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 1.4 2001/06/20 07:43:48 sam 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_motion.h"
  51 #include "video_decoder.h"
  52
  53 #include "vpar_blocks.h"
  54
  55 #include "vpar_headers.h"
  56 #include "vpar_synchro.h"
  57 #include "video_parser.h"
  58 #include "video_fifo.h"
  59
  60 #include "vdec_block.h"
  61
  62 #include "modules.h"
  63 #include "modules_export.h"
  64
  65 /*****************************************************************************
  66  * vdec_InitDecode: initialize video decoder thread
  67  *****************************************************************************/
  68 void _M( vdec_InitDecode ) ( vdec_thread_t *p_vdec )
  69 {
  70     ;
  71 }
  72
  73 /*****************************************************************************
  74  * AddBlock : add a block
  75  *****************************************************************************/
  76 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
  77                                  yuv_data_t * p_data, int i_incr )
  78 {
  79     asm __volatile__ (
  80             "pxor       %%mm7,%%mm7\n\t"
  81
  82             "movq       (%0),%%mm1\n\t"
  83             "movq       %%mm1,%%mm2\n\t"
  84             "punpckhbw  %%mm7,%%mm1\n\t"
  85             "punpcklbw  %%mm7,%%mm2\n\t"
  86             "paddw      (%2),%%mm2\n\t"
  87             "paddw      8(%2),%%mm1\n\t"
  88             "packuswb   %%mm1,%%mm2\n\t"
  89             "movq       %%mm2,(%0)\n\t"
  90             "addl       %3,%0\n\t"
  91
  92             "movq       (%0),%%mm1\n\t"
  93             "movq       %%mm1,%%mm2\n\t"
  94             "punpckhbw  %%mm7,%%mm1\n\t"
  95             "punpcklbw  %%mm7,%%mm2\n\t"
  96             "paddw      16(%2),%%mm2\n\t"
  97             "paddw      24(%2),%%mm1\n\t"
  98             "packuswb   %%mm1,%%mm2\n\t"
  99             "movq       %%mm2,(%0)\n\t"
 100             "addl       %3,%0\n\t"
 101
 102             "movq       (%0),%%mm1\n\t"
 103             "movq       %%mm1,%%mm2\n\t"
 104             "punpckhbw  %%mm7,%%mm1\n\t"
 105             "punpcklbw  %%mm7,%%mm2\n\t"
 106             "paddw      32(%2),%%mm2\n\t"
 107             "paddw      40(%2),%%mm1\n\t"
 108             "packuswb   %%mm1,%%mm2\n\t"
 109             "movq       %%mm2,(%0)\n\t"
 110             "addl       %3,%0\n\t"
 111
 112             "movq       (%0),%%mm1\n\t"
 113             "movq       %%mm1,%%mm2\n\t"
 114             "punpckhbw  %%mm7,%%mm1\n\t"
 115             "punpcklbw  %%mm7,%%mm2\n\t"
 116             "paddw      48(%2),%%mm2\n\t"
 117             "paddw      56(%2),%%mm1\n\t"
 118             "packuswb   %%mm1,%%mm2\n\t"
 119             "movq       %%mm2,(%0)\n\t"
 120             "addl       %3,%0\n\t"
 121
 122             "movq       (%0),%%mm1\n\t"
 123             "movq       %%mm1,%%mm2\n\t"
 124             "punpckhbw  %%mm7,%%mm1\n\t"
 125             "punpcklbw  %%mm7,%%mm2\n\t"
 126             "paddw      64(%2),%%mm2\n\t"
 127             "paddw      72(%2),%%mm1\n\t"
 128             "packuswb   %%mm1,%%mm2\n\t"
 129             "movq       %%mm2,(%0)\n\t"
 130             "addl       %3,%0\n\t"
 131
 132             "movq       (%0),%%mm1\n\t"
 133             "movq       %%mm1,%%mm2\n\t"
 134             "punpckhbw  %%mm7,%%mm1\n\t"
 135             "punpcklbw  %%mm7,%%mm2\n\t"
 136             "paddw      80(%2),%%mm2\n\t"
 137             "paddw      88(%2),%%mm1\n\t"
 138             "packuswb   %%mm1,%%mm2\n\t"
 139             "movq       %%mm2,(%0)\n\t"
 140             "addl       %3,%0\n\t"
 141
 142             "movq       (%0),%%mm1\n\t"
 143             "movq       %%mm1,%%mm2\n\t"
 144             "punpckhbw  %%mm7,%%mm1\n\t"
 145             "punpcklbw  %%mm7,%%mm2\n\t"
 146             "paddw      96(%2),%%mm2\n\t"
 147             "paddw      104(%2),%%mm1\n\t"
 148             "packuswb   %%mm1,%%mm2\n\t"
 149             "movq       %%mm2,(%0)\n\t"
 150             "addl       %3,%0\n\t"
 151
 152             "movq       (%0),%%mm1\n\t"
 153             "movq       %%mm1,%%mm2\n\t"
 154             "punpckhbw  %%mm7,%%mm1\n\t"
 155             "punpcklbw  %%mm7,%%mm2\n\t"
 156             "paddw      112(%2),%%mm2\n\t"
 157             "paddw      120(%2),%%mm1\n\t"
 158             "packuswb   %%mm1,%%mm2\n\t"
 159             "movq       %%mm2,(%0)\n\t"
 160
 161             //"emms"
 162             : "=r" (p_data)
 163             : "0" (p_data), "r" (p_block), "r" (i_incr + 8) );
 164 }
 165
 166 /*****************************************************************************
 167  * CopyBlock : copy a block
 168  *****************************************************************************/
 169 static  __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 170                                    yuv_data_t * p_data, int i_incr )
 171 {
 172     asm __volatile__ (
 173             "movq         (%2),%%mm0\n\t"
 174             "packuswb   8(%2),%%mm0\n\t"
 175             "movq        %%mm0,(%0)\n\t"
 176             "addl           %3,%0\n\t"
 177
 178             "movq        16(%2),%%mm0\n\t"
 179             "packuswb   24(%2),%%mm0\n\t"
 180             "movq        %%mm0,(%0)\n\t"
 181             "addl           %3,%0\n\t"
 182
 183             "movq        32(%2),%%mm0\n\t"
 184             "packuswb   40(%2),%%mm0\n\t"
 185             "movq        %%mm0,(%0)\n\t"
 186             "addl           %3,%0\n\t"
 187
 188             "movq        48(%2),%%mm0\n\t"
 189             "packuswb   56(%2),%%mm0\n\t"
 190             "movq        %%mm0,(%0)\n\t"
 191             "addl           %3,%0\n\t"
 192
 193             "movq        64(%2),%%mm0\n\t"
 194             "packuswb   72(%2),%%mm0\n\t"
 195             "movq        %%mm0,(%0)\n\t"
 196             "addl           %3,%0\n\t"
 197
 198             "movq        80(%2),%%mm0\n\t"
 199             "packuswb   88(%2),%%mm0\n\t"
 200             "movq        %%mm0,(%0)\n\t"
 201             "addl           %3,%0\n\t"
 202
 203             "movq        96(%2),%%mm0\n\t"
 204             "packuswb   104(%2),%%mm0\n\t"
 205             "movq        %%mm0,(%0)\n\t"
 206             "addl           %3,%0\n\t"
 207
 208             "movq        112(%2),%%mm0\n\t"
 209             "packuswb   120(%2),%%mm0\n\t"
 210             "movq        %%mm0,(%0)\n\t"
 211
 212             //"emms"
 213             : "=r" (p_data)
 214             : "0" (p_data), "r" (p_block), "r" (i_incr + 8) );
 215 }
 216
 217 void _M( vdec_DecodeMacroblockC ) ( vdec_thread_t *p_vdec,
 218                                     macroblock_t * p_mb )
 219 {
 220     if( !(p_mb->i_mb_type & MB_INTRA) )
 221     {
 222         /*
 223          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 224          */
 225         if( p_mb->pf_motion == 0 )
 226         {
 227             intf_WarnMsg( 2, "pf_motion set to NULL" );
 228         }
 229         else
 230         {
 231             p_mb->pf_motion( p_mb );
 232         }
 233
 234         DECODEBLOCKSC( AddBlock )
 235     }
 236     else
 237     {
 238         DECODEBLOCKSC( CopyBlock )
 239     }
 240 }
 241
 242 void _M( vdec_DecodeMacroblockBW ) ( vdec_thread_t *p_vdec,
 243                                      macroblock_t * p_mb )
 244 {
 245     if( !(p_mb->i_mb_type & MB_INTRA) )
 246     {
 247         /*
 248          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 249          */
 250         if( p_mb->pf_motion == 0 )
 251         {
 252             intf_WarnMsg( 2, "pf_motion set to NULL" );
 253         }
 254         else
 255         {
 256             p_mb->pf_motion( p_mb );
 257         }
 258
 259         DECODEBLOCKSBW( AddBlock )
 260     }
 261     else
 262     {
 263         DECODEBLOCKSBW( CopyBlock )
 264     }
 265 }
 266