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 FFmpeg.
   6  *
   7  * FFmpeg 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  * FFmpeg 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 FFmpeg; 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 #include "cabac.h"
  36
  37 //FIXME use some macros to avoid duplicating get_cabac (cannot be done yet
  38 //as that would make optimization work hard)
  39 #if HAVE_7REGS && !defined(BROKEN_RELOCATIONS)
  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     __asm__ volatile(
  49         "3:                                     \n\t"
  50
  51         BRANCHLESS_GET_CABAC("%4", "(%1)", "%3", "%w3",
  52                              "%5", "%k0", "%b0",
  53                              "%a11(%6)", "%a12(%6)")
  54
  55         "test $1, %4                            \n\t"
  56         " jz 4f                                 \n\t"
  57         "add  %10, %1                           \n\t"
  58
  59         BRANCHLESS_GET_CABAC("%4", "(%1)", "%3", "%w3",
  60                              "%5", "%k0", "%b0",
  61                              "%a11(%6)", "%a12(%6)")
  62
  63         "sub  %10, %1                           \n\t"
  64         "mov  %2, %0                            \n\t"
  65         "movl %7, %%ecx                         \n\t"
  66         "add  %1, %%"REG_c"                     \n\t"
  67         "movl %%ecx, (%0)                       \n\t"
  68
  69         "test $1, %4                            \n\t"
  70         " jnz 5f                                \n\t"
  71
  72         "add"OPSIZE"  $4, %2                    \n\t"
  73
  74         "4:                                     \n\t"
  75         "add  $1, %1                            \n\t"
  76         "cmp  %8, %1                            \n\t"
  77         " jb 3b                                 \n\t"
  78         "mov  %2, %0                            \n\t"
  79         "movl %7, %%ecx                         \n\t"
  80         "add  %1, %%"REG_c"                     \n\t"
  81         "movl %%ecx, (%0)                       \n\t"
  82         "5:                                     \n\t"
  83         "add  %9, %k0                           \n\t"
  84         "shr $2, %k0                            \n\t"
  85         : "=&q"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
  86           "+&r"(c->low), "=&r"(bit), "+&r"(c->range)
  87         : "r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
  88           "i"(offsetof(CABACContext, bytestream)),
  89           "i"(offsetof(CABACContext, bytestream_end))
  90         : "%"REG_c, "memory"
  91     );
  92     return coeff_count;
  93 }
  94
  95 static int decode_significance_8x8_x86(CABACContext *c,
  96                                        uint8_t *significant_coeff_ctx_base,
  97                                        int *index, uint8_t *last_coeff_ctx_base, const uint8_t *sig_off){
  98     int minusindex= 4-(intptr_t)index;
  99     int bit;
 100     x86_reg coeff_count;
 101     x86_reg last=0;
 102     x86_reg state;
 103     __asm__ volatile(
 104         "mov %1, %6                             \n\t"
 105         "3:                                     \n\t"
 106
 107         "mov %10, %0                            \n\t"
 108         "movzbl (%0, %6), %k6                   \n\t"
 109         "add %9, %6                             \n\t"
 110
 111         BRANCHLESS_GET_CABAC("%4", "(%6)", "%3", "%w3",
 112                              "%5", "%k0", "%b0",
 113                              "%a12(%7)", "%a13(%7)")
 114
 115         "mov %1, %k6                            \n\t"
 116         "test $1, %4                            \n\t"
 117         " jz 4f                                 \n\t"
 118
 119         "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%k6), %k6\n\t"
 120         "add %11, %6                            \n\t"
 121
 122         BRANCHLESS_GET_CABAC("%4", "(%6)", "%3", "%w3",
 123                              "%5", "%k0", "%b0",
 124                              "%a12(%7)", "%a13(%7)")
 125
 126         "mov %2, %0                             \n\t"
 127         "mov %1, %k6                            \n\t"
 128         "movl %k6, (%0)                         \n\t"
 129
 130         "test $1, %4                            \n\t"
 131         " jnz 5f                                \n\t"
 132
 133         "add"OPSIZE"  $4, %2                    \n\t"
 134
 135         "4:                                     \n\t"
 136         "addl $1, %k6                           \n\t"
 137         "mov %k6, %1                            \n\t"
 138         "cmpl $63, %k6                          \n\t"
 139         " jb 3b                                 \n\t"
 140         "mov %2, %0                             \n\t"
 141         "movl %k6, (%0)                         \n\t"
 142         "5:                                     \n\t"
 143         "addl %8, %k0                           \n\t"
 144         "shr $2, %k0                            \n\t"
 145         : "=&q"(coeff_count), "+m"(last), "+m"(index), "+&r"(c->low),
 146           "=&r"(bit), "+&r"(c->range), "=&r"(state)
 147         : "r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base),
 148           "m"(sig_off), "m"(last_coeff_ctx_base),
 149           "i"(offsetof(CABACContext, bytestream)),
 150           "i"(offsetof(CABACContext, bytestream_end))
 151         : "%"REG_c, "memory"
 152     );
 153     return coeff_count;
 154 }
 155 #endif /* HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
 156
 157 #endif /* AVCODEC_X86_H264_I386_H */