diff options
Diffstat (limited to 'silk/x86/VQ_WMat_EC_sse4_1.c')
| -rw-r--r-- | silk/x86/VQ_WMat_EC_sse4_1.c | 189 |
1 files changed, 110 insertions, 79 deletions
diff --git a/silk/x86/VQ_WMat_EC_sse4_1.c b/silk/x86/VQ_WMat_EC_sse4_1.c index 74d6c6d0..2c7d18d0 100644 --- a/silk/x86/VQ_WMat_EC_sse4_1.c +++ b/silk/x86/VQ_WMat_EC_sse4_1.c @@ -1,5 +1,5 @@ -/* Copyright (c) 2014, Cisco Systems, INC - Written by XiangMingZhu WeiZhou MinPeng YanWang +/* 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 @@ -38,105 +38,136 @@ /* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */ void silk_VQ_WMat_EC_sse4_1( opus_int8 *ind, /* O index of best codebook vector */ - opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */ + opus_int32 *res_nrg_Q15, /* O best residual energy */ + opus_int32 *rate_dist_Q8, /* O best total bitrate */ opus_int *gain_Q7, /* O sum of absolute LTP coefficients */ - const opus_int16 *in_Q14, /* I input vector to be quantized */ - const opus_int32 *W_Q18, /* I weighting matrix */ + const opus_int32 *XX_Q17, /* I correlation matrix */ + const opus_int32 *xX_Q17, /* I correlation vector */ const opus_int8 *cb_Q7, /* I codebook */ const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */ const opus_uint8 *cl_Q5, /* I code length for each codebook vector */ - const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */ + const opus_int subfr_len, /* I number of samples per subframe */ const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ - opus_int L /* I number of vectors in codebook */ + const opus_int L /* I number of vectors in codebook */ ) { opus_int k, gain_tmp_Q7; const opus_int8 *cb_row_Q7; - opus_int16 diff_Q14[ 5 ]; - opus_int32 sum1_Q14, sum2_Q16; + opus_int32 neg_xX_Q24[ 5 ]; + opus_int32 sum1_Q15, sum2_Q24; + opus_int32 bits_res_Q8, bits_tot_Q8; + __m128i v_XX_31_Q17, v_XX_42_Q17, v_cb_row_31_Q7, v_cb_row_42_Q7, v_acc1_Q24, v_acc2_Q24; + + /* Negate and convert to new Q domain */ + neg_xX_Q24[ 0 ] = -silk_LSHIFT32( xX_Q17[ 0 ], 7 ); + neg_xX_Q24[ 1 ] = -silk_LSHIFT32( xX_Q17[ 1 ], 7 ); + neg_xX_Q24[ 2 ] = -silk_LSHIFT32( xX_Q17[ 2 ], 7 ); + neg_xX_Q24[ 3 ] = -silk_LSHIFT32( xX_Q17[ 3 ], 7 ); + neg_xX_Q24[ 4 ] = -silk_LSHIFT32( xX_Q17[ 4 ], 7 ); + + v_XX_31_Q17 = _mm_loadu_si128( (__m128i *)(&XX_Q17[ 1 ] ) ); + v_XX_42_Q17 = _mm_shuffle_epi32( v_XX_31_Q17, _MM_SHUFFLE( 0, 3, 2, 1 ) ); - __m128i C_tmp1, C_tmp2, C_tmp3, C_tmp4, C_tmp5; /* Loop over codebook */ - *rate_dist_Q14 = silk_int32_MAX; + *rate_dist_Q8 = silk_int32_MAX; + *res_nrg_Q15 = silk_int32_MAX; cb_row_Q7 = cb_Q7; + /* If things go really bad, at least *ind is set to something safe. */ + *ind = 0; for( k = 0; k < L; k++ ) { + opus_int32 penalty; gain_tmp_Q7 = cb_gain_Q7[k]; - - diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 ); - - C_tmp1 = OP_CVTEPI16_EPI32_M64( &in_Q14[ 1 ] ); - C_tmp2 = OP_CVTEPI8_EPI32_M32( &cb_row_Q7[ 1 ] ); - C_tmp2 = _mm_slli_epi32( C_tmp2, 7 ); - C_tmp1 = _mm_sub_epi32( C_tmp1, C_tmp2 ); - - diff_Q14[ 1 ] = _mm_extract_epi16( C_tmp1, 0 ); - diff_Q14[ 2 ] = _mm_extract_epi16( C_tmp1, 2 ); - diff_Q14[ 3 ] = _mm_extract_epi16( C_tmp1, 4 ); - diff_Q14[ 4 ] = _mm_extract_epi16( C_tmp1, 6 ); - /* Weighted rate */ - sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] ); + /* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */ + sum1_Q15 = SILK_FIX_CONST( 1.001, 15 ); /* Penalty for too large gain */ - sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 ); - - silk_assert( sum1_Q14 >= 0 ); - - /* first row of W_Q18 */ - C_tmp3 = _mm_loadu_si128( (__m128i *)(&W_Q18[ 1 ] ) ); - C_tmp4 = _mm_mul_epi32( C_tmp3, C_tmp1 ); - C_tmp4 = _mm_srli_si128( C_tmp4, 2 ); - - C_tmp1 = _mm_shuffle_epi32( C_tmp1, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* shift right 4 bytes */ - C_tmp3 = _mm_shuffle_epi32( C_tmp3, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* shift right 4 bytes */ - - C_tmp5 = _mm_mul_epi32( C_tmp3, C_tmp1 ); - C_tmp5 = _mm_srli_si128( C_tmp5, 2 ); - - C_tmp5 = _mm_add_epi32( C_tmp4, C_tmp5 ); - C_tmp5 = _mm_slli_epi32( C_tmp5, 1 ); - - C_tmp5 = _mm_add_epi32( C_tmp5, _mm_shuffle_epi32( C_tmp5, _MM_SHUFFLE( 0, 0, 0, 2 ) ) ); - sum2_Q16 = _mm_cvtsi128_si32( C_tmp5 ); - - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 0 ], diff_Q14[ 0 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 0 ] ); - - /* second row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 7 ], diff_Q14[ 2 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 8 ], diff_Q14[ 3 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 9 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 6 ], diff_Q14[ 1 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 1 ] ); - - /* third row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 13 ], diff_Q14[ 3 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 14 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 12 ], diff_Q14[ 2 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 2 ] ); - - /* fourth row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 19 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 18 ], diff_Q14[ 3 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 3 ] ); - - /* last row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 24 ], diff_Q14[ 4 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 4 ] ); - - silk_assert( sum1_Q14 >= 0 ); + penalty = silk_LSHIFT32( silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 11 ); + + /* first row of XX_Q17 */ + v_cb_row_31_Q7 = OP_CVTEPI8_EPI32_M32( &cb_row_Q7[ 1 ] ); + v_cb_row_42_Q7 = _mm_shuffle_epi32( v_cb_row_31_Q7, _MM_SHUFFLE( 0, 3, 2, 1 ) ); + v_cb_row_31_Q7 = _mm_mul_epi32( v_XX_31_Q17, v_cb_row_31_Q7 ); + v_cb_row_42_Q7 = _mm_mul_epi32( v_XX_42_Q17, v_cb_row_42_Q7 ); + v_acc1_Q24 = _mm_add_epi64( v_cb_row_31_Q7, v_cb_row_42_Q7); + v_acc2_Q24 = _mm_shuffle_epi32( v_acc1_Q24, _MM_SHUFFLE( 1, 0, 3, 2 ) ); + v_acc1_Q24 = _mm_add_epi64( v_acc1_Q24, v_acc2_Q24); + sum2_Q24 = _mm_cvtsi128_si32( v_acc1_Q24 ); + sum2_Q24 = silk_ADD32( neg_xX_Q24[ 0 ], sum2_Q24 ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 0 ], cb_row_Q7[ 0 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 0 ] ); + + /* second row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 1 ], XX_Q17[ 7 ], cb_row_Q7[ 2 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 8 ], cb_row_Q7[ 3 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 9 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 6 ], cb_row_Q7[ 1 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 1 ] ); + + /* third row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 2 ], XX_Q17[ 13 ], cb_row_Q7[ 3 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 14 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 12 ], cb_row_Q7[ 2 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 2 ] ); + + /* fourth row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 3 ], XX_Q17[ 19 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 18 ], cb_row_Q7[ 3 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 3 ] ); + + /* last row of XX_Q17 */ + sum2_Q24 = silk_LSHIFT32( neg_xX_Q24[ 4 ], 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 24 ], cb_row_Q7[ 4 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 4 ] ); /* find best */ - if( sum1_Q14 < *rate_dist_Q14 ) { - *rate_dist_Q14 = sum1_Q14; - *ind = (opus_int8)k; - *gain_Q7 = gain_tmp_Q7; + if( sum1_Q15 >= 0 ) { + /* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */ + bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 + penalty) - (15 << 7) ); + /* In the following line we reduce the codelength component by half ("-1"); seems to slightly improve quality */ + bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 ); + if( bits_tot_Q8 <= *rate_dist_Q8 ) { + *rate_dist_Q8 = bits_tot_Q8; + *res_nrg_Q15 = sum1_Q15 + penalty; + *ind = (opus_int8)k; + *gain_Q7 = gain_tmp_Q7; + } } /* Go to next cbk vector */ cb_row_Q7 += LTP_ORDER; } + +#ifdef OPUS_CHECK_ASM + { + opus_int8 ind_c = 0; + opus_int32 res_nrg_Q15_c = 0; + opus_int32 rate_dist_Q8_c = 0; + opus_int gain_Q7_c = 0; + + silk_VQ_WMat_EC_c( + &ind_c, + &res_nrg_Q15_c, + &rate_dist_Q8_c, + &gain_Q7_c, + XX_Q17, + xX_Q17, + cb_Q7, + cb_gain_Q7, + cl_Q5, + subfr_len, + max_gain_Q7, + L + ); + + silk_assert( *ind == ind_c ); + silk_assert( *res_nrg_Q15 == res_nrg_Q15_c ); + silk_assert( *rate_dist_Q8 == rate_dist_Q8_c ); + silk_assert( *gain_Q7 == gain_Q7_c ); + } +#endif } |