VQ_WMat_EC_sse4_1.c (8236B)
1 /* Copyright (c) 2014-2020, Cisco Systems, INC 2 Written by XiangMingZhu WeiZhou MinPeng YanWang FrancisQuiers 3 4 Redistribution and use in source and binary forms, with or without 5 modification, are permitted provided that the following conditions 6 are met: 7 8 - Redistributions of source code must retain the above copyright 9 notice, this list of conditions and the following disclaimer. 10 11 - Redistributions in binary form must reproduce the above copyright 12 notice, this list of conditions and the following disclaimer in the 13 documentation and/or other materials provided with the distribution. 14 15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 16 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 17 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 18 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER 19 OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 20 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 21 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 22 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 23 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 24 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 25 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 */ 27 28 #ifdef HAVE_CONFIG_H 29 #include "config.h" 30 #endif 31 32 #include <xmmintrin.h> 33 #include <emmintrin.h> 34 #include <smmintrin.h> 35 #include "main.h" 36 #include "celt/x86/x86cpu.h" 37 38 /* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */ 39 void silk_VQ_WMat_EC_sse4_1( 40 opus_int8 *ind, /* O index of best codebook vector */ 41 opus_int32 *res_nrg_Q15, /* O best residual energy */ 42 opus_int32 *rate_dist_Q8, /* O best total bitrate */ 43 opus_int *gain_Q7, /* O sum of absolute LTP coefficients */ 44 const opus_int32 *XX_Q17, /* I correlation matrix */ 45 const opus_int32 *xX_Q17, /* I correlation vector */ 46 const opus_int8 *cb_Q7, /* I codebook */ 47 const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */ 48 const opus_uint8 *cl_Q5, /* I code length for each codebook vector */ 49 const opus_int subfr_len, /* I number of samples per subframe */ 50 const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ 51 const opus_int L /* I number of vectors in codebook */ 52 ) 53 { 54 opus_int k, gain_tmp_Q7; 55 const opus_int8 *cb_row_Q7; 56 opus_int32 neg_xX_Q24[ 5 ]; 57 opus_int32 sum1_Q15, sum2_Q24; 58 opus_int32 bits_res_Q8, bits_tot_Q8; 59 __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; 60 61 /* Negate and convert to new Q domain */ 62 neg_xX_Q24[ 0 ] = -silk_LSHIFT32( xX_Q17[ 0 ], 7 ); 63 neg_xX_Q24[ 1 ] = -silk_LSHIFT32( xX_Q17[ 1 ], 7 ); 64 neg_xX_Q24[ 2 ] = -silk_LSHIFT32( xX_Q17[ 2 ], 7 ); 65 neg_xX_Q24[ 3 ] = -silk_LSHIFT32( xX_Q17[ 3 ], 7 ); 66 neg_xX_Q24[ 4 ] = -silk_LSHIFT32( xX_Q17[ 4 ], 7 ); 67 68 v_XX_31_Q17 = _mm_loadu_si128( (__m128i *)(void*)(&XX_Q17[ 1 ] ) ); 69 v_XX_42_Q17 = _mm_shuffle_epi32( v_XX_31_Q17, _MM_SHUFFLE( 0, 3, 2, 1 ) ); 70 71 /* Loop over codebook */ 72 *rate_dist_Q8 = silk_int32_MAX; 73 *res_nrg_Q15 = silk_int32_MAX; 74 cb_row_Q7 = cb_Q7; 75 /* If things go really bad, at least *ind is set to something safe. */ 76 *ind = 0; 77 for( k = 0; k < L; k++ ) { 78 opus_int32 penalty; 79 gain_tmp_Q7 = cb_gain_Q7[k]; 80 /* Weighted rate */ 81 /* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */ 82 sum1_Q15 = SILK_FIX_CONST( 1.001, 15 ); 83 84 /* Penalty for too large gain */ 85 penalty = silk_LSHIFT32( silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 11 ); 86 87 /* first row of XX_Q17 */ 88 v_cb_row_31_Q7 = OP_CVTEPI8_EPI32_M32( &cb_row_Q7[ 1 ] ); 89 v_cb_row_42_Q7 = _mm_shuffle_epi32( v_cb_row_31_Q7, _MM_SHUFFLE( 0, 3, 2, 1 ) ); 90 v_cb_row_31_Q7 = _mm_mul_epi32( v_XX_31_Q17, v_cb_row_31_Q7 ); 91 v_cb_row_42_Q7 = _mm_mul_epi32( v_XX_42_Q17, v_cb_row_42_Q7 ); 92 v_acc1_Q24 = _mm_add_epi64( v_cb_row_31_Q7, v_cb_row_42_Q7); 93 v_acc2_Q24 = _mm_shuffle_epi32( v_acc1_Q24, _MM_SHUFFLE( 1, 0, 3, 2 ) ); 94 v_acc1_Q24 = _mm_add_epi64( v_acc1_Q24, v_acc2_Q24); 95 sum2_Q24 = _mm_cvtsi128_si32( v_acc1_Q24 ); 96 sum2_Q24 = silk_ADD32( neg_xX_Q24[ 0 ], sum2_Q24 ); 97 sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); 98 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 0 ], cb_row_Q7[ 0 ] ); 99 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 0 ] ); 100 101 /* second row of XX_Q17 */ 102 sum2_Q24 = silk_MLA( neg_xX_Q24[ 1 ], XX_Q17[ 7 ], cb_row_Q7[ 2 ] ); 103 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 8 ], cb_row_Q7[ 3 ] ); 104 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 9 ], cb_row_Q7[ 4 ] ); 105 sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); 106 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 6 ], cb_row_Q7[ 1 ] ); 107 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 1 ] ); 108 109 /* third row of XX_Q17 */ 110 sum2_Q24 = silk_MLA( neg_xX_Q24[ 2 ], XX_Q17[ 13 ], cb_row_Q7[ 3 ] ); 111 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 14 ], cb_row_Q7[ 4 ] ); 112 sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); 113 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 12 ], cb_row_Q7[ 2 ] ); 114 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 2 ] ); 115 116 /* fourth row of XX_Q17 */ 117 sum2_Q24 = silk_MLA( neg_xX_Q24[ 3 ], XX_Q17[ 19 ], cb_row_Q7[ 4 ] ); 118 sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); 119 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 18 ], cb_row_Q7[ 3 ] ); 120 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 3 ] ); 121 122 /* last row of XX_Q17 */ 123 sum2_Q24 = silk_LSHIFT32( neg_xX_Q24[ 4 ], 1 ); 124 sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 24 ], cb_row_Q7[ 4 ] ); 125 sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 4 ] ); 126 127 /* find best */ 128 if( sum1_Q15 >= 0 ) { 129 /* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */ 130 bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 + penalty) - (15 << 7) ); 131 /* In the following line we reduce the codelength component by half ("-1"); seems to slightly improve quality */ 132 bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 ); 133 if( bits_tot_Q8 <= *rate_dist_Q8 ) { 134 *rate_dist_Q8 = bits_tot_Q8; 135 *res_nrg_Q15 = sum1_Q15 + penalty; 136 *ind = (opus_int8)k; 137 *gain_Q7 = gain_tmp_Q7; 138 } 139 } 140 141 /* Go to next cbk vector */ 142 cb_row_Q7 += LTP_ORDER; 143 } 144 145 #ifdef OPUS_CHECK_ASM 146 { 147 opus_int8 ind_c = 0; 148 opus_int32 res_nrg_Q15_c = 0; 149 opus_int32 rate_dist_Q8_c = 0; 150 opus_int gain_Q7_c = 0; 151 152 silk_VQ_WMat_EC_c( 153 &ind_c, 154 &res_nrg_Q15_c, 155 &rate_dist_Q8_c, 156 &gain_Q7_c, 157 XX_Q17, 158 xX_Q17, 159 cb_Q7, 160 cb_gain_Q7, 161 cl_Q5, 162 subfr_len, 163 max_gain_Q7, 164 L 165 ); 166 167 silk_assert( *ind == ind_c ); 168 silk_assert( *res_nrg_Q15 == res_nrg_Q15_c ); 169 silk_assert( *rate_dist_Q8 == rate_dist_Q8_c ); 170 silk_assert( *gain_Q7 == gain_Q7_c ); 171 } 172 #endif 173 }