git.sesse.net Git - ffmpeg/blob - libavcodec/x86/h264_cabac.c

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of Libav.
   6  *
   7  * Libav is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * Libav is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with Libav; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG-4 part10 codec.
  25  * non-SIMD x86-specific optimizations for H.264
  26  * @author Michael Niedermayer <michaelni@gmx.at>
  27  */
  28
  29 #include <stddef.h>
  30
  31 #include "libavcodec/cabac.h"
  32 #include "cabac.h"
  33
  34 #if HAVE_INLINE_ASM
  35
  36 //FIXME use some macros to avoid duplicating get_cabac (cannot be done yet
  37 //as that would make optimization work hard)
  38 #if HAVE_7REGS
  39 #define decode_significance decode_significance_x86
  40 static int decode_significance_x86(CABACContext *c, int max_coeff,
  41                                    uint8_t *significant_coeff_ctx_base,
  42                                    int *index, x86_reg last_off){
  43     void *end= significant_coeff_ctx_base + max_coeff - 1;
  44     int minusstart= -(intptr_t)significant_coeff_ctx_base;
  45     int minusindex= 4-(intptr_t)index;
  46     int bit;
  47     x86_reg coeff_count;
  48
  49 #ifdef BROKEN_RELOCATIONS
  50     void *tables;
  51
  52     __asm__ volatile(
  53         "lea   "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
  54         : "=&r"(tables)
  55     );
  56 #endif
  57
  58     __asm__ volatile(
  59         "3:                                     \n\t"
  60
  61         BRANCHLESS_GET_CABAC("%4", "%q4", "(%1)", "%3", "%w3",
  62                              "%5", "%q5", "%k0", "%b0",
  63                              "%c11(%6)", "%c12(%6)",
  64                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
  65                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
  66                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
  67                              "%13")
  68
  69         "test $1, %4                            \n\t"
  70         " jz 4f                                 \n\t"
  71         "add  %10, %1                           \n\t"
  72
  73         BRANCHLESS_GET_CABAC("%4", "%q4", "(%1)", "%3", "%w3",
  74                              "%5", "%q5", "%k0", "%b0",
  75                              "%c11(%6)", "%c12(%6)",
  76                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
  77                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
  78                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
  79                              "%13")
  80
  81         "sub  %10, %1                           \n\t"
  82         "mov  %2, %0                            \n\t"
  83         "movl %7, %%ecx                         \n\t"
  84         "add  %1, %%"FF_REG_c"                  \n\t"
  85         "movl %%ecx, (%0)                       \n\t"
  86
  87         "test $1, %4                            \n\t"
  88         " jnz 5f                                \n\t"
  89
  90         "add"FF_OPSIZE"  $4, %2                 \n\t"
  91
  92         "4:                                     \n\t"
  93         "add  $1, %1                            \n\t"
  94         "cmp  %8, %1                            \n\t"
  95         " jb 3b                                 \n\t"
  96         "mov  %2, %0                            \n\t"
  97         "movl %7, %%ecx                         \n\t"
  98         "add  %1, %%"FF_REG_c"                  \n\t"
  99         "movl %%ecx, (%0)                       \n\t"
 100         "5:                                     \n\t"
 101         "add  %9, %k0                           \n\t"
 102         "shr $2, %k0                            \n\t"
 103         : "=&q"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
 104           "+&r"(c->low), "=&r"(bit), "+&r"(c->range)
 105         : "r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
 106           "i"(offsetof(CABACContext, bytestream)),
 107           "i"(offsetof(CABACContext, bytestream_end))
 108           TABLES_ARG
 109         : "%"FF_REG_c, "memory"
 110     );
 111     return coeff_count;
 112 }
 113
 114 #define decode_significance_8x8 decode_significance_8x8_x86
 115 static int decode_significance_8x8_x86(CABACContext *c,
 116                                        uint8_t *significant_coeff_ctx_base,
 117                                        int *index, uint8_t *last_coeff_ctx_base, const uint8_t *sig_off){
 118     int minusindex= 4-(intptr_t)index;
 119     int bit;
 120     x86_reg coeff_count;
 121     x86_reg last=0;
 122     x86_reg state;
 123
 124 #ifdef BROKEN_RELOCATIONS
 125     void *tables;
 126
 127     __asm__ volatile(
 128         "lea    "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
 129         : "=&r"(tables)
 130     );
 131 #endif
 132
 133     __asm__ volatile(
 134         "mov %1, %6                             \n\t"
 135         "3:                                     \n\t"
 136
 137         "mov %10, %0                            \n\t"
 138         "movzbl (%0, %6), %k6                   \n\t"
 139         "add %9, %6                             \n\t"
 140
 141         BRANCHLESS_GET_CABAC("%4", "%q4", "(%6)", "%3", "%w3",
 142                              "%5", "%q5", "%k0", "%b0",
 143                              "%c12(%7)", "%c13(%7)",
 144                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
 145                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
 146                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
 147                              "%15")
 148
 149         "mov %1, %k6                            \n\t"
 150         "test $1, %4                            \n\t"
 151         " jz 4f                                 \n\t"
 152
 153 #ifdef BROKEN_RELOCATIONS
 154         "movzbl %c14(%15, %q6), %k6\n\t"
 155 #else
 156         "movzbl "MANGLE(ff_h264_cabac_tables)"+%c14(%k6), %k6\n\t"
 157 #endif
 158         "add %11, %6                            \n\t"
 159
 160         BRANCHLESS_GET_CABAC("%4", "%q4", "(%6)", "%3", "%w3",
 161                              "%5", "%q5", "%k0", "%b0",
 162                              "%c12(%7)", "%c13(%7)",
 163                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
 164                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
 165                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
 166                              "%15")
 167
 168         "mov %2, %0                             \n\t"
 169         "mov %1, %k6                            \n\t"
 170         "movl %k6, (%0)                         \n\t"
 171
 172         "test $1, %4                            \n\t"
 173         " jnz 5f                                \n\t"
 174
 175         "add"FF_OPSIZE"  $4, %2                 \n\t"
 176
 177         "4:                                     \n\t"
 178         "addl $1, %k6                           \n\t"
 179         "mov %k6, %1                            \n\t"
 180         "cmpl $63, %k6                          \n\t"
 181         " jb 3b                                 \n\t"
 182         "mov %2, %0                             \n\t"
 183         "movl %k6, (%0)                         \n\t"
 184         "5:                                     \n\t"
 185         "addl %8, %k0                           \n\t"
 186         "shr $2, %k0                            \n\t"
 187         : "=&q"(coeff_count), "+m"(last), "+m"(index), "+&r"(c->low),
 188           "=&r"(bit), "+&r"(c->range), "=&r"(state)
 189         : "r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base),
 190           "m"(sig_off), "m"(last_coeff_ctx_base),
 191           "i"(offsetof(CABACContext, bytestream)),
 192           "i"(offsetof(CABACContext, bytestream_end)),
 193           "i"(H264_LAST_COEFF_FLAG_OFFSET_8x8_OFFSET) TABLES_ARG
 194         : "%"FF_REG_c, "memory"
 195     );
 196     return coeff_count;
 197 }
 198 #endif /* HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
 199
 200 #endif /* HAVE_INLINE_ASM */