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

   1 /*****************************************************************************
   2  * video_decoder.c : video decoder thread
   3  * (c)1999 VideoLAN
   4  *****************************************************************************/
   5
   6 /* ?? passer en terminate/destroy avec les signaux supplémentaires */
   7
   8 /*****************************************************************************
   9  * Preamble
  10  *****************************************************************************/
  11 //#include "vlc.h"
  12
  13 #include <errno.h>
  14 #include <stdlib.h>
  15 #include <stdio.h>
  16 #include <unistd.h>
  17 #include <string.h>
  18 #include <sys/uio.h>
  19
  20 #include "config.h"
  21 #include "common.h"
  22 #include "mtime.h"
  23 #include "vlc_thread.h"
  24
  25 #include "intf_msg.h"
  26 #include "debug.h"                    /* ?? temporaire, requis par netlist.h */
  27
  28 #include "input.h"
  29 #include "input_netlist.h"
  30 #include "decoder_fifo.h"
  31 #include "video.h"
  32 #include "video_output.h"
  33
  34 #include "vdec_idct.h"
  35 #include "video_decoder.h"
  36 #include "vdec_motion.h"
  37
  38 #include "vpar_blocks.h"
  39 #include "vpar_headers.h"
  40 #include "vpar_synchro.h"
  41 #include "video_parser.h"
  42 #include "video_fifo.h"
  43
  44 /*
  45  * Local prototypes
  46  */
  47 #ifdef VDEC_SMP
  48 static int      vdec_InitThread     ( vdec_thread_t *p_vdec );
  49 static void     vdec_DecodeMacroblock( vdec_thread_t *p_vdec,
  50                                        macroblock_t * p_mb );
  51 #endif
  52 static void     RunThread           ( vdec_thread_t *p_vdec );
  53 static void     ErrorThread         ( vdec_thread_t *p_vdec );
  54 static void     EndThread           ( vdec_thread_t *p_vdec );
  55
  56 /*****************************************************************************
  57  * vdec_CreateThread: create a video decoder thread
  58  *****************************************************************************
  59  * This function creates a new video decoder thread, and returns a pointer
  60  * to its description. On error, it returns NULL.
  61  * Following configuration properties are used:
  62  * ??
  63  *****************************************************************************/
  64 vdec_thread_t * vdec_CreateThread( vpar_thread_t *p_vpar /*, int *pi_status */ )
  65 {
  66     vdec_thread_t *     p_vdec;
  67
  68     intf_DbgMsg("vdec debug: creating video decoder thread\n");
  69
  70     /* Allocate the memory needed to store the thread's structure */
  71     if ( (p_vdec = (vdec_thread_t *)malloc( sizeof(vdec_thread_t) )) == NULL )
  72     {
  73         intf_ErrMsg("vdec error: not enough memory for vdec_CreateThread() to create the new thread\n");
  74         return( NULL );
  75     }
  76
  77     /*
  78      * Initialize the thread properties
  79      */
  80     p_vdec->b_die = 0;
  81     p_vdec->b_error = 0;
  82
  83     /*
  84      * Initialize the parser properties
  85      */
  86     p_vdec->p_vpar = p_vpar;
  87
  88     /* Spawn the video decoder thread */
  89     if ( vlc_thread_create(&p_vdec->thread_id, "video decoder",
  90          (vlc_thread_func_t)RunThread, (void *)p_vdec) )
  91     {
  92         intf_ErrMsg("vdec error: can't spawn video decoder thread\n");
  93         free( p_vdec );
  94         return( NULL );
  95     }
  96
  97     intf_DbgMsg("vdec debug: video decoder thread (%p) created\n", p_vdec);
  98     return( p_vdec );
  99 }
 100
 101 /*****************************************************************************
 102  * vdec_DestroyThread: destroy a video decoder thread
 103  *****************************************************************************
 104  * Destroy and terminate thread. This function will return 0 if the thread could
 105  * be destroyed, and non 0 else. The last case probably means that the thread
 106  * was still active, and another try may succeed.
 107  *****************************************************************************/
 108 void vdec_DestroyThread( vdec_thread_t *p_vdec /*, int *pi_status */ )
 109 {
 110     intf_DbgMsg("vdec debug: requesting termination of video decoder thread %p\n", p_vdec);
 111
 112     /* Ask thread to kill itself */
 113     p_vdec->b_die = 1;
 114
 115 #ifdef VDEC_SMP
 116     /* Make sure the decoder thread leaves the vpar_GetMacroblock() function */
 117     vlc_mutex_lock( &(p_vdec->p_vpar->vfifo.lock) );
 118     vlc_cond_signal( &(p_vdec->p_vpar->vfifo.wait) );
 119     vlc_mutex_unlock( &(p_vdec->p_vpar->vfifo.lock) );
 120 #endif
 121
 122     /* Waiting for the decoder thread to exit */
 123     /* Remove this as soon as the "status" flag is implemented */
 124     vlc_thread_join( p_vdec->thread_id );
 125 }
 126
 127 /* following functions are local */
 128
 129 /*****************************************************************************
 130  * vdec_InitThread: initialize video decoder thread
 131  *****************************************************************************
 132  * This function is called from RunThread and performs the second step of the
 133  * initialization. It returns 0 on success. Note that the thread's flag are not
 134  * modified inside this function.
 135  *****************************************************************************/
 136 #ifdef VDEC_SMP
 137 static int vdec_InitThread( vdec_thread_t *p_vdec )
 138 #else
 139 int vdec_InitThread( vdec_thread_t *p_vdec )
 140 #endif
 141 {
 142     int i_dummy;
 143
 144     intf_DbgMsg("vdec debug: initializing video decoder thread %p\n", p_vdec);
 145
 146     /* Initialize other properties */
 147 #ifdef STATS
 148     p_vdec->c_loops = 0;
 149     p_vdec->c_idle_loops = 0;
 150     p_vdec->c_decoded_pictures = 0;
 151     p_vdec->c_decoded_i_pictures = 0;
 152     p_vdec->c_decoded_p_pictures = 0;
 153     p_vdec->c_decoded_b_pictures = 0;
 154 #endif
 155
 156     /* Init crop table */
 157     p_vdec->pi_crop = p_vdec->pi_crop_buf + (VDEC_CROPRANGE >> 1);
 158     for( i_dummy = -(VDEC_CROPRANGE >> 1); i_dummy < 0; i_dummy++ )
 159     {
 160         p_vdec->pi_crop[i_dummy] = 0;
 161     }
 162     for( ; i_dummy < 255; i_dummy ++ )
 163     {
 164         p_vdec->pi_crop[i_dummy] = i_dummy;
 165     }
 166     for( ; i_dummy < (VDEC_CROPRANGE >> 1) -1; i_dummy++ )
 167     {
 168         p_vdec->pi_crop[i_dummy] = 255;
 169     }
 170
 171     /* Mark thread as running and return */
 172     intf_DbgMsg("vdec debug: InitThread(%p) succeeded\n", p_vdec);
 173     return( 0 );
 174 }
 175
 176 /*****************************************************************************
 177  * ErrorThread: RunThread() error loop
 178  *****************************************************************************
 179  * This function is called when an error occured during thread main's loop. The
 180  * thread can still receive feed, but must be ready to terminate as soon as
 181  * possible.
 182  *****************************************************************************/
 183 static void ErrorThread( vdec_thread_t *p_vdec )
 184 {
 185     macroblock_t *       p_mb;
 186
 187     /* Wait until a `die' order */
 188     while( !p_vdec->b_die )
 189     {
 190         p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo );
 191         vpar_DestroyMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 192     }
 193 }
 194
 195 /*****************************************************************************
 196  * EndThread: thread destruction
 197  *****************************************************************************
 198  * This function is called when the thread ends after a sucessfull
 199  * initialization.
 200  *****************************************************************************/
 201 static void EndThread( vdec_thread_t *p_vdec )
 202 {
 203     intf_DbgMsg("vdec debug: EndThread(%p)\n", p_vdec);
 204 }
 205
 206 /*****************************************************************************
 207  * AddBlock : add a block
 208  *****************************************************************************/
 209 #ifndef HAVE_MMX
 210 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 211                                  yuv_data_t * p_data, int i_incr )
 212 {
 213     int i_x, i_y;
 214
 215     for( i_y = 0; i_y < 8; i_y++ )
 216     {
 217         for( i_x = 0; i_x < 8; i_x++ )
 218         {
 219             *p_data = p_vdec->pi_crop[*p_data + *p_block++];
 220             p_data++;
 221         }
 222         p_data += i_incr;
 223     }
 224 }
 225 #else
 226 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 227                                           yuv_data_t * p_data, int i_incr )
 228 {
 229     asm __volatile__ (
 230             "pxor       %%mm7,%%mm7\n\t"
 231
 232             "movq       (%0),%%mm1\n\t"
 233             "movq       %%mm1,%%mm2\n\t"
 234             "punpckhbw  %%mm7,%%mm1\n\t"
 235             "punpcklbw  %%mm7,%%mm2\n\t"
 236             "paddw      (%1),%%mm2\n\t"
 237             "paddw      8(%1),%%mm1\n\t"
 238             "packuswb   %%mm1,%%mm2\n\t"
 239             "movq       %%mm2,(%0)\n\t"
 240             "addl       %2,%0\n\t"
 241
 242             "movq       (%0),%%mm1\n\t"
 243             "movq       %%mm1,%%mm2\n\t"
 244             "punpckhbw  %%mm7,%%mm1\n\t"
 245             "punpcklbw  %%mm7,%%mm2\n\t"
 246             "paddw      16(%1),%%mm2\n\t"
 247             "paddw      24(%1),%%mm1\n\t"
 248             "packuswb   %%mm1,%%mm2\n\t"
 249             "movq       %%mm2,(%0)\n\t"
 250             "addl       %2,%0\n\t"
 251
 252             "movq       (%0),%%mm1\n\t"
 253             "movq       %%mm1,%%mm2\n\t"
 254             "punpckhbw  %%mm7,%%mm1\n\t"
 255             "punpcklbw  %%mm7,%%mm2\n\t"
 256             "paddw      32(%1),%%mm2\n\t"
 257             "paddw      40(%1),%%mm1\n\t"
 258             "packuswb   %%mm1,%%mm2\n\t"
 259             "movq       %%mm2,(%0)\n\t"
 260             "addl       %2,%0\n\t"
 261
 262             "movq       (%0),%%mm1\n\t"
 263             "movq       %%mm1,%%mm2\n\t"
 264             "punpckhbw  %%mm7,%%mm1\n\t"
 265             "punpcklbw  %%mm7,%%mm2\n\t"
 266             "paddw      48(%1),%%mm2\n\t"
 267             "paddw      56(%1),%%mm1\n\t"
 268             "packuswb   %%mm1,%%mm2\n\t"
 269             "movq       %%mm2,(%0)\n\t"
 270             "addl       %2,%0\n\t"
 271
 272             "movq       (%0),%%mm1\n\t"
 273             "movq       %%mm1,%%mm2\n\t"
 274             "punpckhbw  %%mm7,%%mm1\n\t"
 275             "punpcklbw  %%mm7,%%mm2\n\t"
 276             "paddw      64(%1),%%mm2\n\t"
 277             "paddw      72(%1),%%mm1\n\t"
 278             "packuswb   %%mm1,%%mm2\n\t"
 279             "movq       %%mm2,(%0)\n\t"
 280             "addl       %2,%0\n\t"
 281
 282             "movq       (%0),%%mm1\n\t"
 283             "movq       %%mm1,%%mm2\n\t"
 284             "punpckhbw  %%mm7,%%mm1\n\t"
 285             "punpcklbw  %%mm7,%%mm2\n\t"
 286             "paddw      80(%1),%%mm2\n\t"
 287             "paddw      88(%1),%%mm1\n\t"
 288             "packuswb   %%mm1,%%mm2\n\t"
 289             "movq       %%mm2,(%0)\n\t"
 290             "addl       %2,%0\n\t"
 291
 292             "movq       (%0),%%mm1\n\t"
 293             "movq       %%mm1,%%mm2\n\t"
 294             "punpckhbw  %%mm7,%%mm1\n\t"
 295             "punpcklbw  %%mm7,%%mm2\n\t"
 296             "paddw      96(%1),%%mm2\n\t"
 297             "paddw      104(%1),%%mm1\n\t"
 298             "packuswb   %%mm1,%%mm2\n\t"
 299             "movq       %%mm2,(%0)\n\t"
 300             "addl       %2,%0\n\t"
 301
 302             "movq       (%0),%%mm1\n\t"
 303             "movq       %%mm1,%%mm2\n\t"
 304             "punpckhbw  %%mm7,%%mm1\n\t"
 305             "punpcklbw  %%mm7,%%mm2\n\t"
 306             "paddw      112(%1),%%mm2\n\t"
 307             "paddw      120(%1),%%mm1\n\t"
 308             "packuswb   %%mm1,%%mm2\n\t"
 309             "movq       %%mm2,(%0)\n\t"
 310
 311             "emms"
 312              :"+r" (p_data): "r" (p_block),"r" (i_incr+8));
 313 }
 314 #endif
 315
 316
 317 /*****************************************************************************
 318  * CopyBlock : copy a block
 319  *****************************************************************************/
 320 #ifndef HAVE_MMX
 321 static __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 322                                   yuv_data_t * p_data, int i_incr )
 323 {
 324     int i_x, i_y;
 325
 326     for( i_y = 0; i_y < 8; i_y++ )
 327     {
 328         for( i_x = 0; i_x < 8; i_x++ )
 329         {
 330             *p_data++ = p_vdec->pi_crop[*p_block++];
 331         }
 332         p_data += i_incr;
 333     }
 334 }
 335 #else
 336 static  __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 337                                           yuv_data_t * p_data, int i_incr )
 338 {
 339     asm __volatile__ (
 340             "movq         (%1),%%mm0\n\t"
 341             "packuswb   8(%1),%%mm0\n\t"
 342             "movq        %%mm0,(%0)\n\t"
 343             "addl           %2,%0\n\t"
 344
 345             "movq        16(%1),%%mm0\n\t"
 346             "packuswb   24(%1),%%mm0\n\t"
 347             "movq        %%mm0,(%0)\n\t"
 348             "addl           %2,%0\n\t"
 349
 350             "movq        32(%1),%%mm0\n\t"
 351             "packuswb   40(%1),%%mm0\n\t"
 352             "movq        %%mm0,(%0)\n\t"
 353             "addl           %2,%0\n\t"
 354
 355             "movq        48(%1),%%mm0\n\t"
 356             "packuswb   56(%1),%%mm0\n\t"
 357             "movq        %%mm0,(%0)\n\t"
 358             "addl           %2,%0\n\t"
 359
 360             "movq        64(%1),%%mm0\n\t"
 361             "packuswb   72(%1),%%mm0\n\t"
 362             "movq        %%mm0,(%0)\n\t"
 363             "addl           %2,%0\n\t"
 364
 365             "movq        80(%1),%%mm0\n\t"
 366             "packuswb   88(%1),%%mm0\n\t"
 367             "movq        %%mm0,(%0)\n\t"
 368             "addl           %2,%0\n\t"
 369
 370             "movq        96(%1),%%mm0\n\t"
 371             "packuswb   104(%1),%%mm0\n\t"
 372             "movq        %%mm0,(%0)\n\t"
 373             "addl           %2,%0\n\t"
 374
 375             "movq        112(%1),%%mm0\n\t"
 376             "packuswb   120(%1),%%mm0\n\t"
 377             "movq        %%mm0,(%0)\n\t"
 378             "emms"
 379             :"+r" (p_data): "r" (p_block),"r" (i_incr+8));
 380 }
 381 #endif
 382
 383
 384 /*****************************************************************************
 385  * vdec_DecodeMacroblock : decode a macroblock of a picture
 386  *****************************************************************************/
 387 #define DECODEBLOCKS( OPBLOCK )                                         \
 388 {                                                                       \
 389     int             i_b, i_mask;                                        \
 390                                                                         \
 391     i_mask = 1 << (3 + p_mb->i_chroma_nb_blocks);                       \
 392                                                                         \
 393     /* luminance */                                                     \
 394     for( i_b = 0; i_b < 4; i_b++, i_mask >>= 1 )                        \
 395     {                                                                   \
 396         if( p_mb->i_coded_block_pattern & i_mask )                      \
 397         {                                                               \
 398             /*                                                          \
 399              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 400              */                                                         \
 401             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 402                                   p_mb->pi_sparse_pos[i_b] );           \
 403                                                                         \
 404             /*                                                          \
 405              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 406              * section 7.6.8)                                           \
 407              */                                                         \
 408             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 409                      p_mb->p_data[i_b], p_mb->i_addb_l_stride );        \
 410         }                                                               \
 411     }                                                                   \
 412                                                                         \
 413     /* chrominance */                                                   \
 414     for( i_b = 4; i_b < 4 + p_mb->i_chroma_nb_blocks;                   \
 415          i_b++, i_mask >>= 1 )                                          \
 416     {                                                                   \
 417         if( p_mb->i_coded_block_pattern & i_mask )                      \
 418         {                                                               \
 419             /*                                                          \
 420              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 421              */                                                         \
 422             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 423                                   p_mb->pi_sparse_pos[i_b] );           \
 424                                                                         \
 425             /*                                                          \
 426              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 427              * section 7.6.8)                                           \
 428              */                                                         \
 429             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 430                      p_mb->p_data[i_b], p_mb->i_addb_c_stride );        \
 431         }                                                               \
 432     }                                                                   \
 433 }
 434
 435 #ifdef VDEC_SMP
 436 static __inline__ void vdec_DecodeMacroblock( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 437 #else
 438 void vdec_DecodeMacroblock( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 439 #endif
 440 {
 441     if( !(p_mb->i_mb_type & MB_INTRA) )
 442     {
 443         /*
 444          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 445          */
 446         p_mb->pf_motion( p_mb );
 447
 448         DECODEBLOCKS( AddBlock )
 449     }
 450     else
 451     {
 452         DECODEBLOCKS( CopyBlock )
 453     }
 454
 455     /*
 456      * Decoding is finished, release the macroblock and free
 457      * unneeded memory.
 458      */
 459     vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 460 }
 461
 462
 463 /*****************************************************************************
 464  * RunThread: video decoder thread
 465  *****************************************************************************
 466  * Video decoder thread. This function does only return when the thread is
 467  * terminated.
 468  *****************************************************************************/
 469 static void RunThread( vdec_thread_t *p_vdec )
 470 {
 471     intf_DbgMsg("vdec debug: running video decoder thread (%p) (pid == %i)\n",
 472                 p_vdec, getpid());
 473
 474     /*
 475      * Initialize thread and free configuration
 476      */
 477     p_vdec->b_error = vdec_InitThread( p_vdec );
 478     if( p_vdec->b_error )
 479     {
 480         return;
 481     }
 482     p_vdec->b_run = 1;
 483
 484     /*
 485      * Main loop - it is not executed if an error occured during
 486      * initialization
 487      */
 488     while( (!p_vdec->b_die) && (!p_vdec->b_error) )
 489     {
 490         macroblock_t *          p_mb;
 491
 492         if( (p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo )) != NULL )
 493         {
 494             vdec_DecodeMacroblock( p_vdec, p_mb );
 495         }
 496     }
 497
 498     /*
 499      * Error loop
 500      */
 501     if( p_vdec->b_error )
 502     {
 503         ErrorThread( p_vdec );
 504     }
 505
 506     /* End of thread */
 507     EndThread( p_vdec );
 508     p_vdec->b_run = 0;
 509 }