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
  16 #if CONFIG_MULTITHREAD
  17
  18 extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
  19                                          TOKENEXTRA **t, int recon_yoffset,
  20                                          int recon_uvoffset);
  21 extern int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x,
  22                                           TOKENEXTRA **t);
  23 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x);
  24 extern void vp8_build_block_offsets(MACROBLOCK *x);
  25 extern void vp8_setup_block_ptrs(MACROBLOCK *x);
  26
  27 extern void loopfilter_frame(VP8_COMP *cpi, VP8_COMMON *cm);
  28
  29 static THREAD_FUNCTION loopfilter_thread(void *p_data)
  30 {
  31     VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
  32     VP8_COMMON *cm = &cpi->common;
  33
  34     while (1)
  35     {
  36         if (cpi->b_multi_threaded == 0)
  37             break;
  38
  39         if (sem_wait(&cpi->h_event_start_lpf) == 0)
  40         {
  41             if (cpi->b_multi_threaded == FALSE) // we're shutting down
  42                 break;
  43
  44             loopfilter_frame(cpi, cm);
  45
  46             sem_post(&cpi->h_event_end_lpf);
  47         }
  48     }
  49
  50     return 0;
  51 }
  52
  53 static
  54 THREAD_FUNCTION thread_encoding_proc(void *p_data)
  55 {
  56     int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
  57     VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
  58     MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
  59     ENTROPY_CONTEXT_PLANES mb_row_left_context;
  60
  61     const int nsync = cpi->mt_sync_range;
  62     //printf("Started thread %d\n", ithread);
  63
  64     while (1)
  65     {
  66         if (cpi->b_multi_threaded == 0)
  67             break;
  68
  69         //if(WaitForSingleObject(cpi->h_event_mbrencoding[ithread], INFINITE) == WAIT_OBJECT_0)
  70         if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0)
  71         {
  72             VP8_COMMON *cm = &cpi->common;
  73             int mb_row;
  74             MACROBLOCK *x = &mbri->mb;
  75             MACROBLOCKD *xd = &x->e_mbd;
  76             TOKENEXTRA *tp ;
  77
  78             int *segment_counts = mbri->segment_counts;
  79             int *totalrate = &mbri->totalrate;
  80
  81             if (cpi->b_multi_threaded == FALSE) // we're shutting down
  82                 break;
  83
  84             for (mb_row = ithread + 1; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
  85             {
  86
  87                 int i;
  88                 int recon_yoffset, recon_uvoffset;
  89                 int mb_col;
  90                 int ref_fb_idx = cm->lst_fb_idx;
  91                 int dst_fb_idx = cm->new_fb_idx;
  92                 int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
  93                 int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
  94                 int map_index = (mb_row * cm->mb_cols);
  95                 volatile int *last_row_current_mb_col;
  96
  97                 tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
  98
  99                 last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
 100
 101                 // reset above block coeffs
 102                 xd->above_context = cm->above_context;
 103                 xd->left_context = &mb_row_left_context;
 104
 105                 vp8_zero(mb_row_left_context);
 106
 107                 xd->up_available = (mb_row != 0);
 108                 recon_yoffset = (mb_row * recon_y_stride * 16);
 109                 recon_uvoffset = (mb_row * recon_uv_stride * 8);
 110
 111                 cpi->tplist[mb_row].start = tp;
 112
 113                 //printf("Thread mb_row = %d\n", mb_row);
 114
 115                 // Set the mb activity pointer to the start of the row.
 116                 x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
 117                 x->mb_norm_activity_ptr =
 118                     &cpi->mb_norm_activity_map[map_index];
 119
 120                 // for each macroblock col in image
 121                 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
 122                 {
 123                     if ((mb_col & (nsync - 1)) == 0)
 124                     {
 125                         while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != cm->mb_cols - 1)
 126                         {
 127                             x86_pause_hint();
 128                             thread_sleep(0);
 129                         }
 130                     }
 131
 132                     // Distance of Mb to the various image edges.
 133                     // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
 134                     xd->mb_to_left_edge = -((mb_col * 16) << 3);
 135                     xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
 136                     xd->mb_to_top_edge = -((mb_row * 16) << 3);
 137                     xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
 138
 139                     // Set up limit values for motion vectors used to prevent them extending outside the UMV borders
 140                     x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
 141                     x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
 142                     x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
 143                     x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
 144
 145                     xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 146                     xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 147                     xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 148                     xd->left_available = (mb_col != 0);
 149
 150                     x->rddiv = cpi->RDDIV;
 151                     x->rdmult = cpi->RDMULT;
 152
 153                     if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
 154                         vp8_activity_masking(cpi, x);
 155
 156                     // Is segmentation enabled
 157                     // MB level adjutment to quantizer
 158                     if (xd->segmentation_enabled)
 159                     {
 160                         // Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
 161                         if (cpi->segmentation_map[map_index + mb_col] <= 3)
 162                             xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index + mb_col];
 163                         else
 164                             xd->mode_info_context->mbmi.segment_id = 0;
 165
 166                         vp8cx_mb_init_quantizer(cpi, x);
 167                     }
 168                     else
 169                         xd->mode_info_context->mbmi.segment_id = 0; // Set to Segment 0 by default
 170
 171                     x->active_ptr = cpi->active_map + map_index + mb_col;
 172
 173                     if (cm->frame_type == KEY_FRAME)
 174                     {
 175                         *totalrate += vp8cx_encode_intra_macro_block(cpi, x, &tp);
 176 #ifdef MODE_STATS
 177                         y_modes[xd->mbmi.mode] ++;
 178 #endif
 179                     }
 180                     else
 181                     {
 182                         *totalrate += vp8cx_encode_inter_macroblock(cpi, x, &tp, recon_yoffset, recon_uvoffset);
 183
 184 #ifdef MODE_STATS
 185                         inter_y_modes[xd->mbmi.mode] ++;
 186
 187                         if (xd->mbmi.mode == SPLITMV)
 188                         {
 189                             int b;
 190
 191                             for (b = 0; b < xd->mbmi.partition_count; b++)
 192                             {
 193                                 inter_b_modes[x->partition->bmi[b].mode] ++;
 194                             }
 195                         }
 196
 197 #endif
 198
 199                         // Count of last ref frame 0,0 useage
 200                         if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
 201                             cpi->inter_zz_count++;
 202
 203                         // Special case code for cyclic refresh
 204                         // If cyclic update enabled then copy xd->mbmi.segment_id; (which may have been updated based on mode
 205                         // during vp8cx_encode_inter_macroblock()) back into the global sgmentation map
 206                         if (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)
 207                         {
 208                             const MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
 209                             cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
 210
 211                             // 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):
 212                             // 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)
 213                             // else mark it as dirty (1).
 214                             if (mbmi->segment_id)
 215                                 cpi->cyclic_refresh_map[map_index + mb_col] = -1;
 216                             else if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
 217                             {
 218                                 if (cpi->cyclic_refresh_map[map_index + mb_col] == 1)
 219                                     cpi->cyclic_refresh_map[map_index + mb_col] = 0;
 220                             }
 221                             else
 222                                 cpi->cyclic_refresh_map[map_index + mb_col] = 1;
 223
 224                         }
 225                     }
 226                     cpi->tplist[mb_row].stop = tp;
 227
 228                     // Increment pointer into gf useage flags structure.
 229                     x->gf_active_ptr++;
 230
 231                     // Increment the activity mask pointers.
 232                     x->mb_activity_ptr++;
 233                     x->mb_norm_activity_ptr++;
 234
 235                     if(cm->frame_type != INTRA_FRAME)
 236                     {
 237                         if (xd->mode_info_context->mbmi.mode != B_PRED)
 238                         {
 239                             for (i = 0; i < 16; i++)
 240                                 xd->mode_info_context->bmi[i].mv.as_int = xd->block[i].bmi.mv.as_int;
 241                         }else
 242                         {
 243                             for (i = 0; i < 16; i++)
 244                                 xd->mode_info_context->bmi[i].as_mode = xd->block[i].bmi.mode;
 245                         }
 246                     }
 247                     else {
 248                         if(xd->mode_info_context->mbmi.mode != B_PRED)
 249                             for (i = 0; i < 16; i++)
 250                                 xd->mode_info_context->bmi[i].as_mode = xd->block[i].bmi.mode;
 251                     }
 252
 253                     // adjust to the next column of macroblocks
 254                     x->src.y_buffer += 16;
 255                     x->src.u_buffer += 8;
 256                     x->src.v_buffer += 8;
 257
 258                     recon_yoffset += 16;
 259                     recon_uvoffset += 8;
 260
 261                     // Keep track of segment useage
 262                     segment_counts[xd->mode_info_context->mbmi.segment_id]++;
 263
 264                     // skip to next mb
 265                     xd->mode_info_context++;
 266                     x->partition_info++;
 267                     xd->above_context++;
 268
 269                     cpi->mt_current_mb_col[mb_row] = mb_col;
 270                 }
 271
 272                 //extend the recon for intra prediction
 273                 vp8_extend_mb_row(
 274                     &cm->yv12_fb[dst_fb_idx],
 275                     xd->dst.y_buffer + 16,
 276                     xd->dst.u_buffer + 8,
 277                     xd->dst.v_buffer + 8);
 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                     //SetEvent(cpi->h_event_main);
 294                     sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
 295                 }
 296             }
 297         }
 298     }
 299
 300     //printf("exit thread %d\n", ithread);
 301     return 0;
 302 }
 303
 304 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
 305 {
 306
 307     MACROBLOCK *x = mbsrc;
 308     MACROBLOCK *z = mbdst;
 309     int i;
 310
 311     z->ss               = x->ss;
 312     z->ss_count          = x->ss_count;
 313     z->searches_per_step  = x->searches_per_step;
 314     z->errorperbit      = x->errorperbit;
 315
 316     z->sadperbit16      = x->sadperbit16;
 317     z->sadperbit4       = x->sadperbit4;
 318
 319     /*
 320     z->mv_col_min    = x->mv_col_min;
 321     z->mv_col_max    = x->mv_col_max;
 322     z->mv_row_min    = x->mv_row_min;
 323     z->mv_row_max    = x->mv_row_max;
 324     z->vector_range = x->vector_range ;
 325     */
 326
 327     z->vp8_short_fdct4x4     = x->vp8_short_fdct4x4;
 328     z->vp8_short_fdct8x4     = x->vp8_short_fdct8x4;
 329     z->short_walsh4x4    = x->short_walsh4x4;
 330     z->quantize_b        = x->quantize_b;
 331     z->quantize_b_pair   = x->quantize_b_pair;
 332     z->optimize          = x->optimize;
 333
 334     /*
 335     z->mvc              = x->mvc;
 336     z->src.y_buffer      = x->src.y_buffer;
 337     z->src.u_buffer      = x->src.u_buffer;
 338     z->src.v_buffer      = x->src.v_buffer;
 339     */
 340
 341
 342     vpx_memcpy(z->mvcosts,          x->mvcosts,         sizeof(x->mvcosts));
 343     z->mvcost[0] = &z->mvcosts[0][mv_max+1];
 344     z->mvcost[1] = &z->mvcosts[1][mv_max+1];
 345     z->mvsadcost[0] = &z->mvsadcosts[0][mvfp_max+1];
 346     z->mvsadcost[1] = &z->mvsadcosts[1][mvfp_max+1];
 347
 348
 349     vpx_memcpy(z->token_costs,       x->token_costs,      sizeof(x->token_costs));
 350     vpx_memcpy(z->inter_bmode_costs,  x->inter_bmode_costs, sizeof(x->inter_bmode_costs));
 351     //memcpy(z->mvcosts,            x->mvcosts,         sizeof(x->mvcosts));
 352     //memcpy(z->mvcost,         x->mvcost,          sizeof(x->mvcost));
 353     vpx_memcpy(z->mbmode_cost,       x->mbmode_cost,      sizeof(x->mbmode_cost));
 354     vpx_memcpy(z->intra_uv_mode_cost,  x->intra_uv_mode_cost, sizeof(x->intra_uv_mode_cost));
 355     vpx_memcpy(z->bmode_costs,       x->bmode_costs,      sizeof(x->bmode_costs));
 356
 357     for (i = 0; i < 25; i++)
 358     {
 359         z->block[i].quant           = x->block[i].quant;
 360         z->block[i].quant_fast      = x->block[i].quant_fast;
 361         z->block[i].quant_shift     = x->block[i].quant_shift;
 362         z->block[i].zbin            = x->block[i].zbin;
 363         z->block[i].zrun_zbin_boost   = x->block[i].zrun_zbin_boost;
 364         z->block[i].round           = x->block[i].round;
 365         /*
 366         z->block[i].src             = x->block[i].src;
 367         */
 368         z->block[i].src_stride       = x->block[i].src_stride;
 369         z->block[i].force_empty      = x->block[i].force_empty;
 370
 371     }
 372
 373     {
 374         MACROBLOCKD *xd = &x->e_mbd;
 375         MACROBLOCKD *zd = &z->e_mbd;
 376
 377         /*
 378         zd->mode_info_context = xd->mode_info_context;
 379         zd->mode_info        = xd->mode_info;
 380
 381         zd->mode_info_stride  = xd->mode_info_stride;
 382         zd->frame_type       = xd->frame_type;
 383         zd->up_available     = xd->up_available   ;
 384         zd->left_available   = xd->left_available;
 385         zd->left_context     = xd->left_context;
 386         zd->last_frame_dc     = xd->last_frame_dc;
 387         zd->last_frame_dccons = xd->last_frame_dccons;
 388         zd->gold_frame_dc     = xd->gold_frame_dc;
 389         zd->gold_frame_dccons = xd->gold_frame_dccons;
 390         zd->mb_to_left_edge    = xd->mb_to_left_edge;
 391         zd->mb_to_right_edge   = xd->mb_to_right_edge;
 392         zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
 393         zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
 394         zd->gf_active_ptr     = xd->gf_active_ptr;
 395         zd->frames_since_golden       = xd->frames_since_golden;
 396         zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
 397         */
 398         zd->subpixel_predict         = xd->subpixel_predict;
 399         zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
 400         zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
 401         zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
 402         zd->segmentation_enabled     = xd->segmentation_enabled;
 403         zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
 404         vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
 405
 406         for (i = 0; i < 25; i++)
 407         {
 408             zd->block[i].dequant = xd->block[i].dequant;
 409         }
 410     }
 411 }
 412
 413 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
 414                                MACROBLOCK *x,
 415                                MB_ROW_COMP *mbr_ei,
 416                                int mb_row,
 417                                int count
 418                               )
 419 {
 420
 421     VP8_COMMON *const cm = & cpi->common;
 422     MACROBLOCKD *const xd = & x->e_mbd;
 423     int i;
 424     (void) mb_row;
 425
 426     for (i = 0; i < count; i++)
 427     {
 428         MACROBLOCK *mb = & mbr_ei[i].mb;
 429         MACROBLOCKD *mbd = &mb->e_mbd;
 430
 431         mbd->subpixel_predict        = xd->subpixel_predict;
 432         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
 433         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
 434         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
 435 #if CONFIG_RUNTIME_CPU_DETECT
 436         mbd->rtcd                   = xd->rtcd;
 437 #endif
 438         mb->gf_active_ptr            = x->gf_active_ptr;
 439
 440         mb->vector_range             = 32;
 441
 442         vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
 443         mbr_ei[i].totalrate = 0;
 444
 445         mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
 446
 447         mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
 448         mbd->mode_info_stride  = cm->mode_info_stride;
 449
 450         mbd->frame_type = cm->frame_type;
 451
 452         mbd->frames_since_golden = cm->frames_since_golden;
 453         mbd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
 454
 455         mb->src = * cpi->Source;
 456         mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
 457         mbd->dst = cm->yv12_fb[cm->new_fb_idx];
 458
 459         mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
 460         mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
 461         mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
 462
 463         vp8_build_block_offsets(mb);
 464
 465         vp8_setup_block_dptrs(mbd);
 466
 467         vp8_setup_block_ptrs(mb);
 468
 469         mbd->left_context = &cm->left_context;
 470         mb->mvc = cm->fc.mvc;
 471
 472         setup_mbby_copy(&mbr_ei[i].mb, x);
 473
 474     }
 475 }
 476
 477 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
 478 {
 479     const VP8_COMMON * cm = &cpi->common;
 480
 481     cpi->b_multi_threaded = 0;
 482     cpi->encoding_thread_count = 0;
 483
 484     if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
 485     {
 486         int ithread;
 487         int th_count = cpi->oxcf.multi_threaded - 1;
 488
 489         /* don't allocate more threads than cores available */
 490         if (cpi->oxcf.multi_threaded > cm->processor_core_count)
 491             th_count = cm->processor_core_count - 1;
 492
 493         /* we have th_count + 1 (main) threads processing one row each */
 494         /* no point to have more threads than the sync range allows */
 495         if(th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1))
 496         {
 497             th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
 498         }
 499
 500         if(th_count == 0)
 501             return;
 502
 503         CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * th_count));
 504         CHECK_MEM_ERROR(cpi->h_event_start_encoding, vpx_malloc(sizeof(sem_t) * th_count));
 505         CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
 506         vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
 507         CHECK_MEM_ERROR(cpi->en_thread_data,
 508                         vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
 509         CHECK_MEM_ERROR(cpi->mt_current_mb_col,
 510                         vpx_malloc(sizeof(*cpi->mt_current_mb_col) * cm->mb_rows));
 511
 512         sem_init(&cpi->h_event_end_encoding, 0, 0);
 513
 514         cpi->b_multi_threaded = 1;
 515         cpi->encoding_thread_count = th_count;
 516
 517         /*
 518         printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
 519                (cpi->encoding_thread_count +1));
 520         */
 521
 522         for (ithread = 0; ithread < th_count; ithread++)
 523         {
 524             ENCODETHREAD_DATA * ethd = &cpi->en_thread_data[ithread];
 525
 526             sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
 527             ethd->ithread = ithread;
 528             ethd->ptr1 = (void *)cpi;
 529             ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
 530
 531             pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, ethd);
 532         }
 533
 534         {
 535             LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
 536
 537             sem_init(&cpi->h_event_start_lpf, 0, 0);
 538             sem_init(&cpi->h_event_end_lpf, 0, 0);
 539
 540             lpfthd->ptr1 = (void *)cpi;
 541             pthread_create(&cpi->h_filter_thread, 0, loopfilter_thread, lpfthd);
 542         }
 543     }
 544
 545 }
 546
 547 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
 548 {
 549     if (cpi->b_multi_threaded)
 550     {
 551         //shutdown other threads
 552         cpi->b_multi_threaded = 0;
 553         {
 554             int i;
 555
 556             for (i = 0; i < cpi->encoding_thread_count; i++)
 557             {
 558                 //SetEvent(cpi->h_event_mbrencoding[i]);
 559                 sem_post(&cpi->h_event_start_encoding[i]);
 560                 pthread_join(cpi->h_encoding_thread[i], 0);
 561
 562                 sem_destroy(&cpi->h_event_start_encoding[i]);
 563             }
 564
 565             sem_post(&cpi->h_event_start_lpf);
 566             pthread_join(cpi->h_filter_thread, 0);
 567         }
 568
 569         sem_destroy(&cpi->h_event_end_encoding);
 570         sem_destroy(&cpi->h_event_end_lpf);
 571         sem_destroy(&cpi->h_event_start_lpf);
 572
 573         //free thread related resources
 574         vpx_free(cpi->h_event_start_encoding);
 575         vpx_free(cpi->h_encoding_thread);
 576         vpx_free(cpi->mb_row_ei);
 577         vpx_free(cpi->en_thread_data);
 578         vpx_free(cpi->mt_current_mb_col);
 579     }
 580 }
 581 #endif