vp8/encoder/ethreading.c

   1 /*
   2  *  Copyright (c) 2010 The WebM 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 #include "onyx_int.h"
  12 #include "vp8/common/threading.h"
  13 #include "vp8/common/common.h"
  14 #include "vp8/common/extend.h"
  15 #include "bitstream.h"
  16 #include "encodeframe.h"
  17
  18 #if CONFIG_MULTITHREAD
  19
  20 extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
  21                                          TOKENEXTRA **t,
  22                                          int recon_yoffset, int recon_uvoffset,
  23                                          int mb_row, int mb_col);
  24 extern int vp8cx_encode_intra_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
  25                                          TOKENEXTRA **t);
  26 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x, int ok_to_skip);
  27
  28 extern void vp8_loopfilter_frame(VP8_COMP *cpi, VP8_COMMON *cm);
  29
  30 static THREAD_FUNCTION thread_loopfilter(void *p_data)
  31 {
  32     VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
  33     VP8_COMMON *cm = &cpi->common;
  34
  35     while (1)
  36     {
  37         if (cpi->b_multi_threaded == 0)
  38             break;
  39
  40         if (sem_wait(&cpi->h_event_start_lpf) == 0)
  41         {
  42             if (cpi->b_multi_threaded == 0) // we're shutting down
  43                 break;
  44
  45             vp8_loopfilter_frame(cpi, cm);
  46
  47             sem_post(&cpi->h_event_end_lpf);
  48         }
  49     }
  50
  51     return 0;
  52 }
  53
  54 static
  55 THREAD_FUNCTION thread_encoding_proc(void *p_data)
  56 {
  57     int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
  58     VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
  59     MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
  60     ENTROPY_CONTEXT_PLANES mb_row_left_context;
  61
  62     const int nsync = cpi->mt_sync_range;
  63     //printf("Started thread %d\n", ithread);
  64
  65     while (1)
  66     {
  67         if (cpi->b_multi_threaded == 0)
  68             break;
  69
  70         //if(WaitForSingleObject(cpi->h_event_mbrencoding[ithread], INFINITE) == WAIT_OBJECT_0)
  71         if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0)
  72         {
  73             VP8_COMMON *cm = &cpi->common;
  74             int mb_row;
  75             MACROBLOCK *x = &mbri->mb;
  76             MACROBLOCKD *xd = &x->e_mbd;
  77             TOKENEXTRA *tp ;
  78 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
  79             TOKENEXTRA *tp_start = cpi->tok + (1 + ithread) * (16 * 24);
  80             const int num_part = (1 << cm->multi_token_partition);
  81 #endif
  82
  83             int *segment_counts = mbri->segment_counts;
  84             int *totalrate = &mbri->totalrate;
  85
  86             if (cpi->b_multi_threaded == 0) // we're shutting down
  87                 break;
  88
  89             for (mb_row = ithread + 1; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
  90             {
  91
  92                 int recon_yoffset, recon_uvoffset;
  93                 int mb_col;
  94                 int ref_fb_idx = cm->lst_fb_idx;
  95                 int dst_fb_idx = cm->new_fb_idx;
  96                 int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
  97                 int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
  98                 int map_index = (mb_row * cm->mb_cols);
  99                 volatile const int *last_row_current_mb_col;
 100                 volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
 101
 102 #if  (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
 103                 vp8_writer *w = &cpi->bc[1 + (mb_row % num_part)];
 104 #else
 105                 tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
 106                 cpi->tplist[mb_row].start = tp;
 107 #endif
 108
 109                 last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
 110
 111                 // reset above block coeffs
 112                 xd->above_context = cm->above_context;
 113                 xd->left_context = &mb_row_left_context;
 114
 115                 vp8_zero(mb_row_left_context);
 116
 117                 xd->up_available = (mb_row != 0);
 118                 recon_yoffset = (mb_row * recon_y_stride * 16);
 119                 recon_uvoffset = (mb_row * recon_uv_stride * 8);
 120
 121                 // Set the mb activity pointer to the start of the row.
 122                 x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
 123
 124                 // for each macroblock col in image
 125                 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
 126                 {
 127                     *current_mb_col = mb_col - 1;
 128
 129                     if ((mb_col & (nsync - 1)) == 0)
 130                     {
 131                         while (mb_col > (*last_row_current_mb_col - nsync))
 132                         {
 133                             x86_pause_hint();
 134                             thread_sleep(0);
 135                         }
 136                     }
 137
 138 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
 139                     tp = tp_start;
 140 #endif
 141
 142                     // Distance of Mb to the various image edges.
 143                     // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
 144                     xd->mb_to_left_edge = -((mb_col * 16) << 3);
 145                     xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
 146                     xd->mb_to_top_edge = -((mb_row * 16) << 3);
 147                     xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
 148
 149                     // Set up limit values for motion vectors used to prevent them extending outside the UMV borders
 150                     x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
 151                     x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
 152                     x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
 153                     x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
 154
 155                     xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 156                     xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 157                     xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 158                     xd->left_available = (mb_col != 0);
 159
 160                     x->rddiv = cpi->RDDIV;
 161                     x->rdmult = cpi->RDMULT;
 162
 163                     //Copy current mb to a buffer
 164                     vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
 165
 166                     if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
 167                         vp8_activity_masking(cpi, x);
 168
 169                     // Is segmentation enabled
 170                     // MB level adjustment to quantizer
 171                     if (xd->segmentation_enabled)
 172                     {
 173                         // Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
 174                         if (cpi->segmentation_map[map_index + mb_col] <= 3)
 175                             xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index + mb_col];
 176                         else
 177                             xd->mode_info_context->mbmi.segment_id = 0;
 178
 179                         vp8cx_mb_init_quantizer(cpi, x, 1);
 180                     }
 181                     else
 182                         xd->mode_info_context->mbmi.segment_id = 0; // Set to Segment 0 by default
 183
 184                     x->active_ptr = cpi->active_map + map_index + mb_col;
 185
 186                     if (cm->frame_type == KEY_FRAME)
 187                     {
 188                         *totalrate += vp8cx_encode_intra_macroblock(cpi, x, &tp);
 189 #ifdef MODE_STATS
 190                         y_modes[xd->mbmi.mode] ++;
 191 #endif
 192                     }
 193                     else
 194                     {
 195                         *totalrate += vp8cx_encode_inter_macroblock(cpi, x, &tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
 196
 197 #ifdef MODE_STATS
 198                         inter_y_modes[xd->mbmi.mode] ++;
 199
 200                         if (xd->mbmi.mode == SPLITMV)
 201                         {
 202                             int b;
 203
 204                             for (b = 0; b < xd->mbmi.partition_count; b++)
 205                             {
 206                                 inter_b_modes[x->partition->bmi[b].mode] ++;
 207                             }
 208                         }
 209
 210 #endif
 211
 212                         // Count of last ref frame 0,0 usage
 213                         if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
 214                             cpi->inter_zz_count++;
 215
 216                         // Special case code for cyclic refresh
 217                         // If cyclic update enabled then copy xd->mbmi.segment_id; (which may have been updated based on mode
 218                         // during vp8cx_encode_inter_macroblock()) back into the global segmentation map
 219                         if (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)
 220                         {
 221                             const MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
 222                             cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
 223
 224                             // If the block has been refreshed mark it as clean (the magnitude of the -ve influences how long it will be before we consider another refresh):
 225                             // Else if it was coded (last frame 0,0) and has not already been refreshed then mark it as a candidate for cleanup next time (marked 0)
 226                             // else mark it as dirty (1).
 227                             if (mbmi->segment_id)
 228                                 cpi->cyclic_refresh_map[map_index + mb_col] = -1;
 229                             else if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
 230                             {
 231                                 if (cpi->cyclic_refresh_map[map_index + mb_col] == 1)
 232                                     cpi->cyclic_refresh_map[map_index + mb_col] = 0;
 233                             }
 234                             else
 235                                 cpi->cyclic_refresh_map[map_index + mb_col] = 1;
 236
 237                         }
 238                     }
 239
 240 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
 241                     /* pack tokens for this MB */
 242                     {
 243                         int tok_count = tp - tp_start;
 244                         pack_tokens(w, tp_start, tok_count);
 245                     }
 246 #else
 247                     cpi->tplist[mb_row].stop = tp;
 248 #endif
 249                     // Increment pointer into gf usage flags structure.
 250                     x->gf_active_ptr++;
 251
 252                     // Increment the activity mask pointers.
 253                     x->mb_activity_ptr++;
 254
 255                     // adjust to the next column of macroblocks
 256                     x->src.y_buffer += 16;
 257                     x->src.u_buffer += 8;
 258                     x->src.v_buffer += 8;
 259
 260                     recon_yoffset += 16;
 261                     recon_uvoffset += 8;
 262
 263                     // Keep track of segment usage
 264                     segment_counts[xd->mode_info_context->mbmi.segment_id]++;
 265
 266                     // skip to next mb
 267                     xd->mode_info_context++;
 268                     x->partition_info++;
 269                     xd->above_context++;
 270                 }
 271
 272                 vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
 273                                     xd->dst.y_buffer + 16,
 274                                     xd->dst.u_buffer + 8,
 275                                     xd->dst.v_buffer + 8);
 276
 277                 *current_mb_col = mb_col + nsync;
 278
 279                 // this is to account for the border
 280                 xd->mode_info_context++;
 281                 x->partition_info++;
 282
 283                 x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
 284                 x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 285                 x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 286
 287                 xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
 288                 x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
 289                 x->gf_active_ptr   += cm->mb_cols * cpi->encoding_thread_count;
 290
 291                 if (mb_row == cm->mb_rows - 1)
 292                 {
 293                     sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
 294                 }
 295             }
 296         }
 297     }
 298
 299     //printf("exit thread %d\n", ithread);
 300     return 0;
 301 }
 302
 303 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
 304 {
 305
 306     MACROBLOCK *x = mbsrc;
 307     MACROBLOCK *z = mbdst;
 308     int i;
 309
 310     z->ss               = x->ss;
 311     z->ss_count          = x->ss_count;
 312     z->searches_per_step  = x->searches_per_step;
 313     z->errorperbit      = x->errorperbit;
 314
 315     z->sadperbit16      = x->sadperbit16;
 316     z->sadperbit4       = x->sadperbit4;
 317
 318     /*
 319     z->mv_col_min    = x->mv_col_min;
 320     z->mv_col_max    = x->mv_col_max;
 321     z->mv_row_min    = x->mv_row_min;
 322     z->mv_row_max    = x->mv_row_max;
 323     */
 324
 325     z->short_fdct4x4     = x->short_fdct4x4;
 326     z->short_fdct8x4     = x->short_fdct8x4;
 327     z->short_walsh4x4    = x->short_walsh4x4;
 328     z->quantize_b        = x->quantize_b;
 329     z->quantize_b_pair   = x->quantize_b_pair;
 330     z->optimize          = x->optimize;
 331
 332     /*
 333     z->mvc              = x->mvc;
 334     z->src.y_buffer      = x->src.y_buffer;
 335     z->src.u_buffer      = x->src.u_buffer;
 336     z->src.v_buffer      = x->src.v_buffer;
 337     */
 338
 339     z->mvcost[0] =  x->mvcost[0];
 340     z->mvcost[1] =  x->mvcost[1];
 341     z->mvsadcost[0] =  x->mvsadcost[0];
 342     z->mvsadcost[1] =  x->mvsadcost[1];
 343
 344     z->token_costs = x->token_costs;
 345     z->inter_bmode_costs = x->inter_bmode_costs;
 346     z->mbmode_cost = x->mbmode_cost;
 347     z->intra_uv_mode_cost = x->intra_uv_mode_cost;
 348     z->bmode_costs = x->bmode_costs;
 349
 350     for (i = 0; i < 25; i++)
 351     {
 352         z->block[i].quant           = x->block[i].quant;
 353         z->block[i].quant_fast      = x->block[i].quant_fast;
 354         z->block[i].quant_shift     = x->block[i].quant_shift;
 355         z->block[i].zbin            = x->block[i].zbin;
 356         z->block[i].zrun_zbin_boost = x->block[i].zrun_zbin_boost;
 357         z->block[i].round           = x->block[i].round;
 358         z->block[i].src_stride      = x->block[i].src_stride;
 359     }
 360
 361     z->q_index           = x->q_index;
 362     z->act_zbin_adj      = x->act_zbin_adj;
 363     z->last_act_zbin_adj = x->last_act_zbin_adj;
 364
 365     {
 366         MACROBLOCKD *xd = &x->e_mbd;
 367         MACROBLOCKD *zd = &z->e_mbd;
 368
 369         /*
 370         zd->mode_info_context = xd->mode_info_context;
 371         zd->mode_info        = xd->mode_info;
 372
 373         zd->mode_info_stride  = xd->mode_info_stride;
 374         zd->frame_type       = xd->frame_type;
 375         zd->up_available     = xd->up_available   ;
 376         zd->left_available   = xd->left_available;
 377         zd->left_context     = xd->left_context;
 378         zd->last_frame_dc     = xd->last_frame_dc;
 379         zd->last_frame_dccons = xd->last_frame_dccons;
 380         zd->gold_frame_dc     = xd->gold_frame_dc;
 381         zd->gold_frame_dccons = xd->gold_frame_dccons;
 382         zd->mb_to_left_edge    = xd->mb_to_left_edge;
 383         zd->mb_to_right_edge   = xd->mb_to_right_edge;
 384         zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
 385         zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
 386         zd->gf_active_ptr     = xd->gf_active_ptr;
 387         zd->frames_since_golden       = xd->frames_since_golden;
 388         zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
 389         */
 390         zd->subpixel_predict         = xd->subpixel_predict;
 391         zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
 392         zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
 393         zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
 394         zd->segmentation_enabled     = xd->segmentation_enabled;
 395         zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
 396         vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data,
 397                    sizeof(xd->segment_feature_data));
 398
 399         vpx_memcpy(zd->dequant_y1_dc, xd->dequant_y1_dc,
 400                    sizeof(xd->dequant_y1_dc));
 401         vpx_memcpy(zd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
 402         vpx_memcpy(zd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
 403         vpx_memcpy(zd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
 404
 405 #if 1
 406         /*TODO:  Remove dequant from BLOCKD.  This is a temporary solution until
 407          * the quantizer code uses a passed in pointer to the dequant constants.
 408          * This will also require modifications to the x86 and neon assembly.
 409          * */
 410         for (i = 0; i < 16; i++)
 411             zd->block[i].dequant = zd->dequant_y1;
 412         for (i = 16; i < 24; i++)
 413             zd->block[i].dequant = zd->dequant_uv;
 414         zd->block[24].dequant = zd->dequant_y2;
 415 #endif
 416     }
 417 }
 418
 419 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
 420                                MACROBLOCK *x,
 421                                MB_ROW_COMP *mbr_ei,
 422                                int mb_row,
 423                                int count
 424                               )
 425 {
 426
 427     VP8_COMMON *const cm = & cpi->common;
 428     MACROBLOCKD *const xd = & x->e_mbd;
 429     int i;
 430     (void) mb_row;
 431
 432     for (i = 0; i < count; i++)
 433     {
 434         MACROBLOCK *mb = & mbr_ei[i].mb;
 435         MACROBLOCKD *mbd = &mb->e_mbd;
 436
 437         mbd->subpixel_predict        = xd->subpixel_predict;
 438         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
 439         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
 440         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
 441         mb->gf_active_ptr            = x->gf_active_ptr;
 442
 443         vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
 444         mbr_ei[i].totalrate = 0;
 445
 446         mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
 447
 448         mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
 449         mbd->mode_info_stride  = cm->mode_info_stride;
 450
 451         mbd->frame_type = cm->frame_type;
 452
 453         mb->src = * cpi->Source;
 454         mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
 455         mbd->dst = cm->yv12_fb[cm->new_fb_idx];
 456
 457         mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
 458         mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
 459         mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
 460
 461         vp8_build_block_offsets(mb);
 462
 463         vp8_setup_block_dptrs(mbd);
 464
 465         vp8_setup_block_ptrs(mb);
 466
 467         mbd->left_context = &cm->left_context;
 468         mb->mvc = cm->fc.mvc;
 469
 470         setup_mbby_copy(&mbr_ei[i].mb, x);
 471
 472         mbd->fullpixel_mask = 0xffffffff;
 473         if(cm->full_pixel)
 474             mbd->fullpixel_mask = 0xfffffff8;
 475     }
 476 }
 477
 478 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
 479 {
 480     const VP8_COMMON * cm = &cpi->common;
 481
 482     cpi->b_multi_threaded = 0;
 483     cpi->encoding_thread_count = 0;
 484     cpi->b_lpf_running = 0;
 485
 486     if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
 487     {
 488         int ithread;
 489         int th_count = cpi->oxcf.multi_threaded - 1;
 490
 491         /* don't allocate more threads than cores available */
 492         if (cpi->oxcf.multi_threaded > cm->processor_core_count)
 493             th_count = cm->processor_core_count - 1;
 494
 495         /* we have th_count + 1 (main) threads processing one row each */
 496         /* no point to have more threads than the sync range allows */
 497         if(th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1))
 498         {
 499             th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
 500         }
 501
 502         if(th_count == 0)
 503             return;
 504
 505         CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * th_count));
 506         CHECK_MEM_ERROR(cpi->h_event_start_encoding, vpx_malloc(sizeof(sem_t) * th_count));
 507         CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
 508         vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
 509         CHECK_MEM_ERROR(cpi->en_thread_data,
 510                         vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
 511         CHECK_MEM_ERROR(cpi->mt_current_mb_col,
 512                         vpx_malloc(sizeof(*cpi->mt_current_mb_col) * cm->mb_rows));
 513
 514         sem_init(&cpi->h_event_end_encoding, 0, 0);
 515
 516         cpi->b_multi_threaded = 1;
 517         cpi->encoding_thread_count = th_count;
 518
 519         /*
 520         printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
 521                (cpi->encoding_thread_count +1));
 522         */
 523
 524         for (ithread = 0; ithread < th_count; ithread++)
 525         {
 526             ENCODETHREAD_DATA * ethd = &cpi->en_thread_data[ithread];
 527
 528             sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
 529             ethd->ithread = ithread;
 530             ethd->ptr1 = (void *)cpi;
 531             ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
 532
 533             pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, ethd);
 534         }
 535
 536         {
 537             LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
 538
 539             sem_init(&cpi->h_event_start_lpf, 0, 0);
 540             sem_init(&cpi->h_event_end_lpf, 0, 0);
 541
 542             lpfthd->ptr1 = (void *)cpi;
 543             pthread_create(&cpi->h_filter_thread, 0, thread_loopfilter, lpfthd);
 544         }
 545     }
 546
 547 }
 548
 549 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
 550 {
 551     if (cpi->b_multi_threaded)
 552     {
 553         //shutdown other threads
 554         cpi->b_multi_threaded = 0;
 555         {
 556             int i;
 557
 558             for (i = 0; i < cpi->encoding_thread_count; i++)
 559             {
 560                 //SetEvent(cpi->h_event_mbrencoding[i]);
 561                 sem_post(&cpi->h_event_start_encoding[i]);
 562                 pthread_join(cpi->h_encoding_thread[i], 0);
 563
 564                 sem_destroy(&cpi->h_event_start_encoding[i]);
 565             }
 566
 567             sem_post(&cpi->h_event_start_lpf);
 568             pthread_join(cpi->h_filter_thread, 0);
 569         }
 570
 571         sem_destroy(&cpi->h_event_end_encoding);
 572         sem_destroy(&cpi->h_event_end_lpf);
 573         sem_destroy(&cpi->h_event_start_lpf);
 574
 575         //free thread related resources
 576         vpx_free(cpi->h_event_start_encoding);
 577         vpx_free(cpi->h_encoding_thread);
 578         vpx_free(cpi->mb_row_ei);
 579         vpx_free(cpi->en_thread_data);
 580         vpx_free(cpi->mt_current_mb_col);
 581     }
 582 }
 583 #endif