vp8/decoder/threading.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
  12 #if !defined(WIN32) && CONFIG_OS_SUPPORT == 1
  13 # include <unistd.h>
  14 #endif
  15 #include "onyxd_int.h"
  16 #include "vpx_mem/vpx_mem.h"
  17 #include "vp8/common/threading.h"
  18
  19 #include "vp8/common/loopfilter.h"
  20 #include "vp8/common/extend.h"
  21 #include "vpx_ports/vpx_timer.h"
  22 #include "detokenize.h"
  23 #include "vp8/common/reconinter.h"
  24 #include "reconintra_mt.h"
  25
  26 extern void mb_init_dequantizer(VP8D_COMP *pbi, MACROBLOCKD *xd);
  27 extern void clamp_mvs(MACROBLOCKD *xd);
  28 extern void vp8_build_uvmvs(MACROBLOCKD *x, int fullpixel);
  29
  30 #if CONFIG_RUNTIME_CPU_DETECT
  31 #define RTCD_VTABLE(x) (&(pbi)->common.rtcd.x)
  32 #else
  33 #define RTCD_VTABLE(x) NULL
  34 #endif
  35
  36 static void setup_decoding_thread_data(VP8D_COMP *pbi, MACROBLOCKD *xd, MB_ROW_DEC *mbrd, int count)
  37 {
  38     VP8_COMMON *const pc = & pbi->common;
  39     int i, j;
  40
  41     for (i = 0; i < count; i++)
  42     {
  43         MACROBLOCKD *mbd = &mbrd[i].mbd;
  44 #if CONFIG_RUNTIME_CPU_DETECT
  45         mbd->rtcd = xd->rtcd;
  46 #endif
  47         mbd->subpixel_predict        = xd->subpixel_predict;
  48         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
  49         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
  50         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
  51
  52         mbd->mode_info_context = pc->mi   + pc->mode_info_stride * (i + 1);
  53         mbd->mode_info_stride  = pc->mode_info_stride;
  54
  55         mbd->frame_type = pc->frame_type;
  56         mbd->frames_since_golden      = pc->frames_since_golden;
  57         mbd->frames_till_alt_ref_frame  = pc->frames_till_alt_ref_frame;
  58
  59         mbd->pre = pc->yv12_fb[pc->lst_fb_idx];
  60         mbd->dst = pc->yv12_fb[pc->new_fb_idx];
  61
  62         vp8_setup_block_dptrs(mbd);
  63         vp8_build_block_doffsets(mbd);
  64         mbd->segmentation_enabled    = xd->segmentation_enabled;
  65         mbd->mb_segement_abs_delta     = xd->mb_segement_abs_delta;
  66         vpx_memcpy(mbd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
  67
  68         /*signed char ref_lf_deltas[MAX_REF_LF_DELTAS];*/
  69         vpx_memcpy(mbd->ref_lf_deltas, xd->ref_lf_deltas, sizeof(xd->ref_lf_deltas));
  70         /*signed char mode_lf_deltas[MAX_MODE_LF_DELTAS];*/
  71         vpx_memcpy(mbd->mode_lf_deltas, xd->mode_lf_deltas, sizeof(xd->mode_lf_deltas));
  72         /*unsigned char mode_ref_lf_delta_enabled;
  73         unsigned char mode_ref_lf_delta_update;*/
  74         mbd->mode_ref_lf_delta_enabled    = xd->mode_ref_lf_delta_enabled;
  75         mbd->mode_ref_lf_delta_update    = xd->mode_ref_lf_delta_update;
  76
  77         mbd->current_bc = &pbi->bc2;
  78
  79         for (j = 0; j < 25; j++)
  80         {
  81             mbd->block[j].dequant = xd->block[j].dequant;
  82         }
  83     }
  84
  85     for (i=0; i< pc->mb_rows; i++)
  86         pbi->mt_current_mb_col[i]=-1;
  87 }
  88
  89
  90 static void decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd, int mb_row, int mb_col)
  91 {
  92     int eobtotal = 0;
  93     int i, do_clamp = xd->mode_info_context->mbmi.need_to_clamp_mvs;
  94
  95     if (xd->mode_info_context->mbmi.mb_skip_coeff)
  96     {
  97         vp8_reset_mb_tokens_context(xd);
  98     }
  99     else
 100     {
 101         eobtotal = vp8_decode_mb_tokens(pbi, xd);
 102     }
 103
 104     /* Perform temporary clamping of the MV to be used for prediction */
 105     if (do_clamp)
 106     {
 107         clamp_mvs(xd);
 108     }
 109
 110     eobtotal |= (xd->mode_info_context->mbmi.mode == B_PRED ||
 111                   xd->mode_info_context->mbmi.mode == SPLITMV);
 112     if (!eobtotal)
 113     {
 114         /* Special case:  Force the loopfilter to skip when eobtotal and
 115          * mb_skip_coeff are zero.
 116          * */
 117         xd->mode_info_context->mbmi.mb_skip_coeff = 1;
 118
 119         /*mt_skip_recon_mb(pbi, xd, mb_row, mb_col);*/
 120         if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
 121         {
 122             vp8mt_build_intra_predictors_mbuv_s(pbi, xd, mb_row, mb_col);
 123             vp8mt_build_intra_predictors_mby_s(pbi, xd, mb_row, mb_col);
 124         }
 125         else
 126         {
 127             vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer,
 128                                                xd->dst.u_buffer, xd->dst.v_buffer,
 129                                                xd->dst.y_stride, xd->dst.uv_stride);
 130         }
 131         return;
 132     }
 133
 134     if (xd->segmentation_enabled)
 135         mb_init_dequantizer(pbi, xd);
 136
 137     /* do prediction */
 138     if (xd->frame_type == KEY_FRAME  ||  xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
 139     {
 140         vp8mt_build_intra_predictors_mbuv(pbi, xd, mb_row, mb_col);
 141
 142         if (xd->mode_info_context->mbmi.mode != B_PRED)
 143         {
 144             vp8mt_build_intra_predictors_mby(pbi, xd, mb_row, mb_col);
 145         } else {
 146             vp8mt_intra_prediction_down_copy(pbi, xd, mb_row, mb_col);
 147         }
 148     }
 149     else
 150     {
 151         vp8_build_inter_predictors_mb(xd);
 152     }
 153
 154     /* dequantization and idct */
 155     if (xd->mode_info_context->mbmi.mode != B_PRED && xd->mode_info_context->mbmi.mode != SPLITMV)
 156     {
 157         BLOCKD *b = &xd->block[24];
 158         DEQUANT_INVOKE(&pbi->dequant, block)(b);
 159
 160         /* do 2nd order transform on the dc block */
 161         if (xd->eobs[24] > 1)
 162         {
 163             IDCT_INVOKE(RTCD_VTABLE(idct), iwalsh16)(&b->dqcoeff[0], b->diff);
 164             ((int *)b->qcoeff)[0] = 0;
 165             ((int *)b->qcoeff)[1] = 0;
 166             ((int *)b->qcoeff)[2] = 0;
 167             ((int *)b->qcoeff)[3] = 0;
 168             ((int *)b->qcoeff)[4] = 0;
 169             ((int *)b->qcoeff)[5] = 0;
 170             ((int *)b->qcoeff)[6] = 0;
 171             ((int *)b->qcoeff)[7] = 0;
 172         }
 173         else
 174         {
 175             IDCT_INVOKE(RTCD_VTABLE(idct), iwalsh1)(&b->dqcoeff[0], b->diff);
 176             ((int *)b->qcoeff)[0] = 0;
 177         }
 178
 179         DEQUANT_INVOKE (&pbi->dequant, dc_idct_add_y_block)
 180                         (xd->qcoeff, xd->block[0].dequant,
 181                          xd->predictor, xd->dst.y_buffer,
 182                          xd->dst.y_stride, xd->eobs, xd->block[24].diff);
 183     }
 184     else if ((xd->frame_type == KEY_FRAME  ||  xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) && xd->mode_info_context->mbmi.mode == B_PRED)
 185     {
 186         for (i = 0; i < 16; i++)
 187         {
 188             BLOCKD *b = &xd->block[i];
 189             vp8mt_predict_intra4x4(pbi, xd, b->bmi.mode, b->predictor, mb_row, mb_col, i);
 190             if (xd->eobs[i] > 1)
 191             {
 192                 DEQUANT_INVOKE(&pbi->dequant, idct_add)
 193                     (b->qcoeff, b->dequant,  b->predictor,
 194                     *(b->base_dst) + b->dst, 16, b->dst_stride);
 195             }
 196             else
 197             {
 198                 IDCT_INVOKE(RTCD_VTABLE(idct), idct1_scalar_add)
 199                     (b->qcoeff[0] * b->dequant[0], b->predictor,
 200                     *(b->base_dst) + b->dst, 16, b->dst_stride);
 201                 ((int *)b->qcoeff)[0] = 0;
 202             }
 203         }
 204     }
 205     else
 206     {
 207         DEQUANT_INVOKE (&pbi->dequant, idct_add_y_block)
 208                         (xd->qcoeff, xd->block[0].dequant,
 209                          xd->predictor, xd->dst.y_buffer,
 210                          xd->dst.y_stride, xd->eobs);
 211     }
 212
 213     DEQUANT_INVOKE (&pbi->dequant, idct_add_uv_block)
 214                     (xd->qcoeff+16*16, xd->block[16].dequant,
 215                      xd->predictor+16*16, xd->dst.u_buffer, xd->dst.v_buffer,
 216                      xd->dst.uv_stride, xd->eobs+16);
 217 }
 218
 219
 220 static THREAD_FUNCTION thread_decoding_proc(void *p_data)
 221 {
 222     int ithread = ((DECODETHREAD_DATA *)p_data)->ithread;
 223     VP8D_COMP *pbi = (VP8D_COMP *)(((DECODETHREAD_DATA *)p_data)->ptr1);
 224     MB_ROW_DEC *mbrd = (MB_ROW_DEC *)(((DECODETHREAD_DATA *)p_data)->ptr2);
 225     ENTROPY_CONTEXT_PLANES mb_row_left_context;
 226
 227     while (1)
 228     {
 229         if (pbi->b_multithreaded_rd == 0)
 230             break;
 231
 232         /*if(WaitForSingleObject(pbi->h_event_start_decoding[ithread], INFINITE) == WAIT_OBJECT_0)*/
 233         if (sem_wait(&pbi->h_event_start_decoding[ithread]) == 0)
 234         {
 235             if (pbi->b_multithreaded_rd == 0)
 236                 break;
 237             else
 238             {
 239                 VP8_COMMON *pc = &pbi->common;
 240                 MACROBLOCKD *xd = &mbrd->mbd;
 241
 242                 int mb_row;
 243                 int num_part = 1 << pbi->common.multi_token_partition;
 244                 volatile int *last_row_current_mb_col;
 245                 int nsync = pbi->sync_range;
 246
 247                 for (mb_row = ithread+1; mb_row < pc->mb_rows; mb_row += (pbi->decoding_thread_count + 1))
 248                 {
 249                     int i;
 250                     int recon_yoffset, recon_uvoffset;
 251                     int mb_col;
 252                     int ref_fb_idx = pc->lst_fb_idx;
 253                     int dst_fb_idx = pc->new_fb_idx;
 254                     int recon_y_stride = pc->yv12_fb[ref_fb_idx].y_stride;
 255                     int recon_uv_stride = pc->yv12_fb[ref_fb_idx].uv_stride;
 256
 257                     int filter_level;
 258                     loop_filter_info *lfi = pc->lf_info;
 259                     int alt_flt_enabled = xd->segmentation_enabled;
 260                     int Segment;
 261
 262                     pbi->mb_row_di[ithread].mb_row = mb_row;
 263                     pbi->mb_row_di[ithread].mbd.current_bc =  &pbi->mbc[mb_row%num_part];
 264
 265                     last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row -1];
 266
 267                     recon_yoffset = mb_row * recon_y_stride * 16;
 268                     recon_uvoffset = mb_row * recon_uv_stride * 8;
 269                     /* reset above block coeffs */
 270
 271                     xd->above_context = pc->above_context;
 272                     xd->left_context = &mb_row_left_context;
 273                     vpx_memset(&mb_row_left_context, 0, sizeof(mb_row_left_context));
 274                     xd->up_available = (mb_row != 0);
 275
 276                     xd->mb_to_top_edge = -((mb_row * 16)) << 3;
 277                     xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
 278
 279                     for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
 280                     {
 281                         if ((mb_col & (nsync-1)) == 0)
 282                         {
 283                             while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != pc->mb_cols - 1)
 284                             {
 285                                 x86_pause_hint();
 286                                 thread_sleep(0);
 287                             }
 288                         }
 289
 290                         update_blockd_bmi(xd);
 291
 292                         /* Distance of Mb to the various image edges.
 293                          * These are specified to 8th pel as they are always compared to values that are in 1/8th pel units
 294                          */
 295                         xd->mb_to_left_edge = -((mb_col * 16) << 3);
 296                         xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
 297
 298                         xd->dst.y_buffer = pc->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 299                         xd->dst.u_buffer = pc->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 300                         xd->dst.v_buffer = pc->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 301
 302                         xd->left_available = (mb_col != 0);
 303
 304                         /* Select the appropriate reference frame for this MB */
 305                         if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
 306                             ref_fb_idx = pc->lst_fb_idx;
 307                         else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
 308                             ref_fb_idx = pc->gld_fb_idx;
 309                         else
 310                             ref_fb_idx = pc->alt_fb_idx;
 311
 312                         xd->pre.y_buffer = pc->yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
 313                         xd->pre.u_buffer = pc->yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
 314                         xd->pre.v_buffer = pc->yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
 315
 316                         vp8_build_uvmvs(xd, pc->full_pixel);
 317                         decode_macroblock(pbi, xd, mb_row, mb_col);
 318
 319                         if (pbi->common.filter_level)
 320                         {
 321                             int skip_lf;
 322                             if( mb_row != pc->mb_rows-1 )
 323                             {
 324                                 /* Save decoded MB last row data for next-row decoding */
 325                                 vpx_memcpy((pbi->mt_yabove_row[mb_row + 1] + 32 + mb_col*16), (xd->dst.y_buffer + 15 * recon_y_stride), 16);
 326                                 vpx_memcpy((pbi->mt_uabove_row[mb_row + 1] + 16 + mb_col*8), (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
 327                                 vpx_memcpy((pbi->mt_vabove_row[mb_row + 1] + 16 + mb_col*8), (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
 328                             }
 329
 330                             /* save left_col for next MB decoding */
 331                             if(mb_col != pc->mb_cols-1)
 332                             {
 333                                 MODE_INFO *next = xd->mode_info_context +1;
 334
 335                                 if (xd->frame_type == KEY_FRAME  ||  next->mbmi.ref_frame == INTRA_FRAME)
 336                                 {
 337                                     for (i = 0; i < 16; i++)
 338                                         pbi->mt_yleft_col[mb_row][i] = xd->dst.y_buffer [i* recon_y_stride + 15];
 339                                     for (i = 0; i < 8; i++)
 340                                     {
 341                                         pbi->mt_uleft_col[mb_row][i] = xd->dst.u_buffer [i* recon_uv_stride + 7];
 342                                         pbi->mt_vleft_col[mb_row][i] = xd->dst.v_buffer [i* recon_uv_stride + 7];
 343                                     }
 344                                 }
 345                             }
 346
 347                             /* update loopfilter info */
 348                             Segment = (alt_flt_enabled) ? xd->mode_info_context->mbmi.segment_id : 0;
 349                             skip_lf = (xd->mode_info_context->mbmi.mode != B_PRED &&
 350                                             xd->mode_info_context->mbmi.mode != SPLITMV &&
 351                                             xd->mode_info_context->mbmi.mb_skip_coeff);
 352
 353                             filter_level = pbi->mt_baseline_filter_level[Segment];
 354                             /* Distance of Mb to the various image edges.
 355                              * These are specified to 8th pel as they are always compared to values that are in 1/8th pel units
 356                              * Apply any context driven MB level adjustment
 357                              */
 358                             filter_level = vp8_adjust_mb_lf_value(xd, filter_level);
 359
 360                             /* loopfilter on this macroblock. */
 361                             if (filter_level)
 362                             {
 363                                 if (mb_col > 0)
 364                                     pc->lf_mbv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 365
 366                                 if (!skip_lf)
 367                                     pc->lf_bv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 368
 369                                 /* don't apply across umv border */
 370                                 if (mb_row > 0)
 371                                     pc->lf_mbh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 372
 373                                 if (!skip_lf)
 374                                     pc->lf_bh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 375                             }
 376                         }
 377
 378                         recon_yoffset += 16;
 379                         recon_uvoffset += 8;
 380
 381                         ++xd->mode_info_context;  /* next mb */
 382
 383                         xd->above_context++;
 384
 385                         /*pbi->mb_row_di[ithread].current_mb_col = mb_col;*/
 386                         pbi->mt_current_mb_col[mb_row] = mb_col;
 387                     }
 388
 389                     /* adjust to the next row of mbs */
 390                     if (pbi->common.filter_level)
 391                     {
 392                         if(mb_row != pc->mb_rows-1)
 393                         {
 394                             int lasty = pc->yv12_fb[ref_fb_idx].y_width + VP8BORDERINPIXELS;
 395                             int lastuv = (pc->yv12_fb[ref_fb_idx].y_width>>1) + (VP8BORDERINPIXELS>>1);
 396
 397                             for (i = 0; i < 4; i++)
 398                             {
 399                                 pbi->mt_yabove_row[mb_row +1][lasty + i] = pbi->mt_yabove_row[mb_row +1][lasty -1];
 400                                 pbi->mt_uabove_row[mb_row +1][lastuv + i] = pbi->mt_uabove_row[mb_row +1][lastuv -1];
 401                                 pbi->mt_vabove_row[mb_row +1][lastuv + i] = pbi->mt_vabove_row[mb_row +1][lastuv -1];
 402                             }
 403                         }
 404                     } else
 405                         vp8_extend_mb_row(&pc->yv12_fb[dst_fb_idx], xd->dst.y_buffer + 16, xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
 406
 407                     ++xd->mode_info_context;      /* skip prediction column */
 408
 409                     /* since we have multithread */
 410                     xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
 411                 }
 412             }
 413         }
 414         /*  add this to each frame */
 415         if ((mbrd->mb_row == pbi->common.mb_rows-1) || ((mbrd->mb_row == pbi->common.mb_rows-2) && (pbi->common.mb_rows % (pbi->decoding_thread_count+1))==1))
 416         {
 417             /*SetEvent(pbi->h_event_end_decoding);*/
 418             sem_post(&pbi->h_event_end_decoding);
 419         }
 420     }
 421
 422     return 0 ;
 423 }
 424
 425
 426 void vp8_decoder_create_threads(VP8D_COMP *pbi)
 427 {
 428     int core_count = 0;
 429     int ithread;
 430
 431     pbi->b_multithreaded_rd = 0;
 432     pbi->allocated_decoding_thread_count = 0;
 433
 434     /* limit decoding threads to the max number of token partitions */
 435     core_count = (pbi->max_threads > 8) ? 8 : pbi->max_threads;
 436
 437     /* limit decoding threads to the available cores */
 438     if (core_count > pbi->common.processor_core_count)
 439         core_count = pbi->common.processor_core_count;
 440
 441     if (core_count > 1)
 442     {
 443         pbi->b_multithreaded_rd = 1;
 444         pbi->decoding_thread_count = core_count - 1;
 445
 446         CHECK_MEM_ERROR(pbi->h_decoding_thread, vpx_malloc(sizeof(pthread_t) * pbi->decoding_thread_count));
 447         CHECK_MEM_ERROR(pbi->h_event_start_decoding, vpx_malloc(sizeof(sem_t) * pbi->decoding_thread_count));
 448         CHECK_MEM_ERROR(pbi->mb_row_di, vpx_memalign(32, sizeof(MB_ROW_DEC) * pbi->decoding_thread_count));
 449         vpx_memset(pbi->mb_row_di, 0, sizeof(MB_ROW_DEC) * pbi->decoding_thread_count);
 450         CHECK_MEM_ERROR(pbi->de_thread_data, vpx_malloc(sizeof(DECODETHREAD_DATA) * pbi->decoding_thread_count));
 451
 452         for (ithread = 0; ithread < pbi->decoding_thread_count; ithread++)
 453         {
 454             sem_init(&pbi->h_event_start_decoding[ithread], 0, 0);
 455
 456             pbi->de_thread_data[ithread].ithread  = ithread;
 457             pbi->de_thread_data[ithread].ptr1     = (void *)pbi;
 458             pbi->de_thread_data[ithread].ptr2     = (void *) &pbi->mb_row_di[ithread];
 459
 460             pthread_create(&pbi->h_decoding_thread[ithread], 0, thread_decoding_proc, (&pbi->de_thread_data[ithread]));
 461         }
 462
 463         sem_init(&pbi->h_event_end_decoding, 0, 0);
 464
 465         pbi->allocated_decoding_thread_count = pbi->decoding_thread_count;
 466     }
 467 }
 468
 469
 470 void vp8mt_de_alloc_temp_buffers(VP8D_COMP *pbi, int mb_rows)
 471 {
 472     int i;
 473
 474     if (pbi->b_multithreaded_rd)
 475     {
 476             vpx_free(pbi->mt_current_mb_col);
 477             pbi->mt_current_mb_col = NULL ;
 478
 479         /* Free above_row buffers. */
 480         if (pbi->mt_yabove_row)
 481         {
 482             for (i=0; i< mb_rows; i++)
 483             {
 484                     vpx_free(pbi->mt_yabove_row[i]);
 485                     pbi->mt_yabove_row[i] = NULL ;
 486             }
 487             vpx_free(pbi->mt_yabove_row);
 488             pbi->mt_yabove_row = NULL ;
 489         }
 490
 491         if (pbi->mt_uabove_row)
 492         {
 493             for (i=0; i< mb_rows; i++)
 494             {
 495                     vpx_free(pbi->mt_uabove_row[i]);
 496                     pbi->mt_uabove_row[i] = NULL ;
 497             }
 498             vpx_free(pbi->mt_uabove_row);
 499             pbi->mt_uabove_row = NULL ;
 500         }
 501
 502         if (pbi->mt_vabove_row)
 503         {
 504             for (i=0; i< mb_rows; i++)
 505             {
 506                     vpx_free(pbi->mt_vabove_row[i]);
 507                     pbi->mt_vabove_row[i] = NULL ;
 508             }
 509             vpx_free(pbi->mt_vabove_row);
 510             pbi->mt_vabove_row = NULL ;
 511         }
 512
 513         /* Free left_col buffers. */
 514         if (pbi->mt_yleft_col)
 515         {
 516             for (i=0; i< mb_rows; i++)
 517             {
 518                     vpx_free(pbi->mt_yleft_col[i]);
 519                     pbi->mt_yleft_col[i] = NULL ;
 520             }
 521             vpx_free(pbi->mt_yleft_col);
 522             pbi->mt_yleft_col = NULL ;
 523         }
 524
 525         if (pbi->mt_uleft_col)
 526         {
 527             for (i=0; i< mb_rows; i++)
 528             {
 529                     vpx_free(pbi->mt_uleft_col[i]);
 530                     pbi->mt_uleft_col[i] = NULL ;
 531             }
 532             vpx_free(pbi->mt_uleft_col);
 533             pbi->mt_uleft_col = NULL ;
 534         }
 535
 536         if (pbi->mt_vleft_col)
 537         {
 538             for (i=0; i< mb_rows; i++)
 539             {
 540                     vpx_free(pbi->mt_vleft_col[i]);
 541                     pbi->mt_vleft_col[i] = NULL ;
 542             }
 543             vpx_free(pbi->mt_vleft_col);
 544             pbi->mt_vleft_col = NULL ;
 545         }
 546     }
 547 }
 548
 549
 550 void vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows)
 551 {
 552     VP8_COMMON *const pc = & pbi->common;
 553     int i;
 554     int uv_width;
 555
 556     if (pbi->b_multithreaded_rd)
 557     {
 558         vp8mt_de_alloc_temp_buffers(pbi, prev_mb_rows);
 559
 560         /* our internal buffers are always multiples of 16 */
 561         if ((width & 0xf) != 0)
 562             width += 16 - (width & 0xf);
 563
 564         if (width < 640) pbi->sync_range = 1;
 565         else if (width <= 1280) pbi->sync_range = 8;
 566         else if (width <= 2560) pbi->sync_range =16;
 567         else pbi->sync_range = 32;
 568
 569         uv_width = width >>1;
 570
 571         /* Allocate an int for each mb row. */
 572         CHECK_MEM_ERROR(pbi->mt_current_mb_col, vpx_malloc(sizeof(int) * pc->mb_rows));
 573
 574         /* Allocate memory for above_row buffers. */
 575         CHECK_MEM_ERROR(pbi->mt_yabove_row, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 576         for (i=0; i< pc->mb_rows; i++)
 577             CHECK_MEM_ERROR(pbi->mt_yabove_row[i], vpx_calloc(sizeof(unsigned char) * (width + (VP8BORDERINPIXELS<<1)), 1));
 578
 579         CHECK_MEM_ERROR(pbi->mt_uabove_row, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 580         for (i=0; i< pc->mb_rows; i++)
 581             CHECK_MEM_ERROR(pbi->mt_uabove_row[i], vpx_calloc(sizeof(unsigned char) * (uv_width + VP8BORDERINPIXELS), 1));
 582
 583         CHECK_MEM_ERROR(pbi->mt_vabove_row, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 584         for (i=0; i< pc->mb_rows; i++)
 585             CHECK_MEM_ERROR(pbi->mt_vabove_row[i], vpx_calloc(sizeof(unsigned char) * (uv_width + VP8BORDERINPIXELS), 1));
 586
 587         /* Allocate memory for left_col buffers. */
 588         CHECK_MEM_ERROR(pbi->mt_yleft_col, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 589         for (i=0; i< pc->mb_rows; i++)
 590             CHECK_MEM_ERROR(pbi->mt_yleft_col[i], vpx_calloc(sizeof(unsigned char) * 16, 1));
 591
 592         CHECK_MEM_ERROR(pbi->mt_uleft_col, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 593         for (i=0; i< pc->mb_rows; i++)
 594             CHECK_MEM_ERROR(pbi->mt_uleft_col[i], vpx_calloc(sizeof(unsigned char) * 8, 1));
 595
 596         CHECK_MEM_ERROR(pbi->mt_vleft_col, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 597         for (i=0; i< pc->mb_rows; i++)
 598             CHECK_MEM_ERROR(pbi->mt_vleft_col[i], vpx_calloc(sizeof(unsigned char) * 8, 1));
 599     }
 600 }
 601
 602
 603 void vp8_decoder_remove_threads(VP8D_COMP *pbi)
 604 {
 605     /* shutdown MB Decoding thread; */
 606     if (pbi->b_multithreaded_rd)
 607     {
 608         int i;
 609
 610         pbi->b_multithreaded_rd = 0;
 611
 612         /* allow all threads to exit */
 613         for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
 614         {
 615             sem_post(&pbi->h_event_start_decoding[i]);
 616             pthread_join(pbi->h_decoding_thread[i], NULL);
 617         }
 618
 619         for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
 620         {
 621             sem_destroy(&pbi->h_event_start_decoding[i]);
 622         }
 623
 624         sem_destroy(&pbi->h_event_end_decoding);
 625
 626             vpx_free(pbi->h_decoding_thread);
 627             pbi->h_decoding_thread = NULL;
 628
 629             vpx_free(pbi->h_event_start_decoding);
 630             pbi->h_event_start_decoding = NULL;
 631
 632             vpx_free(pbi->mb_row_di);
 633             pbi->mb_row_di = NULL ;
 634
 635             vpx_free(pbi->de_thread_data);
 636             pbi->de_thread_data = NULL;
 637     }
 638 }
 639
 640
 641 static void lpf_init( VP8D_COMP *pbi, int default_filt_lvl)
 642 {
 643     VP8_COMMON *cm  = &pbi->common;
 644     MACROBLOCKD *mbd = &pbi->mb;
 645     /*YV12_BUFFER_CONFIG *post = &cm->new_frame;*/  /*frame_to_show;*/
 646     loop_filter_info *lfi = cm->lf_info;
 647     FRAME_TYPE frame_type = cm->frame_type;
 648
 649     /*int mb_row;
 650     int mb_col;
 651     int baseline_filter_level[MAX_MB_SEGMENTS];*/
 652     int alt_flt_enabled = mbd->segmentation_enabled;
 653
 654     int i;
 655     /*unsigned char *y_ptr, *u_ptr, *v_ptr;*/
 656
 657     /* Note the baseline filter values for each segment */
 658     if (alt_flt_enabled)
 659     {
 660         for (i = 0; i < MAX_MB_SEGMENTS; i++)
 661         {
 662             /* Abs value */
 663             if (mbd->mb_segement_abs_delta == SEGMENT_ABSDATA)
 664                 pbi->mt_baseline_filter_level[i] = mbd->segment_feature_data[MB_LVL_ALT_LF][i];
 665             /* Delta Value */
 666             else
 667             {
 668                 pbi->mt_baseline_filter_level[i] = default_filt_lvl + mbd->segment_feature_data[MB_LVL_ALT_LF][i];
 669                 pbi->mt_baseline_filter_level[i] = (pbi->mt_baseline_filter_level[i] >= 0) ? ((pbi->mt_baseline_filter_level[i] <= MAX_LOOP_FILTER) ? pbi->mt_baseline_filter_level[i] : MAX_LOOP_FILTER) : 0;  /* Clamp to valid range */
 670             }
 671         }
 672     }
 673     else
 674     {
 675         for (i = 0; i < MAX_MB_SEGMENTS; i++)
 676             pbi->mt_baseline_filter_level[i] = default_filt_lvl;
 677     }
 678
 679     /* Initialize the loop filter for this frame. */
 680     if ((cm->last_filter_type != cm->filter_type) || (cm->last_sharpness_level != cm->sharpness_level))
 681         vp8_init_loop_filter(cm);
 682     else if (frame_type != cm->last_frame_type)
 683         vp8_frame_init_loop_filter(lfi, frame_type);
 684 }
 685
 686
 687 void vp8mt_decode_mb_rows( VP8D_COMP *pbi, MACROBLOCKD *xd)
 688 {
 689     int mb_row;
 690     VP8_COMMON *pc = &pbi->common;
 691
 692     int num_part = 1 << pbi->common.multi_token_partition;
 693     int i;
 694     volatile int *last_row_current_mb_col = NULL;
 695     int nsync = pbi->sync_range;
 696
 697     int filter_level;
 698     loop_filter_info *lfi = pc->lf_info;
 699     int alt_flt_enabled = xd->segmentation_enabled;
 700     int Segment;
 701
 702     if(pbi->common.filter_level)
 703     {
 704         /* Set above_row buffer to 127 for decoding first MB row */
 705         vpx_memset(pbi->mt_yabove_row[0] + VP8BORDERINPIXELS-1, 127, pc->yv12_fb[pc->lst_fb_idx].y_width + 5);
 706         vpx_memset(pbi->mt_uabove_row[0] + (VP8BORDERINPIXELS>>1)-1, 127, (pc->yv12_fb[pc->lst_fb_idx].y_width>>1) +5);
 707         vpx_memset(pbi->mt_vabove_row[0] + (VP8BORDERINPIXELS>>1)-1, 127, (pc->yv12_fb[pc->lst_fb_idx].y_width>>1) +5);
 708
 709         for (i=1; i<pc->mb_rows; i++)
 710         {
 711             vpx_memset(pbi->mt_yabove_row[i] + VP8BORDERINPIXELS-1, (unsigned char)129, 1);
 712             vpx_memset(pbi->mt_uabove_row[i] + (VP8BORDERINPIXELS>>1)-1, (unsigned char)129, 1);
 713             vpx_memset(pbi->mt_vabove_row[i] + (VP8BORDERINPIXELS>>1)-1, (unsigned char)129, 1);
 714         }
 715
 716         /* Set left_col to 129 initially */
 717         for (i=0; i<pc->mb_rows; i++)
 718         {
 719             vpx_memset(pbi->mt_yleft_col[i], (unsigned char)129, 16);
 720             vpx_memset(pbi->mt_uleft_col[i], (unsigned char)129, 8);
 721             vpx_memset(pbi->mt_vleft_col[i], (unsigned char)129, 8);
 722         }
 723         lpf_init(pbi, pc->filter_level);
 724     }
 725
 726     setup_decoding_thread_data(pbi, xd, pbi->mb_row_di, pbi->decoding_thread_count);
 727
 728     for (i = 0; i < pbi->decoding_thread_count; i++)
 729         sem_post(&pbi->h_event_start_decoding[i]);
 730
 731     for (mb_row = 0; mb_row < pc->mb_rows; mb_row += (pbi->decoding_thread_count + 1))
 732     {
 733
 734         xd->current_bc = &pbi->mbc[mb_row%num_part];
 735
 736         /* vp8_decode_mb_row(pbi, pc, mb_row, xd); */
 737         {
 738             int i;
 739             int recon_yoffset, recon_uvoffset;
 740             int mb_col;
 741             int ref_fb_idx = pc->lst_fb_idx;
 742             int dst_fb_idx = pc->new_fb_idx;
 743             int recon_y_stride = pc->yv12_fb[ref_fb_idx].y_stride;
 744             int recon_uv_stride = pc->yv12_fb[ref_fb_idx].uv_stride;
 745
 746            /* volatile int *last_row_current_mb_col = NULL; */
 747             if (mb_row > 0)
 748                 last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row -1];
 749
 750             vpx_memset(&pc->left_context, 0, sizeof(pc->left_context));
 751             recon_yoffset = mb_row * recon_y_stride * 16;
 752             recon_uvoffset = mb_row * recon_uv_stride * 8;
 753             /* reset above block coeffs */
 754
 755             xd->above_context = pc->above_context;
 756             xd->up_available = (mb_row != 0);
 757
 758             xd->mb_to_top_edge = -((mb_row * 16)) << 3;
 759             xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
 760
 761             for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
 762             {
 763                 if ( mb_row > 0 && (mb_col & (nsync-1)) == 0){
 764                     while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != pc->mb_cols - 1)
 765                     {
 766                         x86_pause_hint();
 767                         thread_sleep(0);
 768                     }
 769                 }
 770
 771                 update_blockd_bmi(xd);
 772
 773                 /* Distance of Mb to the various image edges.
 774                  * These are specified to 8th pel as they are always compared to values that are in 1/8th pel units
 775                  */
 776                 xd->mb_to_left_edge = -((mb_col * 16) << 3);
 777                 xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
 778
 779                 xd->dst.y_buffer = pc->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 780                 xd->dst.u_buffer = pc->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 781                 xd->dst.v_buffer = pc->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 782
 783                 xd->left_available = (mb_col != 0);
 784
 785                 /* Select the appropriate reference frame for this MB */
 786                 if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
 787                     ref_fb_idx = pc->lst_fb_idx;
 788                 else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
 789                     ref_fb_idx = pc->gld_fb_idx;
 790                 else
 791                     ref_fb_idx = pc->alt_fb_idx;
 792
 793                 xd->pre.y_buffer = pc->yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
 794                 xd->pre.u_buffer = pc->yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
 795                 xd->pre.v_buffer = pc->yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
 796
 797                 if (xd->mode_info_context->mbmi.ref_frame != INTRA_FRAME)
 798                 {
 799                     /* propagate errors from reference frames */
 800                     xd->corrupted |= pc->yv12_fb[ref_fb_idx].corrupted;
 801                 }
 802
 803                 vp8_build_uvmvs(xd, pc->full_pixel);
 804                 decode_macroblock(pbi, xd, mb_row, mb_col);
 805
 806                 /* check if the boolean decoder has suffered an error */
 807                 xd->corrupted |= vp8dx_bool_error(xd->current_bc);
 808
 809                 if (pbi->common.filter_level)
 810                 {
 811                     int skip_lf;
 812                     /* Save decoded MB last row data for next-row decoding */
 813                     if(mb_row != pc->mb_rows-1)
 814                     {
 815                         vpx_memcpy((pbi->mt_yabove_row[mb_row +1] + 32 + mb_col*16), (xd->dst.y_buffer + 15 * recon_y_stride), 16);
 816                         vpx_memcpy((pbi->mt_uabove_row[mb_row +1] + 16 + mb_col*8), (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
 817                         vpx_memcpy((pbi->mt_vabove_row[mb_row +1] + 16 + mb_col*8), (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
 818                     }
 819
 820                     /* save left_col for next MB decoding */
 821                     if(mb_col != pc->mb_cols-1)
 822                     {
 823                         MODE_INFO *next = xd->mode_info_context +1;
 824
 825                         if (xd->frame_type == KEY_FRAME  ||  next->mbmi.ref_frame == INTRA_FRAME)
 826                         {
 827                             for (i = 0; i < 16; i++)
 828                                 pbi->mt_yleft_col[mb_row][i] = xd->dst.y_buffer [i* recon_y_stride + 15];
 829                             for (i = 0; i < 8; i++)
 830                             {
 831                                 pbi->mt_uleft_col[mb_row][i] = xd->dst.u_buffer [i* recon_uv_stride + 7];
 832                                 pbi->mt_vleft_col[mb_row][i] = xd->dst.v_buffer [i* recon_uv_stride + 7];
 833                             }
 834                         }
 835                     }
 836
 837                     /* update loopfilter info */
 838                     Segment = (alt_flt_enabled) ? xd->mode_info_context->mbmi.segment_id : 0;
 839                     skip_lf = (xd->mode_info_context->mbmi.mode != B_PRED &&
 840                                     xd->mode_info_context->mbmi.mode != SPLITMV &&
 841                                     xd->mode_info_context->mbmi.mb_skip_coeff);
 842                     filter_level = pbi->mt_baseline_filter_level[Segment];
 843                     /* Distance of Mb to the various image edges.
 844                      * These are specified to 8th pel as they are always compared to values that are in 1/8th pel units
 845                      * Apply any context driven MB level adjustment
 846                      */
 847                     filter_level = vp8_adjust_mb_lf_value(xd, filter_level);
 848
 849                     /* loopfilter on this macroblock. */
 850                     if (filter_level)
 851                     {
 852                         if (mb_col > 0)
 853                             pc->lf_mbv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 854
 855                         if (!skip_lf)
 856                             pc->lf_bv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 857
 858                         /* don't apply across umv border */
 859                         if (mb_row > 0)
 860                             pc->lf_mbh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 861
 862                         if (!skip_lf)
 863                             pc->lf_bh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 864                     }
 865                 }
 866
 867                 recon_yoffset += 16;
 868                 recon_uvoffset += 8;
 869
 870                 ++xd->mode_info_context;  /* next mb */
 871
 872                 xd->above_context++;
 873
 874                 pbi->mt_current_mb_col[mb_row] = mb_col;
 875             }
 876
 877             /* adjust to the next row of mbs */
 878             if (pbi->common.filter_level)
 879             {
 880                 if(mb_row != pc->mb_rows-1)
 881                 {
 882                     int lasty = pc->yv12_fb[ref_fb_idx].y_width + VP8BORDERINPIXELS;
 883                     int lastuv = (pc->yv12_fb[ref_fb_idx].y_width>>1) + (VP8BORDERINPIXELS>>1);
 884
 885                     for (i = 0; i < 4; i++)
 886                     {
 887                         pbi->mt_yabove_row[mb_row +1][lasty + i] = pbi->mt_yabove_row[mb_row +1][lasty -1];
 888                         pbi->mt_uabove_row[mb_row +1][lastuv + i] = pbi->mt_uabove_row[mb_row +1][lastuv -1];
 889                         pbi->mt_vabove_row[mb_row +1][lastuv + i] = pbi->mt_vabove_row[mb_row +1][lastuv -1];
 890                     }
 891                 }
 892             }else
 893                 vp8_extend_mb_row(&pc->yv12_fb[dst_fb_idx], xd->dst.y_buffer + 16, xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
 894
 895             ++xd->mode_info_context;      /* skip prediction column */
 896         }
 897         xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
 898     }
 899
 900     sem_wait(&pbi->h_event_end_decoding);   /* add back for each frame */
 901 }