cannam@154: /* Copyright (c) 2014, Cisco Systems, INC cannam@154: Written by XiangMingZhu WeiZhou MinPeng YanWang cannam@154: cannam@154: Redistribution and use in source and binary forms, with or without cannam@154: modification, are permitted provided that the following conditions cannam@154: are met: cannam@154: cannam@154: - Redistributions of source code must retain the above copyright cannam@154: notice, this list of conditions and the following disclaimer. cannam@154: cannam@154: - Redistributions in binary form must reproduce the above copyright cannam@154: notice, this list of conditions and the following disclaimer in the cannam@154: documentation and/or other materials provided with the distribution. cannam@154: cannam@154: THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS cannam@154: ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT cannam@154: LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR cannam@154: A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER cannam@154: OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, cannam@154: EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, cannam@154: PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR cannam@154: PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF cannam@154: LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING cannam@154: NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS cannam@154: SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. cannam@154: */ cannam@154: cannam@154: #ifdef HAVE_CONFIG_H cannam@154: #include "config.h" cannam@154: #endif cannam@154: cannam@154: #include cannam@154: #include cannam@154: #include cannam@154: #include "main.h" cannam@154: #include "celt/x86/x86cpu.h" cannam@154: #include "stack_alloc.h" cannam@154: cannam@154: static OPUS_INLINE void silk_nsq_scale_states_sse4_1( cannam@154: const silk_encoder_state *psEncC, /* I Encoder State */ cannam@154: silk_nsq_state *NSQ, /* I/O NSQ state */ cannam@154: const opus_int32 x_Q3[], /* I input in Q3 */ cannam@154: opus_int32 x_sc_Q10[], /* O input scaled with 1/Gain */ cannam@154: const opus_int16 sLTP[], /* I re-whitened LTP state in Q0 */ cannam@154: opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ cannam@154: opus_int subfr, /* I subframe number */ cannam@154: const opus_int LTP_scale_Q14, /* I */ cannam@154: const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */ cannam@154: const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */ cannam@154: const opus_int signal_type /* I Signal type */ cannam@154: ); cannam@154: cannam@154: static OPUS_INLINE void silk_noise_shape_quantizer_10_16_sse4_1( cannam@154: silk_nsq_state *NSQ, /* I/O NSQ state */ cannam@154: opus_int signalType, /* I Signal type */ cannam@154: const opus_int32 x_sc_Q10[], /* I */ cannam@154: opus_int8 pulses[], /* O */ cannam@154: opus_int16 xq[], /* O */ cannam@154: opus_int32 sLTP_Q15[], /* I/O LTP state */ cannam@154: const opus_int16 a_Q12[], /* I Short term prediction coefs */ cannam@154: const opus_int16 b_Q14[], /* I Long term prediction coefs */ cannam@154: const opus_int16 AR_shp_Q13[], /* I Noise shaping AR coefs */ cannam@154: opus_int lag, /* I Pitch lag */ cannam@154: opus_int32 HarmShapeFIRPacked_Q14, /* I */ cannam@154: opus_int Tilt_Q14, /* I Spectral tilt */ cannam@154: opus_int32 LF_shp_Q14, /* I */ cannam@154: opus_int32 Gain_Q16, /* I */ cannam@154: opus_int offset_Q10, /* I */ cannam@154: opus_int length, /* I Input length */ cannam@154: opus_int32 table[][4] /* I */ cannam@154: ); cannam@154: cannam@154: void silk_NSQ_sse4_1( cannam@154: const silk_encoder_state *psEncC, /* I Encoder State */ cannam@154: silk_nsq_state *NSQ, /* I/O NSQ state */ cannam@154: SideInfoIndices *psIndices, /* I/O Quantization Indices */ cannam@154: const opus_int32 x_Q3[], /* I Prefiltered input signal */ cannam@154: opus_int8 pulses[], /* O Quantized pulse signal */ cannam@154: const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ cannam@154: const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ cannam@154: const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ cannam@154: const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ cannam@154: const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ cannam@154: const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ cannam@154: const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */ cannam@154: const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */ cannam@154: const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */ cannam@154: const opus_int LTP_scale_Q14 /* I LTP state scaling */ cannam@154: ) cannam@154: { cannam@154: opus_int k, lag, start_idx, LSF_interpolation_flag; cannam@154: const opus_int16 *A_Q12, *B_Q14, *AR_shp_Q13; cannam@154: opus_int16 *pxq; cannam@154: VARDECL( opus_int32, sLTP_Q15 ); cannam@154: VARDECL( opus_int16, sLTP ); cannam@154: opus_int32 HarmShapeFIRPacked_Q14; cannam@154: opus_int offset_Q10; cannam@154: VARDECL( opus_int32, x_sc_Q10 ); cannam@154: cannam@154: opus_int32 table[ 64 ][ 4 ]; cannam@154: opus_int32 tmp1; cannam@154: opus_int32 q1_Q10, q2_Q10, rd1_Q20, rd2_Q20; cannam@154: cannam@154: SAVE_STACK; cannam@154: cannam@154: NSQ->rand_seed = psIndices->Seed; cannam@154: cannam@154: /* Set unvoiced lag to the previous one, overwrite later for voiced */ cannam@154: lag = NSQ->lagPrev; cannam@154: cannam@154: silk_assert( NSQ->prev_gain_Q16 != 0 ); cannam@154: cannam@154: offset_Q10 = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ]; cannam@154: cannam@154: /* 0 */ cannam@154: q1_Q10 = offset_Q10; cannam@154: q2_Q10 = offset_Q10 + ( 1024 - QUANT_LEVEL_ADJUST_Q10 ); cannam@154: rd1_Q20 = q1_Q10 * Lambda_Q10; cannam@154: rd2_Q20 = q2_Q10 * Lambda_Q10; cannam@154: cannam@154: table[ 32 ][ 0 ] = q1_Q10; cannam@154: table[ 32 ][ 1 ] = q2_Q10; cannam@154: table[ 32 ][ 2 ] = 2 * (q1_Q10 - q2_Q10); cannam@154: table[ 32 ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10); cannam@154: cannam@154: /* -1 */ cannam@154: q1_Q10 = offset_Q10 - ( 1024 - QUANT_LEVEL_ADJUST_Q10 ); cannam@154: q2_Q10 = offset_Q10; cannam@154: rd1_Q20 = - q1_Q10 * Lambda_Q10; cannam@154: rd2_Q20 = q2_Q10 * Lambda_Q10; cannam@154: cannam@154: table[ 31 ][ 0 ] = q1_Q10; cannam@154: table[ 31 ][ 1 ] = q2_Q10; cannam@154: table[ 31 ][ 2 ] = 2 * (q1_Q10 - q2_Q10); cannam@154: table[ 31 ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10); cannam@154: cannam@154: /* > 0 */ cannam@154: for (k = 1; k <= 31; k++) cannam@154: { cannam@154: tmp1 = offset_Q10 + silk_LSHIFT( k, 10 ); cannam@154: cannam@154: q1_Q10 = tmp1 - QUANT_LEVEL_ADJUST_Q10; cannam@154: q2_Q10 = tmp1 - QUANT_LEVEL_ADJUST_Q10 + 1024; cannam@154: rd1_Q20 = q1_Q10 * Lambda_Q10; cannam@154: rd2_Q20 = q2_Q10 * Lambda_Q10; cannam@154: cannam@154: table[ 32 + k ][ 0 ] = q1_Q10; cannam@154: table[ 32 + k ][ 1 ] = q2_Q10; cannam@154: table[ 32 + k ][ 2 ] = 2 * (q1_Q10 - q2_Q10); cannam@154: table[ 32 + k ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10); cannam@154: } cannam@154: cannam@154: /* < -1 */ cannam@154: for (k = -32; k <= -2; k++) cannam@154: { cannam@154: tmp1 = offset_Q10 + silk_LSHIFT( k, 10 ); cannam@154: cannam@154: q1_Q10 = tmp1 + QUANT_LEVEL_ADJUST_Q10; cannam@154: q2_Q10 = tmp1 + QUANT_LEVEL_ADJUST_Q10 + 1024; cannam@154: rd1_Q20 = - q1_Q10 * Lambda_Q10; cannam@154: rd2_Q20 = - q2_Q10 * Lambda_Q10; cannam@154: cannam@154: table[ 32 + k ][ 0 ] = q1_Q10; cannam@154: table[ 32 + k ][ 1 ] = q2_Q10; cannam@154: table[ 32 + k ][ 2 ] = 2 * (q1_Q10 - q2_Q10); cannam@154: table[ 32 + k ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10); cannam@154: } cannam@154: cannam@154: if( psIndices->NLSFInterpCoef_Q2 == 4 ) { cannam@154: LSF_interpolation_flag = 0; cannam@154: } else { cannam@154: LSF_interpolation_flag = 1; cannam@154: } cannam@154: cannam@154: ALLOC( sLTP_Q15, cannam@154: psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); cannam@154: ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 ); cannam@154: ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 ); cannam@154: /* Set up pointers to start of sub frame */ cannam@154: NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length; cannam@154: NSQ->sLTP_buf_idx = psEncC->ltp_mem_length; cannam@154: pxq = &NSQ->xq[ psEncC->ltp_mem_length ]; cannam@154: for( k = 0; k < psEncC->nb_subfr; k++ ) { cannam@154: A_Q12 = &PredCoef_Q12[ (( k >> 1 ) | ( 1 - LSF_interpolation_flag )) * MAX_LPC_ORDER ]; cannam@154: B_Q14 = <PCoef_Q14[ k * LTP_ORDER ]; cannam@154: AR_shp_Q13 = &AR2_Q13[ k * MAX_SHAPE_LPC_ORDER ]; cannam@154: cannam@154: /* Noise shape parameters */ cannam@154: silk_assert( HarmShapeGain_Q14[ k ] >= 0 ); cannam@154: HarmShapeFIRPacked_Q14 = silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 ); cannam@154: HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 ); cannam@154: cannam@154: NSQ->rewhite_flag = 0; cannam@154: if( psIndices->signalType == TYPE_VOICED ) { cannam@154: /* Voiced */ cannam@154: lag = pitchL[ k ]; cannam@154: cannam@154: /* Re-whitening */ cannam@154: if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) { cannam@154: /* Rewhiten with new A coefs */ cannam@154: start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2; cannam@154: celt_assert( start_idx > 0 ); cannam@154: cannam@154: silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ], cannam@154: A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch ); cannam@154: cannam@154: NSQ->rewhite_flag = 1; cannam@154: NSQ->sLTP_buf_idx = psEncC->ltp_mem_length; cannam@154: } cannam@154: } cannam@154: cannam@154: silk_nsq_scale_states_sse4_1( psEncC, NSQ, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType ); cannam@154: cannam@154: if ( opus_likely( ( 10 == psEncC->shapingLPCOrder ) && ( 16 == psEncC->predictLPCOrder) ) ) cannam@154: { cannam@154: silk_noise_shape_quantizer_10_16_sse4_1( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14, cannam@154: AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ], cannam@154: offset_Q10, psEncC->subfr_length, &(table[32]) ); cannam@154: } cannam@154: else cannam@154: { cannam@154: silk_noise_shape_quantizer( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14, cannam@154: AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ], Lambda_Q10, cannam@154: offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->arch ); cannam@154: } cannam@154: cannam@154: x_Q3 += psEncC->subfr_length; cannam@154: pulses += psEncC->subfr_length; cannam@154: pxq += psEncC->subfr_length; cannam@154: } cannam@154: cannam@154: /* Update lagPrev for next frame */ cannam@154: NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; cannam@154: cannam@154: /* Save quantized speech and noise shaping signals */ cannam@154: silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); cannam@154: silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); cannam@154: RESTORE_STACK; cannam@154: } cannam@154: cannam@154: /***********************************/ cannam@154: /* silk_noise_shape_quantizer_10_16 */ cannam@154: /***********************************/ cannam@154: static OPUS_INLINE void silk_noise_shape_quantizer_10_16_sse4_1( cannam@154: silk_nsq_state *NSQ, /* I/O NSQ state */ cannam@154: opus_int signalType, /* I Signal type */ cannam@154: const opus_int32 x_sc_Q10[], /* I */ cannam@154: opus_int8 pulses[], /* O */ cannam@154: opus_int16 xq[], /* O */ cannam@154: opus_int32 sLTP_Q15[], /* I/O LTP state */ cannam@154: const opus_int16 a_Q12[], /* I Short term prediction coefs */ cannam@154: const opus_int16 b_Q14[], /* I Long term prediction coefs */ cannam@154: const opus_int16 AR_shp_Q13[], /* I Noise shaping AR coefs */ cannam@154: opus_int lag, /* I Pitch lag */ cannam@154: opus_int32 HarmShapeFIRPacked_Q14, /* I */ cannam@154: opus_int Tilt_Q14, /* I Spectral tilt */ cannam@154: opus_int32 LF_shp_Q14, /* I */ cannam@154: opus_int32 Gain_Q16, /* I */ cannam@154: opus_int offset_Q10, /* I */ cannam@154: opus_int length, /* I Input length */ cannam@154: opus_int32 table[][4] /* I */ cannam@154: ) cannam@154: { cannam@154: opus_int i; cannam@154: opus_int32 LTP_pred_Q13, LPC_pred_Q10, n_AR_Q12, n_LTP_Q13; cannam@154: opus_int32 n_LF_Q12, r_Q10, q1_Q0, q1_Q10, q2_Q10; cannam@154: opus_int32 exc_Q14, LPC_exc_Q14, xq_Q14, Gain_Q10; cannam@154: opus_int32 tmp1, tmp2, sLF_AR_shp_Q14; cannam@154: opus_int32 *psLPC_Q14, *shp_lag_ptr, *pred_lag_ptr; cannam@154: cannam@154: __m128i xmm_tempa, xmm_tempb; cannam@154: cannam@154: __m128i xmm_one; cannam@154: cannam@154: __m128i psLPC_Q14_hi_01234567, psLPC_Q14_hi_89ABCDEF; cannam@154: __m128i psLPC_Q14_lo_01234567, psLPC_Q14_lo_89ABCDEF; cannam@154: __m128i a_Q12_01234567, a_Q12_89ABCDEF; cannam@154: cannam@154: __m128i sAR2_Q14_hi_76543210, sAR2_Q14_lo_76543210; cannam@154: __m128i AR_shp_Q13_76543210; cannam@154: cannam@154: shp_lag_ptr = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ]; cannam@154: pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ]; cannam@154: Gain_Q10 = silk_RSHIFT( Gain_Q16, 6 ); cannam@154: cannam@154: /* Set up short term AR state */ cannam@154: psLPC_Q14 = &NSQ->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 ]; cannam@154: cannam@154: sLF_AR_shp_Q14 = NSQ->sLF_AR_shp_Q14; cannam@154: xq_Q14 = psLPC_Q14[ 0 ]; cannam@154: LTP_pred_Q13 = 0; cannam@154: cannam@154: /* load a_Q12 */ cannam@154: xmm_one = _mm_set_epi8( 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 ); cannam@154: cannam@154: /* load a_Q12[0] - a_Q12[7] */ cannam@154: a_Q12_01234567 = _mm_loadu_si128( (__m128i *)(&a_Q12[ 0 ] ) ); cannam@154: /* load a_Q12[ 8 ] - a_Q12[ 15 ] */ cannam@154: a_Q12_89ABCDEF = _mm_loadu_si128( (__m128i *)(&a_Q12[ 8 ] ) ); cannam@154: cannam@154: a_Q12_01234567 = _mm_shuffle_epi8( a_Q12_01234567, xmm_one ); cannam@154: a_Q12_89ABCDEF = _mm_shuffle_epi8( a_Q12_89ABCDEF, xmm_one ); cannam@154: cannam@154: /* load AR_shp_Q13 */ cannam@154: AR_shp_Q13_76543210 = _mm_loadu_si128( (__m128i *)(&AR_shp_Q13[0] ) ); cannam@154: cannam@154: /* load psLPC_Q14 */ cannam@154: xmm_one = _mm_set_epi8(15, 14, 11, 10, 7, 6, 3, 2, 13, 12, 9, 8, 5, 4, 1, 0 ); cannam@154: cannam@154: xmm_tempa = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[-16]) ); cannam@154: xmm_tempb = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[-12]) ); cannam@154: cannam@154: xmm_tempa = _mm_shuffle_epi8( xmm_tempa, xmm_one ); cannam@154: xmm_tempb = _mm_shuffle_epi8( xmm_tempb, xmm_one ); cannam@154: cannam@154: psLPC_Q14_hi_89ABCDEF = _mm_unpackhi_epi64( xmm_tempa, xmm_tempb ); cannam@154: psLPC_Q14_lo_89ABCDEF = _mm_unpacklo_epi64( xmm_tempa, xmm_tempb ); cannam@154: cannam@154: xmm_tempa = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -8 ]) ); cannam@154: xmm_tempb = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -4 ]) ); cannam@154: cannam@154: xmm_tempa = _mm_shuffle_epi8( xmm_tempa, xmm_one ); cannam@154: xmm_tempb = _mm_shuffle_epi8( xmm_tempb, xmm_one ); cannam@154: cannam@154: psLPC_Q14_hi_01234567 = _mm_unpackhi_epi64( xmm_tempa, xmm_tempb ); cannam@154: psLPC_Q14_lo_01234567 = _mm_unpacklo_epi64( xmm_tempa, xmm_tempb ); cannam@154: cannam@154: /* load sAR2_Q14 */ cannam@154: xmm_tempa = _mm_loadu_si128( (__m128i *)(&(NSQ->sAR2_Q14[ 0 ]) ) ); cannam@154: xmm_tempb = _mm_loadu_si128( (__m128i *)(&(NSQ->sAR2_Q14[ 4 ]) ) ); cannam@154: cannam@154: xmm_tempa = _mm_shuffle_epi8( xmm_tempa, xmm_one ); cannam@154: xmm_tempb = _mm_shuffle_epi8( xmm_tempb, xmm_one ); cannam@154: cannam@154: sAR2_Q14_hi_76543210 = _mm_unpackhi_epi64( xmm_tempa, xmm_tempb ); cannam@154: sAR2_Q14_lo_76543210 = _mm_unpacklo_epi64( xmm_tempa, xmm_tempb ); cannam@154: cannam@154: /* prepare 1 in 8 * 16bit */ cannam@154: xmm_one = _mm_set1_epi16(1); cannam@154: cannam@154: for( i = 0; i < length; i++ ) cannam@154: { cannam@154: /* Short-term prediction */ cannam@154: __m128i xmm_hi_07, xmm_hi_8F, xmm_lo_07, xmm_lo_8F; cannam@154: cannam@154: /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ cannam@154: LPC_pred_Q10 = 8; /* silk_RSHIFT( predictLPCOrder, 1 ); */ cannam@154: cannam@154: /* shift psLPC_Q14 */ cannam@154: psLPC_Q14_hi_89ABCDEF = _mm_alignr_epi8( psLPC_Q14_hi_01234567, psLPC_Q14_hi_89ABCDEF, 2 ); cannam@154: psLPC_Q14_lo_89ABCDEF = _mm_alignr_epi8( psLPC_Q14_lo_01234567, psLPC_Q14_lo_89ABCDEF, 2 ); cannam@154: cannam@154: psLPC_Q14_hi_01234567 = _mm_srli_si128( psLPC_Q14_hi_01234567, 2 ); cannam@154: psLPC_Q14_lo_01234567 = _mm_srli_si128( psLPC_Q14_lo_01234567, 2 ); cannam@154: cannam@154: psLPC_Q14_hi_01234567 = _mm_insert_epi16( psLPC_Q14_hi_01234567, (xq_Q14 >> 16), 7 ); cannam@154: psLPC_Q14_lo_01234567 = _mm_insert_epi16( psLPC_Q14_lo_01234567, (xq_Q14), 7 ); cannam@154: cannam@154: /* high part, use pmaddwd, results in 4 32-bit */ cannam@154: xmm_hi_07 = _mm_madd_epi16( psLPC_Q14_hi_01234567, a_Q12_01234567 ); cannam@154: xmm_hi_8F = _mm_madd_epi16( psLPC_Q14_hi_89ABCDEF, a_Q12_89ABCDEF ); cannam@154: cannam@154: /* low part, use pmulhw, results in 8 16-bit, note we need simulate unsigned * signed, _mm_srai_epi16(psLPC_Q14_lo_01234567, 15) */ cannam@154: xmm_tempa = _mm_cmpgt_epi16( _mm_setzero_si128(), psLPC_Q14_lo_01234567 ); cannam@154: xmm_tempb = _mm_cmpgt_epi16( _mm_setzero_si128(), psLPC_Q14_lo_89ABCDEF ); cannam@154: cannam@154: xmm_tempa = _mm_and_si128( xmm_tempa, a_Q12_01234567 ); cannam@154: xmm_tempb = _mm_and_si128( xmm_tempb, a_Q12_89ABCDEF ); cannam@154: cannam@154: xmm_lo_07 = _mm_mulhi_epi16( psLPC_Q14_lo_01234567, a_Q12_01234567 ); cannam@154: xmm_lo_8F = _mm_mulhi_epi16( psLPC_Q14_lo_89ABCDEF, a_Q12_89ABCDEF ); cannam@154: cannam@154: xmm_lo_07 = _mm_add_epi16( xmm_lo_07, xmm_tempa ); cannam@154: xmm_lo_8F = _mm_add_epi16( xmm_lo_8F, xmm_tempb ); cannam@154: cannam@154: xmm_lo_07 = _mm_madd_epi16( xmm_lo_07, xmm_one ); cannam@154: xmm_lo_8F = _mm_madd_epi16( xmm_lo_8F, xmm_one ); cannam@154: cannam@154: /* accumulate */ cannam@154: xmm_hi_07 = _mm_add_epi32( xmm_hi_07, xmm_hi_8F ); cannam@154: xmm_lo_07 = _mm_add_epi32( xmm_lo_07, xmm_lo_8F ); cannam@154: cannam@154: xmm_hi_07 = _mm_add_epi32( xmm_hi_07, xmm_lo_07 ); cannam@154: cannam@154: xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_unpackhi_epi64(xmm_hi_07, xmm_hi_07 ) ); cannam@154: xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_shufflelo_epi16(xmm_hi_07, 0x0E ) ); cannam@154: cannam@154: LPC_pred_Q10 += _mm_cvtsi128_si32( xmm_hi_07 ); cannam@154: cannam@154: /* Long-term prediction */ cannam@154: if ( opus_likely( signalType == TYPE_VOICED ) ) { cannam@154: /* Unrolled loop */ cannam@154: /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ cannam@154: LTP_pred_Q13 = 2; cannam@154: { cannam@154: __m128i b_Q14_3210, b_Q14_0123, pred_lag_ptr_0123; cannam@154: cannam@154: b_Q14_3210 = OP_CVTEPI16_EPI32_M64( b_Q14 ); cannam@154: b_Q14_0123 = _mm_shuffle_epi32( b_Q14_3210, 0x1B ); cannam@154: cannam@154: /* loaded: [0] [-1] [-2] [-3] */ cannam@154: pred_lag_ptr_0123 = _mm_loadu_si128( (__m128i *)(&pred_lag_ptr[ -3 ] ) ); cannam@154: /* shuffle to [-3] [-2] [-1] [0] and to new xmm */ cannam@154: xmm_tempa = _mm_shuffle_epi32( pred_lag_ptr_0123, 0x1B ); cannam@154: /*64-bit multiply, a[2] * b[-2], a[0] * b[0] */ cannam@154: xmm_tempa = _mm_mul_epi32( xmm_tempa, b_Q14_3210 ); cannam@154: /* right shift 2 bytes (16 bits), zero extended */ cannam@154: xmm_tempa = _mm_srli_si128( xmm_tempa, 2 ); cannam@154: cannam@154: /* a[1] * b[-1], a[3] * b[-3] */ cannam@154: pred_lag_ptr_0123 = _mm_mul_epi32( pred_lag_ptr_0123, b_Q14_0123 ); cannam@154: pred_lag_ptr_0123 = _mm_srli_si128( pred_lag_ptr_0123, 2 ); cannam@154: cannam@154: pred_lag_ptr_0123 = _mm_add_epi32( pred_lag_ptr_0123, xmm_tempa ); cannam@154: /* equal shift right 8 bytes*/ cannam@154: xmm_tempa = _mm_shuffle_epi32( pred_lag_ptr_0123, _MM_SHUFFLE( 0, 0, 3, 2 ) ); cannam@154: xmm_tempa = _mm_add_epi32( xmm_tempa, pred_lag_ptr_0123 ); cannam@154: cannam@154: LTP_pred_Q13 += _mm_cvtsi128_si32( xmm_tempa ); cannam@154: cannam@154: LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -4 ], b_Q14[ 4 ] ); cannam@154: pred_lag_ptr++; cannam@154: } cannam@154: } cannam@154: cannam@154: /* Noise shape feedback */ cannam@154: NSQ->sAR2_Q14[ 9 ] = NSQ->sAR2_Q14[ 8 ]; cannam@154: NSQ->sAR2_Q14[ 8 ] = _mm_cvtsi128_si32( _mm_srli_si128(_mm_unpackhi_epi16( sAR2_Q14_lo_76543210, sAR2_Q14_hi_76543210 ), 12 ) ); cannam@154: cannam@154: sAR2_Q14_hi_76543210 = _mm_slli_si128( sAR2_Q14_hi_76543210, 2 ); cannam@154: sAR2_Q14_lo_76543210 = _mm_slli_si128( sAR2_Q14_lo_76543210, 2 ); cannam@154: cannam@154: sAR2_Q14_hi_76543210 = _mm_insert_epi16( sAR2_Q14_hi_76543210, (xq_Q14 >> 16), 0 ); cannam@154: sAR2_Q14_lo_76543210 = _mm_insert_epi16( sAR2_Q14_lo_76543210, (xq_Q14), 0 ); cannam@154: cannam@154: /* high part, use pmaddwd, results in 4 32-bit */ cannam@154: xmm_hi_07 = _mm_madd_epi16( sAR2_Q14_hi_76543210, AR_shp_Q13_76543210 ); cannam@154: cannam@154: /* low part, use pmulhw, results in 8 16-bit, note we need simulate unsigned * signed,_mm_srai_epi16(sAR2_Q14_lo_76543210, 15) */ cannam@154: xmm_tempa = _mm_cmpgt_epi16( _mm_setzero_si128(), sAR2_Q14_lo_76543210 ); cannam@154: xmm_tempa = _mm_and_si128( xmm_tempa, AR_shp_Q13_76543210 ); cannam@154: cannam@154: xmm_lo_07 = _mm_mulhi_epi16( sAR2_Q14_lo_76543210, AR_shp_Q13_76543210 ); cannam@154: xmm_lo_07 = _mm_add_epi16( xmm_lo_07, xmm_tempa ); cannam@154: cannam@154: xmm_lo_07 = _mm_madd_epi16( xmm_lo_07, xmm_one ); cannam@154: cannam@154: /* accumulate */ cannam@154: xmm_hi_07 = _mm_add_epi32( xmm_hi_07, xmm_lo_07 ); cannam@154: cannam@154: xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_unpackhi_epi64(xmm_hi_07, xmm_hi_07 ) ); cannam@154: xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_shufflelo_epi16(xmm_hi_07, 0x0E ) ); cannam@154: cannam@154: n_AR_Q12 = 5 + _mm_cvtsi128_si32( xmm_hi_07 ); cannam@154: cannam@154: n_AR_Q12 = silk_SMLAWB( n_AR_Q12, NSQ->sAR2_Q14[ 8 ], AR_shp_Q13[ 8 ] ); cannam@154: n_AR_Q12 = silk_SMLAWB( n_AR_Q12, NSQ->sAR2_Q14[ 9 ], AR_shp_Q13[ 9 ] ); cannam@154: cannam@154: n_AR_Q12 = silk_LSHIFT32( n_AR_Q12, 1 ); /* Q11 -> Q12 */ cannam@154: n_AR_Q12 = silk_SMLAWB( n_AR_Q12, sLF_AR_shp_Q14, Tilt_Q14 ); cannam@154: cannam@154: n_LF_Q12 = silk_SMULWB( NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - 1 ], LF_shp_Q14 ); cannam@154: n_LF_Q12 = silk_SMLAWT( n_LF_Q12, sLF_AR_shp_Q14, LF_shp_Q14 ); cannam@154: cannam@154: silk_assert( lag > 0 || signalType != TYPE_VOICED ); cannam@154: cannam@154: /* Combine prediction and noise shaping signals */ cannam@154: tmp1 = silk_SUB32( silk_LSHIFT32( LPC_pred_Q10, 2 ), n_AR_Q12 ); /* Q12 */ cannam@154: tmp1 = silk_SUB32( tmp1, n_LF_Q12 ); /* Q12 */ cannam@154: if( lag > 0 ) { cannam@154: /* Symmetric, packed FIR coefficients */ cannam@154: n_LTP_Q13 = silk_SMULWB( silk_ADD32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 ); cannam@154: n_LTP_Q13 = silk_SMLAWT( n_LTP_Q13, shp_lag_ptr[ -1 ], HarmShapeFIRPacked_Q14 ); cannam@154: n_LTP_Q13 = silk_LSHIFT( n_LTP_Q13, 1 ); cannam@154: shp_lag_ptr++; cannam@154: cannam@154: tmp2 = silk_SUB32( LTP_pred_Q13, n_LTP_Q13 ); /* Q13 */ cannam@154: tmp1 = silk_ADD_LSHIFT32( tmp2, tmp1, 1 ); /* Q13 */ cannam@154: tmp1 = silk_RSHIFT_ROUND( tmp1, 3 ); /* Q10 */ cannam@154: } else { cannam@154: tmp1 = silk_RSHIFT_ROUND( tmp1, 2 ); /* Q10 */ cannam@154: } cannam@154: cannam@154: r_Q10 = silk_SUB32( x_sc_Q10[ i ], tmp1 ); /* residual error Q10 */ cannam@154: cannam@154: /* Generate dither */ cannam@154: NSQ->rand_seed = silk_RAND( NSQ->rand_seed ); cannam@154: cannam@154: /* Flip sign depending on dither */ cannam@154: tmp2 = -r_Q10; cannam@154: if ( NSQ->rand_seed < 0 ) r_Q10 = tmp2; cannam@154: cannam@154: r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 ); cannam@154: cannam@154: /* Find two quantization level candidates and measure their rate-distortion */ cannam@154: q1_Q10 = silk_SUB32( r_Q10, offset_Q10 ); cannam@154: q1_Q0 = silk_RSHIFT( q1_Q10, 10 ); cannam@154: cannam@154: q1_Q10 = table[q1_Q0][0]; cannam@154: q2_Q10 = table[q1_Q0][1]; cannam@154: cannam@154: if (r_Q10 * table[q1_Q0][2] - table[q1_Q0][3] < 0) cannam@154: { cannam@154: q1_Q10 = q2_Q10; cannam@154: } cannam@154: cannam@154: pulses[ i ] = (opus_int8)silk_RSHIFT_ROUND( q1_Q10, 10 ); cannam@154: cannam@154: /* Excitation */ cannam@154: exc_Q14 = silk_LSHIFT( q1_Q10, 4 ); cannam@154: cannam@154: tmp2 = -exc_Q14; cannam@154: if ( NSQ->rand_seed < 0 ) exc_Q14 = tmp2; cannam@154: cannam@154: /* Add predictions */ cannam@154: LPC_exc_Q14 = silk_ADD_LSHIFT32( exc_Q14, LTP_pred_Q13, 1 ); cannam@154: xq_Q14 = silk_ADD_LSHIFT32( LPC_exc_Q14, LPC_pred_Q10, 4 ); cannam@154: cannam@154: /* Update states */ cannam@154: psLPC_Q14++; cannam@154: *psLPC_Q14 = xq_Q14; cannam@154: sLF_AR_shp_Q14 = silk_SUB_LSHIFT32( xq_Q14, n_AR_Q12, 2 ); cannam@154: cannam@154: NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx ] = silk_SUB_LSHIFT32( sLF_AR_shp_Q14, n_LF_Q12, 2 ); cannam@154: sLTP_Q15[ NSQ->sLTP_buf_idx ] = silk_LSHIFT( LPC_exc_Q14, 1 ); cannam@154: NSQ->sLTP_shp_buf_idx++; cannam@154: NSQ->sLTP_buf_idx++; cannam@154: cannam@154: /* Make dither dependent on quantized signal */ cannam@154: NSQ->rand_seed = silk_ADD32_ovflw( NSQ->rand_seed, pulses[ i ] ); cannam@154: } cannam@154: cannam@154: NSQ->sLF_AR_shp_Q14 = sLF_AR_shp_Q14; cannam@154: cannam@154: /* Scale XQ back to normal level before saving */ cannam@154: psLPC_Q14 = &NSQ->sLPC_Q14[ NSQ_LPC_BUF_LENGTH ]; cannam@154: cannam@154: /* write back sAR2_Q14 */ cannam@154: xmm_tempa = _mm_unpackhi_epi16( sAR2_Q14_lo_76543210, sAR2_Q14_hi_76543210 ); cannam@154: xmm_tempb = _mm_unpacklo_epi16( sAR2_Q14_lo_76543210, sAR2_Q14_hi_76543210 ); cannam@154: _mm_storeu_si128( (__m128i *)(&NSQ->sAR2_Q14[ 4 ]), xmm_tempa ); cannam@154: _mm_storeu_si128( (__m128i *)(&NSQ->sAR2_Q14[ 0 ]), xmm_tempb ); cannam@154: cannam@154: /* xq[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psLPC_Q14[ i ], Gain_Q10 ), 8 ) ); */ cannam@154: { cannam@154: __m128i xmm_Gain_Q10; cannam@154: __m128i xmm_xq_Q14_3210, xmm_xq_Q14_x3x1, xmm_xq_Q14_7654, xmm_xq_Q14_x7x5; cannam@154: cannam@154: /* prepare (1 << 7) in packed 4 32-bits */ cannam@154: xmm_tempa = _mm_set1_epi32( (1 << 7) ); cannam@154: cannam@154: /* prepare Gain_Q10 in packed 4 32-bits */ cannam@154: xmm_Gain_Q10 = _mm_set1_epi32( Gain_Q10 ); cannam@154: cannam@154: /* process xq */ cannam@154: for (i = 0; i < length - 7; i += 8) cannam@154: { cannam@154: xmm_xq_Q14_3210 = _mm_loadu_si128( (__m128i *)(&(psLPC_Q14[ i + 0 ] ) ) ); cannam@154: xmm_xq_Q14_7654 = _mm_loadu_si128( (__m128i *)(&(psLPC_Q14[ i + 4 ] ) ) ); cannam@154: cannam@154: /* equal shift right 4 bytes*/ cannam@154: xmm_xq_Q14_x3x1 = _mm_shuffle_epi32( xmm_xq_Q14_3210, _MM_SHUFFLE( 0, 3, 2, 1 ) ); cannam@154: /* equal shift right 4 bytes*/ cannam@154: xmm_xq_Q14_x7x5 = _mm_shuffle_epi32( xmm_xq_Q14_7654, _MM_SHUFFLE( 0, 3, 2, 1 ) ); cannam@154: cannam@154: xmm_xq_Q14_3210 = _mm_mul_epi32( xmm_xq_Q14_3210, xmm_Gain_Q10 ); cannam@154: xmm_xq_Q14_x3x1 = _mm_mul_epi32( xmm_xq_Q14_x3x1, xmm_Gain_Q10 ); cannam@154: xmm_xq_Q14_7654 = _mm_mul_epi32( xmm_xq_Q14_7654, xmm_Gain_Q10 ); cannam@154: xmm_xq_Q14_x7x5 = _mm_mul_epi32( xmm_xq_Q14_x7x5, xmm_Gain_Q10 ); cannam@154: cannam@154: xmm_xq_Q14_3210 = _mm_srli_epi64( xmm_xq_Q14_3210, 16 ); cannam@154: xmm_xq_Q14_x3x1 = _mm_slli_epi64( xmm_xq_Q14_x3x1, 16 ); cannam@154: xmm_xq_Q14_7654 = _mm_srli_epi64( xmm_xq_Q14_7654, 16 ); cannam@154: xmm_xq_Q14_x7x5 = _mm_slli_epi64( xmm_xq_Q14_x7x5, 16 ); cannam@154: cannam@154: xmm_xq_Q14_3210 = _mm_blend_epi16( xmm_xq_Q14_3210, xmm_xq_Q14_x3x1, 0xCC ); cannam@154: xmm_xq_Q14_7654 = _mm_blend_epi16( xmm_xq_Q14_7654, xmm_xq_Q14_x7x5, 0xCC ); cannam@154: cannam@154: /* silk_RSHIFT_ROUND(xq, 8) */ cannam@154: xmm_xq_Q14_3210 = _mm_add_epi32( xmm_xq_Q14_3210, xmm_tempa ); cannam@154: xmm_xq_Q14_7654 = _mm_add_epi32( xmm_xq_Q14_7654, xmm_tempa ); cannam@154: cannam@154: xmm_xq_Q14_3210 = _mm_srai_epi32( xmm_xq_Q14_3210, 8 ); cannam@154: xmm_xq_Q14_7654 = _mm_srai_epi32( xmm_xq_Q14_7654, 8 ); cannam@154: cannam@154: /* silk_SAT16 */ cannam@154: xmm_xq_Q14_3210 = _mm_packs_epi32( xmm_xq_Q14_3210, xmm_xq_Q14_7654 ); cannam@154: cannam@154: /* save to xq */ cannam@154: _mm_storeu_si128( (__m128i *)(&xq[ i ] ), xmm_xq_Q14_3210 ); cannam@154: } cannam@154: } cannam@154: for ( ; i < length; i++) cannam@154: { cannam@154: xq[i] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psLPC_Q14[ i ], Gain_Q10 ), 8 ) ); cannam@154: } cannam@154: cannam@154: /* Update LPC synth buffer */ cannam@154: silk_memcpy( NSQ->sLPC_Q14, &NSQ->sLPC_Q14[ length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) ); cannam@154: } cannam@154: cannam@154: static OPUS_INLINE void silk_nsq_scale_states_sse4_1( cannam@154: const silk_encoder_state *psEncC, /* I Encoder State */ cannam@154: silk_nsq_state *NSQ, /* I/O NSQ state */ cannam@154: const opus_int32 x_Q3[], /* I input in Q3 */ cannam@154: opus_int32 x_sc_Q10[], /* O input scaled with 1/Gain */ cannam@154: const opus_int16 sLTP[], /* I re-whitened LTP state in Q0 */ cannam@154: opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ cannam@154: opus_int subfr, /* I subframe number */ cannam@154: const opus_int LTP_scale_Q14, /* I */ cannam@154: const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */ cannam@154: const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */ cannam@154: const opus_int signal_type /* I Signal type */ cannam@154: ) cannam@154: { cannam@154: opus_int i, lag; cannam@154: opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23; cannam@154: __m128i xmm_inv_gain_Q23, xmm_x_Q3_x2x0, xmm_x_Q3_x3x1; cannam@154: cannam@154: lag = pitchL[ subfr ]; cannam@154: inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); cannam@154: silk_assert( inv_gain_Q31 != 0 ); cannam@154: cannam@154: /* Calculate gain adjustment factor */ cannam@154: if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { cannam@154: gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); cannam@154: } else { cannam@154: gain_adj_Q16 = (opus_int32)1 << 16; cannam@154: } cannam@154: cannam@154: /* Scale input */ cannam@154: inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 ); cannam@154: cannam@154: /* prepare inv_gain_Q23 in packed 4 32-bits */ cannam@154: xmm_inv_gain_Q23 = _mm_set1_epi32(inv_gain_Q23); cannam@154: cannam@154: for( i = 0; i < psEncC->subfr_length - 3; i += 4 ) { cannam@154: xmm_x_Q3_x2x0 = _mm_loadu_si128( (__m128i *)(&(x_Q3[ i ] ) ) ); cannam@154: cannam@154: /* equal shift right 4 bytes*/ cannam@154: xmm_x_Q3_x3x1 = _mm_shuffle_epi32( xmm_x_Q3_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) ); cannam@154: cannam@154: xmm_x_Q3_x2x0 = _mm_mul_epi32( xmm_x_Q3_x2x0, xmm_inv_gain_Q23 ); cannam@154: xmm_x_Q3_x3x1 = _mm_mul_epi32( xmm_x_Q3_x3x1, xmm_inv_gain_Q23 ); cannam@154: cannam@154: xmm_x_Q3_x2x0 = _mm_srli_epi64( xmm_x_Q3_x2x0, 16 ); cannam@154: xmm_x_Q3_x3x1 = _mm_slli_epi64( xmm_x_Q3_x3x1, 16 ); cannam@154: cannam@154: xmm_x_Q3_x2x0 = _mm_blend_epi16( xmm_x_Q3_x2x0, xmm_x_Q3_x3x1, 0xCC ); cannam@154: cannam@154: _mm_storeu_si128( (__m128i *)(&(x_sc_Q10[ i ] ) ), xmm_x_Q3_x2x0 ); cannam@154: } cannam@154: cannam@154: for( ; i < psEncC->subfr_length; i++ ) { cannam@154: x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 ); cannam@154: } cannam@154: cannam@154: /* Save inverse gain */ cannam@154: NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; cannam@154: cannam@154: /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ cannam@154: if( NSQ->rewhite_flag ) { cannam@154: if( subfr == 0 ) { cannam@154: /* Do LTP downscaling */ cannam@154: inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 ); cannam@154: } cannam@154: for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) { cannam@154: silk_assert( i < MAX_FRAME_LENGTH ); cannam@154: sLTP_Q15[ i ] = silk_SMULWB( inv_gain_Q31, sLTP[ i ] ); cannam@154: } cannam@154: } cannam@154: cannam@154: /* Adjust for changing gain */ cannam@154: if( gain_adj_Q16 != (opus_int32)1 << 16 ) { cannam@154: /* Scale long-term shaping state */ cannam@154: __m128i xmm_gain_adj_Q16, xmm_sLTP_shp_Q14_x2x0, xmm_sLTP_shp_Q14_x3x1; cannam@154: cannam@154: /* prepare gain_adj_Q16 in packed 4 32-bits */ cannam@154: xmm_gain_adj_Q16 = _mm_set1_epi32(gain_adj_Q16); cannam@154: cannam@154: for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 3; i += 4 ) cannam@154: { cannam@154: xmm_sLTP_shp_Q14_x2x0 = _mm_loadu_si128( (__m128i *)(&(NSQ->sLTP_shp_Q14[ i ] ) ) ); cannam@154: /* equal shift right 4 bytes*/ cannam@154: xmm_sLTP_shp_Q14_x3x1 = _mm_shuffle_epi32( xmm_sLTP_shp_Q14_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) ); cannam@154: cannam@154: xmm_sLTP_shp_Q14_x2x0 = _mm_mul_epi32( xmm_sLTP_shp_Q14_x2x0, xmm_gain_adj_Q16 ); cannam@154: xmm_sLTP_shp_Q14_x3x1 = _mm_mul_epi32( xmm_sLTP_shp_Q14_x3x1, xmm_gain_adj_Q16 ); cannam@154: cannam@154: xmm_sLTP_shp_Q14_x2x0 = _mm_srli_epi64( xmm_sLTP_shp_Q14_x2x0, 16 ); cannam@154: xmm_sLTP_shp_Q14_x3x1 = _mm_slli_epi64( xmm_sLTP_shp_Q14_x3x1, 16 ); cannam@154: cannam@154: xmm_sLTP_shp_Q14_x2x0 = _mm_blend_epi16( xmm_sLTP_shp_Q14_x2x0, xmm_sLTP_shp_Q14_x3x1, 0xCC ); cannam@154: cannam@154: _mm_storeu_si128( (__m128i *)(&(NSQ->sLTP_shp_Q14[ i ] ) ), xmm_sLTP_shp_Q14_x2x0 ); cannam@154: } cannam@154: cannam@154: for( ; i < NSQ->sLTP_shp_buf_idx; i++ ) { cannam@154: NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); cannam@154: } cannam@154: cannam@154: /* Scale long-term prediction state */ cannam@154: if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) { cannam@154: for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) { cannam@154: sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] ); cannam@154: } cannam@154: } cannam@154: cannam@154: NSQ->sLF_AR_shp_Q14 = silk_SMULWW( gain_adj_Q16, NSQ->sLF_AR_shp_Q14 ); cannam@154: cannam@154: /* Scale short-term prediction and shaping states */ cannam@154: for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { cannam@154: NSQ->sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLPC_Q14[ i ] ); cannam@154: } cannam@154: for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) { cannam@154: NSQ->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sAR2_Q14[ i ] ); cannam@154: } cannam@154: } cannam@154: }