vp8/encoder/ethreading.c

   1 /*
   2  *  Copyright (c) 2010 The VP8 project authors. All Rights Reserved.
   3  *
   4  *  Use of this source code is governed by a BSD-style license and patent
   5  *  grant that can be found in the LICENSE file in the root of the source
   6  *  tree. All contributing project authors may be found in the AUTHORS
   7  *  file in the root of the source tree.
   8  */
   9
  10
  11 #include "onyx_int.h"
  12 #include "threading.h"
  13 #include "common.h"
  14 #include "extend.h"
  15
  16
  17 extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t, int recon_yoffset, int recon_uvoffset);
  18 extern int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t);
  19 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x);
  20 extern void vp8_build_block_offsets(MACROBLOCK *x);
  21 extern void vp8_setup_block_ptrs(MACROBLOCK *x);
  22
  23 static
  24 THREAD_FUNCTION thread_encoding_proc(void *p_data)
  25 {
  26 #if CONFIG_MULTITHREAD
  27     int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
  28     VP8_COMP *cpi   = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
  29     MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
  30     ENTROPY_CONTEXT mb_row_left_context[4][4];
  31
  32     //printf("Started thread %d\n", ithread);
  33
  34     while (1)
  35     {
  36         if (cpi->b_multi_threaded == 0)
  37             break;
  38
  39         //if(WaitForSingleObject(cpi->h_event_mbrencoding[ithread], INFINITE) == WAIT_OBJECT_0)
  40         if (sem_wait(&cpi->h_event_mbrencoding[ithread]) == 0)
  41         {
  42             if (cpi->b_multi_threaded == FALSE) // we're shutting down
  43                 break;
  44             else
  45             {
  46                 VP8_COMMON *cm      = &cpi->common;
  47                 int mb_row           = mbri->mb_row;
  48                 MACROBLOCK  *x      = &mbri->mb;
  49                 MACROBLOCKD *xd     = &x->e_mbd;
  50                 TOKENEXTRA **tp     = &mbri->tp;
  51                 int *segment_counts  = mbri->segment_counts;
  52                 int *totalrate      = &mbri->totalrate;
  53
  54                 {
  55                     int i;
  56                     int recon_yoffset, recon_uvoffset;
  57                     int mb_col;
  58                     int recon_y_stride = cm->last_frame.y_stride;
  59                     int recon_uv_stride = cm->last_frame.uv_stride;
  60                     volatile int *last_row_current_mb_col;
  61
  62                     if (ithread > 0)
  63                         last_row_current_mb_col = &cpi->mb_row_ei[ithread-1].current_mb_col;
  64                     else
  65                         last_row_current_mb_col = &cpi->current_mb_col_main;
  66
  67                     // reset above block coeffs
  68                     xd->above_context[Y1CONTEXT] = cm->above_context[Y1CONTEXT];
  69                     xd->above_context[UCONTEXT ] = cm->above_context[UCONTEXT ];
  70                     xd->above_context[VCONTEXT ] = cm->above_context[VCONTEXT ];
  71                     xd->above_context[Y2CONTEXT] = cm->above_context[Y2CONTEXT];
  72                     xd->left_context = mb_row_left_context;
  73
  74                     vp8_zero(mb_row_left_context);
  75
  76                     xd->up_available = (mb_row != 0);
  77                     recon_yoffset = (mb_row * recon_y_stride * 16);
  78                     recon_uvoffset = (mb_row * recon_uv_stride * 8);
  79
  80
  81                     cpi->tplist[mb_row].start = *tp;
  82
  83                     //printf("Thread mb_row = %d\n", mb_row);
  84
  85                     // for each macroblock col in image
  86                     for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
  87                     {
  88                         int seg_map_index = (mb_row * cm->mb_cols);
  89
  90                         while (mb_col > (*last_row_current_mb_col - 1) && *last_row_current_mb_col != cm->mb_cols - 1)
  91                         {
  92                             x86_pause_hint();
  93                             thread_sleep(0);
  94                         }
  95
  96                         // Distance of Mb to the various image edges.
  97                         // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
  98                         xd->mb_to_left_edge = -((mb_col * 16) << 3);
  99                         xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
 100                         xd->mb_to_top_edge = -((mb_row * 16) << 3);
 101                         xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
 102
 103                         // Set up limit values for motion vectors used to prevent them extending outside the UMV borders
 104                         x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
 105                         x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
 106                         x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
 107                         x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
 108
 109                         xd->dst.y_buffer = cm->new_frame.y_buffer + recon_yoffset;
 110                         xd->dst.u_buffer = cm->new_frame.u_buffer + recon_uvoffset;
 111                         xd->dst.v_buffer = cm->new_frame.v_buffer + recon_uvoffset;
 112                         xd->left_available = (mb_col != 0);
 113
 114                         // Is segmentation enabled
 115                         // MB level adjutment to quantizer
 116                         if (xd->segmentation_enabled)
 117                         {
 118                             // Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
 119                             if (cpi->segmentation_map[seg_map_index+mb_col] <= 3)
 120                                 xd->mbmi.segment_id = cpi->segmentation_map[seg_map_index+mb_col];
 121                             else
 122                                 xd->mbmi.segment_id = 0;
 123
 124                             vp8cx_mb_init_quantizer(cpi, x);
 125                         }
 126                         else
 127                             xd->mbmi.segment_id = 0;         // Set to Segment 0 by default
 128
 129
 130                         if (cm->frame_type == KEY_FRAME)
 131                         {
 132                             *totalrate += vp8cx_encode_intra_macro_block(cpi, x, tp);
 133 #ifdef MODE_STATS
 134                             y_modes[xd->mbmi.mode] ++;
 135 #endif
 136                         }
 137                         else
 138                         {
 139                             *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset);
 140
 141 #ifdef MODE_STATS
 142                             inter_y_modes[xd->mbmi.mode] ++;
 143
 144                             if (xd->mbmi.mode == SPLITMV)
 145                             {
 146                                 int b;
 147
 148                                 for (b = 0; b < xd->mbmi.partition_count; b++)
 149                                 {
 150                                     inter_b_modes[xd->mbmi.partition_bmi[b].mode] ++;
 151                                 }
 152                             }
 153
 154 #endif
 155
 156                             // Count of last ref frame 0,0 useage
 157                             if ((xd->mbmi.mode == ZEROMV) && (xd->mbmi.ref_frame == LAST_FRAME))
 158                                 cpi->inter_zz_count ++;
 159
 160                         }
 161
 162                         cpi->tplist[mb_row].stop = *tp;
 163
 164                         xd->gf_active_ptr++;      // Increment pointer into gf useage flags structure for next mb
 165
 166                         // store macroblock mode info into context array
 167                         vpx_memcpy(&xd->mode_info_context->mbmi, &xd->mbmi, sizeof(xd->mbmi));
 168
 169                         for (i = 0; i < 16; i++)
 170                             vpx_memcpy(&xd->mode_info_context->bmi[i], &xd->block[i].bmi, sizeof(xd->block[i].bmi));
 171
 172                         // adjust to the next column of macroblocks
 173                         x->src.y_buffer += 16;
 174                         x->src.u_buffer += 8;
 175                         x->src.v_buffer += 8;
 176
 177                         recon_yoffset += 16;
 178                         recon_uvoffset += 8;
 179
 180                         // Keep track of segment useage
 181                         segment_counts[xd->mbmi.segment_id] ++;
 182
 183                         // skip to next mb
 184                         xd->mode_info_context++;
 185
 186                         xd->above_context[Y1CONTEXT] += 4;
 187                         xd->above_context[UCONTEXT ] += 2;
 188                         xd->above_context[VCONTEXT ] += 2;
 189                         xd->above_context[Y2CONTEXT] ++;
 190
 191                         cpi->mb_row_ei[ithread].current_mb_col = mb_col;
 192
 193                     }
 194
 195                     //extend the recon for intra prediction
 196                     vp8_extend_mb_row(
 197                         &cm->new_frame,
 198                         xd->dst.y_buffer + 16,
 199                         xd->dst.u_buffer + 8,
 200                         xd->dst.v_buffer + 8);
 201
 202                     // this is to account for the border
 203                     xd->mode_info_context++;
 204
 205                     x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
 206                     x->src.u_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 207                     x->src.v_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 208
 209                     xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
 210
 211                     if (ithread == (cpi->encoding_thread_count - 1) || mb_row == cm->mb_rows - 1)
 212                     {
 213                         //SetEvent(cpi->h_event_main);
 214                         sem_post(&cpi->h_event_main);
 215                     }
 216
 217                 }
 218
 219             }
 220         }
 221     }
 222
 223 #else
 224     (void) p_data;
 225 #endif
 226
 227     //printf("exit thread %d\n", ithread);
 228     return 0;
 229 }
 230
 231 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
 232 {
 233
 234     MACROBLOCK *x = mbsrc;
 235     MACROBLOCK *z = mbdst;
 236     int i;
 237
 238     z->ss               = x->ss;
 239     z->ss_count          = x->ss_count;
 240     z->searches_per_step  = x->searches_per_step;
 241     z->errorperbit      = x->errorperbit;
 242
 243     z->sadperbit16      = x->sadperbit16;
 244     z->sadperbit4       = x->sadperbit4;
 245     z->errthresh        = x->errthresh;
 246     z->rddiv            = x->rddiv;
 247     z->rdmult           = x->rdmult;
 248
 249     /*
 250     z->mv_col_min    = x->mv_col_min;
 251     z->mv_col_max    = x->mv_col_max;
 252     z->mv_row_min    = x->mv_row_min;
 253     z->mv_row_max    = x->mv_row_max;
 254     z->vector_range = x->vector_range ;
 255     */
 256
 257     z->vp8_short_fdct4x4     = x->vp8_short_fdct4x4;
 258     z->vp8_short_fdct8x4     = x->vp8_short_fdct8x4;
 259     z->short_fdct4x4rd   = x->short_fdct4x4rd;
 260     z->short_fdct8x4rd   = x->short_fdct8x4rd;
 261     z->short_fdct8x4rd   = x->short_fdct8x4rd;
 262     z->vp8_short_fdct4x4_ptr = x->vp8_short_fdct4x4_ptr;
 263     z->short_walsh4x4    = x->short_walsh4x4;
 264     z->quantize_b        = x->quantize_b;
 265     z->quantize_brd      = x->quantize_brd;
 266
 267     /*
 268     z->mvc              = x->mvc;
 269     z->src.y_buffer      = x->src.y_buffer;
 270     z->src.u_buffer      = x->src.u_buffer;
 271     z->src.v_buffer      = x->src.v_buffer;
 272     */
 273
 274
 275     vpx_memcpy(z->mvcosts,          x->mvcosts,         sizeof(x->mvcosts));
 276     z->mvcost[0] = &z->mvcosts[0][mv_max+1];
 277     z->mvcost[1] = &z->mvcosts[1][mv_max+1];
 278     z->mvsadcost[0] = &z->mvsadcosts[0][mv_max+1];
 279     z->mvsadcost[1] = &z->mvsadcosts[1][mv_max+1];
 280
 281
 282     vpx_memcpy(z->token_costs,       x->token_costs,      sizeof(x->token_costs));
 283     vpx_memcpy(z->inter_bmode_costs,  x->inter_bmode_costs, sizeof(x->inter_bmode_costs));
 284     //memcpy(z->mvcosts,            x->mvcosts,         sizeof(x->mvcosts));
 285     //memcpy(z->mvcost,         x->mvcost,          sizeof(x->mvcost));
 286     vpx_memcpy(z->mbmode_cost,       x->mbmode_cost,      sizeof(x->mbmode_cost));
 287     vpx_memcpy(z->intra_uv_mode_cost,  x->intra_uv_mode_cost, sizeof(x->intra_uv_mode_cost));
 288     vpx_memcpy(z->bmode_costs,       x->bmode_costs,      sizeof(x->bmode_costs));
 289
 290     for (i = 0; i < 25; i++)
 291     {
 292         z->block[i].quant           = x->block[i].quant;
 293         z->block[i].zbin            = x->block[i].zbin;
 294         z->block[i].zrun_zbin_boost   = x->block[i].zrun_zbin_boost;
 295         z->block[i].round           = x->block[i].round;
 296         /*
 297         z->block[i].src             = x->block[i].src;
 298         */
 299         z->block[i].src_stride       = x->block[i].src_stride;
 300         z->block[i].force_empty      = x->block[i].force_empty;
 301
 302     }
 303
 304     {
 305         MACROBLOCKD *xd = &x->e_mbd;
 306         MACROBLOCKD *zd = &z->e_mbd;
 307
 308         /*
 309         zd->mode_info_context = xd->mode_info_context;
 310         zd->mode_info        = xd->mode_info;
 311
 312         zd->mode_info_stride  = xd->mode_info_stride;
 313         zd->frame_type       = xd->frame_type;
 314         zd->up_available     = xd->up_available   ;
 315         zd->left_available   = xd->left_available;
 316         zd->left_context     = xd->left_context;
 317         zd->last_frame_dc     = xd->last_frame_dc;
 318         zd->last_frame_dccons = xd->last_frame_dccons;
 319         zd->gold_frame_dc     = xd->gold_frame_dc;
 320         zd->gold_frame_dccons = xd->gold_frame_dccons;
 321         zd->mb_to_left_edge    = xd->mb_to_left_edge;
 322         zd->mb_to_right_edge   = xd->mb_to_right_edge;
 323         zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
 324         zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
 325         zd->gf_active_ptr     = xd->gf_active_ptr;
 326         zd->frames_since_golden       = xd->frames_since_golden;
 327         zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
 328         */
 329         zd->subpixel_predict         = xd->subpixel_predict;
 330         zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
 331         zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
 332         zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
 333         zd->segmentation_enabled     = xd->segmentation_enabled;
 334         zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
 335         vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
 336
 337         /*
 338         memcpy(zd->above_context,        xd->above_context, sizeof(xd->above_context));
 339         memcpy(zd->mb_segment_tree_probs,  xd->mb_segment_tree_probs, sizeof(xd->mb_segment_tree_probs));
 340         memcpy(zd->segment_feature_data,  xd->segment_feature_data, sizeof(xd->segment_feature_data));
 341         */
 342         for (i = 0; i < 25; i++)
 343         {
 344             zd->block[i].dequant = xd->block[i].dequant;
 345         }
 346     }
 347 }
 348
 349
 350 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
 351                                MACROBLOCK *x,
 352                                MB_ROW_COMP *mbr_ei,
 353                                int mb_row,
 354                                int count
 355                               )
 356 {
 357
 358     VP8_COMMON *const cm = & cpi->common;
 359     MACROBLOCKD *const xd = & x->e_mbd;
 360     int i;
 361     (void) mb_row;
 362
 363     for (i = 0; i < count; i++)
 364     {
 365         MACROBLOCK *mb = & mbr_ei[i].mb;
 366         MACROBLOCKD *mbd = &mb->e_mbd;
 367
 368         mbd->subpixel_predict        = xd->subpixel_predict;
 369         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
 370         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
 371         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
 372 #if CONFIG_RUNTIME_CPU_DETECT
 373         mbd->rtcd                   = xd->rtcd;
 374 #endif
 375         mbd->gf_active_ptr            = xd->gf_active_ptr;
 376
 377         mb->vector_range             = 32;
 378
 379         vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
 380         mbr_ei[i].totalrate = 0;
 381
 382         mbd->mode_info        = cm->mi - 1;
 383         mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
 384         mbd->mode_info_stride  = cm->mode_info_stride;
 385
 386         mbd->frame_type = cm->frame_type;
 387
 388         mbd->frames_since_golden = cm->frames_since_golden;
 389         mbd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
 390
 391         mb->src = * cpi->Source;
 392         mbd->pre = cm->last_frame;
 393         mbd->dst = cm->new_frame;
 394
 395         mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
 396         mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
 397         mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
 398
 399
 400         vp8_build_block_offsets(mb);
 401
 402         vp8_setup_block_dptrs(mbd);
 403
 404         vp8_setup_block_ptrs(mb);
 405
 406         mb->rddiv = cpi->RDDIV;
 407         mb->rdmult = cpi->RDMULT;
 408
 409         mbd->mbmi.mode = DC_PRED;
 410         mbd->mbmi.uv_mode = DC_PRED;
 411
 412         mbd->left_context = cm->left_context;
 413         mb->mvc = cm->fc.mvc;
 414
 415         setup_mbby_copy(&mbr_ei[i].mb, x);
 416
 417     }
 418 }
 419
 420
 421 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
 422 {
 423     cpi->b_multi_threaded = 0;
 424
 425     cpi->processor_core_count = 32; //vp8_get_proc_core_count();
 426
 427     CHECK_MEM_ERROR(cpi->tplist, vpx_malloc(sizeof(TOKENLIST) * cpi->common.mb_rows));
 428
 429 #if CONFIG_MULTITHREAD
 430
 431     if (cpi->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
 432     {
 433         int ithread;
 434
 435         if (cpi->oxcf.multi_threaded > cpi->processor_core_count)
 436             cpi->encoding_thread_count = cpi->processor_core_count - 1;
 437         else
 438             cpi->encoding_thread_count = cpi->oxcf.multi_threaded - 1;
 439
 440
 441         CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * cpi->encoding_thread_count));
 442         CHECK_MEM_ERROR(cpi->h_event_mbrencoding, vpx_malloc(sizeof(sem_t) * cpi->encoding_thread_count));
 443         CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count));
 444         vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count);
 445         CHECK_MEM_ERROR(cpi->en_thread_data, vpx_malloc(sizeof(ENCODETHREAD_DATA) * cpi->encoding_thread_count));
 446         //cpi->h_event_main = CreateEvent(NULL, FALSE, FALSE, NULL);
 447         sem_init(&cpi->h_event_main, 0, 0);
 448
 449         cpi->b_multi_threaded = 1;
 450
 451         //printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n", (cpi->encoding_thread_count +1));
 452
 453         for (ithread = 0; ithread < cpi->encoding_thread_count; ithread++)
 454         {
 455             //cpi->h_event_mbrencoding[ithread] = CreateEvent(NULL, FALSE, FALSE, NULL);
 456             sem_init(&cpi->h_event_mbrencoding[ithread], 0, 0);
 457             cpi->en_thread_data[ithread].ithread = ithread;
 458             cpi->en_thread_data[ithread].ptr1 = (void *)cpi;
 459             cpi->en_thread_data[ithread].ptr2 = (void *)&cpi->mb_row_ei[ithread];
 460
 461             //printf(" call begin thread %d \n", ithread);
 462
 463             //cpi->h_encoding_thread[ithread] =   (HANDLE)_beginthreadex(
 464             //  NULL,           // security
 465             //  0,              // stksize
 466             //  thread_encoding_proc,
 467             //  (&cpi->en_thread_data[ithread]),          // Thread data
 468             //  0,
 469             //  NULL);
 470
 471             pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, (&cpi->en_thread_data[ithread]));
 472
 473         }
 474
 475     }
 476
 477 #endif
 478 }
 479
 480 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
 481 {
 482 #if CONFIG_MULTITHREAD
 483
 484     if (cpi->b_multi_threaded)
 485     {
 486         //shutdown other threads
 487         cpi->b_multi_threaded = 0;
 488         {
 489             int i;
 490
 491             for (i = 0; i < cpi->encoding_thread_count; i++)
 492             {
 493                 //SetEvent(cpi->h_event_mbrencoding[i]);
 494                 sem_post(&cpi->h_event_mbrencoding[i]);
 495                 pthread_join(cpi->h_encoding_thread[i], 0);
 496             }
 497
 498             for (i = 0; i < cpi->encoding_thread_count; i++)
 499                 sem_destroy(&cpi->h_event_mbrencoding[i]);
 500         }
 501         //free thread related resources
 502         vpx_free(cpi->h_event_mbrencoding);
 503         vpx_free(cpi->h_encoding_thread);
 504         vpx_free(cpi->mb_row_ei);
 505         vpx_free(cpi->en_thread_data);
 506     }
 507
 508 #endif
 509     vpx_free(cpi->tplist);
 510 }