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 1.1 2001/05/06 04:32:02 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 "config.h"
  37 #include "common.h"
  38 #include "threads.h"
  39 #include "mtime.h"
  40
  41 #include "intf_msg.h"
  42
  43 #include "input_ext-dec.h"
  44
  45 #include "video.h"
  46 #include "video_output.h"
  47
  48 #include "vdec_motion.h"
  49 #include "video_decoder.h"
  50
  51 #include "vpar_blocks.h"
  52
  53 #include "vpar_headers.h"
  54 #include "vpar_synchro.h"
  55 #include "video_parser.h"
  56 #include "video_fifo.h"
  57
  58 #include "vdec_block.h"
  59
  60 /*****************************************************************************
  61  * vdec_Init: initialize video decoder thread
  62  *****************************************************************************/
  63 void _M( vdec_Init ) ( vdec_thread_t *p_vdec )
  64 {
  65     ;
  66 }
  67
  68 /*****************************************************************************
  69  * AddBlock : add a block
  70  *****************************************************************************/
  71 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
  72                                  yuv_data_t * p_data, int i_incr )
  73 {
  74     asm __volatile__ (
  75             "pxor       %%mm7,%%mm7\n\t"
  76
  77             "movq       (%0),%%mm1\n\t"
  78             "movq       %%mm1,%%mm2\n\t"
  79             "punpckhbw  %%mm7,%%mm1\n\t"
  80             "punpcklbw  %%mm7,%%mm2\n\t"
  81             "paddw      (%2),%%mm2\n\t"
  82             "paddw      8(%2),%%mm1\n\t"
  83             "packuswb   %%mm1,%%mm2\n\t"
  84             "movq       %%mm2,(%0)\n\t"
  85             "addl       %3,%0\n\t"
  86
  87             "movq       (%0),%%mm1\n\t"
  88             "movq       %%mm1,%%mm2\n\t"
  89             "punpckhbw  %%mm7,%%mm1\n\t"
  90             "punpcklbw  %%mm7,%%mm2\n\t"
  91             "paddw      16(%2),%%mm2\n\t"
  92             "paddw      24(%2),%%mm1\n\t"
  93             "packuswb   %%mm1,%%mm2\n\t"
  94             "movq       %%mm2,(%0)\n\t"
  95             "addl       %3,%0\n\t"
  96
  97             "movq       (%0),%%mm1\n\t"
  98             "movq       %%mm1,%%mm2\n\t"
  99             "punpckhbw  %%mm7,%%mm1\n\t"
 100             "punpcklbw  %%mm7,%%mm2\n\t"
 101             "paddw      32(%2),%%mm2\n\t"
 102             "paddw      40(%2),%%mm1\n\t"
 103             "packuswb   %%mm1,%%mm2\n\t"
 104             "movq       %%mm2,(%0)\n\t"
 105             "addl       %3,%0\n\t"
 106
 107             "movq       (%0),%%mm1\n\t"
 108             "movq       %%mm1,%%mm2\n\t"
 109             "punpckhbw  %%mm7,%%mm1\n\t"
 110             "punpcklbw  %%mm7,%%mm2\n\t"
 111             "paddw      48(%2),%%mm2\n\t"
 112             "paddw      56(%2),%%mm1\n\t"
 113             "packuswb   %%mm1,%%mm2\n\t"
 114             "movq       %%mm2,(%0)\n\t"
 115             "addl       %3,%0\n\t"
 116
 117             "movq       (%0),%%mm1\n\t"
 118             "movq       %%mm1,%%mm2\n\t"
 119             "punpckhbw  %%mm7,%%mm1\n\t"
 120             "punpcklbw  %%mm7,%%mm2\n\t"
 121             "paddw      64(%2),%%mm2\n\t"
 122             "paddw      72(%2),%%mm1\n\t"
 123             "packuswb   %%mm1,%%mm2\n\t"
 124             "movq       %%mm2,(%0)\n\t"
 125             "addl       %3,%0\n\t"
 126
 127             "movq       (%0),%%mm1\n\t"
 128             "movq       %%mm1,%%mm2\n\t"
 129             "punpckhbw  %%mm7,%%mm1\n\t"
 130             "punpcklbw  %%mm7,%%mm2\n\t"
 131             "paddw      80(%2),%%mm2\n\t"
 132             "paddw      88(%2),%%mm1\n\t"
 133             "packuswb   %%mm1,%%mm2\n\t"
 134             "movq       %%mm2,(%0)\n\t"
 135             "addl       %3,%0\n\t"
 136
 137             "movq       (%0),%%mm1\n\t"
 138             "movq       %%mm1,%%mm2\n\t"
 139             "punpckhbw  %%mm7,%%mm1\n\t"
 140             "punpcklbw  %%mm7,%%mm2\n\t"
 141             "paddw      96(%2),%%mm2\n\t"
 142             "paddw      104(%2),%%mm1\n\t"
 143             "packuswb   %%mm1,%%mm2\n\t"
 144             "movq       %%mm2,(%0)\n\t"
 145             "addl       %3,%0\n\t"
 146
 147             "movq       (%0),%%mm1\n\t"
 148             "movq       %%mm1,%%mm2\n\t"
 149             "punpckhbw  %%mm7,%%mm1\n\t"
 150             "punpcklbw  %%mm7,%%mm2\n\t"
 151             "paddw      112(%2),%%mm2\n\t"
 152             "paddw      120(%2),%%mm1\n\t"
 153             "packuswb   %%mm1,%%mm2\n\t"
 154             "movq       %%mm2,(%0)\n\t"
 155
 156             //"emms"
 157             : "=r" (p_data)
 158             : "0" (p_data), "r" (p_block), "r" (i_incr + 8) );
 159 }
 160
 161 /*****************************************************************************
 162  * CopyBlock : copy a block
 163  *****************************************************************************/
 164 static  __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 165                                    yuv_data_t * p_data, int i_incr )
 166 {
 167     asm __volatile__ (
 168             "movq         (%2),%%mm0\n\t"
 169             "packuswb   8(%2),%%mm0\n\t"
 170             "movq        %%mm0,(%0)\n\t"
 171             "addl           %3,%0\n\t"
 172
 173             "movq        16(%2),%%mm0\n\t"
 174             "packuswb   24(%2),%%mm0\n\t"
 175             "movq        %%mm0,(%0)\n\t"
 176             "addl           %3,%0\n\t"
 177
 178             "movq        32(%2),%%mm0\n\t"
 179             "packuswb   40(%2),%%mm0\n\t"
 180             "movq        %%mm0,(%0)\n\t"
 181             "addl           %3,%0\n\t"
 182
 183             "movq        48(%2),%%mm0\n\t"
 184             "packuswb   56(%2),%%mm0\n\t"
 185             "movq        %%mm0,(%0)\n\t"
 186             "addl           %3,%0\n\t"
 187
 188             "movq        64(%2),%%mm0\n\t"
 189             "packuswb   72(%2),%%mm0\n\t"
 190             "movq        %%mm0,(%0)\n\t"
 191             "addl           %3,%0\n\t"
 192
 193             "movq        80(%2),%%mm0\n\t"
 194             "packuswb   88(%2),%%mm0\n\t"
 195             "movq        %%mm0,(%0)\n\t"
 196             "addl           %3,%0\n\t"
 197
 198             "movq        96(%2),%%mm0\n\t"
 199             "packuswb   104(%2),%%mm0\n\t"
 200             "movq        %%mm0,(%0)\n\t"
 201             "addl           %3,%0\n\t"
 202
 203             "movq        112(%2),%%mm0\n\t"
 204             "packuswb   120(%2),%%mm0\n\t"
 205             "movq        %%mm0,(%0)\n\t"
 206
 207             //"emms"
 208             : "=r" (p_data)
 209             : "0" (p_data), "r" (p_block), "r" (i_incr + 8) );
 210 }
 211
 212 void _M( vdec_DecodeMacroblockC ) ( vdec_thread_t *p_vdec,
 213                                     macroblock_t * p_mb )
 214 {
 215     if( !(p_mb->i_mb_type & MB_INTRA) )
 216     {
 217         /*
 218          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 219          */
 220         if( p_mb->pf_motion == 0 )
 221         {
 222             intf_WarnMsg( 2, "pf_motion set to NULL" );
 223         }
 224         else
 225         {
 226             p_mb->pf_motion( p_mb );
 227         }
 228
 229         DECODEBLOCKSC( AddBlock )
 230     }
 231     else
 232     {
 233         DECODEBLOCKSC( CopyBlock )
 234     }
 235
 236     /*
 237      * Decoding is finished, release the macroblock and free
 238      * unneeded memory.
 239      */
 240     vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 241 }
 242
 243 void _M( vdec_DecodeMacroblockBW ) ( vdec_thread_t *p_vdec,
 244                                      macroblock_t * p_mb )
 245 {
 246     if( !(p_mb->i_mb_type & MB_INTRA) )
 247     {
 248         /*
 249          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 250          */
 251         if( p_mb->pf_motion == 0 )
 252         {
 253             intf_WarnMsg( 2, "pf_motion set to NULL" );
 254         }
 255         else
 256         {
 257             p_mb->pf_motion( p_mb );
 258         }
 259
 260         DECODEBLOCKSBW( AddBlock )
 261     }
 262     else
 263     {
 264         DECODEBLOCKSBW( CopyBlock )
 265     }
 266
 267     /*
 268      * Decoding is finished, release the macroblock and free
 269      * unneeded memory.
 270      */
 271     vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 272 }
 273