git.sesse.net Git - ffmpeg/blob - libavcodec/x86/h264_i386.h

   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 / MPEG4 part10 codec.
  25  * non-MMX i386-specific optimizations for H.264
  26  * @author Michael Niedermayer <michaelni@gmx.at>
  27  */
  28
  29 #ifndef AVCODEC_X86_H264_I386_H
  30 #define AVCODEC_X86_H264_I386_H
  31
  32 #include <stddef.h>
  33
  34 #include "libavcodec/cabac.h"
  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 ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS)
  39 static int decode_significance_x86(CABACContext *c, int max_coeff,
  40                                    uint8_t *significant_coeff_ctx_base,
  41                                    int *index, x86_reg last_off){
  42     void *end= significant_coeff_ctx_base + max_coeff - 1;
  43     int minusstart= -(int)significant_coeff_ctx_base;
  44     int minusindex= 4-(int)index;
  45     int bit;
  46     x86_reg coeff_count;
  47     int low;
  48     __asm__ volatile(
  49         "movl %a10(%5), %%esi                   \n\t"
  50         "movl %a11(%5), %3                      \n\t"
  51
  52         "2:                                     \n\t"
  53
  54         BRANCHLESS_GET_CABAC("%4", "%5", "(%1)", "%3",
  55                              "%w3", "%%esi", "%k0", "%b0", "%a12")
  56
  57         "test $1, %4                            \n\t"
  58         " jz 3f                                 \n\t"
  59         "add  %9, %1                            \n\t"
  60
  61         BRANCHLESS_GET_CABAC("%4", "%5", "(%1)", "%3",
  62                              "%w3", "%%esi", "%k0", "%b0", "%a12")
  63
  64         "sub  %9, %1                            \n\t"
  65         "mov  %2, %0                            \n\t"
  66         "movl %6, %%ecx                         \n\t"
  67         "add  %1, %%"REG_c"                     \n\t"
  68         "movl %%ecx, (%0)                       \n\t"
  69
  70         "test $1, %4                            \n\t"
  71         " jnz 4f                                \n\t"
  72
  73         "add  $4, %0                            \n\t"
  74         "mov  %0, %2                            \n\t"
  75
  76         "3:                                     \n\t"
  77         "add  $1, %1                            \n\t"
  78         "cmp  %7, %1                            \n\t"
  79         " jb 2b                                 \n\t"
  80         "mov  %2, %0                            \n\t"
  81         "movl %6, %%ecx                         \n\t"
  82         "add  %1, %%"REG_c"                     \n\t"
  83         "movl %%ecx, (%0)                       \n\t"
  84         "4:                                     \n\t"
  85         "add  %8, %k0                           \n\t"
  86         "shr $2, %k0                            \n\t"
  87
  88         "movl %%esi, %a10(%5)                   \n\t"
  89         "movl %3, %a11(%5)                      \n\t"
  90         :"=&r"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
  91          "=&r"(low), "=&r"(bit)
  92         :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
  93          "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)),
  94          "i"(offsetof(CABACContext, bytestream))
  95         : "%"REG_c, "%esi", "memory"
  96     );
  97     return coeff_count;
  98 }
  99
 100 static int decode_significance_8x8_x86(CABACContext *c,
 101                                        uint8_t *significant_coeff_ctx_base,
 102                                        int *index, x86_reg last_off, const uint8_t *sig_off){
 103     int minusindex= 4-(int)index;
 104     int bit;
 105     x86_reg coeff_count;
 106     int low;
 107     x86_reg last=0;
 108     __asm__ volatile(
 109         "movl %a10(%5), %%esi                   \n\t"
 110         "movl %a11(%5), %3                      \n\t"
 111
 112         "mov %1, %%"REG_D"                      \n\t"
 113         "2:                                     \n\t"
 114
 115         "mov %8, %0                             \n\t"
 116         "movzbl (%0, %%"REG_D"), %%edi          \n\t"
 117         "add %7, %%"REG_D"                      \n\t"
 118
 119         BRANCHLESS_GET_CABAC("%4", "%5", "(%%"REG_D")", "%3",
 120                              "%w3", "%%esi", "%k0", "%b0", "%a12")
 121
 122         "mov %1, %%edi                          \n\t"
 123         "test $1, %4                            \n\t"
 124         " jz 3f                                 \n\t"
 125
 126         "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t"
 127         "add %7, %%"REG_D"                      \n\t"
 128         "add %9, %%"REG_D"                      \n\t"
 129
 130         BRANCHLESS_GET_CABAC("%4", "%5", "(%%"REG_D")", "%3",
 131                              "%w3", "%%esi", "%k0", "%b0", "%a12")
 132
 133         "mov %2, %0                             \n\t"
 134         "mov %1, %%edi                          \n\t"
 135         "movl %%edi, (%0)                       \n\t"
 136
 137         "test $1, %4                            \n\t"
 138         " jnz 4f                                \n\t"
 139
 140         "add $4, %0                             \n\t"
 141         "mov %0, %2                             \n\t"
 142
 143         "3:                                     \n\t"
 144         "addl $1, %%edi                         \n\t"
 145         "mov %%edi, %1                          \n\t"
 146         "cmpl $63, %%edi                        \n\t"
 147         " jb 2b                                 \n\t"
 148         "mov %2, %0                             \n\t"
 149         "movl %%edi, (%0)                       \n\t"
 150         "4:                                     \n\t"
 151         "addl %6, %k0                           \n\t"
 152         "shr $2, %k0                            \n\t"
 153
 154         "movl %%esi, %a10(%5)                   \n\t"
 155         "movl %3, %a11(%5)                      \n\t"
 156         :"=&r"(coeff_count),"+m"(last), "+m"(index), "=&r"(low), "=&r"(bit)
 157         :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off), "m"(last_off),
 158          "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)),
 159          "i"(offsetof(CABACContext, bytestream))
 160         : "%"REG_c, "%esi", "%"REG_D, "memory"
 161     );
 162     return coeff_count;
 163 }
 164 #endif /* ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
 165
 166 #endif /* AVCODEC_X86_H264_I386_H */