q16; + opus_int32 Tilt_smth_Q16; +} silk_shape_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 */ + + /* 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; + +/************************/ +/* Encoder control FIX */ +/************************/ +typedef struct { + /* Prediction and coding parameters */ + opus_int32 Gains_Q16[ MAX_NB_SUBFR ]; + silk_DWORD_ALIGN opus_int16 PredCoef_Q12[ 2 ][ MAX_LPC_ORDER ]; + opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ]; + opus_int LTP_scale_Q14; + opus_int pitchL[ MAX_NB_SUBFR ]; + + /* Noise shaping parameters */ + /* Testing */ + silk_DWORD_ALIGN opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; + opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ]; /* Packs two int16 coefficients per int32 value */ + opus_int Tilt_Q14[ MAX_NB_SUBFR ]; + opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ]; + opus_int Lambda_Q10; + opus_int input_quality_Q14; + opus_int coding_quality_Q14; + + /* measures */ + opus_int32 predGain_Q16; + opus_int LTPredCodGain_Q7; + opus_int32 ResNrg[ MAX_NB_SUBFR ]; /* Residual energy per subframe */ + opus_int ResNrgQ[ MAX_NB_SUBFR ]; /* Q domain for the residual energy > 0 */ + + /* Parameters for CBR mode */ + opus_int32 GainsUnq_Q16[ MAX_NB_SUBFR ]; + opus_int8 lastGainIndexPrev; +} silk_encoder_control_FIX; + +/************************/ +/* Encoder Super Struct */ +/************************/ +typedef struct { + silk_encoder_state_FIX state_Fxx[ ENCODER_NUM_CHANNELS ]; + stereo_enc_state sStereo; + opus_int32 nBitsUsedLBRR; + opus_int32 nBitsExceeded; + opus_int nChannelsAPI; + opus_int nChannelsInternal; + opus_int nPrevChannelsInternal; + opus_int timeSinceSwitchAllowed_ms; + opus_int allowBandwidthSwitch; + opus_int prev_decode_only_middle; +} silk_encoder; + + +#ifdef __cplusplus +} +#endif + +#endif diff --git a/media/libopus/silk/fixed/vector_ops_FIX.c b/media/libopus/silk/fixed/vector_ops_FIX.c new file mode 100644 index 0000000000..dcf84070a6 --- /dev/null +++ b/media/libopus/silk/fixed/vector_ops_FIX.c @@ -0,0 +1,102 @@ +/*********************************************************************** +Copyright (c) 2006-2011, Skype Limited. All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "SigProc_FIX.h" +#include "pitch.h" + +/* Copy and multiply a vector by a constant */ +void silk_scale_copy_vector16( + opus_int16 *data_out, + const opus_int16 *data_in, + opus_int32 gain_Q16, /* I Gain in Q16 */ + const opus_int dataSize /* I Length */ +) +{ + opus_int i; + opus_int32 tmp32; + + for( i = 0; i < dataSize; i++ ) { + tmp32 = silk_SMULWB( gain_Q16, data_in[ i ] ); + data_out[ i ] = (opus_int16)silk_CHECK_FIT16( tmp32 ); + } +} + +/* Multiply a vector by a constant */ +void silk_scale_vector32_Q26_lshift_18( + opus_int32 *data1, /* I/O Q0/Q18 */ + opus_int32 gain_Q26, /* I Q26 */ + opus_int dataSize /* I length */ +) +{ + opus_int i; + + for( i = 0; i < dataSize; i++ ) { + data1[ i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( silk_SMULL( data1[ i ], gain_Q26 ), 8 ) ); /* OUTPUT: Q18 */ + } +} + +/* sum = for(i=0;i<len;i++)inVec1[i]*inVec2[i]; --- inner product */ +/* Note for ARM asm: */ +/* * inVec1 and inVec2 should be at least 2 byte aligned. */ +/* * len should be positive 16bit integer. */ +/* * only when len>6, memory access can be reduced by half. */ +opus_int32 silk_inner_prod_aligned( + const opus_int16 *const inVec1, /* I input vector 1 */ + const opus_int16 *const inVec2, /* I input vector 2 */ + const opus_int len, /* I vector lengths */ + int arch /* I Run-time architecture */ +) +{ +#ifdef FIXED_POINT + return celt_inner_prod(inVec1, inVec2, len, arch); +#else + opus_int i; + opus_int32 sum = 0; + for( i = 0; i < len; i++ ) { + sum = silk_SMLABB( sum, inVec1[ i ], inVec2[ i ] ); + } + return sum; +#endif +} + +opus_int64 silk_inner_prod16_c( + const opus_int16 *inVec1, /* I input vector 1 */ + const opus_int16 *inVec2, /* I input vector 2 */ + const opus_int len /* I vector lengths */ +) +{ + opus_int i; + opus_int64 sum = 0; + for( i = 0; i < len; i++ ) { + sum = silk_SMLALBB( sum, inVec1[ i ], inVec2[ i ] ); + } + return sum; +} diff --git a/media/libopus/silk/fixed/warped_autocorrelation_FIX.c b/media/libopus/silk/fixed/warped_autocorrelation_FIX.c new file mode 100644 index 0000000000..5c79553bc0 --- /dev/null +++ b/media/libopus/silk/fixed/warped_autocorrelation_FIX.c @@ -0,0 +1,92 @@ +/*********************************************************************** +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" + +#if defined(MIPSr1_ASM) +#include "mips/warped_autocorrelation_FIX_mipsr1.h" +#endif + + +/* Autocorrelations for a warped frequency axis */ +#ifndef OVERRIDE_silk_warped_autocorrelation_FIX_c +void silk_warped_autocorrelation_FIX_c( + opus_int32 *corr, /* O Result [order + 1] */ + opus_int *scale, /* O Scaling of the correlation vector */ + const opus_int16 *input, /* I Input data to correlate */ + const opus_int warping_Q16, /* I Warping coefficient */ + const opus_int length, /* I Length of input */ + const opus_int order /* I Correlation order (even) */ +) +{ + opus_int n, i, lsh; + opus_int32 tmp1_QS, tmp2_QS; + opus_int32 state_QS[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; + opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; + + /* Order must be even */ + celt_assert( ( order & 1 ) == 0 ); + silk_assert( 2 * QS - QC >= 0 ); + + /* Loop over samples */ + for( n = 0; n < length; n++ ) { + tmp1_QS = silk_LSHIFT32( (opus_int32)input[ n ], QS ); + /* Loop over allpass sections */ + for( i = 0; i < order; i += 2 ) { + /* Output of allpass section */ + tmp2_QS = silk_SMLAWB( state_QS[ i ], state_QS[ i + 1 ] - tmp1_QS, warping_Q16 ); + state_QS[ i ] = tmp1_QS; + corr_QC[ i ] += silk_RSHIFT64( silk_SMULL( tmp1_QS, state_QS[ 0 ] ), 2 * QS - QC ); + /* Output of allpass section */ + tmp1_QS = silk_SMLAWB( state_QS[ i + 1 ], state_QS[ i + 2 ] - tmp2_QS, warping_Q16 ); + state_QS[ i + 1 ] = tmp2_QS; + corr_QC[ i + 1 ] += silk_RSHIFT64( silk_SMULL( tmp2_QS, state_QS[ 0 ] ), 2 * QS - QC ); + } + state_QS[ order ] = tmp1_QS; + corr_QC[ order ] += silk_RSHIFT64( silk_SMULL( tmp1_QS, state_QS[ 0 ] ), 2 * QS - QC ); + } + + lsh = silk_CLZ64( corr_QC[ 0 ] ) - 35; + lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC ); + *scale = -( QC + lsh ); + silk_assert( *scale >= -30 && *scale <= 12 ); + if( lsh >= 0 ) { + for( i = 0; i < order + 1; i++ ) { + corr[ i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QC[ i ], lsh ) ); + } + } else { + for( i = 0; i < order + 1; i++ ) { + corr[ i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QC[ i ], -lsh ) ); + } + } + silk_assert( corr_QC[ 0 ] >= 0 ); /* If breaking, decrease QC*/ +} +#endif /* OVERRIDE_silk_warped_autocorrelation_FIX_c */ diff --git a/media/libopus/silk/fixed/x86/burg_modified_FIX_sse4_1.c b/media/libopus/silk/fixed/x86/burg_modified_FIX_sse4_1.c new file mode 100644 index 0000000000..e58bf079e5 --- /dev/null +++ b/media/libopus/silk/fixed/x86/burg_modified_FIX_sse4_1.c @@ -0,0 +1,400 @@ +/* Copyright (c) 2014-2020, Cisco Systems, INC + Written by XiangMingZhu WeiZhou MinPeng YanWang FrancisQuiers + + 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 "SigProc_FIX.h" +#include "define.h" +#include "tuning_parameters.h" +#include "pitch.h" +#include "celt/x86/x86cpu.h" + +#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 MIN_RSHIFTS -16 +#define MAX_RSHIFTS (32 - QA) + +/* Compute reflection coefficients from input signal */ +void silk_burg_modified_sse4_1( + opus_int32 *res_nrg, /* O Residual energy */ + opus_int *res_nrg_Q, /* O Residual energy Q value */ + opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ + const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */ + const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ + const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ + const opus_int nb_subfr, /* I Number of subframes stacked in x */ + const opus_int D, /* I Order */ + int arch /* I Run-time architecture */ +) +{ + opus_int k, n, s, lz, rshifts, reached_max_gain; + opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; + const opus_int16 *x_ptr; + opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; + opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; + opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; + opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; + opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; + opus_int32 xcorr[ SILK_MAX_ORDER_LPC ]; + opus_int64 C0_64; + + __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 ); + + /* Compute autocorrelations, added over subframes */ + C0_64 = silk_inner_prod16( x, x, subfr_length*nb_subfr, arch ); + lz = silk_CLZ64(C0_64); + rshifts = 32 + 1 + N_BITS_HEAD_ROOM - lz; + if (rshifts > MAX_RSHIFTS) rshifts = MAX_RSHIFTS; + if (rshifts < MIN_RSHIFTS) rshifts = MIN_RSHIFTS; + + if (rshifts > 0) { + C0 = (opus_int32)silk_RSHIFT64(C0_64, rshifts ); + } else { + C0 = silk_LSHIFT32((opus_int32)C0_64, -rshifts ); + } + + CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ + silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); + if( rshifts > 0 ) { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + for( n = 1; n < D + 1; n++ ) { + C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( + silk_inner_prod16( x_ptr, x_ptr + n, subfr_length - n, arch ), rshifts ); + } + } + } else { + for( s = 0; s < nb_subfr; s++ ) { + int i; + opus_int32 d; + x_ptr = x + s * subfr_length; + celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch ); + for( n = 1; n < D + 1; n++ ) { + for ( i = n + subfr_length - D, d = 0; i < subfr_length; i++ ) + d = MAC16_16( d, x_ptr[ i ], x_ptr[ i - n ] ); + xcorr[ n - 1 ] += d; + } + for( n = 1; n < D + 1; n++ ) { + C_first_row[ n - 1 ] += silk_LSHIFT32( xcorr[ n - 1 ], -rshifts ); + } + } + } + silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); + + /* Initialize */ + CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ + + invGain_Q30 = (opus_int32)1 << 30; + reached_max_gain = 0; + for( n = 0; n < D; n++ ) { + /* Update first row of correlation matrix (without first element) */ + /* Update last row of correlation matrix (without last element, stored in reversed order) */ + /* Update C * Af */ + /* Update C * flipud(Af) (stored in reversed order) */ + if( rshifts > -2 ) { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts) */ + x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts) */ + tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16) */ + tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16) */ + for( k = 0; k < n; k++ ) { + C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ + C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ + Atmp_QA = Af_QA[ k ]; + tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ + tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16) */ + } + tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts) */ + tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts) */ + for( k = 0; k <= n; k++ ) { + CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift ) */ + CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift ) */ + } + } + } else { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts ) */ + x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts ) */ + tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17 */ + tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17 */ + + X1_3210 = _mm_set1_epi32( x1 ); + X2_3210 = _mm_set1_epi32( x2 ); + TMP1_3210 = _mm_setzero_si128(); + TMP2_3210 = _mm_setzero_si128(); + for( k = 0; k < n - 3; k += 4 ) { + PTR_3210 = OP_CVTEPI16_EPI32_M64( &x_ptr[ n - k - 1 - 3 ] ); + SUBFR_3210 = OP_CVTEPI16_EPI32_M64( &x_ptr[ subfr_length - n + k ] ); + FIRST_3210 = _mm_loadu_si128( (__m128i *)&C_first_row[ k ] ); + PTR_3210 = _mm_shuffle_epi32( PTR_3210, _MM_SHUFFLE( 0, 1, 2, 3 ) ); + LAST_3210 = _mm_loadu_si128( (__m128i *)&C_last_row[ k ] ); + ATMP_3210 = _mm_loadu_si128( (__m128i *)&Af_QA[ k ] ); + + T1_3210 = _mm_mullo_epi32( PTR_3210, X1_3210 ); + T2_3210 = _mm_mullo_epi32( SUBFR_3210, X2_3210 ); + + ATMP_3210 = _mm_srai_epi32( ATMP_3210, 7 ); + ATMP_3210 = _mm_add_epi32( ATMP_3210, CONST1 ); + ATMP_3210 = _mm_srai_epi32( ATMP_3210, 1 ); + + FIRST_3210 = _mm_add_epi32( FIRST_3210, T1_3210 ); + LAST_3210 = _mm_add_epi32( LAST_3210, T2_3210 ); + + PTR_3210 = _mm_mullo_epi32( ATMP_3210, PTR_3210 ); + SUBFR_3210 = _mm_mullo_epi32( ATMP_3210, SUBFR_3210 ); + + _mm_storeu_si128( (__m128i *)&C_first_row[ k ], FIRST_3210 ); + _mm_storeu_si128( (__m128i *)&C_last_row[ k ], LAST_3210 ); + + TMP1_3210 = _mm_add_epi32( TMP1_3210, PTR_3210 ); + TMP2_3210 = _mm_add_epi32( TMP2_3210, SUBFR_3210 ); + } + + TMP1_3210 = _mm_add_epi32( TMP1_3210, _mm_unpackhi_epi64(TMP1_3210, TMP1_3210 ) ); + TMP2_3210 = _mm_add_epi32( TMP2_3210, _mm_unpackhi_epi64(TMP2_3210, TMP2_3210 ) ); + TMP1_3210 = _mm_add_epi32( TMP1_3210, _mm_shufflelo_epi16(TMP1_3210, 0x0E ) ); + TMP2_3210 = _mm_add_epi32( TMP2_3210, _mm_shufflelo_epi16(TMP2_3210, 0x0E ) ); + + tmp1 += _mm_cvtsi128_si32( TMP1_3210 ); + tmp2 += _mm_cvtsi128_si32( TMP2_3210 ); + + for( ; k < n; k++ ) { + C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ + C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ + Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */ + /* We sometimes get overflows in the multiplications (even beyond +/- 2^32), + but they cancel each other and the real result seems to always fit in a 32-bit + signed integer. This was determined experimentally, not theoretically (unfortunately). */ + tmp1 = silk_MLA_ovflw( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */ + tmp2 = silk_MLA_ovflw( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */ + } + + tmp1 = -tmp1; /* Q17 */ + tmp2 = -tmp2; /* Q17 */ + + { + __m128i xmm_tmp1, xmm_tmp2; + __m128i xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1; + __m128i xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1; + + xmm_tmp1 = _mm_set1_epi32( tmp1 ); + xmm_tmp2 = _mm_set1_epi32( tmp2 ); + + for( k = 0; k <= n - 3; k += 4 ) { + xmm_x_ptr_n_k_x2x0 = OP_CVTEPI16_EPI32_M64( &x_ptr[ n - k - 3 ] ); + xmm_x_ptr_sub_x2x0 = OP_CVTEPI16_EPI32_M64( &x_ptr[ subfr_length - n + k - 1 ] ); + + xmm_x_ptr_n_k_x2x0 = _mm_shuffle_epi32( xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE( 0, 1, 2, 3 ) ); + + xmm_x_ptr_n_k_x2x0 = _mm_slli_epi32( xmm_x_ptr_n_k_x2x0, -rshifts - 1 ); + xmm_x_ptr_sub_x2x0 = _mm_slli_epi32( xmm_x_ptr_sub_x2x0, -rshifts - 1 ); + + /* equal shift right 4 bytes, xmm_x_ptr_n_k_x3x1 = _mm_srli_si128(xmm_x_ptr_n_k_x2x0, 4)*/ + xmm_x_ptr_n_k_x3x1 = _mm_shuffle_epi32( xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) ); + xmm_x_ptr_sub_x3x1 = _mm_shuffle_epi32( xmm_x_ptr_sub_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) ); + + xmm_x_ptr_n_k_x2x0 = _mm_mul_epi32( xmm_x_ptr_n_k_x2x0, xmm_tmp1 ); + xmm_x_ptr_n_k_x3x1 = _mm_mul_epi32( xmm_x_ptr_n_k_x3x1, xmm_tmp1 ); + xmm_x_ptr_sub_x2x0 = _mm_mul_epi32( xmm_x_ptr_sub_x2x0, xmm_tmp2 ); + xmm_x_ptr_sub_x3x1 = _mm_mul_epi32( xmm_x_ptr_sub_x3x1, xmm_tmp2 ); + + xmm_x_ptr_n_k_x2x0 = _mm_srli_epi64( xmm_x_ptr_n_k_x2x0, 16 ); + xmm_x_ptr_n_k_x3x1 = _mm_slli_epi64( xmm_x_ptr_n_k_x3x1, 16 ); + xmm_x_ptr_sub_x2x0 = _mm_srli_epi64( xmm_x_ptr_sub_x2x0, 16 ); + xmm_x_ptr_sub_x3x1 = _mm_slli_epi64( xmm_x_ptr_sub_x3x1, 16 ); + + xmm_x_ptr_n_k_x2x0 = _mm_blend_epi16( xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1, 0xCC ); + xmm_x_ptr_sub_x2x0 = _mm_blend_epi16( xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1, 0xCC ); + + X1_3210 = _mm_loadu_si128( (__m128i *)&CAf[ k ] ); + PTR_3210 = _mm_loadu_si128( (__m128i *)&CAb[ k ] ); + + X1_3210 = _mm_add_epi32( X1_3210, xmm_x_ptr_n_k_x2x0 ); + PTR_3210 = _mm_add_epi32( PTR_3210, xmm_x_ptr_sub_x2x0 ); + + _mm_storeu_si128( (__m128i *)&CAf[ k ], X1_3210 ); + _mm_storeu_si128( (__m128i *)&CAb[ k ], PTR_3210 ); + } + + for( ; k <= n; k++ ) { + CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1, + silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift ) */ + CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2, + silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */ + } + } + } + } + + /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ + tmp1 = C_first_row[ n ]; /* Q( -rshifts ) */ + tmp2 = C_last_row[ n ]; /* Q( -rshifts ) */ + num = 0; /* Q( -rshifts ) */ + nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts ) */ + for( k = 0; k < n; k++ ) { + Atmp_QA = Af_QA[ k ]; + lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1; + lz = silk_min( 32 - QA, lz ); + Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz ) */ + + tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ + tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ + num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ + nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ), + Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts ) */ + } + CAf[ n + 1 ] = tmp1; /* Q( -rshifts ) */ + CAb[ n + 1 ] = tmp2; /* Q( -rshifts ) */ + num = silk_ADD32( num, tmp2 ); /* Q( -rshifts ) */ + num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts ) */ + + /* Calculate the next order reflection (parcor) coefficient */ + if( silk_abs( num ) < nrg ) { + rc_Q31 = silk_DIV32_varQ( num, nrg, 31 ); + } else { + rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN; + } + + /* Update inverse prediction gain */ + tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); + tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 ); + if( tmp1 <= minInvGain_Q30 ) { + /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ + tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */ + rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */ + if( rc_Q31 > 0 ) { + /* Newton-Raphson iteration */ + rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */ + rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */ + if( num < 0 ) { + /* Ensure adjusted reflection coefficients has the original sign */ + rc_Q31 = -rc_Q31; + } + } + invGain_Q30 = minInvGain_Q30; + reached_max_gain = 1; + } else { + invGain_Q30 = tmp1; + } + + /* Update the AR coefficients */ + for( k = 0; k < (n + 1) >> 1; k++ ) { + tmp1 = Af_QA[ k ]; /* QA */ + tmp2 = Af_QA[ n - k - 1 ]; /* QA */ + Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA */ + Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA */ + } + Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA */ + + if( reached_max_gain ) { + /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ + for( k = n + 1; k < D; k++ ) { + Af_QA[ k ] = 0; + } + break; + } + + /* Update C * Af and C * Ab */ + for( k = 0; k <= n + 1; k++ ) { + tmp1 = CAf[ k ]; /* Q( -rshifts ) */ + tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts ) */ + CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts ) */ + CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts ) */ + } + } + + if( reached_max_gain ) { + for( k = 0; k < D; k++ ) { + /* Scale coefficients */ + A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); + } + /* Subtract energy of preceding samples from C0 */ + if( rshifts > 0 ) { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16( x_ptr, x_ptr, D, arch ), rshifts ); + } + } else { + for( s = 0; s < nb_subfr; s++ ) { + x_ptr = x + s * subfr_length; + C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D, arch ), -rshifts ); + } + } + /* Approximate residual energy */ + *res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 ); + *res_nrg_Q = -rshifts; + } else { + /* Return residual energy */ + nrg = CAf[ 0 ]; /* Q( -rshifts ) */ + tmp1 = (opus_int32)1 << 16; /* Q16 */ + for( k = 0; k < D; k++ ) { + Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16 */ + nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts ) */ + tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16 */ + A_Q16[ k ] = -Atmp1; + } + *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ), -tmp1 );/* Q( -rshifts ) */ + *res_nrg_Q = -rshifts; + } + +#ifdef OPUS_CHECK_ASM + { + opus_int32 res_nrg_c = 0; + opus_int res_nrg_Q_c = 0; + opus_int32 A_Q16_c[ MAX_LPC_ORDER ] = {0}; + + silk_burg_modified_c( + &res_nrg_c, + &res_nrg_Q_c, + A_Q16_c, + x, + minInvGain_Q30, + subfr_length, + nb_subfr, + D, + 0 + ); + + silk_assert( *res_nrg == res_nrg_c ); + silk_assert( *res_nrg_Q == res_nrg_Q_c ); + silk_assert( !memcmp( A_Q16, A_Q16_c, D * sizeof( *A_Q16 ) ) ); + } +#endif +} diff --git a/media/libopus/silk/fixed/x86/vector_ops_FIX_sse4_1.c b/media/libopus/silk/fixed/x86/vector_ops_FIX_sse4_1.c new file mode 100644 index 0000000000..a46289bbc4 --- /dev/null +++ b/media/libopus/silk/fixed/x86/vector_ops_FIX_sse4_1.c @@ -0,0 +1,93 @@ +/* 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 "SigProc_FIX.h" +#include "pitch.h" +#include "celt/x86/x86cpu.h" + +opus_int64 silk_inner_prod16_sse4_1( + const opus_int16 *inVec1, /* I input vector 1 */ + const opus_int16 *inVec2, /* I input vector 2 */ + const opus_int len /* I vector lengths */ +) +{ + opus_int i, dataSize4; + opus_int64 sum; + + __m128i xmm_prod_20, xmm_prod_31; + __m128i inVec1_3210, acc1; + __m128i inVec2_3210, acc2; + + sum = 0; + dataSize4 = len & ~3; + + acc1 = _mm_setzero_si128(); + acc2 = _mm_setzero_si128(); + + for( i = 0; i < dataSize4; i += 4 ) { + inVec1_3210 = OP_CVTEPI16_EPI32_M64( &inVec1[i + 0] ); + inVec2_3210 = OP_CVTEPI16_EPI32_M64( &inVec2[i + 0] ); + xmm_prod_20 = _mm_mul_epi32( inVec1_3210, inVec2_3210 ); + + inVec1_3210 = _mm_shuffle_epi32( inVec1_3210, _MM_SHUFFLE( 0, 3, 2, 1 ) ); + inVec2_3210 = _mm_shuffle_epi32( inVec2_3210, _MM_SHUFFLE( 0, 3, 2, 1 ) ); + xmm_prod_31 = _mm_mul_epi32( inVec1_3210, inVec2_3210 ); + + acc1 = _mm_add_epi64( acc1, xmm_prod_20 ); + acc2 = _mm_add_epi64( acc2, xmm_prod_31 ); + } + + acc1 = _mm_add_epi64( acc1, acc2 ); + + /* equal shift right 8 bytes */ + acc2 = _mm_shuffle_epi32( acc1, _MM_SHUFFLE( 0, 0, 3, 2 ) ); + acc1 = _mm_add_epi64( acc1, acc2 ); + + _mm_storel_epi64( (__m128i *)&sum, acc1 ); + + for( ; i < len; i++ ) { + sum = silk_SMLALBB( sum, inVec1[ i ], inVec2[ i ] ); + } + +#ifdef OPUS_CHECK_ASM + { + opus_int64 sum_c = silk_inner_prod16_c( inVec1, inVec2, len ); + silk_assert( sum == sum_c ); + } +#endif + + return sum; +} |