FLAC/src/fixed_intrin_ssse3.c

   1 /* libFLAC - Free Lossless Audio Codec library
   2  * Copyright (C) 2000-2009  Josh Coalson
   3  * Copyright (C) 2011-2016  Xiph.Org Foundation
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  *
   9  * - Redistributions of source code must retain the above copyright
  10  * notice, this list of conditions and the following disclaimer.
  11  *
  12  * - Redistributions in binary form must reproduce the above copyright
  13  * notice, this list of conditions and the following disclaimer in the
  14  * documentation and/or other materials provided with the distribution.
  15  *
  16  * - Neither the name of the Xiph.org Foundation nor the names of its
  17  * contributors may be used to endorse or promote products derived from
  18  * this software without specific prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
  24  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
  25  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  26  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
  27  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  28  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
  29  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
  30  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  31  */
  32
  33 #ifdef HAVE_CONFIG_H
  34 #  include <config.h>
  35 #endif
  36
  37 #include "private/cpu.h"
  38
  39 #ifndef FLAC__INTEGER_ONLY_LIBRARY
  40 #ifndef FLAC__NO_ASM
  41 #if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN
  42 #include "private/fixed.h"
  43 #ifdef FLAC__SSSE3_SUPPORTED
  44
  45 #include <tmmintrin.h> /* SSSE3 */
  46 #include <math.h>
  47 #include "private/macros.h"
  48 #include "share/compat.h"
  49 #include "FLAC/assert.h"
  50
  51 #ifdef FLAC__CPU_IA32
  52 #define m128i_to_i64(dest, src) _mm_storel_epi64((__m128i*)&dest, src)
  53 #else
  54 #define m128i_to_i64(dest, src) dest = _mm_cvtsi128_si64(src)
  55 #endif
  56
  57 FLAC__SSE_TARGET("ssse3")
  58 uint32_t FLAC__fixed_compute_best_predictor_intrin_ssse3(const FLAC__int32 data[], uint32_t data_len, float residual_bits_per_sample[FLAC__MAX_FIXED_ORDER + 1])
  59 {
  60         FLAC__uint32 total_error_0, total_error_1, total_error_2, total_error_3, total_error_4;
  61         uint32_t i, order;
  62
  63         __m128i total_err0, total_err1, total_err2;
  64
  65         {
  66                 FLAC__int32 itmp;
  67                 __m128i last_error;
  68
  69                 last_error = _mm_cvtsi32_si128(data[-1]);                                                       // 0   0   0   le0
  70                 itmp = data[-2];
  71                 last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0));
  72                 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp));        // 0   0   le0 le1
  73                 itmp -= data[-3];
  74                 last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0));
  75                 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp));        // 0   le0 le1 le2
  76                 itmp -= data[-3] - data[-4];
  77                 last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0));
  78                 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp));        // le0 le1 le2 le3
  79
  80                 total_err0 = total_err1 = _mm_setzero_si128();
  81                 for(i = 0; i < data_len; i++) {
  82                         __m128i err0, err1;
  83                         err0 = _mm_cvtsi32_si128(data[i]);                                                              // 0   0   0   e0
  84                         err1 = _mm_shuffle_epi32(err0, _MM_SHUFFLE(0,0,0,0));                   // e0  e0  e0  e0
  85 #if 1 /* OPT_SSE */
  86                         err1 = _mm_sub_epi32(err1, last_error);
  87                         last_error = _mm_srli_si128(last_error, 4);                                             // 0   le0 le1 le2
  88                         err1 = _mm_sub_epi32(err1, last_error);
  89                         last_error = _mm_srli_si128(last_error, 4);                                             // 0   0   le0 le1
  90                         err1 = _mm_sub_epi32(err1, last_error);
  91                         last_error = _mm_srli_si128(last_error, 4);                                             // 0   0   0   le0
  92                         err1 = _mm_sub_epi32(err1, last_error);                                                 // e1  e2  e3  e4
  93 #else
  94                         last_error = _mm_add_epi32(last_error, _mm_srli_si128(last_error, 8));  // le0  le1  le2+le0  le3+le1
  95                         last_error = _mm_add_epi32(last_error, _mm_srli_si128(last_error, 4));  // le0  le1+le0  le2+le0+le1  le3+le1+le2+le0
  96                         err1 = _mm_sub_epi32(err1, last_error);                                                 // e1  e2  e3  e4
  97 #endif
  98                         last_error = _mm_alignr_epi8(err0, err1, 4);                                    // e0  e1  e2  e3
  99
 100                         err0 = _mm_abs_epi32(err0);
 101                         err1 = _mm_abs_epi32(err1);
 102
 103                         total_err0 = _mm_add_epi32(total_err0, err0);                                   // 0   0   0   te0
 104                         total_err1 = _mm_add_epi32(total_err1, err1);                                   // te1 te2 te3 te4
 105                 }
 106         }
 107
 108         total_error_0 = _mm_cvtsi128_si32(total_err0);
 109         total_err2 = total_err1;                                                                                        // te1  te2  te3  te4
 110         total_err1 = _mm_srli_si128(total_err1, 8);                                                     //  0    0   te1  te2
 111         total_error_4 = _mm_cvtsi128_si32(total_err2);
 112         total_error_2 = _mm_cvtsi128_si32(total_err1);
 113         total_err2 = _mm_srli_si128(total_err2, 4);                                                     //  0   te1  te2  te3
 114         total_err1 = _mm_srli_si128(total_err1, 4);                                                     //  0    0    0   te1
 115         total_error_3 = _mm_cvtsi128_si32(total_err2);
 116         total_error_1 = _mm_cvtsi128_si32(total_err1);
 117
 118         /* prefer higher order */
 119         if(total_error_0 < flac_min(flac_min(flac_min(total_error_1, total_error_2), total_error_3), total_error_4))
 120                 order = 0;
 121         else if(total_error_1 < flac_min(flac_min(total_error_2, total_error_3), total_error_4))
 122                 order = 1;
 123         else if(total_error_2 < flac_min(total_error_3, total_error_4))
 124                 order = 2;
 125         else if(total_error_3 < total_error_4)
 126                 order = 3;
 127         else
 128                 order = 4;
 129
 130         /* Estimate the expected number of bits per residual signal sample. */
 131         /* 'total_error*' is linearly related to the variance of the residual */
 132         /* signal, so we use it directly to compute E(|x|) */
 133         FLAC__ASSERT(data_len > 0 || total_error_0 == 0);
 134         FLAC__ASSERT(data_len > 0 || total_error_1 == 0);
 135         FLAC__ASSERT(data_len > 0 || total_error_2 == 0);
 136         FLAC__ASSERT(data_len > 0 || total_error_3 == 0);
 137         FLAC__ASSERT(data_len > 0 || total_error_4 == 0);
 138
 139         residual_bits_per_sample[0] = (float)((total_error_0 > 0) ? log(M_LN2 * (double)total_error_0 / (double)data_len) / M_LN2 : 0.0);
 140         residual_bits_per_sample[1] = (float)((total_error_1 > 0) ? log(M_LN2 * (double)total_error_1 / (double)data_len) / M_LN2 : 0.0);
 141         residual_bits_per_sample[2] = (float)((total_error_2 > 0) ? log(M_LN2 * (double)total_error_2 / (double)data_len) / M_LN2 : 0.0);
 142         residual_bits_per_sample[3] = (float)((total_error_3 > 0) ? log(M_LN2 * (double)total_error_3 / (double)data_len) / M_LN2 : 0.0);
 143         residual_bits_per_sample[4] = (float)((total_error_4 > 0) ? log(M_LN2 * (double)total_error_4 / (double)data_len) / M_LN2 : 0.0);
 144
 145         return order;
 146 }
 147
 148 FLAC__SSE_TARGET("ssse3")
 149 uint32_t FLAC__fixed_compute_best_predictor_wide_intrin_ssse3(const FLAC__int32 data[], uint32_t data_len, float residual_bits_per_sample[FLAC__MAX_FIXED_ORDER + 1])
 150 {
 151         FLAC__uint64 total_error_0, total_error_1, total_error_2, total_error_3, total_error_4;
 152         uint32_t i, order;
 153
 154         __m128i total_err0, total_err1, total_err3;
 155
 156         {
 157                 FLAC__int32 itmp;
 158                 __m128i last_error, zero = _mm_setzero_si128();
 159
 160                 last_error = _mm_cvtsi32_si128(data[-1]);                                                       // 0   0   0   le0
 161                 itmp = data[-2];
 162                 last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0));
 163                 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp));        // 0   0   le0 le1
 164                 itmp -= data[-3];
 165                 last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0));
 166                 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp));        // 0   le0 le1 le2
 167                 itmp -= data[-3] - data[-4];
 168                 last_error = _mm_shuffle_epi32(last_error, _MM_SHUFFLE(2,1,0,0));
 169                 last_error = _mm_sub_epi32(last_error, _mm_cvtsi32_si128(itmp));        // le0 le1 le2 le3
 170
 171                 total_err0 = total_err1 = total_err3 = _mm_setzero_si128();
 172                 for(i = 0; i < data_len; i++) {
 173                         __m128i err0, err1;
 174                         err0 = _mm_cvtsi32_si128(data[i]);                                                              // 0   0   0   e0
 175                         err1 = _mm_shuffle_epi32(err0, _MM_SHUFFLE(0,0,0,0));                   // e0  e0  e0  e0
 176 #if 1 /* OPT_SSE */
 177                         err1 = _mm_sub_epi32(err1, last_error);
 178                         last_error = _mm_srli_si128(last_error, 4);                                             // 0   le0 le1 le2
 179                         err1 = _mm_sub_epi32(err1, last_error);
 180                         last_error = _mm_srli_si128(last_error, 4);                                             // 0   0   le0 le1
 181                         err1 = _mm_sub_epi32(err1, last_error);
 182                         last_error = _mm_srli_si128(last_error, 4);                                             // 0   0   0   le0
 183                         err1 = _mm_sub_epi32(err1, last_error);                                                 // e1  e2  e3  e4
 184 #else
 185                         last_error = _mm_add_epi32(last_error, _mm_srli_si128(last_error, 8));  // le0  le1  le2+le0  le3+le1
 186                         last_error = _mm_add_epi32(last_error, _mm_srli_si128(last_error, 4));  // le0  le1+le0  le2+le0+le1  le3+le1+le2+le0
 187                         err1 = _mm_sub_epi32(err1, last_error);                                                 // e1  e2  e3  e4
 188 #endif
 189                         last_error = _mm_alignr_epi8(err0, err1, 4);                                    // e0  e1  e2  e3
 190
 191                         err0 = _mm_abs_epi32(err0);
 192                         err1 = _mm_abs_epi32(err1);                                                                             // |e1| |e2| |e3| |e4|
 193
 194                         total_err0 = _mm_add_epi64(total_err0, err0);                                   //        0       te0
 195                         err0 = _mm_unpacklo_epi32(err1, zero);                                                  //   0  |e3|   0  |e4|
 196                         err1 = _mm_unpackhi_epi32(err1, zero);                                                  //   0  |e1|   0  |e2|
 197                         total_err3 = _mm_add_epi64(total_err3, err0);                                   //       te3      te4
 198                         total_err1 = _mm_add_epi64(total_err1, err1);                                   //       te1      te2
 199                 }
 200         }
 201
 202         m128i_to_i64(total_error_0, total_err0);
 203         m128i_to_i64(total_error_4, total_err3);
 204         m128i_to_i64(total_error_2, total_err1);
 205         total_err3 = _mm_srli_si128(total_err3, 8);                                                     //         0      te3
 206         total_err1 = _mm_srli_si128(total_err1, 8);                                                     //         0      te1
 207         m128i_to_i64(total_error_3, total_err3);
 208         m128i_to_i64(total_error_1, total_err1);
 209
 210         /* prefer higher order */
 211         if(total_error_0 < flac_min(flac_min(flac_min(total_error_1, total_error_2), total_error_3), total_error_4))
 212                 order = 0;
 213         else if(total_error_1 < flac_min(flac_min(total_error_2, total_error_3), total_error_4))
 214                 order = 1;
 215         else if(total_error_2 < flac_min(total_error_3, total_error_4))
 216                 order = 2;
 217         else if(total_error_3 < total_error_4)
 218                 order = 3;
 219         else
 220                 order = 4;
 221
 222         /* Estimate the expected number of bits per residual signal sample. */
 223         /* 'total_error*' is linearly related to the variance of the residual */
 224         /* signal, so we use it directly to compute E(|x|) */
 225         FLAC__ASSERT(data_len > 0 || total_error_0 == 0);
 226         FLAC__ASSERT(data_len > 0 || total_error_1 == 0);
 227         FLAC__ASSERT(data_len > 0 || total_error_2 == 0);
 228         FLAC__ASSERT(data_len > 0 || total_error_3 == 0);
 229         FLAC__ASSERT(data_len > 0 || total_error_4 == 0);
 230
 231         residual_bits_per_sample[0] = (float)((total_error_0 > 0) ? log(M_LN2 * (double)total_error_0 / (double)data_len) / M_LN2 : 0.0);
 232         residual_bits_per_sample[1] = (float)((total_error_1 > 0) ? log(M_LN2 * (double)total_error_1 / (double)data_len) / M_LN2 : 0.0);
 233         residual_bits_per_sample[2] = (float)((total_error_2 > 0) ? log(M_LN2 * (double)total_error_2 / (double)data_len) / M_LN2 : 0.0);
 234         residual_bits_per_sample[3] = (float)((total_error_3 > 0) ? log(M_LN2 * (double)total_error_3 / (double)data_len) / M_LN2 : 0.0);
 235         residual_bits_per_sample[4] = (float)((total_error_4 > 0) ? log(M_LN2 * (double)total_error_4 / (double)data_len) / M_LN2 : 0.0);
 236
 237         return order;
 238 }
 239
 240 #endif /* FLAC__SSSE3_SUPPORTED */
 241 #endif /* (FLAC__CPU_IA32 || FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN */
 242 #endif /* FLAC__NO_ASM */
 243 #endif /* FLAC__INTEGER_ONLY_LIBRARY */