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 #if CONFIG_MULTITHREAD
  28
  29 extern void loopfilter_frame(VP8_COMP *cpi, VP8_COMMON *cm);
  30
  31 static THREAD_FUNCTION loopfilter_thread(void *p_data)
  32 {
  33     VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
  34     VP8_COMMON *cm = &cpi->common;
  35
  36     while (1)
  37     {
  38         if (cpi->b_multi_threaded == 0)
  39             break;
  40
  41         if (sem_wait(&cpi->h_event_start_lpf) == 0)
  42         {
  43             if (cpi->b_multi_threaded == FALSE) // we're shutting down
  44                 break;
  45
  46             loopfilter_frame(cpi, cm);
  47
  48             sem_post(&cpi->h_event_end_lpf);
  49         }
  50     }
  51
  52     return 0;
  53 }
  54 #endif
  55
  56 static
  57 THREAD_FUNCTION thread_encoding_proc(void *p_data)
  58 {
  59     int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
  60     VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
  61     MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
  62     ENTROPY_CONTEXT_PLANES mb_row_left_context;
  63
  64     const int nsync = cpi->mt_sync_range;
  65     //printf("Started thread %d\n", ithread);
  66
  67     while (1)
  68     {
  69         if (cpi->b_multi_threaded == 0)
  70             break;
  71
  72         //if(WaitForSingleObject(cpi->h_event_mbrencoding[ithread], INFINITE) == WAIT_OBJECT_0)
  73         if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0)
  74         {
  75             VP8_COMMON *cm = &cpi->common;
  76             int mb_row;
  77             MACROBLOCK *x = &mbri->mb;
  78             MACROBLOCKD *xd = &x->e_mbd;
  79             TOKENEXTRA *tp ;
  80
  81             int *segment_counts = mbri->segment_counts;
  82             int *totalrate = &mbri->totalrate;
  83
  84             if (cpi->b_multi_threaded == FALSE) // we're shutting down
  85                 break;
  86
  87             for (mb_row = ithread + 1; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
  88             {
  89
  90                 int i;
  91                 int recon_yoffset, recon_uvoffset;
  92                 int mb_col;
  93                 int ref_fb_idx = cm->lst_fb_idx;
  94                 int dst_fb_idx = cm->new_fb_idx;
  95                 int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
  96                 int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
  97                 volatile int *last_row_current_mb_col;
  98                 INT64 activity_sum = 0;
  99
 100                 tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
 101
 102                 last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
 103
 104                 // reset above block coeffs
 105                 xd->above_context = cm->above_context;
 106                 xd->left_context = &mb_row_left_context;
 107
 108                 vp8_zero(mb_row_left_context);
 109
 110                 xd->up_available = (mb_row != 0);
 111                 recon_yoffset = (mb_row * recon_y_stride * 16);
 112                 recon_uvoffset = (mb_row * recon_uv_stride * 8);
 113
 114                 cpi->tplist[mb_row].start = tp;
 115
 116                 //printf("Thread mb_row = %d\n", mb_row);
 117
 118                 // for each macroblock col in image
 119                 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
 120                 {
 121                     int seg_map_index = (mb_row * cm->mb_cols);
 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                         activity_sum += 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[seg_map_index + mb_col] <= 3)
 162                             xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[seg_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 + seg_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[seg_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[seg_map_index + mb_col] = -1;
 216                             else if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
 217                             {
 218                                 if (cpi->cyclic_refresh_map[seg_map_index + mb_col] == 1)
 219                                     cpi->cyclic_refresh_map[seg_map_index + mb_col] = 0;
 220                             }
 221                             else
 222                                 cpi->cyclic_refresh_map[seg_map_index + mb_col] = 1;
 223
 224                         }
 225                     }
 226                     cpi->tplist[mb_row].stop = tp;
 227
 228                     x->gf_active_ptr++; // Increment pointer into gf useage flags structure for next mb
 229
 230                     for (i = 0; i < 16; i++)
 231                         vpx_memcpy(&xd->mode_info_context->bmi[i], &xd->block[i].bmi, sizeof(xd->block[i].bmi));
 232
 233                     // adjust to the next column of macroblocks
 234                     x->src.y_buffer += 16;
 235                     x->src.u_buffer += 8;
 236                     x->src.v_buffer += 8;
 237
 238                     recon_yoffset += 16;
 239                     recon_uvoffset += 8;
 240
 241                     // Keep track of segment useage
 242                     segment_counts[xd->mode_info_context->mbmi.segment_id]++;
 243
 244                     // skip to next mb
 245                     xd->mode_info_context++;
 246                     x->partition_info++;
 247                     xd->above_context++;
 248
 249                     cpi->mt_current_mb_col[mb_row] = mb_col;
 250                 }
 251
 252                 //extend the recon for intra prediction
 253                 vp8_extend_mb_row(
 254                     &cm->yv12_fb[dst_fb_idx],
 255                     xd->dst.y_buffer + 16,
 256                     xd->dst.u_buffer + 8,
 257                     xd->dst.v_buffer + 8);
 258
 259                 // this is to account for the border
 260                 xd->mode_info_context++;
 261                 x->partition_info++;
 262                 x->activity_sum += activity_sum;
 263
 264                 x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
 265                 x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 266                 x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 267
 268                 xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
 269                 x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
 270
 271                 if (mb_row == cm->mb_rows - 1)
 272                 {
 273                     //SetEvent(cpi->h_event_main);
 274                     sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
 275                 }
 276             }
 277         }
 278     }
 279
 280     //printf("exit thread %d\n", ithread);
 281     return 0;
 282 }
 283
 284 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
 285 {
 286
 287     MACROBLOCK *x = mbsrc;
 288     MACROBLOCK *z = mbdst;
 289     int i;
 290
 291     z->ss               = x->ss;
 292     z->ss_count          = x->ss_count;
 293     z->searches_per_step  = x->searches_per_step;
 294     z->errorperbit      = x->errorperbit;
 295
 296     z->sadperbit16      = x->sadperbit16;
 297     z->sadperbit4       = x->sadperbit4;
 298     z->errthresh        = x->errthresh;
 299
 300     /*
 301     z->mv_col_min    = x->mv_col_min;
 302     z->mv_col_max    = x->mv_col_max;
 303     z->mv_row_min    = x->mv_row_min;
 304     z->mv_row_max    = x->mv_row_max;
 305     z->vector_range = x->vector_range ;
 306     */
 307
 308     z->vp8_short_fdct4x4     = x->vp8_short_fdct4x4;
 309     z->vp8_short_fdct8x4     = x->vp8_short_fdct8x4;
 310     z->short_walsh4x4    = x->short_walsh4x4;
 311     z->quantize_b        = x->quantize_b;
 312     z->optimize          = x->optimize;
 313
 314     /*
 315     z->mvc              = x->mvc;
 316     z->src.y_buffer      = x->src.y_buffer;
 317     z->src.u_buffer      = x->src.u_buffer;
 318     z->src.v_buffer      = x->src.v_buffer;
 319     */
 320
 321
 322     vpx_memcpy(z->mvcosts,          x->mvcosts,         sizeof(x->mvcosts));
 323     z->mvcost[0] = &z->mvcosts[0][mv_max+1];
 324     z->mvcost[1] = &z->mvcosts[1][mv_max+1];
 325     z->mvsadcost[0] = &z->mvsadcosts[0][mv_max+1];
 326     z->mvsadcost[1] = &z->mvsadcosts[1][mv_max+1];
 327
 328
 329     vpx_memcpy(z->token_costs,       x->token_costs,      sizeof(x->token_costs));
 330     vpx_memcpy(z->inter_bmode_costs,  x->inter_bmode_costs, sizeof(x->inter_bmode_costs));
 331     //memcpy(z->mvcosts,            x->mvcosts,         sizeof(x->mvcosts));
 332     //memcpy(z->mvcost,         x->mvcost,          sizeof(x->mvcost));
 333     vpx_memcpy(z->mbmode_cost,       x->mbmode_cost,      sizeof(x->mbmode_cost));
 334     vpx_memcpy(z->intra_uv_mode_cost,  x->intra_uv_mode_cost, sizeof(x->intra_uv_mode_cost));
 335     vpx_memcpy(z->bmode_costs,       x->bmode_costs,      sizeof(x->bmode_costs));
 336
 337     for (i = 0; i < 25; i++)
 338     {
 339         z->block[i].quant           = x->block[i].quant;
 340         z->block[i].quant_fast      = x->block[i].quant_fast;
 341         z->block[i].quant_shift     = x->block[i].quant_shift;
 342         z->block[i].zbin            = x->block[i].zbin;
 343         z->block[i].zrun_zbin_boost   = x->block[i].zrun_zbin_boost;
 344         z->block[i].round           = x->block[i].round;
 345         /*
 346         z->block[i].src             = x->block[i].src;
 347         */
 348         z->block[i].src_stride       = x->block[i].src_stride;
 349         z->block[i].force_empty      = x->block[i].force_empty;
 350
 351     }
 352
 353     {
 354         MACROBLOCKD *xd = &x->e_mbd;
 355         MACROBLOCKD *zd = &z->e_mbd;
 356
 357         /*
 358         zd->mode_info_context = xd->mode_info_context;
 359         zd->mode_info        = xd->mode_info;
 360
 361         zd->mode_info_stride  = xd->mode_info_stride;
 362         zd->frame_type       = xd->frame_type;
 363         zd->up_available     = xd->up_available   ;
 364         zd->left_available   = xd->left_available;
 365         zd->left_context     = xd->left_context;
 366         zd->last_frame_dc     = xd->last_frame_dc;
 367         zd->last_frame_dccons = xd->last_frame_dccons;
 368         zd->gold_frame_dc     = xd->gold_frame_dc;
 369         zd->gold_frame_dccons = xd->gold_frame_dccons;
 370         zd->mb_to_left_edge    = xd->mb_to_left_edge;
 371         zd->mb_to_right_edge   = xd->mb_to_right_edge;
 372         zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
 373         zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
 374         zd->gf_active_ptr     = xd->gf_active_ptr;
 375         zd->frames_since_golden       = xd->frames_since_golden;
 376         zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
 377         */
 378         zd->subpixel_predict         = xd->subpixel_predict;
 379         zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
 380         zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
 381         zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
 382         zd->segmentation_enabled     = xd->segmentation_enabled;
 383         zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
 384         vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
 385
 386         for (i = 0; i < 25; i++)
 387         {
 388             zd->block[i].dequant = xd->block[i].dequant;
 389         }
 390     }
 391 }
 392
 393 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
 394                                MACROBLOCK *x,
 395                                MB_ROW_COMP *mbr_ei,
 396                                int mb_row,
 397                                int count
 398                               )
 399 {
 400
 401     VP8_COMMON *const cm = & cpi->common;
 402     MACROBLOCKD *const xd = & x->e_mbd;
 403     int i;
 404     (void) mb_row;
 405
 406     for (i = 0; i < count; i++)
 407     {
 408         MACROBLOCK *mb = & mbr_ei[i].mb;
 409         MACROBLOCKD *mbd = &mb->e_mbd;
 410
 411         mbd->subpixel_predict        = xd->subpixel_predict;
 412         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
 413         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
 414         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
 415 #if CONFIG_RUNTIME_CPU_DETECT
 416         mbd->rtcd                   = xd->rtcd;
 417 #endif
 418         mb->gf_active_ptr            = x->gf_active_ptr;
 419
 420         mb->vector_range             = 32;
 421
 422         vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
 423         mbr_ei[i].totalrate = 0;
 424
 425         mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
 426
 427         mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
 428         mbd->mode_info_stride  = cm->mode_info_stride;
 429
 430         mbd->frame_type = cm->frame_type;
 431
 432         mbd->frames_since_golden = cm->frames_since_golden;
 433         mbd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
 434
 435         mb->src = * cpi->Source;
 436         mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
 437         mbd->dst = cm->yv12_fb[cm->new_fb_idx];
 438
 439         mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
 440         mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
 441         mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
 442
 443         vp8_build_block_offsets(mb);
 444
 445         vp8_setup_block_dptrs(mbd);
 446
 447         vp8_setup_block_ptrs(mb);
 448
 449         mb->activity_sum = 0;
 450
 451         mbd->left_context = &cm->left_context;
 452         mb->mvc = cm->fc.mvc;
 453
 454         setup_mbby_copy(&mbr_ei[i].mb, x);
 455
 456     }
 457 }
 458
 459 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
 460 {
 461     cpi->b_multi_threaded = 0;
 462
 463     cpi->processor_core_count = 32; //vp8_get_proc_core_count();
 464
 465     if (cpi->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
 466     {
 467         int ithread;
 468
 469         if (cpi->oxcf.multi_threaded > cpi->processor_core_count)
 470             cpi->encoding_thread_count = cpi->processor_core_count - 1;
 471         else
 472             cpi->encoding_thread_count = cpi->oxcf.multi_threaded - 1;
 473
 474         CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * cpi->encoding_thread_count));
 475         CHECK_MEM_ERROR(cpi->h_event_start_encoding, vpx_malloc(sizeof(sem_t) * cpi->encoding_thread_count));
 476         CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count));
 477         vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count);
 478         CHECK_MEM_ERROR(cpi->en_thread_data, vpx_malloc(sizeof(ENCODETHREAD_DATA) * cpi->encoding_thread_count));
 479         CHECK_MEM_ERROR(cpi->mt_current_mb_col, vpx_malloc(sizeof(*cpi->mt_current_mb_col) * cpi->common.mb_rows));
 480
 481         //cpi->h_event_main = CreateEvent(NULL, FALSE, FALSE, NULL);
 482         sem_init(&cpi->h_event_end_encoding, 0, 0);
 483
 484         cpi->b_multi_threaded = 1;
 485
 486         //printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n", (cpi->encoding_thread_count +1));
 487
 488         for (ithread = 0; ithread < cpi->encoding_thread_count; ithread++)
 489         {
 490             ENCODETHREAD_DATA * ethd = &cpi->en_thread_data[ithread];
 491
 492             //cpi->h_event_mbrencoding[ithread] = CreateEvent(NULL, FALSE, FALSE, NULL);
 493             sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
 494             ethd->ithread = ithread;
 495             ethd->ptr1 = (void *)cpi;
 496             ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
 497
 498             //printf(" call begin thread %d \n", ithread);
 499
 500             //cpi->h_encoding_thread[ithread] =   (HANDLE)_beginthreadex(
 501             //  NULL,           // security
 502             //  0,              // stksize
 503             //  thread_encoding_proc,
 504             //  (&cpi->en_thread_data[ithread]),          // Thread data
 505             //  0,
 506             //  NULL);
 507
 508             pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, ethd);
 509         }
 510
 511         {
 512             LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
 513
 514             sem_init(&cpi->h_event_start_lpf, 0, 0);
 515             sem_init(&cpi->h_event_end_lpf, 0, 0);
 516
 517             lpfthd->ptr1 = (void *)cpi;
 518             pthread_create(&cpi->h_filter_thread, 0, loopfilter_thread, lpfthd);
 519         }
 520     }
 521
 522 }
 523
 524 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
 525 {
 526     if (cpi->b_multi_threaded)
 527     {
 528         //shutdown other threads
 529         cpi->b_multi_threaded = 0;
 530         {
 531             int i;
 532
 533             for (i = 0; i < cpi->encoding_thread_count; i++)
 534             {
 535                 //SetEvent(cpi->h_event_mbrencoding[i]);
 536                 sem_post(&cpi->h_event_start_encoding[i]);
 537                 pthread_join(cpi->h_encoding_thread[i], 0);
 538
 539                 sem_destroy(&cpi->h_event_start_encoding[i]);
 540             }
 541
 542             sem_post(&cpi->h_event_start_lpf);
 543             pthread_join(cpi->h_filter_thread, 0);
 544         }
 545
 546         sem_destroy(&cpi->h_event_end_encoding);
 547         sem_destroy(&cpi->h_event_end_lpf);
 548         sem_destroy(&cpi->h_event_start_lpf);
 549
 550         //free thread related resources
 551         vpx_free(cpi->h_event_start_encoding);
 552         vpx_free(cpi->h_encoding_thread);
 553         vpx_free(cpi->mb_row_ei);
 554         vpx_free(cpi->en_thread_data);
 555         vpx_free(cpi->mt_current_mb_col);
 556     }
 557 }
 558 #endif