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
 118                 // for each macroblock col in image
 119                 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
 120                 {
 121                     if ((mb_col & (nsync - 1)) == 0)
 122                     {
 123                         while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != cm->mb_cols - 1)
 124                         {
 125                             x86_pause_hint();
 126                             thread_sleep(0);
 127                         }
 128                     }
 129
 130                     // Distance of Mb to the various image edges.
 131                     // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
 132                     xd->mb_to_left_edge = -((mb_col * 16) << 3);
 133                     xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
 134                     xd->mb_to_top_edge = -((mb_row * 16) << 3);
 135                     xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
 136
 137                     // Set up limit values for motion vectors used to prevent them extending outside the UMV borders
 138                     x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
 139                     x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
 140                     x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
 141                     x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
 142
 143                     xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 144                     xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 145                     xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 146                     xd->left_available = (mb_col != 0);
 147
 148                     x->rddiv = cpi->RDDIV;
 149                     x->rdmult = cpi->RDMULT;
 150
 151                     if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
 152                         vp8_activity_masking(cpi, x);
 153
 154                     // Is segmentation enabled
 155                     // MB level adjutment to quantizer
 156                     if (xd->segmentation_enabled)
 157                     {
 158                         // Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
 159                         if (cpi->segmentation_map[map_index + mb_col] <= 3)
 160                             xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index + mb_col];
 161                         else
 162                             xd->mode_info_context->mbmi.segment_id = 0;
 163
 164                         vp8cx_mb_init_quantizer(cpi, x);
 165                     }
 166                     else
 167                         xd->mode_info_context->mbmi.segment_id = 0; // Set to Segment 0 by default
 168
 169                     x->active_ptr = cpi->active_map + map_index + mb_col;
 170
 171                     if (cm->frame_type == KEY_FRAME)
 172                     {
 173                         *totalrate += vp8cx_encode_intra_macro_block(cpi, x, &tp);
 174 #ifdef MODE_STATS
 175                         y_modes[xd->mbmi.mode] ++;
 176 #endif
 177                     }
 178                     else
 179                     {
 180                         *totalrate += vp8cx_encode_inter_macroblock(cpi, x, &tp, recon_yoffset, recon_uvoffset);
 181
 182 #ifdef MODE_STATS
 183                         inter_y_modes[xd->mbmi.mode] ++;
 184
 185                         if (xd->mbmi.mode == SPLITMV)
 186                         {
 187                             int b;
 188
 189                             for (b = 0; b < xd->mbmi.partition_count; b++)
 190                             {
 191                                 inter_b_modes[x->partition->bmi[b].mode] ++;
 192                             }
 193                         }
 194
 195 #endif
 196
 197                         // Count of last ref frame 0,0 useage
 198                         if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
 199                             cpi->inter_zz_count++;
 200
 201                         // Special case code for cyclic refresh
 202                         // If cyclic update enabled then copy xd->mbmi.segment_id; (which may have been updated based on mode
 203                         // during vp8cx_encode_inter_macroblock()) back into the global sgmentation map
 204                         if (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)
 205                         {
 206                             const MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
 207                             cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
 208
 209                             // 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):
 210                             // 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)
 211                             // else mark it as dirty (1).
 212                             if (mbmi->segment_id)
 213                                 cpi->cyclic_refresh_map[map_index + mb_col] = -1;
 214                             else if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
 215                             {
 216                                 if (cpi->cyclic_refresh_map[map_index + mb_col] == 1)
 217                                     cpi->cyclic_refresh_map[map_index + mb_col] = 0;
 218                             }
 219                             else
 220                                 cpi->cyclic_refresh_map[map_index + mb_col] = 1;
 221
 222                         }
 223                     }
 224                     cpi->tplist[mb_row].stop = tp;
 225
 226                     // Increment pointer into gf useage flags structure.
 227                     x->gf_active_ptr++;
 228
 229                     // Increment the activity mask pointers.
 230                     x->mb_activity_ptr++;
 231
 232                     /* save the block info */
 233                     for (i = 0; i < 16; i++)
 234                         xd->mode_info_context->bmi[i] = xd->block[i].bmi;
 235
 236                     // adjust to the next column of macroblocks
 237                     x->src.y_buffer += 16;
 238                     x->src.u_buffer += 8;
 239                     x->src.v_buffer += 8;
 240
 241                     recon_yoffset += 16;
 242                     recon_uvoffset += 8;
 243
 244                     // Keep track of segment useage
 245                     segment_counts[xd->mode_info_context->mbmi.segment_id]++;
 246
 247                     // skip to next mb
 248                     xd->mode_info_context++;
 249                     x->partition_info++;
 250                     xd->above_context++;
 251
 252                     cpi->mt_current_mb_col[mb_row] = mb_col;
 253                 }
 254
 255                 //extend the recon for intra prediction
 256                 vp8_extend_mb_row(
 257                     &cm->yv12_fb[dst_fb_idx],
 258                     xd->dst.y_buffer + 16,
 259                     xd->dst.u_buffer + 8,
 260                     xd->dst.v_buffer + 8);
 261
 262                 // this is to account for the border
 263                 xd->mode_info_context++;
 264                 x->partition_info++;
 265
 266                 x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
 267                 x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 268                 x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 269
 270                 xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
 271                 x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
 272                 x->gf_active_ptr   += cm->mb_cols * cpi->encoding_thread_count;
 273
 274                 if (mb_row == cm->mb_rows - 1)
 275                 {
 276                     //SetEvent(cpi->h_event_main);
 277                     sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
 278                 }
 279             }
 280         }
 281     }
 282
 283     //printf("exit thread %d\n", ithread);
 284     return 0;
 285 }
 286
 287 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
 288 {
 289
 290     MACROBLOCK *x = mbsrc;
 291     MACROBLOCK *z = mbdst;
 292     int i;
 293
 294     z->ss               = x->ss;
 295     z->ss_count          = x->ss_count;
 296     z->searches_per_step  = x->searches_per_step;
 297     z->errorperbit      = x->errorperbit;
 298
 299     z->sadperbit16      = x->sadperbit16;
 300     z->sadperbit4       = x->sadperbit4;
 301
 302     /*
 303     z->mv_col_min    = x->mv_col_min;
 304     z->mv_col_max    = x->mv_col_max;
 305     z->mv_row_min    = x->mv_row_min;
 306     z->mv_row_max    = x->mv_row_max;
 307     z->vector_range = x->vector_range ;
 308     */
 309
 310     z->vp8_short_fdct4x4     = x->vp8_short_fdct4x4;
 311     z->vp8_short_fdct8x4     = x->vp8_short_fdct8x4;
 312     z->short_walsh4x4    = x->short_walsh4x4;
 313     z->quantize_b        = x->quantize_b;
 314     z->quantize_b_pair   = x->quantize_b_pair;
 315     z->optimize          = x->optimize;
 316
 317     /*
 318     z->mvc              = x->mvc;
 319     z->src.y_buffer      = x->src.y_buffer;
 320     z->src.u_buffer      = x->src.u_buffer;
 321     z->src.v_buffer      = x->src.v_buffer;
 322     */
 323
 324
 325     vpx_memcpy(z->mvcosts,          x->mvcosts,         sizeof(x->mvcosts));
 326     z->mvcost[0] = &z->mvcosts[0][mv_max+1];
 327     z->mvcost[1] = &z->mvcosts[1][mv_max+1];
 328     z->mvsadcost[0] = &z->mvsadcosts[0][mvfp_max+1];
 329     z->mvsadcost[1] = &z->mvsadcosts[1][mvfp_max+1];
 330
 331
 332     vpx_memcpy(z->token_costs,       x->token_costs,      sizeof(x->token_costs));
 333     vpx_memcpy(z->inter_bmode_costs,  x->inter_bmode_costs, sizeof(x->inter_bmode_costs));
 334     //memcpy(z->mvcosts,            x->mvcosts,         sizeof(x->mvcosts));
 335     //memcpy(z->mvcost,         x->mvcost,          sizeof(x->mvcost));
 336     vpx_memcpy(z->mbmode_cost,       x->mbmode_cost,      sizeof(x->mbmode_cost));
 337     vpx_memcpy(z->intra_uv_mode_cost,  x->intra_uv_mode_cost, sizeof(x->intra_uv_mode_cost));
 338     vpx_memcpy(z->bmode_costs,       x->bmode_costs,      sizeof(x->bmode_costs));
 339
 340     for (i = 0; i < 25; i++)
 341     {
 342         z->block[i].quant           = x->block[i].quant;
 343         z->block[i].quant_fast      = x->block[i].quant_fast;
 344         z->block[i].quant_shift     = x->block[i].quant_shift;
 345         z->block[i].zbin            = x->block[i].zbin;
 346         z->block[i].zrun_zbin_boost   = x->block[i].zrun_zbin_boost;
 347         z->block[i].round           = x->block[i].round;
 348         /*
 349         z->block[i].src             = x->block[i].src;
 350         */
 351         z->block[i].src_stride       = x->block[i].src_stride;
 352         z->block[i].force_empty      = x->block[i].force_empty;
 353
 354     }
 355
 356     {
 357         MACROBLOCKD *xd = &x->e_mbd;
 358         MACROBLOCKD *zd = &z->e_mbd;
 359
 360         /*
 361         zd->mode_info_context = xd->mode_info_context;
 362         zd->mode_info        = xd->mode_info;
 363
 364         zd->mode_info_stride  = xd->mode_info_stride;
 365         zd->frame_type       = xd->frame_type;
 366         zd->up_available     = xd->up_available   ;
 367         zd->left_available   = xd->left_available;
 368         zd->left_context     = xd->left_context;
 369         zd->last_frame_dc     = xd->last_frame_dc;
 370         zd->last_frame_dccons = xd->last_frame_dccons;
 371         zd->gold_frame_dc     = xd->gold_frame_dc;
 372         zd->gold_frame_dccons = xd->gold_frame_dccons;
 373         zd->mb_to_left_edge    = xd->mb_to_left_edge;
 374         zd->mb_to_right_edge   = xd->mb_to_right_edge;
 375         zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
 376         zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
 377         zd->gf_active_ptr     = xd->gf_active_ptr;
 378         zd->frames_since_golden       = xd->frames_since_golden;
 379         zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
 380         */
 381         zd->subpixel_predict         = xd->subpixel_predict;
 382         zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
 383         zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
 384         zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
 385         zd->segmentation_enabled     = xd->segmentation_enabled;
 386         zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
 387         vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
 388
 389         for (i = 0; i < 25; i++)
 390         {
 391             zd->block[i].dequant = xd->block[i].dequant;
 392         }
 393     }
 394 }
 395
 396 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
 397                                MACROBLOCK *x,
 398                                MB_ROW_COMP *mbr_ei,
 399                                int mb_row,
 400                                int count
 401                               )
 402 {
 403
 404     VP8_COMMON *const cm = & cpi->common;
 405     MACROBLOCKD *const xd = & x->e_mbd;
 406     int i;
 407     (void) mb_row;
 408
 409     for (i = 0; i < count; i++)
 410     {
 411         MACROBLOCK *mb = & mbr_ei[i].mb;
 412         MACROBLOCKD *mbd = &mb->e_mbd;
 413
 414         mbd->subpixel_predict        = xd->subpixel_predict;
 415         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
 416         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
 417         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
 418 #if CONFIG_RUNTIME_CPU_DETECT
 419         mbd->rtcd                   = xd->rtcd;
 420 #endif
 421         mb->gf_active_ptr            = x->gf_active_ptr;
 422
 423         mb->vector_range             = 32;
 424
 425         vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
 426         mbr_ei[i].totalrate = 0;
 427
 428         mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
 429
 430         mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
 431         mbd->mode_info_stride  = cm->mode_info_stride;
 432
 433         mbd->frame_type = cm->frame_type;
 434
 435         mbd->frames_since_golden = cm->frames_since_golden;
 436         mbd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
 437
 438         mb->src = * cpi->Source;
 439         mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
 440         mbd->dst = cm->yv12_fb[cm->new_fb_idx];
 441
 442         mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
 443         mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
 444         mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
 445
 446         vp8_build_block_offsets(mb);
 447
 448         vp8_setup_block_dptrs(mbd);
 449
 450         vp8_setup_block_ptrs(mb);
 451
 452         mbd->left_context = &cm->left_context;
 453         mb->mvc = cm->fc.mvc;
 454
 455         setup_mbby_copy(&mbr_ei[i].mb, x);
 456
 457     }
 458 }
 459
 460 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
 461 {
 462     const VP8_COMMON * cm = &cpi->common;
 463
 464     cpi->b_multi_threaded = 0;
 465     cpi->encoding_thread_count = 0;
 466
 467     if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
 468     {
 469         int ithread;
 470         int th_count = cpi->oxcf.multi_threaded - 1;
 471
 472         /* don't allocate more threads than cores available */
 473         if (cpi->oxcf.multi_threaded > cm->processor_core_count)
 474             th_count = cm->processor_core_count - 1;
 475
 476         /* we have th_count + 1 (main) threads processing one row each */
 477         /* no point to have more threads than the sync range allows */
 478         if(th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1))
 479         {
 480             th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
 481         }
 482
 483         if(th_count == 0)
 484             return;
 485
 486         CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * th_count));
 487         CHECK_MEM_ERROR(cpi->h_event_start_encoding, vpx_malloc(sizeof(sem_t) * th_count));
 488         CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
 489         vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
 490         CHECK_MEM_ERROR(cpi->en_thread_data,
 491                         vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
 492         CHECK_MEM_ERROR(cpi->mt_current_mb_col,
 493                         vpx_malloc(sizeof(*cpi->mt_current_mb_col) * cm->mb_rows));
 494
 495         sem_init(&cpi->h_event_end_encoding, 0, 0);
 496
 497         cpi->b_multi_threaded = 1;
 498         cpi->encoding_thread_count = th_count;
 499
 500         /*
 501         printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
 502                (cpi->encoding_thread_count +1));
 503         */
 504
 505         for (ithread = 0; ithread < th_count; ithread++)
 506         {
 507             ENCODETHREAD_DATA * ethd = &cpi->en_thread_data[ithread];
 508
 509             sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
 510             ethd->ithread = ithread;
 511             ethd->ptr1 = (void *)cpi;
 512             ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
 513
 514             pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, ethd);
 515         }
 516
 517         {
 518             LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
 519
 520             sem_init(&cpi->h_event_start_lpf, 0, 0);
 521             sem_init(&cpi->h_event_end_lpf, 0, 0);
 522
 523             lpfthd->ptr1 = (void *)cpi;
 524             pthread_create(&cpi->h_filter_thread, 0, loopfilter_thread, lpfthd);
 525         }
 526     }
 527
 528 }
 529
 530 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
 531 {
 532     if (cpi->b_multi_threaded)
 533     {
 534         //shutdown other threads
 535         cpi->b_multi_threaded = 0;
 536         {
 537             int i;
 538
 539             for (i = 0; i < cpi->encoding_thread_count; i++)
 540             {
 541                 //SetEvent(cpi->h_event_mbrencoding[i]);
 542                 sem_post(&cpi->h_event_start_encoding[i]);
 543                 pthread_join(cpi->h_encoding_thread[i], 0);
 544
 545                 sem_destroy(&cpi->h_event_start_encoding[i]);
 546             }
 547
 548             sem_post(&cpi->h_event_start_lpf);
 549             pthread_join(cpi->h_filter_thread, 0);
 550         }
 551
 552         sem_destroy(&cpi->h_event_end_encoding);
 553         sem_destroy(&cpi->h_event_end_lpf);
 554         sem_destroy(&cpi->h_event_start_lpf);
 555
 556         //free thread related resources
 557         vpx_free(cpi->h_event_start_encoding);
 558         vpx_free(cpi->h_encoding_thread);
 559         vpx_free(cpi->mb_row_ei);
 560         vpx_free(cpi->en_thread_data);
 561         vpx_free(cpi->mt_current_mb_col);
 562     }
 563 }
 564 #endif