granicus.if.org Git - libvpx/blob - 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 #ifndef WIN32
  13 # include <unistd.h>
  14 #endif
  15 #ifdef __APPLE__
  16 #include <mach/mach_init.h>
  17 #endif
  18 #include "onyxd_int.h"
  19 #include "vpx_mem/vpx_mem.h"
  20 #include "threading.h"
  21
  22 #include "loopfilter.h"
  23 #include "extend.h"
  24 #include "vpx_ports/vpx_timer.h"
  25 #include "detokenize.h"
  26 #include "reconinter.h"
  27 #include "reconintra_mt.h"
  28
  29 extern void mb_init_dequantizer(VP8D_COMP *pbi, MACROBLOCKD *xd);
  30 extern void clamp_mvs(MACROBLOCKD *xd);
  31 extern void vp8_build_uvmvs(MACROBLOCKD *x, int fullpixel);
  32
  33 #if CONFIG_RUNTIME_CPU_DETECT
  34 #define RTCD_VTABLE(x) (&(pbi)->common.rtcd.x)
  35 #else
  36 #define RTCD_VTABLE(x) NULL
  37 #endif
  38
  39 void vp8_setup_decoding_thread_data(VP8D_COMP *pbi, MACROBLOCKD *xd, MB_ROW_DEC *mbrd, int count)
  40 {
  41 #if CONFIG_MULTITHREAD
  42     VP8_COMMON *const pc = & pbi->common;
  43     int i, j;
  44
  45     for (i = 0; i < count; i++)
  46     {
  47         MACROBLOCKD *mbd = &mbrd[i].mbd;
  48 #if CONFIG_RUNTIME_CPU_DETECT
  49         mbd->rtcd = xd->rtcd;
  50 #endif
  51         mbd->subpixel_predict        = xd->subpixel_predict;
  52         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
  53         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
  54         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
  55
  56         mbd->mode_info_context = pc->mi   + pc->mode_info_stride * (i + 1);
  57         mbd->mode_info_stride  = pc->mode_info_stride;
  58
  59         mbd->frame_type = pc->frame_type;
  60         mbd->frames_since_golden      = pc->frames_since_golden;
  61         mbd->frames_till_alt_ref_frame  = pc->frames_till_alt_ref_frame;
  62
  63         mbd->pre = pc->yv12_fb[pc->lst_fb_idx];
  64         mbd->dst = pc->yv12_fb[pc->new_fb_idx];
  65
  66         vp8_setup_block_dptrs(mbd);
  67         vp8_build_block_doffsets(mbd);
  68         mbd->segmentation_enabled    = xd->segmentation_enabled;
  69         mbd->mb_segement_abs_delta     = xd->mb_segement_abs_delta;
  70         vpx_memcpy(mbd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
  71
  72         //signed char ref_lf_deltas[MAX_REF_LF_DELTAS];
  73         vpx_memcpy(mbd->ref_lf_deltas, xd->ref_lf_deltas, sizeof(xd->ref_lf_deltas));
  74         //signed char mode_lf_deltas[MAX_MODE_LF_DELTAS];
  75         vpx_memcpy(mbd->mode_lf_deltas, xd->mode_lf_deltas, sizeof(xd->mode_lf_deltas));
  76         //unsigned char mode_ref_lf_delta_enabled;
  77         //unsigned char mode_ref_lf_delta_update;
  78         mbd->mode_ref_lf_delta_enabled    = xd->mode_ref_lf_delta_enabled;
  79         mbd->mode_ref_lf_delta_update    = xd->mode_ref_lf_delta_update;
  80
  81         mbd->current_bc = &pbi->bc2;
  82
  83         for (j = 0; j < 25; j++)
  84         {
  85             mbd->block[j].dequant = xd->block[j].dequant;
  86         }
  87     }
  88
  89     for (i=0; i< pc->mb_rows; i++)
  90         pbi->mt_current_mb_col[i]=-1;
  91 #else
  92     (void) pbi;
  93     (void) xd;
  94     (void) mbrd;
  95     (void) count;
  96 #endif
  97 }
  98
  99
 100 void vp8mt_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd, int mb_row, int mb_col)
 101 {
 102 #if CONFIG_MULTITHREAD
 103     int eobtotal = 0;
 104     int i, do_clamp = xd->mode_info_context->mbmi.need_to_clamp_mvs;
 105     VP8_COMMON *pc = &pbi->common;
 106
 107     if (xd->mode_info_context->mbmi.mb_skip_coeff)
 108     {
 109         vp8_reset_mb_tokens_context(xd);
 110     }
 111     else
 112     {
 113         eobtotal = vp8_decode_mb_tokens(pbi, xd);
 114     }
 115
 116     // Perform temporary clamping of the MV to be used for prediction
 117     if (do_clamp)
 118     {
 119         clamp_mvs(xd);
 120     }
 121
 122     xd->mode_info_context->mbmi.dc_diff = 1;
 123
 124     if (xd->mode_info_context->mbmi.mode != B_PRED && xd->mode_info_context->mbmi.mode != SPLITMV && eobtotal == 0)
 125     {
 126         xd->mode_info_context->mbmi.dc_diff = 0;
 127
 128         //mt_skip_recon_mb(pbi, xd, mb_row, mb_col);
 129         if (xd->frame_type == KEY_FRAME  ||  xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
 130         {
 131             vp8mt_build_intra_predictors_mbuv_s(pbi, xd, mb_row, mb_col);
 132             vp8mt_build_intra_predictors_mby_s(pbi, xd, mb_row, mb_col);
 133         }
 134         else
 135         {
 136             vp8_build_inter_predictors_mb_s(xd);
 137         }
 138         return;
 139     }
 140
 141     if (xd->segmentation_enabled)
 142         mb_init_dequantizer(pbi, xd);
 143
 144     // do prediction
 145     if (xd->frame_type == KEY_FRAME  ||  xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
 146     {
 147         vp8mt_build_intra_predictors_mbuv(pbi, xd, mb_row, mb_col);
 148
 149         if (xd->mode_info_context->mbmi.mode != B_PRED)
 150         {
 151             vp8mt_build_intra_predictors_mby(pbi, xd, mb_row, mb_col);
 152         } else {
 153             vp8mt_intra_prediction_down_copy(pbi, xd, mb_row, mb_col);
 154         }
 155     }
 156     else
 157     {
 158         vp8_build_inter_predictors_mb(xd);
 159     }
 160
 161     // dequantization and idct
 162     if (xd->mode_info_context->mbmi.mode != B_PRED && xd->mode_info_context->mbmi.mode != SPLITMV)
 163     {
 164         BLOCKD *b = &xd->block[24];
 165         DEQUANT_INVOKE(&pbi->dequant, block)(b);
 166
 167         // do 2nd order transform on the dc block
 168         if (xd->eobs[24] > 1)
 169         {
 170             IDCT_INVOKE(RTCD_VTABLE(idct), iwalsh16)(&b->dqcoeff[0], b->diff);
 171             ((int *)b->qcoeff)[0] = 0;
 172             ((int *)b->qcoeff)[1] = 0;
 173             ((int *)b->qcoeff)[2] = 0;
 174             ((int *)b->qcoeff)[3] = 0;
 175             ((int *)b->qcoeff)[4] = 0;
 176             ((int *)b->qcoeff)[5] = 0;
 177             ((int *)b->qcoeff)[6] = 0;
 178             ((int *)b->qcoeff)[7] = 0;
 179         }
 180         else
 181         {
 182             IDCT_INVOKE(RTCD_VTABLE(idct), iwalsh1)(&b->dqcoeff[0], b->diff);
 183             ((int *)b->qcoeff)[0] = 0;
 184         }
 185
 186         DEQUANT_INVOKE (&pbi->dequant, dc_idct_add_y_block)
 187                         (xd->qcoeff, &xd->block[0].dequant[0][0],
 188                          xd->predictor, xd->dst.y_buffer,
 189                          xd->dst.y_stride, xd->eobs, xd->block[24].diff);
 190     }
 191     else if ((xd->frame_type == KEY_FRAME  ||  xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) && xd->mode_info_context->mbmi.mode == B_PRED)
 192     {
 193         for (i = 0; i < 16; i++)
 194         {
 195             BLOCKD *b = &xd->block[i];
 196             vp8mt_predict_intra4x4(pbi, xd, b->bmi.mode, b->predictor, mb_row, mb_col, i);
 197
 198             if (xd->eobs[i] > 1)
 199             {
 200                 DEQUANT_INVOKE(&pbi->dequant, idct_add)
 201                     (b->qcoeff, &b->dequant[0][0],  b->predictor,
 202                     *(b->base_dst) + b->dst, 16, b->dst_stride);
 203             }
 204             else
 205             {
 206                 IDCT_INVOKE(RTCD_VTABLE(idct), idct1_scalar_add)
 207                     (b->qcoeff[0] * b->dequant[0][0], b->predictor,
 208                     *(b->base_dst) + b->dst, 16, b->dst_stride);
 209                 ((int *)b->qcoeff)[0] = 0;
 210             }
 211         }
 212     }
 213     else
 214     {
 215         DEQUANT_INVOKE (&pbi->dequant, idct_add_y_block)
 216                         (xd->qcoeff, &xd->block[0].dequant[0][0],
 217                          xd->predictor, xd->dst.y_buffer,
 218                          xd->dst.y_stride, xd->eobs);
 219     }
 220
 221     DEQUANT_INVOKE (&pbi->dequant, idct_add_uv_block)
 222                     (xd->qcoeff+16*16, &xd->block[16].dequant[0][0],
 223                      xd->predictor+16*16, xd->dst.u_buffer, xd->dst.v_buffer,
 224                      xd->dst.uv_stride, xd->eobs+16);
 225 #else
 226     (void) pbi;
 227     (void) xd;
 228     (void) mb_row;
 229     (void) mb_col;
 230 #endif
 231 }
 232
 233
 234 THREAD_FUNCTION vp8_thread_decoding_proc(void *p_data)
 235 {
 236 #if CONFIG_MULTITHREAD
 237     int ithread = ((DECODETHREAD_DATA *)p_data)->ithread;
 238     VP8D_COMP *pbi = (VP8D_COMP *)(((DECODETHREAD_DATA *)p_data)->ptr1);
 239     MB_ROW_DEC *mbrd = (MB_ROW_DEC *)(((DECODETHREAD_DATA *)p_data)->ptr2);
 240     ENTROPY_CONTEXT_PLANES mb_row_left_context;
 241
 242     while (1)
 243     {
 244         if (pbi->b_multithreaded_rd == 0)
 245             break;
 246
 247         //if(WaitForSingleObject(pbi->h_event_start_decoding[ithread], INFINITE) == WAIT_OBJECT_0)
 248         if (sem_wait(&pbi->h_event_start_decoding[ithread]) == 0)
 249         {
 250             if (pbi->b_multithreaded_rd == 0)
 251                 break;
 252             else
 253             {
 254                 VP8_COMMON *pc = &pbi->common;
 255                 MACROBLOCKD *xd = &mbrd->mbd;
 256
 257                 int mb_row;
 258                 int num_part = 1 << pbi->common.multi_token_partition;
 259                 volatile int *last_row_current_mb_col;
 260                 int nsync = pbi->sync_range;
 261
 262                 for (mb_row = ithread+1; mb_row < pc->mb_rows; mb_row += (pbi->decoding_thread_count + 1))
 263                 {
 264                     int i;
 265                     int recon_yoffset, recon_uvoffset;
 266                     int mb_col;
 267                     int ref_fb_idx = pc->lst_fb_idx;
 268                     int dst_fb_idx = pc->new_fb_idx;
 269                     int recon_y_stride = pc->yv12_fb[ref_fb_idx].y_stride;
 270                     int recon_uv_stride = pc->yv12_fb[ref_fb_idx].uv_stride;
 271
 272                     int filter_level;
 273                     loop_filter_info *lfi = pc->lf_info;
 274                     int alt_flt_enabled = xd->segmentation_enabled;
 275                     int Segment;
 276
 277                     pbi->mb_row_di[ithread].mb_row = mb_row;
 278                     pbi->mb_row_di[ithread].mbd.current_bc =  &pbi->mbc[mb_row%num_part];
 279
 280                     last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row -1];
 281
 282                     recon_yoffset = mb_row * recon_y_stride * 16;
 283                     recon_uvoffset = mb_row * recon_uv_stride * 8;
 284                     // reset above block coeffs
 285
 286                     xd->above_context = pc->above_context;
 287                     xd->left_context = &mb_row_left_context;
 288                     vpx_memset(&mb_row_left_context, 0, sizeof(mb_row_left_context));
 289                     xd->up_available = (mb_row != 0);
 290
 291                     xd->mb_to_top_edge = -((mb_row * 16)) << 3;
 292                     xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
 293
 294                     for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
 295                     {
 296                         if ((mb_col & (nsync-1)) == 0)
 297                         {
 298                             while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != pc->mb_cols - 1)
 299                             {
 300                                 x86_pause_hint();
 301                                 thread_sleep(0);
 302                             }
 303                         }
 304
 305                         if (xd->mode_info_context->mbmi.mode == SPLITMV || xd->mode_info_context->mbmi.mode == B_PRED)
 306                         {
 307                             for (i = 0; i < 16; i++)
 308                             {
 309                                 BLOCKD *d = &xd->block[i];
 310                                 vpx_memcpy(&d->bmi, &xd->mode_info_context->bmi[i], sizeof(B_MODE_INFO));
 311                             }
 312                         }
 313
 314                         if(pbi->common.filter_level)
 315                         {
 316                             //update loopfilter info
 317                             Segment = (alt_flt_enabled) ? xd->mode_info_context->mbmi.segment_id : 0;
 318                             filter_level = pbi->mt_baseline_filter_level[Segment];
 319                             // Distance of Mb to the various image edges.
 320                             // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
 321                             // Apply any context driven MB level adjustment
 322                             vp8_adjust_mb_lf_value(xd, &filter_level);
 323                         }
 324
 325                         // Distance of Mb to the various image edges.
 326                         // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
 327                         xd->mb_to_left_edge = -((mb_col * 16) << 3);
 328                         xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
 329
 330                         xd->dst.y_buffer = pc->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 331                         xd->dst.u_buffer = pc->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 332                         xd->dst.v_buffer = pc->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 333
 334                         xd->left_available = (mb_col != 0);
 335
 336                         // Select the appropriate reference frame for this MB
 337                         if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
 338                             ref_fb_idx = pc->lst_fb_idx;
 339                         else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
 340                             ref_fb_idx = pc->gld_fb_idx;
 341                         else
 342                             ref_fb_idx = pc->alt_fb_idx;
 343
 344                         xd->pre.y_buffer = pc->yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
 345                         xd->pre.u_buffer = pc->yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
 346                         xd->pre.v_buffer = pc->yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
 347
 348                         vp8_build_uvmvs(xd, pc->full_pixel);
 349                         vp8mt_decode_macroblock(pbi, xd, mb_row, mb_col);
 350
 351                         if (pbi->common.filter_level)
 352                         {
 353                             if( mb_row != pc->mb_rows-1 )
 354                             {
 355                                 //Save decoded MB last row data for next-row decoding
 356                                 vpx_memcpy((pbi->mt_yabove_row[mb_row + 1] + 32 + mb_col*16), (xd->dst.y_buffer + 15 * recon_y_stride), 16);
 357                                 vpx_memcpy((pbi->mt_uabove_row[mb_row + 1] + 16 + mb_col*8), (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
 358                                 vpx_memcpy((pbi->mt_vabove_row[mb_row + 1] + 16 + mb_col*8), (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
 359                             }
 360
 361                             //save left_col for next MB decoding
 362                             if(mb_col != pc->mb_cols-1)
 363                             {
 364                                 MODE_INFO *next = xd->mode_info_context +1;
 365
 366                                 if (xd->frame_type == KEY_FRAME  ||  next->mbmi.ref_frame == INTRA_FRAME)
 367                                 {
 368                                     for (i = 0; i < 16; i++)
 369                                         pbi->mt_yleft_col[mb_row][i] = xd->dst.y_buffer [i* recon_y_stride + 15];
 370                                     for (i = 0; i < 8; i++)
 371                                     {
 372                                         pbi->mt_uleft_col[mb_row][i] = xd->dst.u_buffer [i* recon_uv_stride + 7];
 373                                         pbi->mt_vleft_col[mb_row][i] = xd->dst.v_buffer [i* recon_uv_stride + 7];
 374                                     }
 375                                 }
 376                             }
 377
 378                           // loopfilter on this macroblock.
 379                             if (filter_level)
 380                             {
 381                                 if (mb_col > 0)
 382                                     pc->lf_mbv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level], pc->simpler_lpf);
 383
 384                                 if (xd->mode_info_context->mbmi.dc_diff > 0)
 385                                     pc->lf_bv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level], pc->simpler_lpf);
 386
 387                                 // don't apply across umv border
 388                                 if (mb_row > 0)
 389                                     pc->lf_mbh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level], pc->simpler_lpf);
 390
 391                                 if (xd->mode_info_context->mbmi.dc_diff > 0)
 392                                     pc->lf_bh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level], pc->simpler_lpf);
 393                             }
 394                         }
 395
 396                         recon_yoffset += 16;
 397                         recon_uvoffset += 8;
 398
 399                         ++xd->mode_info_context;  /* next mb */
 400
 401                         xd->above_context++;
 402
 403                         //pbi->mb_row_di[ithread].current_mb_col = mb_col;
 404                         pbi->mt_current_mb_col[mb_row] = mb_col;
 405                     }
 406
 407                     // adjust to the next row of mbs
 408                     if (pbi->common.filter_level)
 409                     {
 410                         if(mb_row != pc->mb_rows-1)
 411                         {
 412                             int lasty = pc->yv12_fb[ref_fb_idx].y_width + VP8BORDERINPIXELS;
 413                             int lastuv = (pc->yv12_fb[ref_fb_idx].y_width>>1) + (VP8BORDERINPIXELS>>1);
 414
 415                             for (i = 0; i < 4; i++)
 416                             {
 417                                 pbi->mt_yabove_row[mb_row +1][lasty + i] = pbi->mt_yabove_row[mb_row +1][lasty -1];
 418                                 pbi->mt_uabove_row[mb_row +1][lastuv + i] = pbi->mt_uabove_row[mb_row +1][lastuv -1];
 419                                 pbi->mt_vabove_row[mb_row +1][lastuv + i] = pbi->mt_vabove_row[mb_row +1][lastuv -1];
 420                             }
 421                         }
 422                     } else
 423                         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);
 424
 425                     ++xd->mode_info_context;      /* skip prediction column */
 426
 427                     // since we have multithread
 428                     xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
 429                 }
 430             }
 431         }
 432         //  add this to each frame
 433         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))
 434         {
 435             //SetEvent(pbi->h_event_end_decoding);
 436             sem_post(&pbi->h_event_end_decoding);
 437         }
 438     }
 439 #else
 440     (void) p_data;
 441 #endif
 442
 443     return 0 ;
 444 }
 445
 446
 447 void vp8_decoder_create_threads(VP8D_COMP *pbi)
 448 {
 449 #if CONFIG_MULTITHREAD
 450     int core_count = 0;
 451     int ithread;
 452     int i;
 453
 454     pbi->b_multithreaded_rd = 0;
 455     pbi->allocated_decoding_thread_count = 0;
 456     core_count = (pbi->max_threads > 16) ? 16 : pbi->max_threads;
 457
 458     if (core_count > 1)
 459     {
 460         pbi->b_multithreaded_rd = 1;
 461         pbi->decoding_thread_count = core_count -1;
 462
 463         CHECK_MEM_ERROR(pbi->h_decoding_thread, vpx_malloc(sizeof(pthread_t) * pbi->decoding_thread_count));
 464         CHECK_MEM_ERROR(pbi->h_event_start_decoding, vpx_malloc(sizeof(sem_t) * pbi->decoding_thread_count));
 465         CHECK_MEM_ERROR(pbi->mb_row_di, vpx_memalign(32, sizeof(MB_ROW_DEC) * pbi->decoding_thread_count));
 466         vpx_memset(pbi->mb_row_di, 0, sizeof(MB_ROW_DEC) * pbi->decoding_thread_count);
 467         CHECK_MEM_ERROR(pbi->de_thread_data, vpx_malloc(sizeof(DECODETHREAD_DATA) * pbi->decoding_thread_count));
 468
 469         for (ithread = 0; ithread < pbi->decoding_thread_count; ithread++)
 470         {
 471             sem_init(&pbi->h_event_start_decoding[ithread], 0, 0);
 472
 473             pbi->de_thread_data[ithread].ithread  = ithread;
 474             pbi->de_thread_data[ithread].ptr1     = (void *)pbi;
 475             pbi->de_thread_data[ithread].ptr2     = (void *) &pbi->mb_row_di[ithread];
 476
 477             pthread_create(&pbi->h_decoding_thread[ithread], 0, vp8_thread_decoding_proc, (&pbi->de_thread_data[ithread]));
 478         }
 479
 480         sem_init(&pbi->h_event_end_decoding, 0, 0);
 481
 482         pbi->allocated_decoding_thread_count = pbi->decoding_thread_count;
 483     }
 484
 485 #else
 486     (void) pbi;
 487 #endif
 488 }
 489
 490
 491 void vp8mt_de_alloc_temp_buffers(VP8D_COMP *pbi, int mb_rows)
 492 {
 493 #if CONFIG_MULTITHREAD
 494     VP8_COMMON *const pc = & pbi->common;
 495     int i;
 496
 497     if (pbi->b_multithreaded_rd)
 498     {
 499         if (pbi->mt_current_mb_col)
 500         {
 501             vpx_free(pbi->mt_current_mb_col);
 502             pbi->mt_current_mb_col = NULL ;
 503         }
 504
 505         // Free above_row buffers.
 506         if (pbi->mt_yabove_row)
 507         {
 508             for (i=0; i< mb_rows; i++)
 509             {
 510                 if (pbi->mt_yabove_row[i])
 511                 {
 512                     vpx_free(pbi->mt_yabove_row[i]);
 513                     pbi->mt_yabove_row[i] = NULL ;
 514                 }
 515             }
 516             vpx_free(pbi->mt_yabove_row);
 517             pbi->mt_yabove_row = NULL ;
 518         }
 519
 520         if (pbi->mt_uabove_row)
 521         {
 522             for (i=0; i< mb_rows; i++)
 523             {
 524                 if (pbi->mt_uabove_row[i])
 525                 {
 526                     vpx_free(pbi->mt_uabove_row[i]);
 527                     pbi->mt_uabove_row[i] = NULL ;
 528                 }
 529             }
 530             vpx_free(pbi->mt_uabove_row);
 531             pbi->mt_uabove_row = NULL ;
 532         }
 533
 534         if (pbi->mt_vabove_row)
 535         {
 536             for (i=0; i< mb_rows; i++)
 537             {
 538                 if (pbi->mt_vabove_row[i])
 539                 {
 540                     vpx_free(pbi->mt_vabove_row[i]);
 541                     pbi->mt_vabove_row[i] = NULL ;
 542                 }
 543             }
 544             vpx_free(pbi->mt_vabove_row);
 545             pbi->mt_vabove_row = NULL ;
 546         }
 547
 548         // Free left_col buffers.
 549         if (pbi->mt_yleft_col)
 550         {
 551             for (i=0; i< mb_rows; i++)
 552             {
 553                 if (pbi->mt_yleft_col[i])
 554                 {
 555                     vpx_free(pbi->mt_yleft_col[i]);
 556                     pbi->mt_yleft_col[i] = NULL ;
 557                 }
 558             }
 559             vpx_free(pbi->mt_yleft_col);
 560             pbi->mt_yleft_col = NULL ;
 561         }
 562
 563         if (pbi->mt_uleft_col)
 564         {
 565             for (i=0; i< mb_rows; i++)
 566             {
 567                 if (pbi->mt_uleft_col[i])
 568                 {
 569                     vpx_free(pbi->mt_uleft_col[i]);
 570                     pbi->mt_uleft_col[i] = NULL ;
 571                 }
 572             }
 573             vpx_free(pbi->mt_uleft_col);
 574             pbi->mt_uleft_col = NULL ;
 575         }
 576
 577         if (pbi->mt_vleft_col)
 578         {
 579             for (i=0; i< mb_rows; i++)
 580             {
 581                 if (pbi->mt_vleft_col[i])
 582                 {
 583                     vpx_free(pbi->mt_vleft_col[i]);
 584                     pbi->mt_vleft_col[i] = NULL ;
 585                 }
 586             }
 587             vpx_free(pbi->mt_vleft_col);
 588             pbi->mt_vleft_col = NULL ;
 589         }
 590     }
 591 #else
 592     (void) pbi;
 593 #endif
 594 }
 595
 596
 597 int vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows)
 598 {
 599 #if CONFIG_MULTITHREAD
 600     VP8_COMMON *const pc = & pbi->common;
 601     int i;
 602     int uv_width;
 603
 604     if (pbi->b_multithreaded_rd)
 605     {
 606         vp8mt_de_alloc_temp_buffers(pbi, prev_mb_rows);
 607
 608         // our internal buffers are always multiples of 16
 609         if ((width & 0xf) != 0)
 610             width += 16 - (width & 0xf);
 611
 612         if (width < 640) pbi->sync_range = 1;
 613         else if (width <= 1280) pbi->sync_range = 8;
 614         else if (width <= 2560) pbi->sync_range =16;
 615         else pbi->sync_range = 32;
 616
 617         uv_width = width >>1;
 618
 619         // Allocate an int for each mb row.
 620         CHECK_MEM_ERROR(pbi->mt_current_mb_col, vpx_malloc(sizeof(int) * pc->mb_rows));
 621
 622         // Allocate memory for above_row buffers.
 623         CHECK_MEM_ERROR(pbi->mt_yabove_row, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 624         for (i=0; i< pc->mb_rows; i++)
 625             CHECK_MEM_ERROR(pbi->mt_yabove_row[i], vpx_calloc(sizeof(unsigned char) * (width + (VP8BORDERINPIXELS<<1)), 1));
 626
 627         CHECK_MEM_ERROR(pbi->mt_uabove_row, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 628         for (i=0; i< pc->mb_rows; i++)
 629             CHECK_MEM_ERROR(pbi->mt_uabove_row[i], vpx_calloc(sizeof(unsigned char) * (uv_width + VP8BORDERINPIXELS), 1));
 630
 631         CHECK_MEM_ERROR(pbi->mt_vabove_row, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 632         for (i=0; i< pc->mb_rows; i++)
 633             CHECK_MEM_ERROR(pbi->mt_vabove_row[i], vpx_calloc(sizeof(unsigned char) * (uv_width + VP8BORDERINPIXELS), 1));
 634
 635         // Allocate memory for left_col buffers.
 636         CHECK_MEM_ERROR(pbi->mt_yleft_col, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 637         for (i=0; i< pc->mb_rows; i++)
 638             CHECK_MEM_ERROR(pbi->mt_yleft_col[i], vpx_calloc(sizeof(unsigned char) * 16, 1));
 639
 640         CHECK_MEM_ERROR(pbi->mt_uleft_col, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 641         for (i=0; i< pc->mb_rows; i++)
 642             CHECK_MEM_ERROR(pbi->mt_uleft_col[i], vpx_calloc(sizeof(unsigned char) * 8, 1));
 643
 644         CHECK_MEM_ERROR(pbi->mt_vleft_col, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 645         for (i=0; i< pc->mb_rows; i++)
 646             CHECK_MEM_ERROR(pbi->mt_vleft_col[i], vpx_calloc(sizeof(unsigned char) * 8, 1));
 647     }
 648     return 0;
 649 #else
 650     (void) pbi;
 651     (void) width;
 652 #endif
 653 }
 654
 655
 656 void vp8_decoder_remove_threads(VP8D_COMP *pbi)
 657 {
 658 #if CONFIG_MULTITHREAD
 659
 660     //shutdown MB Decoding thread;
 661     if (pbi->b_multithreaded_rd)
 662     {
 663         int i;
 664
 665         pbi->b_multithreaded_rd = 0;
 666
 667         // allow all threads to exit
 668         for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
 669         {
 670             sem_post(&pbi->h_event_start_decoding[i]);
 671             pthread_join(pbi->h_decoding_thread[i], NULL);
 672         }
 673
 674         for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
 675         {
 676             sem_destroy(&pbi->h_event_start_decoding[i]);
 677         }
 678
 679         sem_destroy(&pbi->h_event_end_decoding);
 680
 681         if (pbi->h_decoding_thread)
 682         {
 683             vpx_free(pbi->h_decoding_thread);
 684             pbi->h_decoding_thread = NULL;
 685         }
 686
 687         if (pbi->h_event_start_decoding)
 688         {
 689             vpx_free(pbi->h_event_start_decoding);
 690             pbi->h_event_start_decoding = NULL;
 691         }
 692
 693         if (pbi->mb_row_di)
 694         {
 695             vpx_free(pbi->mb_row_di);
 696             pbi->mb_row_di = NULL ;
 697         }
 698
 699         if (pbi->de_thread_data)
 700         {
 701             vpx_free(pbi->de_thread_data);
 702             pbi->de_thread_data = NULL;
 703         }
 704     }
 705 #else
 706     (void) pbi;
 707 #endif
 708 }
 709
 710
 711 void vp8mt_lpf_init( VP8D_COMP *pbi, int default_filt_lvl)
 712 {
 713 #if CONFIG_MULTITHREAD
 714     VP8_COMMON *cm  = &pbi->common;
 715     MACROBLOCKD *mbd = &pbi->mb;
 716     //YV12_BUFFER_CONFIG *post = &cm->new_frame;  //frame_to_show;
 717     loop_filter_info *lfi = cm->lf_info;
 718     int frame_type = cm->frame_type;
 719
 720     //int mb_row;
 721     //int mb_col;
 722     //int baseline_filter_level[MAX_MB_SEGMENTS];
 723     int filter_level;
 724     int alt_flt_enabled = mbd->segmentation_enabled;
 725
 726     int i;
 727     //unsigned char *y_ptr, *u_ptr, *v_ptr;
 728
 729     // Note the baseline filter values for each segment
 730     if (alt_flt_enabled)
 731     {
 732         for (i = 0; i < MAX_MB_SEGMENTS; i++)
 733         {
 734             // Abs value
 735             if (mbd->mb_segement_abs_delta == SEGMENT_ABSDATA)
 736                 pbi->mt_baseline_filter_level[i] = mbd->segment_feature_data[MB_LVL_ALT_LF][i];
 737             // Delta Value
 738             else
 739             {
 740                 pbi->mt_baseline_filter_level[i] = default_filt_lvl + mbd->segment_feature_data[MB_LVL_ALT_LF][i];
 741                 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
 742             }
 743         }
 744     }
 745     else
 746     {
 747         for (i = 0; i < MAX_MB_SEGMENTS; i++)
 748             pbi->mt_baseline_filter_level[i] = default_filt_lvl;
 749     }
 750
 751     // Initialize the loop filter for this frame.
 752     if ((cm->last_filter_type != cm->filter_type) || (cm->last_sharpness_level != cm->sharpness_level))
 753         vp8_init_loop_filter(cm);
 754     else if (frame_type != cm->last_frame_type)
 755         vp8_frame_init_loop_filter(lfi, frame_type);
 756 #else
 757     (void) pbi;
 758     (void) default_filt_lvl;
 759 #endif
 760 }
 761
 762
 763 void vp8mt_decode_mb_rows( VP8D_COMP *pbi, MACROBLOCKD *xd)
 764 {
 765 #if CONFIG_MULTITHREAD
 766     int mb_row;
 767     VP8_COMMON *pc = &pbi->common;
 768
 769     int ibc = 0;
 770     int num_part = 1 << pbi->common.multi_token_partition;
 771     int i, j;
 772     volatile int *last_row_current_mb_col = NULL;
 773     int nsync = pbi->sync_range;
 774
 775     int filter_level;
 776     loop_filter_info *lfi = pc->lf_info;
 777     int alt_flt_enabled = xd->segmentation_enabled;
 778     int Segment;
 779
 780     if(pbi->common.filter_level)
 781     {
 782         //Set above_row buffer to 127 for decoding first MB row
 783         vpx_memset(pbi->mt_yabove_row[0] + VP8BORDERINPIXELS-1, 127, pc->yv12_fb[pc->lst_fb_idx].y_width + 5);
 784         vpx_memset(pbi->mt_uabove_row[0] + (VP8BORDERINPIXELS>>1)-1, 127, (pc->yv12_fb[pc->lst_fb_idx].y_width>>1) +5);
 785         vpx_memset(pbi->mt_vabove_row[0] + (VP8BORDERINPIXELS>>1)-1, 127, (pc->yv12_fb[pc->lst_fb_idx].y_width>>1) +5);
 786
 787         for (i=1; i<pc->mb_rows; i++)
 788         {
 789             vpx_memset(pbi->mt_yabove_row[i] + VP8BORDERINPIXELS-1, (unsigned char)129, 1);
 790             vpx_memset(pbi->mt_uabove_row[i] + (VP8BORDERINPIXELS>>1)-1, (unsigned char)129, 1);
 791             vpx_memset(pbi->mt_vabove_row[i] + (VP8BORDERINPIXELS>>1)-1, (unsigned char)129, 1);
 792         }
 793
 794         //Set left_col to 129 initially
 795         for (i=0; i<pc->mb_rows; i++)
 796         {
 797             vpx_memset(pbi->mt_yleft_col[i], (unsigned char)129, 16);
 798             vpx_memset(pbi->mt_uleft_col[i], (unsigned char)129, 8);
 799             vpx_memset(pbi->mt_vleft_col[i], (unsigned char)129, 8);
 800         }
 801         vp8mt_lpf_init(pbi, pc->filter_level);
 802     }
 803
 804     vp8_setup_decoding_thread_data(pbi, xd, pbi->mb_row_di, pbi->decoding_thread_count);
 805
 806     for (i = 0; i < pbi->decoding_thread_count; i++)
 807         sem_post(&pbi->h_event_start_decoding[i]);
 808
 809     for (mb_row = 0; mb_row < pc->mb_rows; mb_row += (pbi->decoding_thread_count + 1))
 810     {
 811         int i;
 812
 813         xd->current_bc = &pbi->mbc[mb_row%num_part];
 814
 815         //vp8_decode_mb_row(pbi, pc, mb_row, xd);
 816         {
 817             int i;
 818             int recon_yoffset, recon_uvoffset;
 819             int mb_col;
 820             int ref_fb_idx = pc->lst_fb_idx;
 821             int dst_fb_idx = pc->new_fb_idx;
 822             int recon_y_stride = pc->yv12_fb[ref_fb_idx].y_stride;
 823             int recon_uv_stride = pc->yv12_fb[ref_fb_idx].uv_stride;
 824
 825            // volatile int *last_row_current_mb_col = NULL;
 826             if (mb_row > 0)
 827                 last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row -1];
 828
 829             vpx_memset(&pc->left_context, 0, sizeof(pc->left_context));
 830             recon_yoffset = mb_row * recon_y_stride * 16;
 831             recon_uvoffset = mb_row * recon_uv_stride * 8;
 832             // reset above block coeffs
 833
 834             xd->above_context = pc->above_context;
 835             xd->up_available = (mb_row != 0);
 836
 837             xd->mb_to_top_edge = -((mb_row * 16)) << 3;
 838             xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
 839
 840             for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
 841             {
 842                 if ( mb_row > 0 && (mb_col & (nsync-1)) == 0){
 843                     while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != pc->mb_cols - 1)
 844                     {
 845                         x86_pause_hint();
 846                         thread_sleep(0);
 847                     }
 848                 }
 849
 850                 if (xd->mode_info_context->mbmi.mode == SPLITMV || xd->mode_info_context->mbmi.mode == B_PRED)
 851                 {
 852                     for (i = 0; i < 16; i++)
 853                     {
 854                         BLOCKD *d = &xd->block[i];
 855                         vpx_memcpy(&d->bmi, &xd->mode_info_context->bmi[i], sizeof(B_MODE_INFO));
 856                     }
 857                 }
 858
 859                 if(pbi->common.filter_level)
 860                 {
 861                     //update loopfilter info
 862                     Segment = (alt_flt_enabled) ? xd->mode_info_context->mbmi.segment_id : 0;
 863                     filter_level = pbi->mt_baseline_filter_level[Segment];
 864                     // Distance of Mb to the various image edges.
 865                     // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
 866                     // Apply any context driven MB level adjustment
 867                     vp8_adjust_mb_lf_value(xd, &filter_level);
 868                 }
 869
 870                 // Distance of Mb to the various image edges.
 871                 // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
 872                 xd->mb_to_left_edge = -((mb_col * 16) << 3);
 873                 xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
 874
 875                 xd->dst.y_buffer = pc->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 876                 xd->dst.u_buffer = pc->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 877                 xd->dst.v_buffer = pc->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 878
 879                 xd->left_available = (mb_col != 0);
 880
 881                 // Select the appropriate reference frame for this MB
 882                 if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
 883                     ref_fb_idx = pc->lst_fb_idx;
 884                 else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
 885                     ref_fb_idx = pc->gld_fb_idx;
 886                 else
 887                     ref_fb_idx = pc->alt_fb_idx;
 888
 889                 xd->pre.y_buffer = pc->yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
 890                 xd->pre.u_buffer = pc->yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
 891                 xd->pre.v_buffer = pc->yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
 892
 893                 vp8_build_uvmvs(xd, pc->full_pixel);
 894                 vp8mt_decode_macroblock(pbi, xd, mb_row, mb_col);
 895
 896                 if (pbi->common.filter_level)
 897                 {
 898                     //Save decoded MB last row data for next-row decoding
 899                     if(mb_row != pc->mb_rows-1)
 900                     {
 901                         vpx_memcpy((pbi->mt_yabove_row[mb_row +1] + 32 + mb_col*16), (xd->dst.y_buffer + 15 * recon_y_stride), 16);
 902                         vpx_memcpy((pbi->mt_uabove_row[mb_row +1] + 16 + mb_col*8), (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
 903                         vpx_memcpy((pbi->mt_vabove_row[mb_row +1] + 16 + mb_col*8), (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
 904                     }
 905
 906                     //save left_col for next MB decoding
 907                     if(mb_col != pc->mb_cols-1)
 908                     {
 909                         MODE_INFO *next = xd->mode_info_context +1;
 910
 911                         if (xd->frame_type == KEY_FRAME  ||  next->mbmi.ref_frame == INTRA_FRAME)
 912                         {
 913                             for (i = 0; i < 16; i++)
 914                                 pbi->mt_yleft_col[mb_row][i] = xd->dst.y_buffer [i* recon_y_stride + 15];
 915                             for (i = 0; i < 8; i++)
 916                             {
 917                                 pbi->mt_uleft_col[mb_row][i] = xd->dst.u_buffer [i* recon_uv_stride + 7];
 918                                 pbi->mt_vleft_col[mb_row][i] = xd->dst.v_buffer [i* recon_uv_stride + 7];
 919                             }
 920                         }
 921                     }
 922
 923                     // loopfilter on this macroblock.
 924                     if (filter_level)
 925                     {
 926                         if (mb_col > 0)
 927                             pc->lf_mbv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level], pc->simpler_lpf);
 928
 929                         if (xd->mode_info_context->mbmi.dc_diff > 0)
 930                             pc->lf_bv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level], pc->simpler_lpf);
 931
 932                         // don't apply across umv border
 933                         if (mb_row > 0)
 934                             pc->lf_mbh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level], pc->simpler_lpf);
 935
 936                         if (xd->mode_info_context->mbmi.dc_diff > 0)
 937                             pc->lf_bh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level], pc->simpler_lpf);
 938                     }
 939                 }
 940
 941                 recon_yoffset += 16;
 942                 recon_uvoffset += 8;
 943
 944                 ++xd->mode_info_context;  /* next mb */
 945
 946                 xd->above_context++;
 947
 948                 pbi->mt_current_mb_col[mb_row] = mb_col;
 949             }
 950
 951             // adjust to the next row of mbs
 952             if (pbi->common.filter_level)
 953             {
 954                 if(mb_row != pc->mb_rows-1)
 955                 {
 956                     int lasty = pc->yv12_fb[ref_fb_idx].y_width + VP8BORDERINPIXELS;
 957                     int lastuv = (pc->yv12_fb[ref_fb_idx].y_width>>1) + (VP8BORDERINPIXELS>>1);
 958
 959                     for (i = 0; i < 4; i++)
 960                     {
 961                         pbi->mt_yabove_row[mb_row +1][lasty + i] = pbi->mt_yabove_row[mb_row +1][lasty -1];
 962                         pbi->mt_uabove_row[mb_row +1][lastuv + i] = pbi->mt_uabove_row[mb_row +1][lastuv -1];
 963                         pbi->mt_vabove_row[mb_row +1][lastuv + i] = pbi->mt_vabove_row[mb_row +1][lastuv -1];
 964                     }
 965                 }
 966             }else
 967                 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);
 968
 969             ++xd->mode_info_context;      /* skip prediction column */
 970         }
 971         xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
 972     }
 973
 974     sem_wait(&pbi->h_event_end_decoding);   // add back for each frame
 975 #else
 976     (void) pbi;
 977     (void) xd;
 978 #endif
 979 }