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 GNU
  16  * General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public
  19  * License along with this program; if not, write to the
  20  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  21  * Boston, MA 02111-1307, USA.
  22  *****************************************************************************/
  23
  24 /* FIXME: passer en terminate/destroy avec les signaux supplémentaires ?? */
  25
  26 /*****************************************************************************
  27  * Preamble
  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 "threads.h"
  35 #include "config.h"
  36 #include "common.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 static void     vdec_DecodeMacroblock( vdec_thread_t *p_vdec,
  63                                        macroblock_t * p_mb );
  64 #endif
  65 static void     RunThread           ( vdec_thread_t *p_vdec );
  66 static void     ErrorThread         ( vdec_thread_t *p_vdec );
  67 static void     EndThread           ( vdec_thread_t *p_vdec );
  68
  69 /*****************************************************************************
  70  * vdec_CreateThread: create a video decoder thread
  71  *****************************************************************************
  72  * This function creates a new video decoder thread, and returns a pointer
  73  * to its description. On error, it returns NULL.
  74  * Following configuration properties are used:
  75  * XXX??
  76  *****************************************************************************/
  77 vdec_thread_t * vdec_CreateThread( vpar_thread_t *p_vpar /*, int *pi_status */ )
  78 {
  79     vdec_thread_t *     p_vdec;
  80
  81     intf_DbgMsg("vdec debug: creating video decoder thread\n");
  82
  83     /* Allocate the memory needed to store the thread's structure */
  84     if ( (p_vdec = (vdec_thread_t *)malloc( sizeof(vdec_thread_t) )) == NULL )
  85     {
  86         intf_ErrMsg("vdec error: not enough memory for vdec_CreateThread() to create the new thread\n");
  87         return( NULL );
  88     }
  89
  90     /*
  91      * Initialize the thread properties
  92      */
  93     p_vdec->b_die = 0;
  94     p_vdec->b_error = 0;
  95
  96     /*
  97      * Initialize the parser properties
  98      */
  99     p_vdec->p_vpar = p_vpar;
 100
 101     /* Spawn the video decoder thread */
 102     if ( vlc_thread_create(&p_vdec->thread_id, "video decoder",
 103          (vlc_thread_func_t)RunThread, (void *)p_vdec) )
 104     {
 105         intf_ErrMsg("vdec error: can't spawn video decoder thread\n");
 106         free( p_vdec );
 107         return( NULL );
 108     }
 109
 110     intf_DbgMsg("vdec debug: video decoder thread (%p) created\n", p_vdec);
 111     return( p_vdec );
 112 }
 113
 114 /*****************************************************************************
 115  * vdec_DestroyThread: destroy a video decoder thread
 116  *****************************************************************************
 117  * Destroy and terminate thread. This function will return 0 if the thread could
 118  * be destroyed, and non 0 else. The last case probably means that the thread
 119  * was still active, and another try may succeed.
 120  *****************************************************************************/
 121 void vdec_DestroyThread( vdec_thread_t *p_vdec /*, int *pi_status */ )
 122 {
 123     intf_DbgMsg("vdec debug: requesting termination of video decoder thread %p\n", p_vdec);
 124
 125     /* Ask thread to kill itself */
 126     p_vdec->b_die = 1;
 127
 128 #ifdef VDEC_SMP
 129     /* Make sure the decoder thread leaves the vpar_GetMacroblock() function */
 130     vlc_mutex_lock( &(p_vdec->p_vpar->vfifo.lock) );
 131     vlc_cond_signal( &(p_vdec->p_vpar->vfifo.wait) );
 132     vlc_mutex_unlock( &(p_vdec->p_vpar->vfifo.lock) );
 133 #endif
 134
 135     /* Waiting for the decoder thread to exit */
 136     /* Remove this as soon as the "status" flag is implemented */
 137     vlc_thread_join( p_vdec->thread_id );
 138 }
 139
 140 /* following functions are local */
 141
 142 /*****************************************************************************
 143  * vdec_InitThread: initialize video decoder thread
 144  *****************************************************************************
 145  * This function is called from RunThread and performs the second step of the
 146  * initialization. It returns 0 on success. Note that the thread's flag are not
 147  * modified inside this function.
 148  *****************************************************************************/
 149 #ifdef VDEC_SMP
 150 static int vdec_InitThread( vdec_thread_t *p_vdec )
 151 #else
 152 int vdec_InitThread( vdec_thread_t *p_vdec )
 153 #endif
 154 {
 155     int i_dummy;
 156
 157     intf_DbgMsg("vdec debug: initializing video decoder thread %p\n", p_vdec);
 158
 159     /* Initialize other properties */
 160 #ifdef STATS
 161     p_vdec->c_loops = 0;
 162     p_vdec->c_idle_loops = 0;
 163     p_vdec->c_decoded_pictures = 0;
 164     p_vdec->c_decoded_i_pictures = 0;
 165     p_vdec->c_decoded_p_pictures = 0;
 166     p_vdec->c_decoded_b_pictures = 0;
 167 #endif
 168
 169     /* Init crop table */
 170     p_vdec->pi_crop = p_vdec->pi_crop_buf + (VDEC_CROPRANGE >> 1);
 171     for( i_dummy = -(VDEC_CROPRANGE >> 1); i_dummy < 0; i_dummy++ )
 172     {
 173         p_vdec->pi_crop[i_dummy] = 0;
 174     }
 175     for( ; i_dummy < 255; i_dummy ++ )
 176     {
 177         p_vdec->pi_crop[i_dummy] = i_dummy;
 178     }
 179     for( ; i_dummy < (VDEC_CROPRANGE >> 1) -1; i_dummy++ )
 180     {
 181         p_vdec->pi_crop[i_dummy] = 255;
 182     }
 183
 184     /* Mark thread as running and return */
 185     intf_DbgMsg("vdec debug: InitThread(%p) succeeded\n", p_vdec);
 186     return( 0 );
 187 }
 188
 189 /*****************************************************************************
 190  * ErrorThread: RunThread() error loop
 191  *****************************************************************************
 192  * This function is called when an error occured during thread main's loop. The
 193  * thread can still receive feed, but must be ready to terminate as soon as
 194  * possible.
 195  *****************************************************************************/
 196 static void ErrorThread( vdec_thread_t *p_vdec )
 197 {
 198     macroblock_t *       p_mb;
 199
 200     /* Wait until a `die' order */
 201     while( !p_vdec->b_die )
 202     {
 203         p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo );
 204         vpar_DestroyMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 205     }
 206 }
 207
 208 /*****************************************************************************
 209  * EndThread: thread destruction
 210  *****************************************************************************
 211  * This function is called when the thread ends after a sucessfull
 212  * initialization.
 213  *****************************************************************************/
 214 static void EndThread( vdec_thread_t *p_vdec )
 215 {
 216     intf_DbgMsg("vdec debug: EndThread(%p)\n", p_vdec);
 217 }
 218
 219 /*****************************************************************************
 220  * AddBlock : add a block
 221  *****************************************************************************/
 222 #ifndef HAVE_MMX
 223 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 224                                  yuv_data_t * p_data, int i_incr )
 225 {
 226     int i_x, i_y;
 227
 228     for( i_y = 0; i_y < 8; i_y++ )
 229     {
 230         for( i_x = 0; i_x < 8; i_x++ )
 231         {
 232             *p_data = p_vdec->pi_crop[*p_data + *p_block++];
 233             p_data++;
 234         }
 235         p_data += i_incr;
 236     }
 237 }
 238 #else
 239 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 240                                           yuv_data_t * p_data, int i_incr )
 241 {
 242     asm __volatile__ (
 243             "pxor       %%mm7,%%mm7\n\t"
 244
 245             "movq       (%0),%%mm1\n\t"
 246             "movq       %%mm1,%%mm2\n\t"
 247             "punpckhbw  %%mm7,%%mm1\n\t"
 248             "punpcklbw  %%mm7,%%mm2\n\t"
 249             "paddw      (%1),%%mm2\n\t"
 250             "paddw      8(%1),%%mm1\n\t"
 251             "packuswb   %%mm1,%%mm2\n\t"
 252             "movq       %%mm2,(%0)\n\t"
 253             "addl       %2,%0\n\t"
 254
 255             "movq       (%0),%%mm1\n\t"
 256             "movq       %%mm1,%%mm2\n\t"
 257             "punpckhbw  %%mm7,%%mm1\n\t"
 258             "punpcklbw  %%mm7,%%mm2\n\t"
 259             "paddw      16(%1),%%mm2\n\t"
 260             "paddw      24(%1),%%mm1\n\t"
 261             "packuswb   %%mm1,%%mm2\n\t"
 262             "movq       %%mm2,(%0)\n\t"
 263             "addl       %2,%0\n\t"
 264
 265             "movq       (%0),%%mm1\n\t"
 266             "movq       %%mm1,%%mm2\n\t"
 267             "punpckhbw  %%mm7,%%mm1\n\t"
 268             "punpcklbw  %%mm7,%%mm2\n\t"
 269             "paddw      32(%1),%%mm2\n\t"
 270             "paddw      40(%1),%%mm1\n\t"
 271             "packuswb   %%mm1,%%mm2\n\t"
 272             "movq       %%mm2,(%0)\n\t"
 273             "addl       %2,%0\n\t"
 274
 275             "movq       (%0),%%mm1\n\t"
 276             "movq       %%mm1,%%mm2\n\t"
 277             "punpckhbw  %%mm7,%%mm1\n\t"
 278             "punpcklbw  %%mm7,%%mm2\n\t"
 279             "paddw      48(%1),%%mm2\n\t"
 280             "paddw      56(%1),%%mm1\n\t"
 281             "packuswb   %%mm1,%%mm2\n\t"
 282             "movq       %%mm2,(%0)\n\t"
 283             "addl       %2,%0\n\t"
 284
 285             "movq       (%0),%%mm1\n\t"
 286             "movq       %%mm1,%%mm2\n\t"
 287             "punpckhbw  %%mm7,%%mm1\n\t"
 288             "punpcklbw  %%mm7,%%mm2\n\t"
 289             "paddw      64(%1),%%mm2\n\t"
 290             "paddw      72(%1),%%mm1\n\t"
 291             "packuswb   %%mm1,%%mm2\n\t"
 292             "movq       %%mm2,(%0)\n\t"
 293             "addl       %2,%0\n\t"
 294
 295             "movq       (%0),%%mm1\n\t"
 296             "movq       %%mm1,%%mm2\n\t"
 297             "punpckhbw  %%mm7,%%mm1\n\t"
 298             "punpcklbw  %%mm7,%%mm2\n\t"
 299             "paddw      80(%1),%%mm2\n\t"
 300             "paddw      88(%1),%%mm1\n\t"
 301             "packuswb   %%mm1,%%mm2\n\t"
 302             "movq       %%mm2,(%0)\n\t"
 303             "addl       %2,%0\n\t"
 304
 305             "movq       (%0),%%mm1\n\t"
 306             "movq       %%mm1,%%mm2\n\t"
 307             "punpckhbw  %%mm7,%%mm1\n\t"
 308             "punpcklbw  %%mm7,%%mm2\n\t"
 309             "paddw      96(%1),%%mm2\n\t"
 310             "paddw      104(%1),%%mm1\n\t"
 311             "packuswb   %%mm1,%%mm2\n\t"
 312             "movq       %%mm2,(%0)\n\t"
 313             "addl       %2,%0\n\t"
 314
 315             "movq       (%0),%%mm1\n\t"
 316             "movq       %%mm1,%%mm2\n\t"
 317             "punpckhbw  %%mm7,%%mm1\n\t"
 318             "punpcklbw  %%mm7,%%mm2\n\t"
 319             "paddw      112(%1),%%mm2\n\t"
 320             "paddw      120(%1),%%mm1\n\t"
 321             "packuswb   %%mm1,%%mm2\n\t"
 322             "movq       %%mm2,(%0)\n\t"
 323
 324             "emms"
 325              :"+r" (p_data): "r" (p_block),"r" (i_incr+8));
 326 }
 327 #endif
 328
 329
 330 /*****************************************************************************
 331  * CopyBlock : copy a block
 332  *****************************************************************************/
 333 #ifndef HAVE_MMX
 334 static __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 335                                   yuv_data_t * p_data, int i_incr )
 336 {
 337     int i_x, i_y;
 338
 339     for( i_y = 0; i_y < 8; i_y++ )
 340     {
 341         for( i_x = 0; i_x < 8; i_x++ )
 342         {
 343             *p_data++ = p_vdec->pi_crop[*p_block++];
 344         }
 345         p_data += i_incr;
 346     }
 347 }
 348 #else
 349 static  __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 350                                           yuv_data_t * p_data, int i_incr )
 351 {
 352     asm __volatile__ (
 353             "movq         (%1),%%mm0\n\t"
 354             "packuswb   8(%1),%%mm0\n\t"
 355             "movq        %%mm0,(%0)\n\t"
 356             "addl           %2,%0\n\t"
 357
 358             "movq        16(%1),%%mm0\n\t"
 359             "packuswb   24(%1),%%mm0\n\t"
 360             "movq        %%mm0,(%0)\n\t"
 361             "addl           %2,%0\n\t"
 362
 363             "movq        32(%1),%%mm0\n\t"
 364             "packuswb   40(%1),%%mm0\n\t"
 365             "movq        %%mm0,(%0)\n\t"
 366             "addl           %2,%0\n\t"
 367
 368             "movq        48(%1),%%mm0\n\t"
 369             "packuswb   56(%1),%%mm0\n\t"
 370             "movq        %%mm0,(%0)\n\t"
 371             "addl           %2,%0\n\t"
 372
 373             "movq        64(%1),%%mm0\n\t"
 374             "packuswb   72(%1),%%mm0\n\t"
 375             "movq        %%mm0,(%0)\n\t"
 376             "addl           %2,%0\n\t"
 377
 378             "movq        80(%1),%%mm0\n\t"
 379             "packuswb   88(%1),%%mm0\n\t"
 380             "movq        %%mm0,(%0)\n\t"
 381             "addl           %2,%0\n\t"
 382
 383             "movq        96(%1),%%mm0\n\t"
 384             "packuswb   104(%1),%%mm0\n\t"
 385             "movq        %%mm0,(%0)\n\t"
 386             "addl           %2,%0\n\t"
 387
 388             "movq        112(%1),%%mm0\n\t"
 389             "packuswb   120(%1),%%mm0\n\t"
 390             "movq        %%mm0,(%0)\n\t"
 391             "emms"
 392             :"+r" (p_data): "r" (p_block),"r" (i_incr+8));
 393 }
 394 #endif
 395
 396
 397 /*****************************************************************************
 398  * vdec_DecodeMacroblock : decode a macroblock of a picture
 399  *****************************************************************************/
 400 #define DECODEBLOCKS( OPBLOCK )                                         \
 401 {                                                                       \
 402     int             i_b, i_mask;                                        \
 403                                                                         \
 404     i_mask = 1 << (3 + p_mb->i_chroma_nb_blocks);                       \
 405                                                                         \
 406     /* luminance */                                                     \
 407     for( i_b = 0; i_b < 4; i_b++, i_mask >>= 1 )                        \
 408     {                                                                   \
 409         if( p_mb->i_coded_block_pattern & i_mask )                      \
 410         {                                                               \
 411             /*                                                          \
 412              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 413              */                                                         \
 414             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 415                                   p_mb->pi_sparse_pos[i_b] );           \
 416                                                                         \
 417             /*                                                          \
 418              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 419              * section 7.6.8)                                           \
 420              */                                                         \
 421             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 422                      p_mb->p_data[i_b], p_mb->i_addb_l_stride );        \
 423         }                                                               \
 424     }                                                                   \
 425                                                                         \
 426     /* chrominance */                                                   \
 427     for( i_b = 4; i_b < 4 + p_mb->i_chroma_nb_blocks;                   \
 428          i_b++, i_mask >>= 1 )                                          \
 429     {                                                                   \
 430         if( p_mb->i_coded_block_pattern & i_mask )                      \
 431         {                                                               \
 432             /*                                                          \
 433              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 434              */                                                         \
 435             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 436                                   p_mb->pi_sparse_pos[i_b] );           \
 437                                                                         \
 438             /*                                                          \
 439              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 440              * section 7.6.8)                                           \
 441              */                                                         \
 442             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 443                      p_mb->p_data[i_b], p_mb->i_addb_c_stride );        \
 444         }                                                               \
 445     }                                                                   \
 446 }
 447
 448 #ifdef VDEC_SMP
 449 static __inline__ void vdec_DecodeMacroblock( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 450 #else
 451 void vdec_DecodeMacroblock( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 452 #endif
 453 {
 454     if( !(p_mb->i_mb_type & MB_INTRA) )
 455     {
 456         /*
 457          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 458          */
 459         if( p_mb->pf_motion == 0 )
 460         {
 461             intf_ErrMsg( "vdec error: pf_motion set to NULL\n" );
 462         }
 463         else
 464         {
 465             p_mb->pf_motion( p_mb );
 466         }
 467
 468         DECODEBLOCKS( AddBlock )
 469     }
 470     else
 471     {
 472         DECODEBLOCKS( CopyBlock )
 473     }
 474
 475     /*
 476      * Decoding is finished, release the macroblock and free
 477      * unneeded memory.
 478      */
 479     vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 480 }
 481
 482
 483 /*****************************************************************************
 484  * RunThread: video decoder thread
 485  *****************************************************************************
 486  * Video decoder thread. This function does only return when the thread is
 487  * terminated.
 488  *****************************************************************************/
 489 static void RunThread( vdec_thread_t *p_vdec )
 490 {
 491     intf_DbgMsg("vdec debug: running video decoder thread (%p) (pid == %i)\n",
 492                 p_vdec, getpid());
 493
 494     /*
 495      * Initialize thread and free configuration
 496      */
 497     p_vdec->b_error = vdec_InitThread( p_vdec );
 498     if( p_vdec->b_error )
 499     {
 500         return;
 501     }
 502     p_vdec->b_run = 1;
 503
 504     /*
 505      * Main loop - it is not executed if an error occured during
 506      * initialization
 507      */
 508     while( (!p_vdec->b_die) && (!p_vdec->b_error) )
 509     {
 510         macroblock_t *          p_mb;
 511
 512         if( (p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo )) != NULL )
 513         {
 514             vdec_DecodeMacroblock( p_vdec, p_mb );
 515         }
 516     }
 517
 518     /*
 519      * Error loop
 520      */
 521     if( p_vdec->b_error )
 522     {
 523         ErrorThread( p_vdec );
 524     }
 525
 526     /* End of thread */
 527     EndThread( p_vdec );
 528     p_vdec->b_run = 0;
 529 }