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
   5  *  that can be found in the LICENSE file in the root of the source
   6  *  tree. An additional intellectual property rights grant can be found
   7  *  in the file PATENTS.  All contributing project authors may
   8  *  be found in the AUTHORS file in the root of the source tree.
   9  */
  10
  11
  12 #include "onyx_int.h"
  13 #include "threading.h"
  14 #include "common.h"
  15 #include "extend.h"
  16
  17
  18 extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t, int recon_yoffset, int recon_uvoffset);
  19 extern int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t);
  20 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x);
  21 extern void vp8_build_block_offsets(MACROBLOCK *x);
  22 extern void vp8_setup_block_ptrs(MACROBLOCK *x);
  23
  24 static
  25 THREAD_FUNCTION thread_encoding_proc(void *p_data)
  26 {
  27 #if CONFIG_MULTITHREAD
  28     int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
  29     VP8_COMP *cpi   = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
  30     MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
  31     ENTROPY_CONTEXT mb_row_left_context[4][4];
  32
  33     //printf("Started thread %d\n", ithread);
  34
  35     while (1)
  36     {
  37         if (cpi->b_multi_threaded == 0)
  38             break;
  39
  40         //if(WaitForSingleObject(cpi->h_event_mbrencoding[ithread], INFINITE) == WAIT_OBJECT_0)
  41         if (sem_wait(&cpi->h_event_mbrencoding[ithread]) == 0)
  42         {
  43             if (cpi->b_multi_threaded == FALSE) // we're shutting down
  44                 break;
  45             else
  46             {
  47                 VP8_COMMON *cm      = &cpi->common;
  48                 int mb_row           = mbri->mb_row;
  49                 MACROBLOCK  *x      = &mbri->mb;
  50                 MACROBLOCKD *xd     = &x->e_mbd;
  51                 TOKENEXTRA **tp     = &mbri->tp;
  52                 int *segment_counts  = mbri->segment_counts;
  53                 int *totalrate      = &mbri->totalrate;
  54
  55                 {
  56                     int i;
  57                     int recon_yoffset, recon_uvoffset;
  58                     int mb_col;
  59                     int ref_fb_idx = cm->lst_fb_idx;
  60                     int dst_fb_idx = cm->new_fb_idx;
  61                     int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
  62                     int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
  63                     volatile int *last_row_current_mb_col;
  64
  65                     if (ithread > 0)
  66                         last_row_current_mb_col = &cpi->mb_row_ei[ithread-1].current_mb_col;
  67                     else
  68                         last_row_current_mb_col = &cpi->current_mb_col_main;
  69
  70                     // reset above block coeffs
  71                     xd->above_context[Y1CONTEXT] = cm->above_context[Y1CONTEXT];
  72                     xd->above_context[UCONTEXT ] = cm->above_context[UCONTEXT ];
  73                     xd->above_context[VCONTEXT ] = cm->above_context[VCONTEXT ];
  74                     xd->above_context[Y2CONTEXT] = cm->above_context[Y2CONTEXT];
  75                     xd->left_context = mb_row_left_context;
  76
  77                     vp8_zero(mb_row_left_context);
  78
  79                     xd->up_available = (mb_row != 0);
  80                     recon_yoffset = (mb_row * recon_y_stride * 16);
  81                     recon_uvoffset = (mb_row * recon_uv_stride * 8);
  82
  83
  84                     cpi->tplist[mb_row].start = *tp;
  85
  86                     //printf("Thread mb_row = %d\n", mb_row);
  87
  88                     // for each macroblock col in image
  89                     for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
  90                     {
  91                         int seg_map_index = (mb_row * cm->mb_cols);
  92
  93                         while (mb_col > (*last_row_current_mb_col - 1) && *last_row_current_mb_col != cm->mb_cols - 1)
  94                         {
  95                             x86_pause_hint();
  96                             thread_sleep(0);
  97                         }
  98
  99                         // Distance of Mb to the various image edges.
 100                         // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
 101                         xd->mb_to_left_edge = -((mb_col * 16) << 3);
 102                         xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
 103                         xd->mb_to_top_edge = -((mb_row * 16) << 3);
 104                         xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
 105
 106                         // Set up limit values for motion vectors used to prevent them extending outside the UMV borders
 107                         x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
 108                         x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
 109                         x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
 110                         x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
 111
 112                         xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 113                         xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 114                         xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 115                         xd->left_available = (mb_col != 0);
 116
 117                         // Is segmentation enabled
 118                         // MB level adjutment to quantizer
 119                         if (xd->segmentation_enabled)
 120                         {
 121                             // Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
 122                             if (cpi->segmentation_map[seg_map_index+mb_col] <= 3)
 123                                 xd->mbmi.segment_id = cpi->segmentation_map[seg_map_index+mb_col];
 124                             else
 125                                 xd->mbmi.segment_id = 0;
 126
 127                             vp8cx_mb_init_quantizer(cpi, x);
 128                         }
 129                         else
 130                             xd->mbmi.segment_id = 0;         // Set to Segment 0 by default
 131
 132
 133                         if (cm->frame_type == KEY_FRAME)
 134                         {
 135                             *totalrate += vp8cx_encode_intra_macro_block(cpi, x, tp);
 136 #ifdef MODE_STATS
 137                             y_modes[xd->mbmi.mode] ++;
 138 #endif
 139                         }
 140                         else
 141                         {
 142                             *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset);
 143
 144 #ifdef MODE_STATS
 145                             inter_y_modes[xd->mbmi.mode] ++;
 146
 147                             if (xd->mbmi.mode == SPLITMV)
 148                             {
 149                                 int b;
 150
 151                                 for (b = 0; b < xd->mbmi.partition_count; b++)
 152                                 {
 153                                     inter_b_modes[xd->mbmi.partition_bmi[b].mode] ++;
 154                                 }
 155                             }
 156
 157 #endif
 158
 159                             // Count of last ref frame 0,0 useage
 160                             if ((xd->mbmi.mode == ZEROMV) && (xd->mbmi.ref_frame == LAST_FRAME))
 161                                 cpi->inter_zz_count ++;
 162
 163                         }
 164
 165                         cpi->tplist[mb_row].stop = *tp;
 166
 167                         xd->gf_active_ptr++;      // Increment pointer into gf useage flags structure for next mb
 168
 169                         // store macroblock mode info into context array
 170                         vpx_memcpy(&xd->mode_info_context->mbmi, &xd->mbmi, sizeof(xd->mbmi));
 171
 172                         for (i = 0; i < 16; i++)
 173                             vpx_memcpy(&xd->mode_info_context->bmi[i], &xd->block[i].bmi, sizeof(xd->block[i].bmi));
 174
 175                         // adjust to the next column of macroblocks
 176                         x->src.y_buffer += 16;
 177                         x->src.u_buffer += 8;
 178                         x->src.v_buffer += 8;
 179
 180                         recon_yoffset += 16;
 181                         recon_uvoffset += 8;
 182
 183                         // Keep track of segment useage
 184                         segment_counts[xd->mbmi.segment_id] ++;
 185
 186                         // skip to next mb
 187                         xd->mode_info_context++;
 188
 189                         xd->above_context[Y1CONTEXT] += 4;
 190                         xd->above_context[UCONTEXT ] += 2;
 191                         xd->above_context[VCONTEXT ] += 2;
 192                         xd->above_context[Y2CONTEXT] ++;
 193
 194                         cpi->mb_row_ei[ithread].current_mb_col = mb_col;
 195
 196                     }
 197
 198                     //extend the recon for intra prediction
 199                     vp8_extend_mb_row(
 200                         &cm->yv12_fb[dst_fb_idx],
 201                         xd->dst.y_buffer + 16,
 202                         xd->dst.u_buffer + 8,
 203                         xd->dst.v_buffer + 8);
 204
 205                     // this is to account for the border
 206                     xd->mode_info_context++;
 207
 208                     x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
 209                     x->src.u_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 210                     x->src.v_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 211
 212                     xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
 213
 214                     if (ithread == (cpi->encoding_thread_count - 1) || mb_row == cm->mb_rows - 1)
 215                     {
 216                         //SetEvent(cpi->h_event_main);
 217                         sem_post(&cpi->h_event_main);
 218                     }
 219
 220                 }
 221
 222             }
 223         }
 224     }
 225
 226 #else
 227     (void) p_data;
 228 #endif
 229
 230     //printf("exit thread %d\n", ithread);
 231     return 0;
 232 }
 233
 234 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
 235 {
 236
 237     MACROBLOCK *x = mbsrc;
 238     MACROBLOCK *z = mbdst;
 239     int i;
 240
 241     z->ss               = x->ss;
 242     z->ss_count          = x->ss_count;
 243     z->searches_per_step  = x->searches_per_step;
 244     z->errorperbit      = x->errorperbit;
 245
 246     z->sadperbit16      = x->sadperbit16;
 247     z->sadperbit4       = x->sadperbit4;
 248     z->errthresh        = x->errthresh;
 249     z->rddiv            = x->rddiv;
 250     z->rdmult           = x->rdmult;
 251
 252     /*
 253     z->mv_col_min    = x->mv_col_min;
 254     z->mv_col_max    = x->mv_col_max;
 255     z->mv_row_min    = x->mv_row_min;
 256     z->mv_row_max    = x->mv_row_max;
 257     z->vector_range = x->vector_range ;
 258     */
 259
 260     z->vp8_short_fdct4x4     = x->vp8_short_fdct4x4;
 261     z->vp8_short_fdct8x4     = x->vp8_short_fdct8x4;
 262     z->short_walsh4x4    = x->short_walsh4x4;
 263     z->quantize_b        = x->quantize_b;
 264
 265     /*
 266     z->mvc              = x->mvc;
 267     z->src.y_buffer      = x->src.y_buffer;
 268     z->src.u_buffer      = x->src.u_buffer;
 269     z->src.v_buffer      = x->src.v_buffer;
 270     */
 271
 272
 273     vpx_memcpy(z->mvcosts,          x->mvcosts,         sizeof(x->mvcosts));
 274     z->mvcost[0] = &z->mvcosts[0][mv_max+1];
 275     z->mvcost[1] = &z->mvcosts[1][mv_max+1];
 276     z->mvsadcost[0] = &z->mvsadcosts[0][mv_max+1];
 277     z->mvsadcost[1] = &z->mvsadcosts[1][mv_max+1];
 278
 279
 280     vpx_memcpy(z->token_costs,       x->token_costs,      sizeof(x->token_costs));
 281     vpx_memcpy(z->inter_bmode_costs,  x->inter_bmode_costs, sizeof(x->inter_bmode_costs));
 282     //memcpy(z->mvcosts,            x->mvcosts,         sizeof(x->mvcosts));
 283     //memcpy(z->mvcost,         x->mvcost,          sizeof(x->mvcost));
 284     vpx_memcpy(z->mbmode_cost,       x->mbmode_cost,      sizeof(x->mbmode_cost));
 285     vpx_memcpy(z->intra_uv_mode_cost,  x->intra_uv_mode_cost, sizeof(x->intra_uv_mode_cost));
 286     vpx_memcpy(z->bmode_costs,       x->bmode_costs,      sizeof(x->bmode_costs));
 287
 288     for (i = 0; i < 25; i++)
 289     {
 290         z->block[i].quant           = x->block[i].quant;
 291         z->block[i].quant_shift     = x->block[i].quant_shift;
 292         z->block[i].zbin            = x->block[i].zbin;
 293         z->block[i].zrun_zbin_boost   = x->block[i].zrun_zbin_boost;
 294         z->block[i].round           = x->block[i].round;
 295         /*
 296         z->block[i].src             = x->block[i].src;
 297         */
 298         z->block[i].src_stride       = x->block[i].src_stride;
 299         z->block[i].force_empty      = x->block[i].force_empty;
 300
 301     }
 302
 303     {
 304         MACROBLOCKD *xd = &x->e_mbd;
 305         MACROBLOCKD *zd = &z->e_mbd;
 306
 307         /*
 308         zd->mode_info_context = xd->mode_info_context;
 309         zd->mode_info        = xd->mode_info;
 310
 311         zd->mode_info_stride  = xd->mode_info_stride;
 312         zd->frame_type       = xd->frame_type;
 313         zd->up_available     = xd->up_available   ;
 314         zd->left_available   = xd->left_available;
 315         zd->left_context     = xd->left_context;
 316         zd->last_frame_dc     = xd->last_frame_dc;
 317         zd->last_frame_dccons = xd->last_frame_dccons;
 318         zd->gold_frame_dc     = xd->gold_frame_dc;
 319         zd->gold_frame_dccons = xd->gold_frame_dccons;
 320         zd->mb_to_left_edge    = xd->mb_to_left_edge;
 321         zd->mb_to_right_edge   = xd->mb_to_right_edge;
 322         zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
 323         zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
 324         zd->gf_active_ptr     = xd->gf_active_ptr;
 325         zd->frames_since_golden       = xd->frames_since_golden;
 326         zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
 327         */
 328         zd->subpixel_predict         = xd->subpixel_predict;
 329         zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
 330         zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
 331         zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
 332         zd->segmentation_enabled     = xd->segmentation_enabled;
 333         zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
 334         vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
 335
 336         /*
 337         memcpy(zd->above_context,        xd->above_context, sizeof(xd->above_context));
 338         memcpy(zd->mb_segment_tree_probs,  xd->mb_segment_tree_probs, sizeof(xd->mb_segment_tree_probs));
 339         memcpy(zd->segment_feature_data,  xd->segment_feature_data, sizeof(xd->segment_feature_data));
 340         */
 341         for (i = 0; i < 25; i++)
 342         {
 343             zd->block[i].dequant = xd->block[i].dequant;
 344         }
 345     }
 346 }
 347
 348
 349 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
 350                                MACROBLOCK *x,
 351                                MB_ROW_COMP *mbr_ei,
 352                                int mb_row,
 353                                int count
 354                               )
 355 {
 356
 357     VP8_COMMON *const cm = & cpi->common;
 358     MACROBLOCKD *const xd = & x->e_mbd;
 359     int i;
 360     (void) mb_row;
 361
 362     for (i = 0; i < count; i++)
 363     {
 364         MACROBLOCK *mb = & mbr_ei[i].mb;
 365         MACROBLOCKD *mbd = &mb->e_mbd;
 366
 367         mbd->subpixel_predict        = xd->subpixel_predict;
 368         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
 369         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
 370         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
 371 #if CONFIG_RUNTIME_CPU_DETECT
 372         mbd->rtcd                   = xd->rtcd;
 373 #endif
 374         mbd->gf_active_ptr            = xd->gf_active_ptr;
 375
 376         mb->vector_range             = 32;
 377
 378         vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
 379         mbr_ei[i].totalrate = 0;
 380
 381         mbd->mode_info        = cm->mi - 1;
 382         mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
 383         mbd->mode_info_stride  = cm->mode_info_stride;
 384
 385         mbd->frame_type = cm->frame_type;
 386
 387         mbd->frames_since_golden = cm->frames_since_golden;
 388         mbd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
 389
 390         mb->src = * cpi->Source;
 391         mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
 392         mbd->dst = cm->yv12_fb[cm->new_fb_idx];
 393
 394         mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
 395         mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
 396         mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
 397
 398
 399         vp8_build_block_offsets(mb);
 400
 401         vp8_setup_block_dptrs(mbd);
 402
 403         vp8_setup_block_ptrs(mb);
 404
 405         mb->rddiv = cpi->RDDIV;
 406         mb->rdmult = cpi->RDMULT;
 407
 408         mbd->mbmi.mode = DC_PRED;
 409         mbd->mbmi.uv_mode = DC_PRED;
 410
 411         mbd->left_context = cm->left_context;
 412         mb->mvc = cm->fc.mvc;
 413
 414         setup_mbby_copy(&mbr_ei[i].mb, x);
 415
 416     }
 417 }
 418
 419
 420 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
 421 {
 422     cpi->b_multi_threaded = 0;
 423
 424     cpi->processor_core_count = 32; //vp8_get_proc_core_count();
 425
 426     CHECK_MEM_ERROR(cpi->tplist, vpx_malloc(sizeof(TOKENLIST) * cpi->common.mb_rows));
 427
 428 #if CONFIG_MULTITHREAD
 429
 430     if (cpi->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
 431     {
 432         int ithread;
 433
 434         if (cpi->oxcf.multi_threaded > cpi->processor_core_count)
 435             cpi->encoding_thread_count = cpi->processor_core_count - 1;
 436         else
 437             cpi->encoding_thread_count = cpi->oxcf.multi_threaded - 1;
 438
 439
 440         CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * cpi->encoding_thread_count));
 441         CHECK_MEM_ERROR(cpi->h_event_mbrencoding, vpx_malloc(sizeof(sem_t) * cpi->encoding_thread_count));
 442         CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count));
 443         vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count);
 444         CHECK_MEM_ERROR(cpi->en_thread_data, vpx_malloc(sizeof(ENCODETHREAD_DATA) * cpi->encoding_thread_count));
 445         //cpi->h_event_main = CreateEvent(NULL, FALSE, FALSE, NULL);
 446         sem_init(&cpi->h_event_main, 0, 0);
 447
 448         cpi->b_multi_threaded = 1;
 449
 450         //printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n", (cpi->encoding_thread_count +1));
 451
 452         for (ithread = 0; ithread < cpi->encoding_thread_count; ithread++)
 453         {
 454             //cpi->h_event_mbrencoding[ithread] = CreateEvent(NULL, FALSE, FALSE, NULL);
 455             sem_init(&cpi->h_event_mbrencoding[ithread], 0, 0);
 456             cpi->en_thread_data[ithread].ithread = ithread;
 457             cpi->en_thread_data[ithread].ptr1 = (void *)cpi;
 458             cpi->en_thread_data[ithread].ptr2 = (void *)&cpi->mb_row_ei[ithread];
 459
 460             //printf(" call begin thread %d \n", ithread);
 461
 462             //cpi->h_encoding_thread[ithread] =   (HANDLE)_beginthreadex(
 463             //  NULL,           // security
 464             //  0,              // stksize
 465             //  thread_encoding_proc,
 466             //  (&cpi->en_thread_data[ithread]),          // Thread data
 467             //  0,
 468             //  NULL);
 469
 470             pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, (&cpi->en_thread_data[ithread]));
 471
 472         }
 473
 474     }
 475
 476 #endif
 477 }
 478
 479 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
 480 {
 481 #if CONFIG_MULTITHREAD
 482
 483     if (cpi->b_multi_threaded)
 484     {
 485         //shutdown other threads
 486         cpi->b_multi_threaded = 0;
 487         {
 488             int i;
 489
 490             for (i = 0; i < cpi->encoding_thread_count; i++)
 491             {
 492                 //SetEvent(cpi->h_event_mbrencoding[i]);
 493                 sem_post(&cpi->h_event_mbrencoding[i]);
 494                 pthread_join(cpi->h_encoding_thread[i], 0);
 495             }
 496
 497             for (i = 0; i < cpi->encoding_thread_count; i++)
 498                 sem_destroy(&cpi->h_event_mbrencoding[i]);
 499         }
 500         //free thread related resources
 501         vpx_free(cpi->h_event_mbrencoding);
 502         vpx_free(cpi->h_encoding_thread);
 503         vpx_free(cpi->mb_row_ei);
 504         vpx_free(cpi->en_thread_data);
 505     }
 506
 507 #endif
 508     vpx_free(cpi->tplist);
 509 }