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:
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  21  *****************************************************************************/
  22
  23 /* FIXME: passer en terminate/destroy avec les signaux supplémentaires ?? */
  24
  25 /*****************************************************************************
  26  * Preamble
  27  *****************************************************************************/
  28 #include "defs.h"
  29
  30 #include <stdlib.h>                                                /* free() */
  31 #include <unistd.h>                                              /* getpid() */
  32 #include <sys/types.h>                        /* on BSD, uio.h needs types.h */
  33 #include <sys/uio.h>                                          /* for input.h */
  34
  35 #include "config.h"
  36 #include "common.h"
  37 #include "threads.h"
  38 #include "mtime.h"
  39 #include "plugins.h"
  40
  41 #include "intf_msg.h"
  42
  43 #include "input.h"
  44 #include "decoder_fifo.h"
  45 #include "video.h"
  46 #include "video_output.h"
  47
  48 #include "vdec_idct.h"
  49 #include "video_decoder.h"
  50 #include "vdec_motion.h"
  51
  52 #include "vpar_blocks.h"
  53 #include "vpar_headers.h"
  54 #include "vpar_synchro.h"
  55 #include "video_parser.h"
  56 #include "video_fifo.h"
  57
  58 /*
  59  * Local prototypes
  60  */
  61 #ifdef VDEC_SMP
  62 static int      vdec_InitThread     ( vdec_thread_t *p_vdec );
  63 static void     vdec_DecodeMacroblock( vdec_thread_t *p_vdec,
  64                                        macroblock_t * p_mb );
  65 #endif
  66 static void     RunThread           ( vdec_thread_t *p_vdec );
  67 static void     ErrorThread         ( vdec_thread_t *p_vdec );
  68 static void     EndThread           ( vdec_thread_t *p_vdec );
  69
  70 /*****************************************************************************
  71  * vdec_CreateThread: create a video decoder thread
  72  *****************************************************************************
  73  * This function creates a new video decoder thread, and returns a pointer
  74  * to its description. On error, it returns NULL.
  75  * Following configuration properties are used:
  76  * XXX??
  77  *****************************************************************************/
  78 vdec_thread_t * vdec_CreateThread( vpar_thread_t *p_vpar /*, int *pi_status */ )
  79 {
  80     vdec_thread_t *     p_vdec;
  81
  82     intf_DbgMsg("vdec debug: creating video decoder thread\n");
  83
  84     /* Allocate the memory needed to store the thread's structure */
  85     if ( (p_vdec = (vdec_thread_t *)malloc( sizeof(vdec_thread_t) )) == NULL )
  86     {
  87         intf_ErrMsg("vdec error: not enough memory for vdec_CreateThread() to create the new thread\n");
  88         return( NULL );
  89     }
  90
  91     /*
  92      * Initialize the thread properties
  93      */
  94     p_vdec->b_die = 0;
  95     p_vdec->b_error = 0;
  96
  97     /*
  98      * Initialize the parser properties
  99      */
 100     p_vdec->p_vpar = p_vpar;
 101
 102     /* Spawn the video decoder thread */
 103     if ( vlc_thread_create(&p_vdec->thread_id, "video decoder",
 104          (vlc_thread_func_t)RunThread, (void *)p_vdec) )
 105     {
 106         intf_ErrMsg("vdec error: can't spawn video decoder thread\n");
 107         free( p_vdec );
 108         return( NULL );
 109     }
 110
 111     intf_DbgMsg("vdec debug: video decoder thread (%p) created\n", p_vdec);
 112     return( p_vdec );
 113 }
 114
 115 /*****************************************************************************
 116  * vdec_DestroyThread: destroy a video decoder thread
 117  *****************************************************************************
 118  * Destroy and terminate thread. This function will return 0 if the thread could
 119  * be destroyed, and non 0 else. The last case probably means that the thread
 120  * was still active, and another try may succeed.
 121  *****************************************************************************/
 122 void vdec_DestroyThread( vdec_thread_t *p_vdec /*, int *pi_status */ )
 123 {
 124     intf_DbgMsg("vdec debug: requesting termination of video decoder thread %p\n", p_vdec);
 125
 126     /* Ask thread to kill itself */
 127     p_vdec->b_die = 1;
 128
 129 #ifdef VDEC_SMP
 130     /* Make sure the decoder thread leaves the vpar_GetMacroblock() function */
 131     vlc_mutex_lock( &(p_vdec->p_vpar->vfifo.lock) );
 132     vlc_cond_signal( &(p_vdec->p_vpar->vfifo.wait) );
 133     vlc_mutex_unlock( &(p_vdec->p_vpar->vfifo.lock) );
 134 #endif
 135
 136     /* Waiting for the decoder thread to exit */
 137     /* Remove this as soon as the "status" flag is implemented */
 138     vlc_thread_join( p_vdec->thread_id );
 139 }
 140
 141 /* following functions are local */
 142
 143 /*****************************************************************************
 144  * vdec_InitThread: initialize video decoder thread
 145  *****************************************************************************
 146  * This function is called from RunThread and performs the second step of the
 147  * initialization. It returns 0 on success. Note that the thread's flag are not
 148  * modified inside this function.
 149  *****************************************************************************/
 150 #ifdef VDEC_SMP
 151 static int vdec_InitThread( vdec_thread_t *p_vdec )
 152 #else
 153 int vdec_InitThread( vdec_thread_t *p_vdec )
 154 #endif
 155 {
 156     int i_dummy;
 157
 158     intf_DbgMsg("vdec debug: initializing video decoder thread %p\n", p_vdec);
 159
 160     /* Initialize other properties */
 161 #ifdef STATS
 162     p_vdec->c_loops = 0;
 163     p_vdec->c_idle_loops = 0;
 164     p_vdec->c_decoded_pictures = 0;
 165     p_vdec->c_decoded_i_pictures = 0;
 166     p_vdec->c_decoded_p_pictures = 0;
 167     p_vdec->c_decoded_b_pictures = 0;
 168 #endif
 169
 170     /* Init crop table */
 171     p_vdec->pi_crop = p_vdec->pi_crop_buf + (VDEC_CROPRANGE >> 1);
 172     for( i_dummy = -(VDEC_CROPRANGE >> 1); i_dummy < 0; i_dummy++ )
 173     {
 174         p_vdec->pi_crop[i_dummy] = 0;
 175     }
 176     for( ; i_dummy < 255; i_dummy ++ )
 177     {
 178         p_vdec->pi_crop[i_dummy] = i_dummy;
 179     }
 180     for( ; i_dummy < (VDEC_CROPRANGE >> 1) -1; i_dummy++ )
 181     {
 182         p_vdec->pi_crop[i_dummy] = 255;
 183     }
 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 DECODEBLOCKS( 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 #ifdef VDEC_SMP
 450 static __inline__ void vdec_DecodeMacroblock( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 451 #else
 452 void vdec_DecodeMacroblock( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 453 #endif
 454 {
 455     if( !(p_mb->i_mb_type & MB_INTRA) )
 456     {
 457         /*
 458          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 459          */
 460         if( p_mb->pf_motion == 0 )
 461         {
 462             intf_ErrMsg( "vdec error: pf_motion set to NULL\n" );
 463         }
 464         else
 465         {
 466             p_mb->pf_motion( p_mb );
 467         }
 468
 469         DECODEBLOCKS( AddBlock )
 470     }
 471     else
 472     {
 473         DECODEBLOCKS( CopyBlock )
 474     }
 475
 476     /*
 477      * Decoding is finished, release the macroblock and free
 478      * unneeded memory.
 479      */
 480     vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 481 }
 482
 483
 484 /*****************************************************************************
 485  * RunThread: video decoder thread
 486  *****************************************************************************
 487  * Video decoder thread. This function does only return when the thread is
 488  * terminated.
 489  *****************************************************************************/
 490 static void RunThread( vdec_thread_t *p_vdec )
 491 {
 492     intf_DbgMsg("vdec debug: running video decoder thread (%p) (pid == %i)\n",
 493                 p_vdec, getpid());
 494
 495     /*
 496      * Initialize thread and free configuration
 497      */
 498     p_vdec->b_error = vdec_InitThread( p_vdec );
 499     if( p_vdec->b_error )
 500     {
 501         return;
 502     }
 503     p_vdec->b_run = 1;
 504
 505     /*
 506      * Main loop - it is not executed if an error occured during
 507      * initialization
 508      */
 509     while( (!p_vdec->b_die) && (!p_vdec->b_error) )
 510     {
 511         macroblock_t *          p_mb;
 512
 513         if( (p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo )) != NULL )
 514         {
 515             vdec_DecodeMacroblock( p_vdec, p_mb );
 516         }
 517     }
 518
 519     /*
 520      * Error loop
 521      */
 522     if( p_vdec->b_error )
 523     {
 524         ErrorThread( p_vdec );
 525     }
 526
 527     /* End of thread */
 528     EndThread( p_vdec );
 529     p_vdec->b_run = 0;
 530 }