git.sesse.net Git - vlc/blob - src/video_decoder/video_decoder.c

   1 /*****************************************************************************
   2  * video_decoder.c : video decoder thread
   3  *****************************************************************************
   4  * Copyright (C) 1999, 2000 VideoLAN
   5  *
   6  * Authors: Christophe Massiot <massiot@via.ecp.fr>
   7  *          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 /*****************************************************************************
  25  * Preamble
  26  *****************************************************************************/
  27 #include "defs.h"
  28
  29 #include <stdlib.h>                                                /* free() */
  30 #include <unistd.h>                                              /* getpid() */
  31 #include <sys/types.h>                        /* on BSD, uio.h needs types.h */
  32 #include <sys/uio.h>                                          /* for input.h */
  33
  34 #include "config.h"
  35 #include "common.h"
  36 #include "threads.h"
  37 #include "mtime.h"
  38 #include "plugins.h"
  39
  40 #include "intf_msg.h"
  41
  42 #include "input.h"
  43 #include "decoder_fifo.h"
  44 #include "video.h"
  45 #include "video_output.h"
  46
  47 #include "vdec_idct.h"
  48 #include "video_decoder.h"
  49 #include "vdec_motion.h"
  50
  51 #include "vpar_blocks.h"
  52 #include "vpar_headers.h"
  53 #include "vpar_synchro.h"
  54 #include "video_parser.h"
  55 #include "video_fifo.h"
  56
  57 /*
  58  * Local prototypes
  59  */
  60 #ifdef VDEC_SMP
  61 static int      vdec_InitThread     ( vdec_thread_t *p_vdec );
  62 #endif
  63 static void     RunThread           ( vdec_thread_t *p_vdec );
  64 static void     ErrorThread         ( vdec_thread_t *p_vdec );
  65 static void     EndThread           ( vdec_thread_t *p_vdec );
  66
  67 /*****************************************************************************
  68  * vdec_CreateThread: create a video decoder thread
  69  *****************************************************************************
  70  * This function creates a new video decoder thread, and returns a pointer
  71  * to its description. On error, it returns NULL.
  72  * Following configuration properties are used:
  73  * XXX??
  74  *****************************************************************************/
  75 vdec_thread_t * vdec_CreateThread( vpar_thread_t *p_vpar /*, int *pi_status */ )
  76 {
  77     vdec_thread_t *     p_vdec;
  78
  79     intf_DbgMsg("vdec debug: creating video decoder thread\n");
  80
  81     /* Allocate the memory needed to store the thread's structure */
  82     if ( (p_vdec = (vdec_thread_t *)malloc( sizeof(vdec_thread_t) )) == NULL )
  83     {
  84         intf_ErrMsg("vdec error: not enough memory for vdec_CreateThread() to create the new thread\n");
  85         return( NULL );
  86     }
  87
  88     /*
  89      * Initialize the thread properties
  90      */
  91     p_vdec->b_die = 0;
  92     p_vdec->b_error = 0;
  93
  94     /*
  95      * Initialize the parser properties
  96      */
  97     p_vdec->p_vpar = p_vpar;
  98
  99     /* Spawn the video decoder thread */
 100     if ( vlc_thread_create(&p_vdec->thread_id, "video decoder",
 101          (vlc_thread_func_t)RunThread, (void *)p_vdec) )
 102     {
 103         intf_ErrMsg("vdec error: can't spawn video decoder thread\n");
 104         free( p_vdec );
 105         return( NULL );
 106     }
 107
 108     intf_DbgMsg("vdec debug: video decoder thread (%p) created\n", p_vdec);
 109     return( p_vdec );
 110 }
 111
 112 /*****************************************************************************
 113  * vdec_DestroyThread: destroy a video decoder thread
 114  *****************************************************************************
 115  * Destroy and terminate thread. This function will return 0 if the thread could
 116  * be destroyed, and non 0 else. The last case probably means that the thread
 117  * was still active, and another try may succeed.
 118  *****************************************************************************/
 119 void vdec_DestroyThread( vdec_thread_t *p_vdec /*, int *pi_status */ )
 120 {
 121     intf_DbgMsg("vdec debug: requesting termination of video decoder thread %p\n", p_vdec);
 122
 123     /* Ask thread to kill itself */
 124     p_vdec->b_die = 1;
 125
 126 #ifdef VDEC_SMP
 127     /* Make sure the decoder thread leaves the vpar_GetMacroblock() function */
 128     vlc_mutex_lock( &(p_vdec->p_vpar->vfifo.lock) );
 129     vlc_cond_signal( &(p_vdec->p_vpar->vfifo.wait) );
 130     vlc_mutex_unlock( &(p_vdec->p_vpar->vfifo.lock) );
 131 #endif
 132
 133     /* Waiting for the decoder thread to exit */
 134     /* Remove this as soon as the "status" flag is implemented */
 135     vlc_thread_join( p_vdec->thread_id );
 136 }
 137
 138 /* following functions are local */
 139
 140 /*****************************************************************************
 141  * vdec_InitThread: initialize video decoder thread
 142  *****************************************************************************
 143  * This function is called from RunThread and performs the second step of the
 144  * initialization. It returns 0 on success. Note that the thread's flag are not
 145  * modified inside this function.
 146  *****************************************************************************/
 147 #ifdef VDEC_SMP
 148 static int vdec_InitThread( vdec_thread_t *p_vdec )
 149 #else
 150 int vdec_InitThread( vdec_thread_t *p_vdec )
 151 #endif
 152 {
 153 #ifndef HAVE_MMX
 154     int i_dummy;
 155 #endif
 156
 157     intf_DbgMsg("vdec debug: initializing video decoder thread %p\n", p_vdec);
 158
 159 #ifndef HAVE_MMX
 160     /* Init crop table */
 161     p_vdec->pi_crop = p_vdec->pi_crop_buf + (VDEC_CROPRANGE >> 1);
 162     for( i_dummy = -(VDEC_CROPRANGE >> 1); i_dummy < 0; i_dummy++ )
 163     {
 164         p_vdec->pi_crop[i_dummy] = 0;
 165     }
 166     for( ; i_dummy < 255; i_dummy ++ )
 167     {
 168         p_vdec->pi_crop[i_dummy] = i_dummy;
 169     }
 170     for( ; i_dummy < (VDEC_CROPRANGE >> 1) -1; i_dummy++ )
 171     {
 172         p_vdec->pi_crop[i_dummy] = 255;
 173     }
 174 #endif
 175
 176 #ifdef VDEC_SMP
 177     /* Re-nice ourself */
 178     if( nice(VDEC_NICE) == -1 )
 179     {
 180         intf_WarnMsg( 2, "vdec warning : couldn't nice() (%s)\n",
 181                       strerror(errno) );
 182     }
 183 #endif
 184
 185     /* Mark thread as running and return */
 186     intf_DbgMsg("vdec debug: InitThread(%p) succeeded\n", p_vdec);
 187     return( 0 );
 188 }
 189
 190 /*****************************************************************************
 191  * ErrorThread: RunThread() error loop
 192  *****************************************************************************
 193  * This function is called when an error occured during thread main's loop. The
 194  * thread can still receive feed, but must be ready to terminate as soon as
 195  * possible.
 196  *****************************************************************************/
 197 static void ErrorThread( vdec_thread_t *p_vdec )
 198 {
 199     macroblock_t *       p_mb;
 200
 201     /* Wait until a `die' order */
 202     while( !p_vdec->b_die )
 203     {
 204         p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo );
 205         vpar_DestroyMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 206     }
 207 }
 208
 209 /*****************************************************************************
 210  * EndThread: thread destruction
 211  *****************************************************************************
 212  * This function is called when the thread ends after a sucessful
 213  * initialization.
 214  *****************************************************************************/
 215 static void EndThread( vdec_thread_t *p_vdec )
 216 {
 217     intf_DbgMsg("vdec debug: EndThread(%p)\n", p_vdec);
 218 }
 219
 220 /*****************************************************************************
 221  * AddBlock : add a block
 222  *****************************************************************************/
 223 #ifndef HAVE_MMX
 224 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 225                                  yuv_data_t * p_data, int i_incr )
 226 {
 227     int i_x, i_y;
 228
 229     for( i_y = 0; i_y < 8; i_y++ )
 230     {
 231         for( i_x = 0; i_x < 8; i_x++ )
 232         {
 233             *p_data = p_vdec->pi_crop[*p_data + *p_block++];
 234             p_data++;
 235         }
 236         p_data += i_incr;
 237     }
 238 }
 239 #else
 240 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 241                                           yuv_data_t * p_data, int i_incr )
 242 {
 243     asm __volatile__ (
 244             "pxor       %%mm7,%%mm7\n\t"
 245
 246             "movq       (%0),%%mm1\n\t"
 247             "movq       %%mm1,%%mm2\n\t"
 248             "punpckhbw  %%mm7,%%mm1\n\t"
 249             "punpcklbw  %%mm7,%%mm2\n\t"
 250             "paddw      (%1),%%mm2\n\t"
 251             "paddw      8(%1),%%mm1\n\t"
 252             "packuswb   %%mm1,%%mm2\n\t"
 253             "movq       %%mm2,(%0)\n\t"
 254             "addl       %2,%0\n\t"
 255
 256             "movq       (%0),%%mm1\n\t"
 257             "movq       %%mm1,%%mm2\n\t"
 258             "punpckhbw  %%mm7,%%mm1\n\t"
 259             "punpcklbw  %%mm7,%%mm2\n\t"
 260             "paddw      16(%1),%%mm2\n\t"
 261             "paddw      24(%1),%%mm1\n\t"
 262             "packuswb   %%mm1,%%mm2\n\t"
 263             "movq       %%mm2,(%0)\n\t"
 264             "addl       %2,%0\n\t"
 265
 266             "movq       (%0),%%mm1\n\t"
 267             "movq       %%mm1,%%mm2\n\t"
 268             "punpckhbw  %%mm7,%%mm1\n\t"
 269             "punpcklbw  %%mm7,%%mm2\n\t"
 270             "paddw      32(%1),%%mm2\n\t"
 271             "paddw      40(%1),%%mm1\n\t"
 272             "packuswb   %%mm1,%%mm2\n\t"
 273             "movq       %%mm2,(%0)\n\t"
 274             "addl       %2,%0\n\t"
 275
 276             "movq       (%0),%%mm1\n\t"
 277             "movq       %%mm1,%%mm2\n\t"
 278             "punpckhbw  %%mm7,%%mm1\n\t"
 279             "punpcklbw  %%mm7,%%mm2\n\t"
 280             "paddw      48(%1),%%mm2\n\t"
 281             "paddw      56(%1),%%mm1\n\t"
 282             "packuswb   %%mm1,%%mm2\n\t"
 283             "movq       %%mm2,(%0)\n\t"
 284             "addl       %2,%0\n\t"
 285
 286             "movq       (%0),%%mm1\n\t"
 287             "movq       %%mm1,%%mm2\n\t"
 288             "punpckhbw  %%mm7,%%mm1\n\t"
 289             "punpcklbw  %%mm7,%%mm2\n\t"
 290             "paddw      64(%1),%%mm2\n\t"
 291             "paddw      72(%1),%%mm1\n\t"
 292             "packuswb   %%mm1,%%mm2\n\t"
 293             "movq       %%mm2,(%0)\n\t"
 294             "addl       %2,%0\n\t"
 295
 296             "movq       (%0),%%mm1\n\t"
 297             "movq       %%mm1,%%mm2\n\t"
 298             "punpckhbw  %%mm7,%%mm1\n\t"
 299             "punpcklbw  %%mm7,%%mm2\n\t"
 300             "paddw      80(%1),%%mm2\n\t"
 301             "paddw      88(%1),%%mm1\n\t"
 302             "packuswb   %%mm1,%%mm2\n\t"
 303             "movq       %%mm2,(%0)\n\t"
 304             "addl       %2,%0\n\t"
 305
 306             "movq       (%0),%%mm1\n\t"
 307             "movq       %%mm1,%%mm2\n\t"
 308             "punpckhbw  %%mm7,%%mm1\n\t"
 309             "punpcklbw  %%mm7,%%mm2\n\t"
 310             "paddw      96(%1),%%mm2\n\t"
 311             "paddw      104(%1),%%mm1\n\t"
 312             "packuswb   %%mm1,%%mm2\n\t"
 313             "movq       %%mm2,(%0)\n\t"
 314             "addl       %2,%0\n\t"
 315
 316             "movq       (%0),%%mm1\n\t"
 317             "movq       %%mm1,%%mm2\n\t"
 318             "punpckhbw  %%mm7,%%mm1\n\t"
 319             "punpcklbw  %%mm7,%%mm2\n\t"
 320             "paddw      112(%1),%%mm2\n\t"
 321             "paddw      120(%1),%%mm1\n\t"
 322             "packuswb   %%mm1,%%mm2\n\t"
 323             "movq       %%mm2,(%0)\n\t"
 324
 325             //"emms"
 326             :"+r" (p_data): "r" (p_block),"r" (i_incr+8));
 327 }
 328 #endif
 329
 330
 331 /*****************************************************************************
 332  * CopyBlock : copy a block
 333  *****************************************************************************/
 334 #ifndef HAVE_MMX
 335 static __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 336                                   yuv_data_t * p_data, int i_incr )
 337 {
 338     int i_x, i_y;
 339
 340     for( i_y = 0; i_y < 8; i_y++ )
 341     {
 342         for( i_x = 0; i_x < 8; i_x++ )
 343         {
 344             *p_data++ = p_vdec->pi_crop[*p_block++];
 345         }
 346         p_data += i_incr;
 347     }
 348 }
 349 #else
 350 static  __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 351                                           yuv_data_t * p_data, int i_incr )
 352 {
 353     asm __volatile__ (
 354             "movq         (%1),%%mm0\n\t"
 355             "packuswb   8(%1),%%mm0\n\t"
 356             "movq        %%mm0,(%0)\n\t"
 357             "addl           %2,%0\n\t"
 358
 359             "movq        16(%1),%%mm0\n\t"
 360             "packuswb   24(%1),%%mm0\n\t"
 361             "movq        %%mm0,(%0)\n\t"
 362             "addl           %2,%0\n\t"
 363
 364             "movq        32(%1),%%mm0\n\t"
 365             "packuswb   40(%1),%%mm0\n\t"
 366             "movq        %%mm0,(%0)\n\t"
 367             "addl           %2,%0\n\t"
 368
 369             "movq        48(%1),%%mm0\n\t"
 370             "packuswb   56(%1),%%mm0\n\t"
 371             "movq        %%mm0,(%0)\n\t"
 372             "addl           %2,%0\n\t"
 373
 374             "movq        64(%1),%%mm0\n\t"
 375             "packuswb   72(%1),%%mm0\n\t"
 376             "movq        %%mm0,(%0)\n\t"
 377             "addl           %2,%0\n\t"
 378
 379             "movq        80(%1),%%mm0\n\t"
 380             "packuswb   88(%1),%%mm0\n\t"
 381             "movq        %%mm0,(%0)\n\t"
 382             "addl           %2,%0\n\t"
 383
 384             "movq        96(%1),%%mm0\n\t"
 385             "packuswb   104(%1),%%mm0\n\t"
 386             "movq        %%mm0,(%0)\n\t"
 387             "addl           %2,%0\n\t"
 388
 389             "movq        112(%1),%%mm0\n\t"
 390             "packuswb   120(%1),%%mm0\n\t"
 391             "movq        %%mm0,(%0)\n\t"
 392             //"emms"
 393             :"+r" (p_data): "r" (p_block),"r" (i_incr+8));
 394 }
 395 #endif
 396
 397
 398 /*****************************************************************************
 399  * vdec_DecodeMacroblock : decode a macroblock of a picture
 400  *****************************************************************************/
 401 #define DECODEBLOCKSC( OPBLOCK )                                        \
 402 {                                                                       \
 403     int             i_b, i_mask;                                        \
 404                                                                         \
 405     i_mask = 1 << (3 + p_mb->i_chroma_nb_blocks);                       \
 406                                                                         \
 407     /* luminance */                                                     \
 408     for( i_b = 0; i_b < 4; i_b++, i_mask >>= 1 )                        \
 409     {                                                                   \
 410         if( p_mb->i_coded_block_pattern & i_mask )                      \
 411         {                                                               \
 412             /*                                                          \
 413              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 414              */                                                         \
 415             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 416                                   p_mb->pi_sparse_pos[i_b] );           \
 417                                                                         \
 418             /*                                                          \
 419              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 420              * section 7.6.8)                                           \
 421              */                                                         \
 422             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 423                      p_mb->p_data[i_b], p_mb->i_addb_l_stride );        \
 424         }                                                               \
 425     }                                                                   \
 426                                                                         \
 427     /* chrominance */                                                   \
 428     for( i_b = 4; i_b < 4 + p_mb->i_chroma_nb_blocks;                   \
 429          i_b++, i_mask >>= 1 )                                          \
 430     {                                                                   \
 431         if( p_mb->i_coded_block_pattern & i_mask )                      \
 432         {                                                               \
 433             /*                                                          \
 434              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 435              */                                                         \
 436             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 437                                   p_mb->pi_sparse_pos[i_b] );           \
 438                                                                         \
 439             /*                                                          \
 440              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 441              * section 7.6.8)                                           \
 442              */                                                         \
 443             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 444                      p_mb->p_data[i_b], p_mb->i_addb_c_stride );        \
 445         }                                                               \
 446     }                                                                   \
 447 }
 448
 449 #define DECODEBLOCKSBW( OPBLOCK )                                       \
 450 {                                                                       \
 451     int             i_b, i_mask;                                        \
 452                                                                         \
 453     i_mask = 1 << (3 + p_mb->i_chroma_nb_blocks);                       \
 454                                                                         \
 455     /* luminance */                                                     \
 456     for( i_b = 0; i_b < 4; i_b++, i_mask >>= 1 )                        \
 457     {                                                                   \
 458         if( p_mb->i_coded_block_pattern & i_mask )                      \
 459         {                                                               \
 460             /*                                                          \
 461              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 462              */                                                         \
 463             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 464                                   p_mb->pi_sparse_pos[i_b] );           \
 465                                                                         \
 466             /*                                                          \
 467              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 468              * section 7.6.8)                                           \
 469              */                                                         \
 470             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 471                      p_mb->p_data[i_b], p_mb->i_addb_l_stride );        \
 472         }                                                               \
 473     }                                                                   \
 474 }
 475
 476 void vdec_DecodeMacroblockC ( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 477 {
 478     if( !(p_mb->i_mb_type & MB_INTRA) )
 479     {
 480         /*
 481          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 482          */
 483         if( p_mb->pf_motion == 0 )
 484         {
 485             intf_ErrMsg( "vdec error: pf_motion set to NULL\n" );
 486         }
 487         else
 488         {
 489             p_mb->pf_motion( p_mb );
 490         }
 491
 492         DECODEBLOCKSC( AddBlock )
 493     }
 494     else
 495     {
 496         DECODEBLOCKSC( CopyBlock )
 497     }
 498
 499     /*
 500      * Decoding is finished, release the macroblock and free
 501      * unneeded memory.
 502      */
 503     vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 504 }
 505
 506 void vdec_DecodeMacroblockBW ( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 507 {
 508     if( !(p_mb->i_mb_type & MB_INTRA) )
 509     {
 510         /*
 511          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 512          */
 513         if( p_mb->pf_motion == 0 )
 514         {
 515             intf_ErrMsg( "vdec error: pf_motion set to NULL\n" );
 516         }
 517         else
 518         {
 519             p_mb->pf_motion( p_mb );
 520         }
 521
 522         DECODEBLOCKSBW( AddBlock )
 523     }
 524     else
 525     {
 526         DECODEBLOCKSBW( CopyBlock )
 527     }
 528
 529     /*
 530      * Decoding is finished, release the macroblock and free
 531      * unneeded memory.
 532      */
 533     vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 534 }
 535
 536
 537
 538 /*****************************************************************************
 539  * RunThread: video decoder thread
 540  *****************************************************************************
 541  * Video decoder thread. This function does only return when the thread is
 542  * terminated.
 543  *****************************************************************************/
 544 static void RunThread( vdec_thread_t *p_vdec )
 545 {
 546     intf_DbgMsg("vdec debug: running video decoder thread (%p) (pid == %i)\n",
 547                 p_vdec, getpid());
 548
 549     /*
 550      * Initialize thread and free configuration
 551      */
 552     p_vdec->b_error = vdec_InitThread( p_vdec );
 553     if( p_vdec->b_error )
 554     {
 555         return;
 556     }
 557     p_vdec->b_run = 1;
 558
 559     /*
 560      * Main loop - it is not executed if an error occured during
 561      * initialization
 562      */
 563     while( (!p_vdec->b_die) && (!p_vdec->b_error) )
 564     {
 565         macroblock_t *          p_mb;
 566
 567         if( (p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo )) != NULL )
 568         {
 569             p_vdec->p_vpar->p_vout->vdec_DecodeMacroblock ( p_vdec, p_mb );
 570         }
 571     }
 572
 573     /*
 574      * Error loop
 575      */
 576     if( p_vdec->b_error )
 577     {
 578         ErrorThread( p_vdec );
 579     }
 580
 581     /* End of thread */
 582     EndThread( p_vdec );
 583     p_vdec->b_run = 0;
 584 }