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  * $Id: video_decoder.c,v 1.48 2001/04/06 09:15:48 sam Exp $
   6  *
   7  * Authors: Christophe Massiot <massiot@via.ecp.fr>
   8  *          Gaël Hendryckx <jimmy@via.ecp.fr>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with this program; if not, write to the Free Software
  22  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  23  *****************************************************************************/
  24
  25 /*****************************************************************************
  26  * Preamble
  27  *****************************************************************************/
  28 #include "defs.h"
  29
  30 #include <unistd.h>                                              /* getpid() */
  31
  32 #include <stdlib.h>                                                /* free() */
  33 #include <string.h>                                    /* memcpy(), memset() */
  34 #include <errno.h>                                                  /* errno */
  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 "stream_control.h"
  44 #include "input_ext-dec.h"
  45
  46 #include "video.h"
  47 #include "video_output.h"
  48
  49 #include "vdec_motion.h"
  50 #include "vdec_idct.h"
  51 #include "video_decoder.h"
  52
  53 #include "vpar_blocks.h"
  54 #include "vpar_headers.h"
  55 #include "vpar_synchro.h"
  56 #include "video_parser.h"
  57 #include "video_fifo.h"
  58
  59 /*
  60  * Local prototypes
  61  */
  62 #ifdef VDEC_SMP
  63 static int      vdec_InitThread     ( vdec_thread_t *p_vdec );
  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");
  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");
  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");
 106         free( p_vdec );
 107         return( NULL );
 108     }
 109
 110     intf_DbgMsg("vdec debug: video decoder thread (%p) created", 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", 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 #ifndef HAVE_MMX
 156     int i_dummy;
 157 #endif
 158
 159     intf_DbgMsg("vdec debug: initializing video decoder thread %p", p_vdec);
 160
 161 #ifndef HAVE_MMX
 162     /* Init crop table */
 163     p_vdec->pi_crop = p_vdec->pi_crop_buf + (VDEC_CROPRANGE >> 1);
 164     for( i_dummy = -(VDEC_CROPRANGE >> 1); i_dummy < 0; i_dummy++ )
 165     {
 166         p_vdec->pi_crop[i_dummy] = 0;
 167     }
 168     for( ; i_dummy < 255; i_dummy ++ )
 169     {
 170         p_vdec->pi_crop[i_dummy] = i_dummy;
 171     }
 172     for( ; i_dummy < (VDEC_CROPRANGE >> 1) -1; i_dummy++ )
 173     {
 174         p_vdec->pi_crop[i_dummy] = 255;
 175     }
 176 #endif
 177
 178 #ifdef VDEC_SMP
 179     /* Re-nice ourself */
 180     if( nice(VDEC_NICE) == -1 )
 181     {
 182         intf_WarnMsg( 2, "vdec warning : couldn't nice() (%s)",
 183                       strerror(errno) );
 184     }
 185 #endif
 186
 187     /* Mark thread as running and return */
 188     intf_DbgMsg("vdec debug: InitThread(%p) succeeded", p_vdec);
 189     return( 0 );
 190 }
 191
 192 /*****************************************************************************
 193  * ErrorThread: RunThread() error loop
 194  *****************************************************************************
 195  * This function is called when an error occured during thread main's loop. The
 196  * thread can still receive feed, but must be ready to terminate as soon as
 197  * possible.
 198  *****************************************************************************/
 199 static void ErrorThread( vdec_thread_t *p_vdec )
 200 {
 201     macroblock_t *       p_mb;
 202
 203     /* Wait until a `die' order */
 204     while( !p_vdec->b_die )
 205     {
 206         p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo );
 207         vpar_DestroyMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 208     }
 209 }
 210
 211 /*****************************************************************************
 212  * EndThread: thread destruction
 213  *****************************************************************************
 214  * This function is called when the thread ends after a sucessful
 215  * initialization.
 216  *****************************************************************************/
 217 static void EndThread( vdec_thread_t *p_vdec )
 218 {
 219     intf_DbgMsg("vdec debug: EndThread(%p)", p_vdec);
 220 }
 221
 222 /*****************************************************************************
 223  * AddBlock : add a block
 224  *****************************************************************************/
 225 #ifndef HAVE_MMX
 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     int i_x, i_y;
 230
 231     for( i_y = 0; i_y < 8; i_y++ )
 232     {
 233         for( i_x = 0; i_x < 8; i_x++ )
 234         {
 235             *p_data = p_vdec->pi_crop[*p_data + *p_block++];
 236             p_data++;
 237         }
 238         p_data += i_incr;
 239     }
 240 }
 241 #else
 242 static __inline__ void AddBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 243                                  yuv_data_t * p_data, int i_incr )
 244 {
 245     asm __volatile__ (
 246             "pxor       %%mm7,%%mm7\n\t"
 247
 248             "movq       (%0),%%mm1\n\t"
 249             "movq       %%mm1,%%mm2\n\t"
 250             "punpckhbw  %%mm7,%%mm1\n\t"
 251             "punpcklbw  %%mm7,%%mm2\n\t"
 252             "paddw      (%2),%%mm2\n\t"
 253             "paddw      8(%2),%%mm1\n\t"
 254             "packuswb   %%mm1,%%mm2\n\t"
 255             "movq       %%mm2,(%0)\n\t"
 256             "addl       %3,%0\n\t"
 257
 258             "movq       (%0),%%mm1\n\t"
 259             "movq       %%mm1,%%mm2\n\t"
 260             "punpckhbw  %%mm7,%%mm1\n\t"
 261             "punpcklbw  %%mm7,%%mm2\n\t"
 262             "paddw      16(%2),%%mm2\n\t"
 263             "paddw      24(%2),%%mm1\n\t"
 264             "packuswb   %%mm1,%%mm2\n\t"
 265             "movq       %%mm2,(%0)\n\t"
 266             "addl       %3,%0\n\t"
 267
 268             "movq       (%0),%%mm1\n\t"
 269             "movq       %%mm1,%%mm2\n\t"
 270             "punpckhbw  %%mm7,%%mm1\n\t"
 271             "punpcklbw  %%mm7,%%mm2\n\t"
 272             "paddw      32(%2),%%mm2\n\t"
 273             "paddw      40(%2),%%mm1\n\t"
 274             "packuswb   %%mm1,%%mm2\n\t"
 275             "movq       %%mm2,(%0)\n\t"
 276             "addl       %3,%0\n\t"
 277
 278             "movq       (%0),%%mm1\n\t"
 279             "movq       %%mm1,%%mm2\n\t"
 280             "punpckhbw  %%mm7,%%mm1\n\t"
 281             "punpcklbw  %%mm7,%%mm2\n\t"
 282             "paddw      48(%2),%%mm2\n\t"
 283             "paddw      56(%2),%%mm1\n\t"
 284             "packuswb   %%mm1,%%mm2\n\t"
 285             "movq       %%mm2,(%0)\n\t"
 286             "addl       %3,%0\n\t"
 287
 288             "movq       (%0),%%mm1\n\t"
 289             "movq       %%mm1,%%mm2\n\t"
 290             "punpckhbw  %%mm7,%%mm1\n\t"
 291             "punpcklbw  %%mm7,%%mm2\n\t"
 292             "paddw      64(%2),%%mm2\n\t"
 293             "paddw      72(%2),%%mm1\n\t"
 294             "packuswb   %%mm1,%%mm2\n\t"
 295             "movq       %%mm2,(%0)\n\t"
 296             "addl       %3,%0\n\t"
 297
 298             "movq       (%0),%%mm1\n\t"
 299             "movq       %%mm1,%%mm2\n\t"
 300             "punpckhbw  %%mm7,%%mm1\n\t"
 301             "punpcklbw  %%mm7,%%mm2\n\t"
 302             "paddw      80(%2),%%mm2\n\t"
 303             "paddw      88(%2),%%mm1\n\t"
 304             "packuswb   %%mm1,%%mm2\n\t"
 305             "movq       %%mm2,(%0)\n\t"
 306             "addl       %3,%0\n\t"
 307
 308             "movq       (%0),%%mm1\n\t"
 309             "movq       %%mm1,%%mm2\n\t"
 310             "punpckhbw  %%mm7,%%mm1\n\t"
 311             "punpcklbw  %%mm7,%%mm2\n\t"
 312             "paddw      96(%2),%%mm2\n\t"
 313             "paddw      104(%2),%%mm1\n\t"
 314             "packuswb   %%mm1,%%mm2\n\t"
 315             "movq       %%mm2,(%0)\n\t"
 316             "addl       %3,%0\n\t"
 317
 318             "movq       (%0),%%mm1\n\t"
 319             "movq       %%mm1,%%mm2\n\t"
 320             "punpckhbw  %%mm7,%%mm1\n\t"
 321             "punpcklbw  %%mm7,%%mm2\n\t"
 322             "paddw      112(%2),%%mm2\n\t"
 323             "paddw      120(%2),%%mm1\n\t"
 324             "packuswb   %%mm1,%%mm2\n\t"
 325             "movq       %%mm2,(%0)\n\t"
 326
 327             //"emms"
 328             : "=r" (p_data)
 329             : "0" (p_data), "r" (p_block), "r" (i_incr + 8) );
 330 }
 331 #endif
 332
 333
 334 /*****************************************************************************
 335  * CopyBlock : copy a block
 336  *****************************************************************************/
 337 #ifndef HAVE_MMX
 338 static __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 339                                   yuv_data_t * p_data, int i_incr )
 340 {
 341     int i_x, i_y;
 342
 343     for( i_y = 0; i_y < 8; i_y++ )
 344     {
 345         for( i_x = 0; i_x < 8; i_x++ )
 346         {
 347             *p_data++ = p_vdec->pi_crop[*p_block++];
 348         }
 349         p_data += i_incr;
 350     }
 351 }
 352 #else
 353 static  __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
 354                                    yuv_data_t * p_data, int i_incr )
 355 {
 356     asm __volatile__ (
 357             "movq         (%2),%%mm0\n\t"
 358             "packuswb   8(%2),%%mm0\n\t"
 359             "movq        %%mm0,(%0)\n\t"
 360             "addl           %3,%0\n\t"
 361
 362             "movq        16(%2),%%mm0\n\t"
 363             "packuswb   24(%2),%%mm0\n\t"
 364             "movq        %%mm0,(%0)\n\t"
 365             "addl           %3,%0\n\t"
 366
 367             "movq        32(%2),%%mm0\n\t"
 368             "packuswb   40(%2),%%mm0\n\t"
 369             "movq        %%mm0,(%0)\n\t"
 370             "addl           %3,%0\n\t"
 371
 372             "movq        48(%2),%%mm0\n\t"
 373             "packuswb   56(%2),%%mm0\n\t"
 374             "movq        %%mm0,(%0)\n\t"
 375             "addl           %3,%0\n\t"
 376
 377             "movq        64(%2),%%mm0\n\t"
 378             "packuswb   72(%2),%%mm0\n\t"
 379             "movq        %%mm0,(%0)\n\t"
 380             "addl           %3,%0\n\t"
 381
 382             "movq        80(%2),%%mm0\n\t"
 383             "packuswb   88(%2),%%mm0\n\t"
 384             "movq        %%mm0,(%0)\n\t"
 385             "addl           %3,%0\n\t"
 386
 387             "movq        96(%2),%%mm0\n\t"
 388             "packuswb   104(%2),%%mm0\n\t"
 389             "movq        %%mm0,(%0)\n\t"
 390             "addl           %3,%0\n\t"
 391
 392             "movq        112(%2),%%mm0\n\t"
 393             "packuswb   120(%2),%%mm0\n\t"
 394             "movq        %%mm0,(%0)\n\t"
 395
 396             //"emms"
 397             : "=r" (p_data)
 398             : "0" (p_data), "r" (p_block), "r" (i_incr + 8) );
 399 }
 400 #endif
 401
 402
 403 /*****************************************************************************
 404  * vdec_DecodeMacroblock : decode a macroblock of a picture
 405  *****************************************************************************/
 406 #define DECODEBLOCKSC( OPBLOCK )                                        \
 407 {                                                                       \
 408     int             i_b, i_mask;                                        \
 409                                                                         \
 410     i_mask = 1 << (3 + p_mb->i_chroma_nb_blocks);                       \
 411                                                                         \
 412     /* luminance */                                                     \
 413     for( i_b = 0; i_b < 4; i_b++, i_mask >>= 1 )                        \
 414     {                                                                   \
 415         if( p_mb->i_coded_block_pattern & i_mask )                      \
 416         {                                                               \
 417             /*                                                          \
 418              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 419              */                                                         \
 420             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 421                                   p_mb->pi_sparse_pos[i_b] );           \
 422                                                                         \
 423             /*                                                          \
 424              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 425              * section 7.6.8)                                           \
 426              */                                                         \
 427             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 428                      p_mb->p_data[i_b], p_mb->i_addb_l_stride );        \
 429         }                                                               \
 430     }                                                                   \
 431                                                                         \
 432     /* chrominance */                                                   \
 433     for( i_b = 4; i_b < 4 + p_mb->i_chroma_nb_blocks;                   \
 434          i_b++, i_mask >>= 1 )                                          \
 435     {                                                                   \
 436         if( p_mb->i_coded_block_pattern & i_mask )                      \
 437         {                                                               \
 438             /*                                                          \
 439              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 440              */                                                         \
 441             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 442                                   p_mb->pi_sparse_pos[i_b] );           \
 443                                                                         \
 444             /*                                                          \
 445              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 446              * section 7.6.8)                                           \
 447              */                                                         \
 448             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 449                      p_mb->p_data[i_b], p_mb->i_addb_c_stride );        \
 450         }                                                               \
 451     }                                                                   \
 452 }
 453
 454 #define DECODEBLOCKSBW( OPBLOCK )                                       \
 455 {                                                                       \
 456     int             i_b, i_mask;                                        \
 457                                                                         \
 458     i_mask = 1 << (3 + p_mb->i_chroma_nb_blocks);                       \
 459                                                                         \
 460     /* luminance */                                                     \
 461     for( i_b = 0; i_b < 4; i_b++, i_mask >>= 1 )                        \
 462     {                                                                   \
 463         if( p_mb->i_coded_block_pattern & i_mask )                      \
 464         {                                                               \
 465             /*                                                          \
 466              * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
 467              */                                                         \
 468             (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
 469                                   p_mb->pi_sparse_pos[i_b] );           \
 470                                                                         \
 471             /*                                                          \
 472              * Adding prediction and coefficient data (ISO/IEC 13818-2  \
 473              * section 7.6.8)                                           \
 474              */                                                         \
 475             OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
 476                      p_mb->p_data[i_b], p_mb->i_addb_l_stride );        \
 477         }                                                               \
 478     }                                                                   \
 479 }
 480
 481 void vdec_DecodeMacroblockC ( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 482 {
 483     if( !(p_mb->i_mb_type & MB_INTRA) )
 484     {
 485         /*
 486          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 487          */
 488         if( p_mb->pf_motion == 0 )
 489         {
 490             intf_WarnMsg( 2, "pf_motion set to NULL" );
 491         }
 492         else
 493         {
 494             p_mb->pf_motion( p_mb );
 495         }
 496
 497         DECODEBLOCKSC( AddBlock )
 498     }
 499     else
 500     {
 501         DECODEBLOCKSC( CopyBlock )
 502     }
 503
 504     /*
 505      * Decoding is finished, release the macroblock and free
 506      * unneeded memory.
 507      */
 508     vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 509 }
 510
 511 void vdec_DecodeMacroblockBW ( vdec_thread_t *p_vdec, macroblock_t * p_mb )
 512 {
 513     if( !(p_mb->i_mb_type & MB_INTRA) )
 514     {
 515         /*
 516          * Motion Compensation (ISO/IEC 13818-2 section 7.6)
 517          */
 518         if( p_mb->pf_motion == 0 )
 519         {
 520             intf_WarnMsg( 2, "pf_motion set to NULL" );
 521         }
 522         else
 523         {
 524             p_mb->pf_motion( p_mb );
 525         }
 526
 527         DECODEBLOCKSBW( AddBlock )
 528     }
 529     else
 530     {
 531         DECODEBLOCKSBW( CopyBlock )
 532     }
 533
 534     /*
 535      * Decoding is finished, release the macroblock and free
 536      * unneeded memory.
 537      */
 538     vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
 539 }
 540
 541
 542
 543 /*****************************************************************************
 544  * RunThread: video decoder thread
 545  *****************************************************************************
 546  * Video decoder thread. This function does only return when the thread is
 547  * terminated.
 548  *****************************************************************************/
 549 static void RunThread( vdec_thread_t *p_vdec )
 550 {
 551     intf_DbgMsg("vdec debug: running video decoder thread (%p) (pid == %i)",
 552                 p_vdec, getpid());
 553
 554     /*
 555      * Initialize thread and free configuration
 556      */
 557     p_vdec->b_error = vdec_InitThread( p_vdec );
 558     if( p_vdec->b_error )
 559     {
 560         return;
 561     }
 562     p_vdec->b_run = 1;
 563
 564     /*
 565      * Main loop - it is not executed if an error occured during
 566      * initialization
 567      */
 568     while( (!p_vdec->b_die) && (!p_vdec->b_error) )
 569     {
 570         macroblock_t *          p_mb;
 571
 572         if( (p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo )) != NULL )
 573         {
 574             vdec_DecodeMacroblockC ( p_vdec, p_mb );
 575         }
 576     }
 577
 578     /*
 579      * Error loop
 580      */
 581     if( p_vdec->b_error )
 582     {
 583         ErrorThread( p_vdec );
 584     }
 585
 586     /* End of thread */
 587     EndThread( p_vdec );
 588     p_vdec->b_run = 0;
 589 }