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     VP8_COMMON *pc = &pbi->common;
  95
  96     if (xd->mode_info_context->mbmi.mb_skip_coeff)
  97     {
  98         vp8_reset_mb_tokens_context(xd);
  99     }
 100     else
 101     {
 102         eobtotal = vp8_decode_mb_tokens(pbi, xd);
 103     }
 104
 105     /* Perform temporary clamping of the MV to be used for prediction */
 106     if (do_clamp)
 107     {
 108         clamp_mvs(xd);
 109     }
 110
 111     eobtotal |= (xd->mode_info_context->mbmi.mode == B_PRED ||
 112                   xd->mode_info_context->mbmi.mode == SPLITMV);
 113     if (!eobtotal)
 114     {
 115         /* Special case:  Force the loopfilter to skip when eobtotal and
 116          * mb_skip_coeff are zero.
 117          * */
 118         xd->mode_info_context->mbmi.mb_skip_coeff = 1;
 119
 120         /*mt_skip_recon_mb(pbi, xd, mb_row, mb_col);*/
 121         if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
 122         {
 123             vp8mt_build_intra_predictors_mbuv_s(pbi, xd, mb_row, mb_col);
 124             vp8mt_build_intra_predictors_mby_s(pbi, xd, mb_row, mb_col);
 125         }
 126         else
 127         {
 128             vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer,
 129                                                xd->dst.u_buffer, xd->dst.v_buffer,
 130                                                xd->dst.y_stride, xd->dst.uv_stride);
 131         }
 132         return;
 133     }
 134
 135     if (xd->segmentation_enabled)
 136         mb_init_dequantizer(pbi, xd);
 137
 138     /* do prediction */
 139     if (xd->frame_type == KEY_FRAME  ||  xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
 140     {
 141         vp8mt_build_intra_predictors_mbuv(pbi, xd, mb_row, mb_col);
 142
 143         if (xd->mode_info_context->mbmi.mode != B_PRED)
 144         {
 145             vp8mt_build_intra_predictors_mby(pbi, xd, mb_row, mb_col);
 146         } else {
 147             vp8mt_intra_prediction_down_copy(pbi, xd, mb_row, mb_col);
 148         }
 149     }
 150     else
 151     {
 152         vp8_build_inter_predictors_mb(xd);
 153     }
 154
 155     /* dequantization and idct */
 156     if (xd->mode_info_context->mbmi.mode != B_PRED && xd->mode_info_context->mbmi.mode != SPLITMV)
 157     {
 158         BLOCKD *b = &xd->block[24];
 159         DEQUANT_INVOKE(&pbi->dequant, block)(b);
 160
 161         /* do 2nd order transform on the dc block */
 162         if (xd->eobs[24] > 1)
 163         {
 164             IDCT_INVOKE(RTCD_VTABLE(idct), iwalsh16)(&b->dqcoeff[0], b->diff);
 165             ((int *)b->qcoeff)[0] = 0;
 166             ((int *)b->qcoeff)[1] = 0;
 167             ((int *)b->qcoeff)[2] = 0;
 168             ((int *)b->qcoeff)[3] = 0;
 169             ((int *)b->qcoeff)[4] = 0;
 170             ((int *)b->qcoeff)[5] = 0;
 171             ((int *)b->qcoeff)[6] = 0;
 172             ((int *)b->qcoeff)[7] = 0;
 173         }
 174         else
 175         {
 176             IDCT_INVOKE(RTCD_VTABLE(idct), iwalsh1)(&b->dqcoeff[0], b->diff);
 177             ((int *)b->qcoeff)[0] = 0;
 178         }
 179
 180         DEQUANT_INVOKE (&pbi->dequant, dc_idct_add_y_block)
 181                         (xd->qcoeff, xd->block[0].dequant,
 182                          xd->predictor, xd->dst.y_buffer,
 183                          xd->dst.y_stride, xd->eobs, xd->block[24].diff);
 184     }
 185     else if ((xd->frame_type == KEY_FRAME  ||  xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) && xd->mode_info_context->mbmi.mode == B_PRED)
 186     {
 187         for (i = 0; i < 16; i++)
 188         {
 189             BLOCKD *b = &xd->block[i];
 190             vp8mt_predict_intra4x4(pbi, xd, b->bmi.mode, b->predictor, mb_row, mb_col, i);
 191
 192             if (xd->eobs[i] > 1)
 193             {
 194                 DEQUANT_INVOKE(&pbi->dequant, idct_add)
 195                     (b->qcoeff, b->dequant,  b->predictor,
 196                     *(b->base_dst) + b->dst, 16, b->dst_stride);
 197             }
 198             else
 199             {
 200                 IDCT_INVOKE(RTCD_VTABLE(idct), idct1_scalar_add)
 201                     (b->qcoeff[0] * b->dequant[0], b->predictor,
 202                     *(b->base_dst) + b->dst, 16, b->dst_stride);
 203                 ((int *)b->qcoeff)[0] = 0;
 204             }
 205         }
 206     }
 207     else
 208     {
 209         DEQUANT_INVOKE (&pbi->dequant, idct_add_y_block)
 210                         (xd->qcoeff, xd->block[0].dequant,
 211                          xd->predictor, xd->dst.y_buffer,
 212                          xd->dst.y_stride, xd->eobs);
 213     }
 214
 215     DEQUANT_INVOKE (&pbi->dequant, idct_add_uv_block)
 216                     (xd->qcoeff+16*16, xd->block[16].dequant,
 217                      xd->predictor+16*16, xd->dst.u_buffer, xd->dst.v_buffer,
 218                      xd->dst.uv_stride, xd->eobs+16);
 219 }
 220
 221
 222 static THREAD_FUNCTION thread_decoding_proc(void *p_data)
 223 {
 224     int ithread = ((DECODETHREAD_DATA *)p_data)->ithread;
 225     VP8D_COMP *pbi = (VP8D_COMP *)(((DECODETHREAD_DATA *)p_data)->ptr1);
 226     MB_ROW_DEC *mbrd = (MB_ROW_DEC *)(((DECODETHREAD_DATA *)p_data)->ptr2);
 227     ENTROPY_CONTEXT_PLANES mb_row_left_context;
 228
 229     while (1)
 230     {
 231         if (pbi->b_multithreaded_rd == 0)
 232             break;
 233
 234         /*if(WaitForSingleObject(pbi->h_event_start_decoding[ithread], INFINITE) == WAIT_OBJECT_0)*/
 235         if (sem_wait(&pbi->h_event_start_decoding[ithread]) == 0)
 236         {
 237             if (pbi->b_multithreaded_rd == 0)
 238                 break;
 239             else
 240             {
 241                 VP8_COMMON *pc = &pbi->common;
 242                 MACROBLOCKD *xd = &mbrd->mbd;
 243
 244                 int mb_row;
 245                 int num_part = 1 << pbi->common.multi_token_partition;
 246                 volatile int *last_row_current_mb_col;
 247                 int nsync = pbi->sync_range;
 248
 249                 for (mb_row = ithread+1; mb_row < pc->mb_rows; mb_row += (pbi->decoding_thread_count + 1))
 250                 {
 251                     int i;
 252                     int recon_yoffset, recon_uvoffset;
 253                     int mb_col;
 254                     int ref_fb_idx = pc->lst_fb_idx;
 255                     int dst_fb_idx = pc->new_fb_idx;
 256                     int recon_y_stride = pc->yv12_fb[ref_fb_idx].y_stride;
 257                     int recon_uv_stride = pc->yv12_fb[ref_fb_idx].uv_stride;
 258
 259                     int filter_level;
 260                     loop_filter_info *lfi = pc->lf_info;
 261                     int alt_flt_enabled = xd->segmentation_enabled;
 262                     int Segment;
 263
 264                     pbi->mb_row_di[ithread].mb_row = mb_row;
 265                     pbi->mb_row_di[ithread].mbd.current_bc =  &pbi->mbc[mb_row%num_part];
 266
 267                     last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row -1];
 268
 269                     recon_yoffset = mb_row * recon_y_stride * 16;
 270                     recon_uvoffset = mb_row * recon_uv_stride * 8;
 271                     /* reset above block coeffs */
 272
 273                     xd->above_context = pc->above_context;
 274                     xd->left_context = &mb_row_left_context;
 275                     vpx_memset(&mb_row_left_context, 0, sizeof(mb_row_left_context));
 276                     xd->up_available = (mb_row != 0);
 277
 278                     xd->mb_to_top_edge = -((mb_row * 16)) << 3;
 279                     xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
 280
 281                     for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
 282                     {
 283                         if ((mb_col & (nsync-1)) == 0)
 284                         {
 285                             while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != pc->mb_cols - 1)
 286                             {
 287                                 x86_pause_hint();
 288                                 thread_sleep(0);
 289                             }
 290                         }
 291
 292                         if (xd->mode_info_context->mbmi.mode == SPLITMV || xd->mode_info_context->mbmi.mode == B_PRED)
 293                         {
 294                             for (i = 0; i < 16; i++)
 295                             {
 296                                 BLOCKD *d = &xd->block[i];
 297                                 vpx_memcpy(&d->bmi, &xd->mode_info_context->bmi[i], sizeof(B_MODE_INFO));
 298                             }
 299                         }
 300
 301                         /* Distance of Mb to the various image edges.
 302                          * These are specified to 8th pel as they are always compared to values that are in 1/8th pel units
 303                          */
 304                         xd->mb_to_left_edge = -((mb_col * 16) << 3);
 305                         xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
 306
 307                         xd->dst.y_buffer = pc->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 308                         xd->dst.u_buffer = pc->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 309                         xd->dst.v_buffer = pc->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 310
 311                         xd->left_available = (mb_col != 0);
 312
 313                         /* Select the appropriate reference frame for this MB */
 314                         if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
 315                             ref_fb_idx = pc->lst_fb_idx;
 316                         else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
 317                             ref_fb_idx = pc->gld_fb_idx;
 318                         else
 319                             ref_fb_idx = pc->alt_fb_idx;
 320
 321                         xd->pre.y_buffer = pc->yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
 322                         xd->pre.u_buffer = pc->yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
 323                         xd->pre.v_buffer = pc->yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
 324
 325                         vp8_build_uvmvs(xd, pc->full_pixel);
 326                         decode_macroblock(pbi, xd, mb_row, mb_col);
 327
 328                         if (pbi->common.filter_level)
 329                         {
 330                             int skip_lf;
 331                             if( mb_row != pc->mb_rows-1 )
 332                             {
 333                                 /* Save decoded MB last row data for next-row decoding */
 334                                 vpx_memcpy((pbi->mt_yabove_row[mb_row + 1] + 32 + mb_col*16), (xd->dst.y_buffer + 15 * recon_y_stride), 16);
 335                                 vpx_memcpy((pbi->mt_uabove_row[mb_row + 1] + 16 + mb_col*8), (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
 336                                 vpx_memcpy((pbi->mt_vabove_row[mb_row + 1] + 16 + mb_col*8), (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
 337                             }
 338
 339                             /* save left_col for next MB decoding */
 340                             if(mb_col != pc->mb_cols-1)
 341                             {
 342                                 MODE_INFO *next = xd->mode_info_context +1;
 343
 344                                 if (xd->frame_type == KEY_FRAME  ||  next->mbmi.ref_frame == INTRA_FRAME)
 345                                 {
 346                                     for (i = 0; i < 16; i++)
 347                                         pbi->mt_yleft_col[mb_row][i] = xd->dst.y_buffer [i* recon_y_stride + 15];
 348                                     for (i = 0; i < 8; i++)
 349                                     {
 350                                         pbi->mt_uleft_col[mb_row][i] = xd->dst.u_buffer [i* recon_uv_stride + 7];
 351                                         pbi->mt_vleft_col[mb_row][i] = xd->dst.v_buffer [i* recon_uv_stride + 7];
 352                                     }
 353                                 }
 354                             }
 355
 356                             /* update loopfilter info */
 357                             Segment = (alt_flt_enabled) ? xd->mode_info_context->mbmi.segment_id : 0;
 358                             skip_lf = (xd->mode_info_context->mbmi.mode != B_PRED &&
 359                                             xd->mode_info_context->mbmi.mode != SPLITMV &&
 360                                             xd->mode_info_context->mbmi.mb_skip_coeff);
 361
 362                             filter_level = pbi->mt_baseline_filter_level[Segment];
 363                             /* Distance of Mb to the various image edges.
 364                              * These are specified to 8th pel as they are always compared to values that are in 1/8th pel units
 365                              * Apply any context driven MB level adjustment
 366                              */
 367                             filter_level = vp8_adjust_mb_lf_value(xd, filter_level);
 368
 369                             /* loopfilter on this macroblock. */
 370                             if (filter_level)
 371                             {
 372                                 if (mb_col > 0)
 373                                     pc->lf_mbv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 374
 375                                 if (!skip_lf)
 376                                     pc->lf_bv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 377
 378                                 /* don't apply across umv border */
 379                                 if (mb_row > 0)
 380                                     pc->lf_mbh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 381
 382                                 if (!skip_lf)
 383                                     pc->lf_bh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 384                             }
 385                         }
 386
 387                         recon_yoffset += 16;
 388                         recon_uvoffset += 8;
 389
 390                         ++xd->mode_info_context;  /* next mb */
 391
 392                         xd->above_context++;
 393
 394                         /*pbi->mb_row_di[ithread].current_mb_col = mb_col;*/
 395                         pbi->mt_current_mb_col[mb_row] = mb_col;
 396                     }
 397
 398                     /* adjust to the next row of mbs */
 399                     if (pbi->common.filter_level)
 400                     {
 401                         if(mb_row != pc->mb_rows-1)
 402                         {
 403                             int lasty = pc->yv12_fb[ref_fb_idx].y_width + VP8BORDERINPIXELS;
 404                             int lastuv = (pc->yv12_fb[ref_fb_idx].y_width>>1) + (VP8BORDERINPIXELS>>1);
 405
 406                             for (i = 0; i < 4; i++)
 407                             {
 408                                 pbi->mt_yabove_row[mb_row +1][lasty + i] = pbi->mt_yabove_row[mb_row +1][lasty -1];
 409                                 pbi->mt_uabove_row[mb_row +1][lastuv + i] = pbi->mt_uabove_row[mb_row +1][lastuv -1];
 410                                 pbi->mt_vabove_row[mb_row +1][lastuv + i] = pbi->mt_vabove_row[mb_row +1][lastuv -1];
 411                             }
 412                         }
 413                     } else
 414                         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);
 415
 416                     ++xd->mode_info_context;      /* skip prediction column */
 417
 418                     /* since we have multithread */
 419                     xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
 420                 }
 421             }
 422         }
 423         /*  add this to each frame */
 424         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))
 425         {
 426             /*SetEvent(pbi->h_event_end_decoding);*/
 427             sem_post(&pbi->h_event_end_decoding);
 428         }
 429     }
 430
 431     return 0 ;
 432 }
 433
 434
 435 void vp8_decoder_create_threads(VP8D_COMP *pbi)
 436 {
 437     int core_count = 0;
 438     int ithread;
 439
 440     pbi->b_multithreaded_rd = 0;
 441     pbi->allocated_decoding_thread_count = 0;
 442
 443     /* limit decoding threads to the max number of token partitions */
 444     core_count = (pbi->max_threads > 8) ? 8 : pbi->max_threads;
 445
 446     /* limit decoding threads to the available cores */
 447     if (core_count > pbi->common.processor_core_count)
 448         core_count = pbi->common.processor_core_count;
 449
 450     if (core_count > 1)
 451     {
 452         pbi->b_multithreaded_rd = 1;
 453         pbi->decoding_thread_count = core_count - 1;
 454
 455         CHECK_MEM_ERROR(pbi->h_decoding_thread, vpx_malloc(sizeof(pthread_t) * pbi->decoding_thread_count));
 456         CHECK_MEM_ERROR(pbi->h_event_start_decoding, vpx_malloc(sizeof(sem_t) * pbi->decoding_thread_count));
 457         CHECK_MEM_ERROR(pbi->mb_row_di, vpx_memalign(32, sizeof(MB_ROW_DEC) * pbi->decoding_thread_count));
 458         vpx_memset(pbi->mb_row_di, 0, sizeof(MB_ROW_DEC) * pbi->decoding_thread_count);
 459         CHECK_MEM_ERROR(pbi->de_thread_data, vpx_malloc(sizeof(DECODETHREAD_DATA) * pbi->decoding_thread_count));
 460
 461         for (ithread = 0; ithread < pbi->decoding_thread_count; ithread++)
 462         {
 463             sem_init(&pbi->h_event_start_decoding[ithread], 0, 0);
 464
 465             pbi->de_thread_data[ithread].ithread  = ithread;
 466             pbi->de_thread_data[ithread].ptr1     = (void *)pbi;
 467             pbi->de_thread_data[ithread].ptr2     = (void *) &pbi->mb_row_di[ithread];
 468
 469             pthread_create(&pbi->h_decoding_thread[ithread], 0, thread_decoding_proc, (&pbi->de_thread_data[ithread]));
 470         }
 471
 472         sem_init(&pbi->h_event_end_decoding, 0, 0);
 473
 474         pbi->allocated_decoding_thread_count = pbi->decoding_thread_count;
 475     }
 476 }
 477
 478
 479 void vp8mt_de_alloc_temp_buffers(VP8D_COMP *pbi, int mb_rows)
 480 {
 481     VP8_COMMON *const pc = & pbi->common;
 482     int i;
 483
 484     if (pbi->b_multithreaded_rd)
 485     {
 486             vpx_free(pbi->mt_current_mb_col);
 487             pbi->mt_current_mb_col = NULL ;
 488
 489         /* Free above_row buffers. */
 490         if (pbi->mt_yabove_row)
 491         {
 492             for (i=0; i< mb_rows; i++)
 493             {
 494                     vpx_free(pbi->mt_yabove_row[i]);
 495                     pbi->mt_yabove_row[i] = NULL ;
 496             }
 497             vpx_free(pbi->mt_yabove_row);
 498             pbi->mt_yabove_row = NULL ;
 499         }
 500
 501         if (pbi->mt_uabove_row)
 502         {
 503             for (i=0; i< mb_rows; i++)
 504             {
 505                     vpx_free(pbi->mt_uabove_row[i]);
 506                     pbi->mt_uabove_row[i] = NULL ;
 507             }
 508             vpx_free(pbi->mt_uabove_row);
 509             pbi->mt_uabove_row = NULL ;
 510         }
 511
 512         if (pbi->mt_vabove_row)
 513         {
 514             for (i=0; i< mb_rows; i++)
 515             {
 516                     vpx_free(pbi->mt_vabove_row[i]);
 517                     pbi->mt_vabove_row[i] = NULL ;
 518             }
 519             vpx_free(pbi->mt_vabove_row);
 520             pbi->mt_vabove_row = NULL ;
 521         }
 522
 523         /* Free left_col buffers. */
 524         if (pbi->mt_yleft_col)
 525         {
 526             for (i=0; i< mb_rows; i++)
 527             {
 528                     vpx_free(pbi->mt_yleft_col[i]);
 529                     pbi->mt_yleft_col[i] = NULL ;
 530             }
 531             vpx_free(pbi->mt_yleft_col);
 532             pbi->mt_yleft_col = NULL ;
 533         }
 534
 535         if (pbi->mt_uleft_col)
 536         {
 537             for (i=0; i< mb_rows; i++)
 538             {
 539                     vpx_free(pbi->mt_uleft_col[i]);
 540                     pbi->mt_uleft_col[i] = NULL ;
 541             }
 542             vpx_free(pbi->mt_uleft_col);
 543             pbi->mt_uleft_col = NULL ;
 544         }
 545
 546         if (pbi->mt_vleft_col)
 547         {
 548             for (i=0; i< mb_rows; i++)
 549             {
 550                     vpx_free(pbi->mt_vleft_col[i]);
 551                     pbi->mt_vleft_col[i] = NULL ;
 552             }
 553             vpx_free(pbi->mt_vleft_col);
 554             pbi->mt_vleft_col = NULL ;
 555         }
 556     }
 557 }
 558
 559
 560 void vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows)
 561 {
 562     VP8_COMMON *const pc = & pbi->common;
 563     int i;
 564     int uv_width;
 565
 566     if (pbi->b_multithreaded_rd)
 567     {
 568         vp8mt_de_alloc_temp_buffers(pbi, prev_mb_rows);
 569
 570         /* our internal buffers are always multiples of 16 */
 571         if ((width & 0xf) != 0)
 572             width += 16 - (width & 0xf);
 573
 574         if (width < 640) pbi->sync_range = 1;
 575         else if (width <= 1280) pbi->sync_range = 8;
 576         else if (width <= 2560) pbi->sync_range =16;
 577         else pbi->sync_range = 32;
 578
 579         uv_width = width >>1;
 580
 581         /* Allocate an int for each mb row. */
 582         CHECK_MEM_ERROR(pbi->mt_current_mb_col, vpx_malloc(sizeof(int) * pc->mb_rows));
 583
 584         /* Allocate memory for above_row buffers. */
 585         CHECK_MEM_ERROR(pbi->mt_yabove_row, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 586         for (i=0; i< pc->mb_rows; i++)
 587             CHECK_MEM_ERROR(pbi->mt_yabove_row[i], vpx_calloc(sizeof(unsigned char) * (width + (VP8BORDERINPIXELS<<1)), 1));
 588
 589         CHECK_MEM_ERROR(pbi->mt_uabove_row, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 590         for (i=0; i< pc->mb_rows; i++)
 591             CHECK_MEM_ERROR(pbi->mt_uabove_row[i], vpx_calloc(sizeof(unsigned char) * (uv_width + VP8BORDERINPIXELS), 1));
 592
 593         CHECK_MEM_ERROR(pbi->mt_vabove_row, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 594         for (i=0; i< pc->mb_rows; i++)
 595             CHECK_MEM_ERROR(pbi->mt_vabove_row[i], vpx_calloc(sizeof(unsigned char) * (uv_width + VP8BORDERINPIXELS), 1));
 596
 597         /* Allocate memory for left_col buffers. */
 598         CHECK_MEM_ERROR(pbi->mt_yleft_col, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 599         for (i=0; i< pc->mb_rows; i++)
 600             CHECK_MEM_ERROR(pbi->mt_yleft_col[i], vpx_calloc(sizeof(unsigned char) * 16, 1));
 601
 602         CHECK_MEM_ERROR(pbi->mt_uleft_col, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 603         for (i=0; i< pc->mb_rows; i++)
 604             CHECK_MEM_ERROR(pbi->mt_uleft_col[i], vpx_calloc(sizeof(unsigned char) * 8, 1));
 605
 606         CHECK_MEM_ERROR(pbi->mt_vleft_col, vpx_malloc(sizeof(unsigned char *) * pc->mb_rows));
 607         for (i=0; i< pc->mb_rows; i++)
 608             CHECK_MEM_ERROR(pbi->mt_vleft_col[i], vpx_calloc(sizeof(unsigned char) * 8, 1));
 609     }
 610 }
 611
 612
 613 void vp8_decoder_remove_threads(VP8D_COMP *pbi)
 614 {
 615     /* shutdown MB Decoding thread; */
 616     if (pbi->b_multithreaded_rd)
 617     {
 618         int i;
 619
 620         pbi->b_multithreaded_rd = 0;
 621
 622         /* allow all threads to exit */
 623         for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
 624         {
 625             sem_post(&pbi->h_event_start_decoding[i]);
 626             pthread_join(pbi->h_decoding_thread[i], NULL);
 627         }
 628
 629         for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
 630         {
 631             sem_destroy(&pbi->h_event_start_decoding[i]);
 632         }
 633
 634         sem_destroy(&pbi->h_event_end_decoding);
 635
 636             vpx_free(pbi->h_decoding_thread);
 637             pbi->h_decoding_thread = NULL;
 638
 639             vpx_free(pbi->h_event_start_decoding);
 640             pbi->h_event_start_decoding = NULL;
 641
 642             vpx_free(pbi->mb_row_di);
 643             pbi->mb_row_di = NULL ;
 644
 645             vpx_free(pbi->de_thread_data);
 646             pbi->de_thread_data = NULL;
 647     }
 648 }
 649
 650
 651 static void lpf_init( VP8D_COMP *pbi, int default_filt_lvl)
 652 {
 653     VP8_COMMON *cm  = &pbi->common;
 654     MACROBLOCKD *mbd = &pbi->mb;
 655     /*YV12_BUFFER_CONFIG *post = &cm->new_frame;*/  /*frame_to_show;*/
 656     loop_filter_info *lfi = cm->lf_info;
 657     FRAME_TYPE frame_type = cm->frame_type;
 658
 659     /*int mb_row;
 660     int mb_col;
 661     int baseline_filter_level[MAX_MB_SEGMENTS];*/
 662     int alt_flt_enabled = mbd->segmentation_enabled;
 663
 664     int i;
 665     /*unsigned char *y_ptr, *u_ptr, *v_ptr;*/
 666
 667     /* Note the baseline filter values for each segment */
 668     if (alt_flt_enabled)
 669     {
 670         for (i = 0; i < MAX_MB_SEGMENTS; i++)
 671         {
 672             /* Abs value */
 673             if (mbd->mb_segement_abs_delta == SEGMENT_ABSDATA)
 674                 pbi->mt_baseline_filter_level[i] = mbd->segment_feature_data[MB_LVL_ALT_LF][i];
 675             /* Delta Value */
 676             else
 677             {
 678                 pbi->mt_baseline_filter_level[i] = default_filt_lvl + mbd->segment_feature_data[MB_LVL_ALT_LF][i];
 679                 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 */
 680             }
 681         }
 682     }
 683     else
 684     {
 685         for (i = 0; i < MAX_MB_SEGMENTS; i++)
 686             pbi->mt_baseline_filter_level[i] = default_filt_lvl;
 687     }
 688
 689     /* Initialize the loop filter for this frame. */
 690     if ((cm->last_filter_type != cm->filter_type) || (cm->last_sharpness_level != cm->sharpness_level))
 691         vp8_init_loop_filter(cm);
 692     else if (frame_type != cm->last_frame_type)
 693         vp8_frame_init_loop_filter(lfi, frame_type);
 694 }
 695
 696
 697 void vp8mt_decode_mb_rows( VP8D_COMP *pbi, MACROBLOCKD *xd)
 698 {
 699     int mb_row;
 700     VP8_COMMON *pc = &pbi->common;
 701
 702     int ibc = 0;
 703     int num_part = 1 << pbi->common.multi_token_partition;
 704     int i;
 705     volatile int *last_row_current_mb_col = NULL;
 706     int nsync = pbi->sync_range;
 707
 708     int filter_level;
 709     loop_filter_info *lfi = pc->lf_info;
 710     int alt_flt_enabled = xd->segmentation_enabled;
 711     int Segment;
 712
 713     if(pbi->common.filter_level)
 714     {
 715         /* Set above_row buffer to 127 for decoding first MB row */
 716         vpx_memset(pbi->mt_yabove_row[0] + VP8BORDERINPIXELS-1, 127, pc->yv12_fb[pc->lst_fb_idx].y_width + 5);
 717         vpx_memset(pbi->mt_uabove_row[0] + (VP8BORDERINPIXELS>>1)-1, 127, (pc->yv12_fb[pc->lst_fb_idx].y_width>>1) +5);
 718         vpx_memset(pbi->mt_vabove_row[0] + (VP8BORDERINPIXELS>>1)-1, 127, (pc->yv12_fb[pc->lst_fb_idx].y_width>>1) +5);
 719
 720         for (i=1; i<pc->mb_rows; i++)
 721         {
 722             vpx_memset(pbi->mt_yabove_row[i] + VP8BORDERINPIXELS-1, (unsigned char)129, 1);
 723             vpx_memset(pbi->mt_uabove_row[i] + (VP8BORDERINPIXELS>>1)-1, (unsigned char)129, 1);
 724             vpx_memset(pbi->mt_vabove_row[i] + (VP8BORDERINPIXELS>>1)-1, (unsigned char)129, 1);
 725         }
 726
 727         /* Set left_col to 129 initially */
 728         for (i=0; i<pc->mb_rows; i++)
 729         {
 730             vpx_memset(pbi->mt_yleft_col[i], (unsigned char)129, 16);
 731             vpx_memset(pbi->mt_uleft_col[i], (unsigned char)129, 8);
 732             vpx_memset(pbi->mt_vleft_col[i], (unsigned char)129, 8);
 733         }
 734         lpf_init(pbi, pc->filter_level);
 735     }
 736
 737     setup_decoding_thread_data(pbi, xd, pbi->mb_row_di, pbi->decoding_thread_count);
 738
 739     for (i = 0; i < pbi->decoding_thread_count; i++)
 740         sem_post(&pbi->h_event_start_decoding[i]);
 741
 742     for (mb_row = 0; mb_row < pc->mb_rows; mb_row += (pbi->decoding_thread_count + 1))
 743     {
 744
 745         xd->current_bc = &pbi->mbc[mb_row%num_part];
 746
 747         /* vp8_decode_mb_row(pbi, pc, mb_row, xd); */
 748         {
 749             int i;
 750             int recon_yoffset, recon_uvoffset;
 751             int mb_col;
 752             int ref_fb_idx = pc->lst_fb_idx;
 753             int dst_fb_idx = pc->new_fb_idx;
 754             int recon_y_stride = pc->yv12_fb[ref_fb_idx].y_stride;
 755             int recon_uv_stride = pc->yv12_fb[ref_fb_idx].uv_stride;
 756
 757            /* volatile int *last_row_current_mb_col = NULL; */
 758             if (mb_row > 0)
 759                 last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row -1];
 760
 761             vpx_memset(&pc->left_context, 0, sizeof(pc->left_context));
 762             recon_yoffset = mb_row * recon_y_stride * 16;
 763             recon_uvoffset = mb_row * recon_uv_stride * 8;
 764             /* reset above block coeffs */
 765
 766             xd->above_context = pc->above_context;
 767             xd->up_available = (mb_row != 0);
 768
 769             xd->mb_to_top_edge = -((mb_row * 16)) << 3;
 770             xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
 771
 772             for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
 773             {
 774                 if ( mb_row > 0 && (mb_col & (nsync-1)) == 0){
 775                     while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != pc->mb_cols - 1)
 776                     {
 777                         x86_pause_hint();
 778                         thread_sleep(0);
 779                     }
 780                 }
 781
 782                 if (xd->mode_info_context->mbmi.mode == SPLITMV || xd->mode_info_context->mbmi.mode == B_PRED)
 783                 {
 784                     for (i = 0; i < 16; i++)
 785                     {
 786                         BLOCKD *d = &xd->block[i];
 787                         vpx_memcpy(&d->bmi, &xd->mode_info_context->bmi[i], sizeof(B_MODE_INFO));
 788                     }
 789                 }
 790
 791                 /* Distance of Mb to the various image edges.
 792                  * These are specified to 8th pel as they are always compared to values that are in 1/8th pel units
 793                  */
 794                 xd->mb_to_left_edge = -((mb_col * 16) << 3);
 795                 xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
 796
 797                 xd->dst.y_buffer = pc->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 798                 xd->dst.u_buffer = pc->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 799                 xd->dst.v_buffer = pc->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 800
 801                 xd->left_available = (mb_col != 0);
 802
 803                 /* Select the appropriate reference frame for this MB */
 804                 if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
 805                     ref_fb_idx = pc->lst_fb_idx;
 806                 else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
 807                     ref_fb_idx = pc->gld_fb_idx;
 808                 else
 809                     ref_fb_idx = pc->alt_fb_idx;
 810
 811                 xd->pre.y_buffer = pc->yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
 812                 xd->pre.u_buffer = pc->yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
 813                 xd->pre.v_buffer = pc->yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
 814
 815                 if (xd->mode_info_context->mbmi.ref_frame != INTRA_FRAME)
 816                 {
 817                     /* propagate errors from reference frames */
 818                     xd->corrupted |= pc->yv12_fb[ref_fb_idx].corrupted;
 819                 }
 820
 821                 vp8_build_uvmvs(xd, pc->full_pixel);
 822                 decode_macroblock(pbi, xd, mb_row, mb_col);
 823
 824                 /* check if the boolean decoder has suffered an error */
 825                 xd->corrupted |= vp8dx_bool_error(xd->current_bc);
 826
 827                 if (pbi->common.filter_level)
 828                 {
 829                     int skip_lf;
 830                     /* Save decoded MB last row data for next-row decoding */
 831                     if(mb_row != pc->mb_rows-1)
 832                     {
 833                         vpx_memcpy((pbi->mt_yabove_row[mb_row +1] + 32 + mb_col*16), (xd->dst.y_buffer + 15 * recon_y_stride), 16);
 834                         vpx_memcpy((pbi->mt_uabove_row[mb_row +1] + 16 + mb_col*8), (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
 835                         vpx_memcpy((pbi->mt_vabove_row[mb_row +1] + 16 + mb_col*8), (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
 836                     }
 837
 838                     /* save left_col for next MB decoding */
 839                     if(mb_col != pc->mb_cols-1)
 840                     {
 841                         MODE_INFO *next = xd->mode_info_context +1;
 842
 843                         if (xd->frame_type == KEY_FRAME  ||  next->mbmi.ref_frame == INTRA_FRAME)
 844                         {
 845                             for (i = 0; i < 16; i++)
 846                                 pbi->mt_yleft_col[mb_row][i] = xd->dst.y_buffer [i* recon_y_stride + 15];
 847                             for (i = 0; i < 8; i++)
 848                             {
 849                                 pbi->mt_uleft_col[mb_row][i] = xd->dst.u_buffer [i* recon_uv_stride + 7];
 850                                 pbi->mt_vleft_col[mb_row][i] = xd->dst.v_buffer [i* recon_uv_stride + 7];
 851                             }
 852                         }
 853                     }
 854
 855                     /* update loopfilter info */
 856                     Segment = (alt_flt_enabled) ? xd->mode_info_context->mbmi.segment_id : 0;
 857                     skip_lf = (xd->mode_info_context->mbmi.mode != B_PRED &&
 858                                     xd->mode_info_context->mbmi.mode != SPLITMV &&
 859                                     xd->mode_info_context->mbmi.mb_skip_coeff);
 860                     filter_level = pbi->mt_baseline_filter_level[Segment];
 861                     /* Distance of Mb to the various image edges.
 862                      * These are specified to 8th pel as they are always compared to values that are in 1/8th pel units
 863                      * Apply any context driven MB level adjustment
 864                      */
 865                     filter_level = vp8_adjust_mb_lf_value(xd, filter_level);
 866
 867                     /* loopfilter on this macroblock. */
 868                     if (filter_level)
 869                     {
 870                         if (mb_col > 0)
 871                             pc->lf_mbv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 872
 873                         if (!skip_lf)
 874                             pc->lf_bv(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 875
 876                         /* don't apply across umv border */
 877                         if (mb_row > 0)
 878                             pc->lf_mbh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 879
 880                         if (!skip_lf)
 881                             pc->lf_bh(xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi[filter_level]);
 882                     }
 883                 }
 884
 885                 recon_yoffset += 16;
 886                 recon_uvoffset += 8;
 887
 888                 ++xd->mode_info_context;  /* next mb */
 889
 890                 xd->above_context++;
 891
 892                 pbi->mt_current_mb_col[mb_row] = mb_col;
 893             }
 894
 895             /* adjust to the next row of mbs */
 896             if (pbi->common.filter_level)
 897             {
 898                 if(mb_row != pc->mb_rows-1)
 899                 {
 900                     int lasty = pc->yv12_fb[ref_fb_idx].y_width + VP8BORDERINPIXELS;
 901                     int lastuv = (pc->yv12_fb[ref_fb_idx].y_width>>1) + (VP8BORDERINPIXELS>>1);
 902
 903                     for (i = 0; i < 4; i++)
 904                     {
 905                         pbi->mt_yabove_row[mb_row +1][lasty + i] = pbi->mt_yabove_row[mb_row +1][lasty -1];
 906                         pbi->mt_uabove_row[mb_row +1][lastuv + i] = pbi->mt_uabove_row[mb_row +1][lastuv -1];
 907                         pbi->mt_vabove_row[mb_row +1][lastuv + i] = pbi->mt_vabove_row[mb_row +1][lastuv -1];
 908                     }
 909                 }
 910             }else
 911                 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);
 912
 913             ++xd->mode_info_context;      /* skip prediction column */
 914         }
 915         xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
 916     }
 917
 918     sem_wait(&pbi->h_event_end_decoding);   /* add back for each frame */
 919 }