git.sesse.net Git - vlc/blob - plugins/idct/vdec_block_mmx.c

   1 /*****************************************************************************
   2  * vdec_block_mmx.c: Macroblock copy functions in MMX assembly
   3  *****************************************************************************
   4  * Copyright (C) 1999, 2000, 2001 VideoLAN
   5  * $Id: vdec_block_mmx.c,v 1.5 2001/07/17 09:48:07 massiot Exp $
   6  *
   7  * Authors: Gaël Hendryckx <jimmy@via.ecp.fr>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software
  21  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  22  *****************************************************************************/
  23
  24 /* MODULE_NAME defined in Makefile together with -DBUILTIN */
  25 #ifdef BUILTIN
  26 #   include "modules_inner.h"
  27 #else
  28 #   define _M( foo ) foo
  29 #endif
  30
  31 /*****************************************************************************
  32  * Preamble
  33  *****************************************************************************/
  34 #include "defs.h"
  35
  36 #include <string.h>
  37
  38 #include "config.h"
  39 #include "common.h"
  40 #include "threads.h"
  41 #include "mtime.h"
  42
  43 #include "intf_msg.h"
  44
  45 #include "input_ext-dec.h"
  46
  47 #include "video.h"
  48 #include "video_output.h"
  49
  50 #include "vdec_ext-plugins.h"
  51
  52 #include "vdec_block.h"
  53
  54 #include "modules.h"
  55 #include "modules_export.h"
  56
  57 /*****************************************************************************
  58  * vdec_InitDecode: initialize video decoder thread
  59  *****************************************************************************/
  60 void _M( vdec_InitDecode ) ( vdec_thread_t *p_vdec )
  61 {
  62     ;
  63 }
  64
  65 /*****************************************************************************
  66  * AddBlock : add a block
  67  *****************************************************************************/
  68 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
  69                                  yuv_data_t * p_data, int i_incr )
  70 {
  71     asm __volatile__ (
  72             "pxor       %%mm7,%%mm7\n\t"
  73
  74             "movq       (%0),%%mm1\n\t"
  75             "movq       %%mm1,%%mm2\n\t"
  76             "punpckhbw  %%mm7,%%mm1\n\t"
  77             "punpcklbw  %%mm7,%%mm2\n\t"
  78             "paddw      (%2),%%mm2\n\t"
  79             "paddw      8(%2),%%mm1\n\t"
  80             "packuswb   %%mm1,%%mm2\n\t"
  81             "movq       %%mm2,(%0)\n\t"
  82             "addl       %3,%0\n\t"
  83
  84             "movq       (%0),%%mm1\n\t"
  85             "movq       %%mm1,%%mm2\n\t"
  86             "punpckhbw  %%mm7,%%mm1\n\t"
  87             "punpcklbw  %%mm7,%%mm2\n\t"
  88             "paddw      16(%2),%%mm2\n\t"
  89             "paddw      24(%2),%%mm1\n\t"
  90             "packuswb   %%mm1,%%mm2\n\t"
  91             "movq       %%mm2,(%0)\n\t"
  92             "addl       %3,%0\n\t"
  93
  94             "movq       (%0),%%mm1\n\t"
  95             "movq       %%mm1,%%mm2\n\t"
  96             "punpckhbw  %%mm7,%%mm1\n\t"
  97             "punpcklbw  %%mm7,%%mm2\n\t"
  98             "paddw      32(%2),%%mm2\n\t"
  99             "paddw      40(%2),%%mm1\n\t"
 100             "packuswb   %%mm1,%%mm2\n\t"
 101             "movq       %%mm2,(%0)\n\t"
 102             "addl       %3,%0\n\t"
 103
 104             "movq       (%0),%%mm1\n\t"
 105             "movq       %%mm1,%%mm2\n\t"
 106             "punpckhbw  %%mm7,%%mm1\n\t"
 107             "punpcklbw  %%mm7,%%mm2\n\t"
 108             "paddw      48(%2),%%mm2\n\t"
 109             "paddw      56(%2),%%mm1\n\t"
 110             "packuswb   %%mm1,%%mm2\n\t"
 111             "movq       %%mm2,(%0)\n\t"
 112             "addl       %3,%0\n\t"
 113
 114             "movq       (%0),%%mm1\n\t"
 115             "movq       %%mm1,%%mm2\n\t"
 116             "punpckhbw  %%mm7,%%mm1\n\t"
 117             "punpcklbw  %%mm7,%%mm2\n\t"
 118             "paddw      64(%2),%%mm2\n\t"
 119             "paddw      72(%2),%%mm1\n\t"
 120             "packuswb   %%mm1,%%mm2\n\t"
 121             "movq       %%mm2,(%0)\n\t"
 122             "addl       %3,%0\n\t"
 123
 124             "movq       (%0),%%mm1\n\t"
 125             "movq       %%mm1,%%mm2\n\t"
 126             "punpckhbw  %%mm7,%%mm1\n\t"
 127             "punpcklbw  %%mm7,%%mm2\n\t"
 128             "paddw      80(%2),%%mm2\n\t"
 129             "paddw      88(%2),%%mm1\n\t"
 130             "packuswb   %%mm1,%%mm2\n\t"
 131             "movq       %%mm2,(%0)\n\t"
 132             "addl       %3,%0\n\t"
 133
 134             "movq       (%0),%%mm1\n\t"
 135             "movq       %%mm1,%%mm2\n\t"
 136             "punpckhbw  %%mm7,%%mm1\n\t"
 137             "punpcklbw  %%mm7,%%mm2\n\t"
 138             "paddw      96(%2),%%mm2\n\t"
 139             "paddw      104(%2),%%mm1\n\t"
 140             "packuswb   %%mm1,%%mm2\n\t"
 141             "movq       %%mm2,(%0)\n\t"
 142             "addl       %3,%0\n\t"
 143
 144             "movq       (%0),%%mm1\n\t"
 145             "movq       %%mm1,%%mm2\n\t"
 146             "punpckhbw  %%mm7,%%mm1\n\t"
 147             "punpcklbw  %%mm7,%%mm2\n\t"
 148             "paddw      112(%2),%%mm2\n\t"
 149             "paddw      120(%2),%%mm1\n\t"
 150             "packuswb   %%mm1,%%mm2\n\t"
 151             "movq       %%mm2,(%0)\n\t"
 152
 153             //"emms"
 154             : "=r" (p_data)
 155             : "0" (p_data), "r" (p_block), "r" (i_incr + 8) );
 156 }
 157
 158 /*****************************************************************************
 159  * CopyBlock : copy a block
 160  *****************************************************************************/
 161 static  __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 162                                    yuv_data_t * p_data, int i_incr )
 163 {
 164     asm __volatile__ (
 165             "movq         (%2),%%mm0\n\t"
 166             "packuswb   8(%2),%%mm0\n\t"
 167             "movq        %%mm0,(%0)\n\t"
 168             "addl           %3,%0\n\t"
 169
 170             "movq        16(%2),%%mm0\n\t"
 171             "packuswb   24(%2),%%mm0\n\t"
 172             "movq        %%mm0,(%0)\n\t"
 173             "addl           %3,%0\n\t"
 174
 175             "movq        32(%2),%%mm0\n\t"
 176             "packuswb   40(%2),%%mm0\n\t"
 177             "movq        %%mm0,(%0)\n\t"
 178             "addl           %3,%0\n\t"
 179
 180             "movq        48(%2),%%mm0\n\t"
 181             "packuswb   56(%2),%%mm0\n\t"
 182             "movq        %%mm0,(%0)\n\t"
 183             "addl           %3,%0\n\t"
 184
 185             "movq        64(%2),%%mm0\n\t"
 186             "packuswb   72(%2),%%mm0\n\t"
 187             "movq        %%mm0,(%0)\n\t"
 188             "addl           %3,%0\n\t"
 189
 190             "movq        80(%2),%%mm0\n\t"
 191             "packuswb   88(%2),%%mm0\n\t"
 192             "movq        %%mm0,(%0)\n\t"
 193             "addl           %3,%0\n\t"
 194
 195             "movq        96(%2),%%mm0\n\t"
 196             "packuswb   104(%2),%%mm0\n\t"
 197             "movq        %%mm0,(%0)\n\t"
 198             "addl           %3,%0\n\t"
 199
 200             "movq        112(%2),%%mm0\n\t"
 201             "packuswb   120(%2),%%mm0\n\t"
 202             "movq        %%mm0,(%0)\n\t"
 203
 204             //"emms"
 205             : "=r" (p_data)
 206             : "0" (p_data), "r" (p_block), "r" (i_incr + 8) );
 207 }
 208
 209 void _M( vdec_DecodeMacroblockC ) ( vdec_thread_t *p_vdec,
 210                                     macroblock_t * p_mb )
 211 {
 212     if( !(p_mb->i_mb_type & MB_INTRA) )
 213     {
 214         /*
 215          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 216          */
 217         if( p_mb->pf_motion == 0 )
 218         {
 219             intf_WarnMsg( 2, "pf_motion set to NULL" );
 220         }
 221         else
 222         {
 223             p_mb->pf_motion( p_mb );
 224         }
 225
 226         DECODEBLOCKSC( AddBlock )
 227     }
 228     else
 229     {
 230         DECODEBLOCKSC( CopyBlock )
 231     }
 232 }
 233
 234 void _M( vdec_DecodeMacroblockBW ) ( vdec_thread_t *p_vdec,
 235                                      macroblock_t * p_mb )
 236 {
 237     if( !(p_mb->i_mb_type & MB_INTRA) )
 238     {
 239         /*
 240          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 241          */
 242         if( p_mb->pf_motion == 0 )
 243         {
 244             intf_WarnMsg( 2, "pf_motion set to NULL" );
 245         }
 246         else
 247         {
 248             p_mb->pf_motion( p_mb );
 249         }
 250
 251         DECODEBLOCKSBW( AddBlock )
 252     }
 253     else
 254     {
 255         DECODEBLOCKSBW( CopyBlock )
 256     }
 257 }
 258