vp8/encoder/ethreading.c

   1 /*
   2  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
   3  *
   4  *  Use of this source code is governed by a BSD-style license
   5  *  that can be found in the LICENSE file in the root of the source
   6  *  tree. An additional intellectual property rights grant can be found
   7  *  in the file PATENTS.  All contributing project authors may
   8  *  be found in the AUTHORS file in the root of the source tree.
   9  */
  10
  11
  12 #include "onyx_int.h"
  13 #include "threading.h"
  14 #include "common.h"
  15 #include "extend.h"
  16
  17
  18 extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t, int recon_yoffset, int recon_uvoffset);
  19 extern int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t);
  20 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x);
  21 extern void vp8_build_block_offsets(MACROBLOCK *x);
  22 extern void vp8_setup_block_ptrs(MACROBLOCK *x);
  23
  24 static
  25 THREAD_FUNCTION thread_encoding_proc(void *p_data)
  26 {
  27 #if CONFIG_MULTITHREAD
  28     int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
  29     VP8_COMP *cpi   = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
  30     MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
  31     ENTROPY_CONTEXT_PLANES mb_row_left_context;
  32
  33     //printf("Started thread %d\n", ithread);
  34
  35     while (1)
  36     {
  37         if (cpi->b_multi_threaded == 0)
  38             break;
  39
  40         //if(WaitForSingleObject(cpi->h_event_mbrencoding[ithread], INFINITE) == WAIT_OBJECT_0)
  41         if (sem_wait(&cpi->h_event_mbrencoding[ithread]) == 0)
  42         {
  43             if (cpi->b_multi_threaded == FALSE) // we're shutting down
  44                 break;
  45             else
  46             {
  47                 VP8_COMMON *cm      = &cpi->common;
  48                 int mb_row           = mbri->mb_row;
  49                 MACROBLOCK  *x      = &mbri->mb;
  50                 MACROBLOCKD *xd     = &x->e_mbd;
  51                 TOKENEXTRA **tp     = &mbri->tp;
  52                 int *segment_counts  = mbri->segment_counts;
  53                 int *totalrate      = &mbri->totalrate;
  54
  55                 {
  56                     int i;
  57                     int recon_yoffset, recon_uvoffset;
  58                     int mb_col;
  59                     int ref_fb_idx = cm->lst_fb_idx;
  60                     int dst_fb_idx = cm->new_fb_idx;
  61                     int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
  62                     int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
  63                     volatile int *last_row_current_mb_col;
  64                     INT64 activity_sum = 0;
  65
  66                     if (ithread > 0)
  67                         last_row_current_mb_col = &cpi->mb_row_ei[ithread-1].current_mb_col;
  68                     else
  69                         last_row_current_mb_col = &cpi->current_mb_col_main;
  70
  71                     // reset above block coeffs
  72                     xd->above_context = cm->above_context;
  73                     xd->left_context = &mb_row_left_context;
  74
  75                     vp8_zero(mb_row_left_context);
  76
  77                     xd->up_available = (mb_row != 0);
  78                     recon_yoffset = (mb_row * recon_y_stride * 16);
  79                     recon_uvoffset = (mb_row * recon_uv_stride * 8);
  80
  81
  82                     cpi->tplist[mb_row].start = *tp;
  83
  84                     //printf("Thread mb_row = %d\n", mb_row);
  85
  86                     // for each macroblock col in image
  87                     for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
  88                     {
  89                         int seg_map_index = (mb_row * cm->mb_cols);
  90
  91                         while (mb_col > (*last_row_current_mb_col - 1) && *last_row_current_mb_col != cm->mb_cols - 1)
  92                         {
  93                             x86_pause_hint();
  94                             thread_sleep(0);
  95                         }
  96
  97                         // Distance of Mb to the various image edges.
  98                         // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
  99                         xd->mb_to_left_edge = -((mb_col * 16) << 3);
 100                         xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
 101                         xd->mb_to_top_edge = -((mb_row * 16) << 3);
 102                         xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
 103
 104                         // Set up limit values for motion vectors used to prevent them extending outside the UMV borders
 105                         x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
 106                         x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
 107                         x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
 108                         x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
 109
 110                         xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
 111                         xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
 112                         xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
 113                         xd->left_available = (mb_col != 0);
 114
 115                         x->rddiv = cpi->RDDIV;
 116                         x->rdmult = cpi->RDMULT;
 117
 118                         activity_sum += vp8_activity_masking(cpi, x);
 119
 120                         // Is segmentation enabled
 121                         // MB level adjutment to quantizer
 122                         if (xd->segmentation_enabled)
 123                         {
 124                             // Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
 125                             if (cpi->segmentation_map[seg_map_index+mb_col] <= 3)
 126                                 xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[seg_map_index+mb_col];
 127                             else
 128                                 xd->mode_info_context->mbmi.segment_id = 0;
 129
 130                             vp8cx_mb_init_quantizer(cpi, x);
 131                         }
 132                         else
 133                             xd->mode_info_context->mbmi.segment_id = 0;         // Set to Segment 0 by default
 134
 135                         x->active_ptr = cpi->active_map + seg_map_index + mb_col;
 136
 137                         if (cm->frame_type == KEY_FRAME)
 138                         {
 139                             *totalrate += vp8cx_encode_intra_macro_block(cpi, x, tp);
 140 #ifdef MODE_STATS
 141                             y_modes[xd->mbmi.mode] ++;
 142 #endif
 143                         }
 144                         else
 145                         {
 146                             *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset);
 147
 148 #ifdef MODE_STATS
 149                             inter_y_modes[xd->mbmi.mode] ++;
 150
 151                             if (xd->mbmi.mode == SPLITMV)
 152                             {
 153                                 int b;
 154
 155                                 for (b = 0; b < xd->mbmi.partition_count; b++)
 156                                 {
 157                                     inter_b_modes[x->partition->bmi[b].mode] ++;
 158                                 }
 159                             }
 160
 161 #endif
 162
 163                             // Count of last ref frame 0,0 useage
 164                             if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
 165                                 cpi->inter_zz_count ++;
 166
 167                             // Special case code for cyclic refresh
 168                             // If cyclic update enabled then copy xd->mbmi.segment_id; (which may have been updated based on mode
 169                             // during vp8cx_encode_inter_macroblock()) back into the global sgmentation map
 170                             if (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)
 171                             {
 172                                 cpi->segmentation_map[seg_map_index+mb_col] = xd->mode_info_context->mbmi.segment_id;
 173
 174                                 // If the block has been refreshed mark it as clean (the magnitude of the -ve influences how long it will be before we consider another refresh):
 175                                 // Else if it was coded (last frame 0,0) and has not already been refreshed then mark it as a candidate for cleanup next time (marked 0)
 176                                 // else mark it as dirty (1).
 177                                 if (xd->mode_info_context->mbmi.segment_id)
 178                                     cpi->cyclic_refresh_map[seg_map_index+mb_col] = -1;
 179                                 else if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
 180                                 {
 181                                     if (cpi->cyclic_refresh_map[seg_map_index+mb_col] == 1)
 182                                         cpi->cyclic_refresh_map[seg_map_index+mb_col] = 0;
 183                                 }
 184                                 else
 185                                     cpi->cyclic_refresh_map[seg_map_index+mb_col] = 1;
 186
 187                             }
 188                         }
 189                         cpi->tplist[mb_row].stop = *tp;
 190
 191                         x->gf_active_ptr++;      // Increment pointer into gf useage flags structure for next mb
 192
 193                         for (i = 0; i < 16; i++)
 194                             vpx_memcpy(&xd->mode_info_context->bmi[i], &xd->block[i].bmi, sizeof(xd->block[i].bmi));
 195
 196                         // adjust to the next column of macroblocks
 197                         x->src.y_buffer += 16;
 198                         x->src.u_buffer += 8;
 199                         x->src.v_buffer += 8;
 200
 201                         recon_yoffset += 16;
 202                         recon_uvoffset += 8;
 203
 204                         // Keep track of segment useage
 205                         segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
 206
 207                         // skip to next mb
 208                         xd->mode_info_context++;
 209                         x->partition_info++;
 210
 211                         xd->above_context++;
 212
 213                         cpi->mb_row_ei[ithread].current_mb_col = mb_col;
 214
 215                     }
 216
 217                     //extend the recon for intra prediction
 218                     vp8_extend_mb_row(
 219                         &cm->yv12_fb[dst_fb_idx],
 220                         xd->dst.y_buffer + 16,
 221                         xd->dst.u_buffer + 8,
 222                         xd->dst.v_buffer + 8);
 223
 224                     // this is to account for the border
 225                     xd->mode_info_context++;
 226                     x->partition_info++;
 227                     x->activity_sum += activity_sum;
 228
 229                     x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
 230                     x->src.u_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 231                     x->src.v_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 232
 233                     xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
 234                     x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
 235
 236                     if (ithread == (cpi->encoding_thread_count - 1) || mb_row == cm->mb_rows - 1)
 237                     {
 238                         //SetEvent(cpi->h_event_main);
 239                         sem_post(&cpi->h_event_main);
 240                     }
 241
 242                 }
 243
 244             }
 245         }
 246     }
 247
 248 #else
 249     (void) p_data;
 250 #endif
 251
 252     //printf("exit thread %d\n", ithread);
 253     return 0;
 254 }
 255
 256 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
 257 {
 258
 259     MACROBLOCK *x = mbsrc;
 260     MACROBLOCK *z = mbdst;
 261     int i;
 262
 263     z->ss               = x->ss;
 264     z->ss_count          = x->ss_count;
 265     z->searches_per_step  = x->searches_per_step;
 266     z->errorperbit      = x->errorperbit;
 267
 268     z->sadperbit16      = x->sadperbit16;
 269     z->sadperbit4       = x->sadperbit4;
 270     z->errthresh        = x->errthresh;
 271
 272     /*
 273     z->mv_col_min    = x->mv_col_min;
 274     z->mv_col_max    = x->mv_col_max;
 275     z->mv_row_min    = x->mv_row_min;
 276     z->mv_row_max    = x->mv_row_max;
 277     z->vector_range = x->vector_range ;
 278     */
 279
 280     z->vp8_short_fdct4x4     = x->vp8_short_fdct4x4;
 281     z->vp8_short_fdct8x4     = x->vp8_short_fdct8x4;
 282     z->short_walsh4x4    = x->short_walsh4x4;
 283     z->quantize_b        = x->quantize_b;
 284
 285     /*
 286     z->mvc              = x->mvc;
 287     z->src.y_buffer      = x->src.y_buffer;
 288     z->src.u_buffer      = x->src.u_buffer;
 289     z->src.v_buffer      = x->src.v_buffer;
 290     */
 291
 292
 293     vpx_memcpy(z->mvcosts,          x->mvcosts,         sizeof(x->mvcosts));
 294     z->mvcost[0] = &z->mvcosts[0][mv_max+1];
 295     z->mvcost[1] = &z->mvcosts[1][mv_max+1];
 296     z->mvsadcost[0] = &z->mvsadcosts[0][mv_max+1];
 297     z->mvsadcost[1] = &z->mvsadcosts[1][mv_max+1];
 298
 299
 300     vpx_memcpy(z->token_costs,       x->token_costs,      sizeof(x->token_costs));
 301     vpx_memcpy(z->inter_bmode_costs,  x->inter_bmode_costs, sizeof(x->inter_bmode_costs));
 302     //memcpy(z->mvcosts,            x->mvcosts,         sizeof(x->mvcosts));
 303     //memcpy(z->mvcost,         x->mvcost,          sizeof(x->mvcost));
 304     vpx_memcpy(z->mbmode_cost,       x->mbmode_cost,      sizeof(x->mbmode_cost));
 305     vpx_memcpy(z->intra_uv_mode_cost,  x->intra_uv_mode_cost, sizeof(x->intra_uv_mode_cost));
 306     vpx_memcpy(z->bmode_costs,       x->bmode_costs,      sizeof(x->bmode_costs));
 307
 308     for (i = 0; i < 25; i++)
 309     {
 310         z->block[i].quant           = x->block[i].quant;
 311         z->block[i].quant_shift     = x->block[i].quant_shift;
 312         z->block[i].zbin            = x->block[i].zbin;
 313         z->block[i].zrun_zbin_boost   = x->block[i].zrun_zbin_boost;
 314         z->block[i].round           = x->block[i].round;
 315         /*
 316         z->block[i].src             = x->block[i].src;
 317         */
 318         z->block[i].src_stride       = x->block[i].src_stride;
 319         z->block[i].force_empty      = x->block[i].force_empty;
 320
 321     }
 322
 323     {
 324         MACROBLOCKD *xd = &x->e_mbd;
 325         MACROBLOCKD *zd = &z->e_mbd;
 326
 327         /*
 328         zd->mode_info_context = xd->mode_info_context;
 329         zd->mode_info        = xd->mode_info;
 330
 331         zd->mode_info_stride  = xd->mode_info_stride;
 332         zd->frame_type       = xd->frame_type;
 333         zd->up_available     = xd->up_available   ;
 334         zd->left_available   = xd->left_available;
 335         zd->left_context     = xd->left_context;
 336         zd->last_frame_dc     = xd->last_frame_dc;
 337         zd->last_frame_dccons = xd->last_frame_dccons;
 338         zd->gold_frame_dc     = xd->gold_frame_dc;
 339         zd->gold_frame_dccons = xd->gold_frame_dccons;
 340         zd->mb_to_left_edge    = xd->mb_to_left_edge;
 341         zd->mb_to_right_edge   = xd->mb_to_right_edge;
 342         zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
 343         zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
 344         zd->gf_active_ptr     = xd->gf_active_ptr;
 345         zd->frames_since_golden       = xd->frames_since_golden;
 346         zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
 347         */
 348         zd->subpixel_predict         = xd->subpixel_predict;
 349         zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
 350         zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
 351         zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
 352         zd->segmentation_enabled     = xd->segmentation_enabled;
 353         zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
 354         vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
 355
 356         for (i = 0; i < 25; i++)
 357         {
 358             zd->block[i].dequant = xd->block[i].dequant;
 359         }
 360     }
 361 }
 362
 363
 364 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
 365                                MACROBLOCK *x,
 366                                MB_ROW_COMP *mbr_ei,
 367                                int mb_row,
 368                                int count
 369                               )
 370 {
 371
 372     VP8_COMMON *const cm = & cpi->common;
 373     MACROBLOCKD *const xd = & x->e_mbd;
 374     int i;
 375     (void) mb_row;
 376
 377     for (i = 0; i < count; i++)
 378     {
 379         MACROBLOCK *mb = & mbr_ei[i].mb;
 380         MACROBLOCKD *mbd = &mb->e_mbd;
 381
 382         mbd->subpixel_predict        = xd->subpixel_predict;
 383         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
 384         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
 385         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
 386 #if CONFIG_RUNTIME_CPU_DETECT
 387         mbd->rtcd                   = xd->rtcd;
 388 #endif
 389         mb->gf_active_ptr            = x->gf_active_ptr;
 390
 391         mb->vector_range             = 32;
 392
 393         vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
 394         mbr_ei[i].totalrate = 0;
 395
 396         mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
 397
 398         mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
 399         mbd->mode_info_stride  = cm->mode_info_stride;
 400
 401         mbd->frame_type = cm->frame_type;
 402
 403         mbd->frames_since_golden = cm->frames_since_golden;
 404         mbd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
 405
 406         mb->src = * cpi->Source;
 407         mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
 408         mbd->dst = cm->yv12_fb[cm->new_fb_idx];
 409
 410         mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
 411         mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
 412         mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
 413
 414
 415         vp8_build_block_offsets(mb);
 416
 417         vp8_setup_block_dptrs(mbd);
 418
 419         vp8_setup_block_ptrs(mb);
 420
 421         mb->activity_sum = 0;
 422
 423         mbd->left_context = &cm->left_context;
 424         mb->mvc = cm->fc.mvc;
 425
 426         setup_mbby_copy(&mbr_ei[i].mb, x);
 427
 428     }
 429 }
 430
 431
 432 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
 433 {
 434     cpi->b_multi_threaded = 0;
 435
 436     cpi->processor_core_count = 32; //vp8_get_proc_core_count();
 437
 438     CHECK_MEM_ERROR(cpi->tplist, vpx_malloc(sizeof(TOKENLIST) * cpi->common.mb_rows));
 439
 440 #if CONFIG_MULTITHREAD
 441
 442     if (cpi->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
 443     {
 444         int ithread;
 445
 446         if (cpi->oxcf.multi_threaded > cpi->processor_core_count)
 447             cpi->encoding_thread_count = cpi->processor_core_count - 1;
 448         else
 449             cpi->encoding_thread_count = cpi->oxcf.multi_threaded - 1;
 450
 451
 452         CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * cpi->encoding_thread_count));
 453         CHECK_MEM_ERROR(cpi->h_event_mbrencoding, vpx_malloc(sizeof(sem_t) * cpi->encoding_thread_count));
 454         CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count));
 455         vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count);
 456         CHECK_MEM_ERROR(cpi->en_thread_data, vpx_malloc(sizeof(ENCODETHREAD_DATA) * cpi->encoding_thread_count));
 457         //cpi->h_event_main = CreateEvent(NULL, FALSE, FALSE, NULL);
 458         sem_init(&cpi->h_event_main, 0, 0);
 459
 460         cpi->b_multi_threaded = 1;
 461
 462         //printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n", (cpi->encoding_thread_count +1));
 463
 464         for (ithread = 0; ithread < cpi->encoding_thread_count; ithread++)
 465         {
 466             //cpi->h_event_mbrencoding[ithread] = CreateEvent(NULL, FALSE, FALSE, NULL);
 467             sem_init(&cpi->h_event_mbrencoding[ithread], 0, 0);
 468             cpi->en_thread_data[ithread].ithread = ithread;
 469             cpi->en_thread_data[ithread].ptr1 = (void *)cpi;
 470             cpi->en_thread_data[ithread].ptr2 = (void *)&cpi->mb_row_ei[ithread];
 471
 472             //printf(" call begin thread %d \n", ithread);
 473
 474             //cpi->h_encoding_thread[ithread] =   (HANDLE)_beginthreadex(
 475             //  NULL,           // security
 476             //  0,              // stksize
 477             //  thread_encoding_proc,
 478             //  (&cpi->en_thread_data[ithread]),          // Thread data
 479             //  0,
 480             //  NULL);
 481
 482             pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, (&cpi->en_thread_data[ithread]));
 483
 484         }
 485
 486     }
 487
 488 #endif
 489 }
 490
 491 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
 492 {
 493 #if CONFIG_MULTITHREAD
 494
 495     if (cpi->b_multi_threaded)
 496     {
 497         //shutdown other threads
 498         cpi->b_multi_threaded = 0;
 499         {
 500             int i;
 501
 502             for (i = 0; i < cpi->encoding_thread_count; i++)
 503             {
 504                 //SetEvent(cpi->h_event_mbrencoding[i]);
 505                 sem_post(&cpi->h_event_mbrencoding[i]);
 506                 pthread_join(cpi->h_encoding_thread[i], 0);
 507             }
 508
 509             for (i = 0; i < cpi->encoding_thread_count; i++)
 510                 sem_destroy(&cpi->h_event_mbrencoding[i]);
 511         }
 512         //free thread related resources
 513         vpx_free(cpi->h_event_mbrencoding);
 514         vpx_free(cpi->h_encoding_thread);
 515         vpx_free(cpi->mb_row_ei);
 516         vpx_free(cpi->en_thread_data);
 517     }
 518
 519 #endif
 520     vpx_free(cpi->tplist);
 521 }