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
 136                         if (cm->frame_type == KEY_FRAME)
 137                         {
 138                             *totalrate += vp8cx_encode_intra_macro_block(cpi, x, tp);
 139 #ifdef MODE_STATS
 140                             y_modes[xd->mbmi.mode] ++;
 141 #endif
 142                         }
 143                         else
 144                         {
 145                             *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset);
 146
 147 #ifdef MODE_STATS
 148                             inter_y_modes[xd->mbmi.mode] ++;
 149
 150                             if (xd->mbmi.mode == SPLITMV)
 151                             {
 152                                 int b;
 153
 154                                 for (b = 0; b < xd->mbmi.partition_count; b++)
 155                                 {
 156                                     inter_b_modes[x->partition->bmi[b].mode] ++;
 157                                 }
 158                             }
 159
 160 #endif
 161
 162                             // Count of last ref frame 0,0 useage
 163                             if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
 164                                 cpi->inter_zz_count ++;
 165
 166                         }
 167
 168                         cpi->tplist[mb_row].stop = *tp;
 169
 170                         x->gf_active_ptr++;      // Increment pointer into gf useage flags structure for next mb
 171
 172                         for (i = 0; i < 16; i++)
 173                             vpx_memcpy(&xd->mode_info_context->bmi[i], &xd->block[i].bmi, sizeof(xd->block[i].bmi));
 174
 175                         // adjust to the next column of macroblocks
 176                         x->src.y_buffer += 16;
 177                         x->src.u_buffer += 8;
 178                         x->src.v_buffer += 8;
 179
 180                         recon_yoffset += 16;
 181                         recon_uvoffset += 8;
 182
 183                         // Keep track of segment useage
 184                         segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
 185
 186                         // skip to next mb
 187                         xd->mode_info_context++;
 188                         x->partition_info++;
 189
 190                         xd->above_context++;
 191
 192                         cpi->mb_row_ei[ithread].current_mb_col = mb_col;
 193
 194                     }
 195
 196                     //extend the recon for intra prediction
 197                     vp8_extend_mb_row(
 198                         &cm->yv12_fb[dst_fb_idx],
 199                         xd->dst.y_buffer + 16,
 200                         xd->dst.u_buffer + 8,
 201                         xd->dst.v_buffer + 8);
 202
 203                     // this is to account for the border
 204                     xd->mode_info_context++;
 205                     x->partition_info++;
 206                     x->activity_sum += activity_sum;
 207
 208                     x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
 209                     x->src.u_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 210                     x->src.v_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
 211
 212                     xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
 213                     x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
 214
 215                     if (ithread == (cpi->encoding_thread_count - 1) || mb_row == cm->mb_rows - 1)
 216                     {
 217                         //SetEvent(cpi->h_event_main);
 218                         sem_post(&cpi->h_event_main);
 219                     }
 220
 221                 }
 222
 223             }
 224         }
 225     }
 226
 227 #else
 228     (void) p_data;
 229 #endif
 230
 231     //printf("exit thread %d\n", ithread);
 232     return 0;
 233 }
 234
 235 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
 236 {
 237
 238     MACROBLOCK *x = mbsrc;
 239     MACROBLOCK *z = mbdst;
 240     int i;
 241
 242     z->ss               = x->ss;
 243     z->ss_count          = x->ss_count;
 244     z->searches_per_step  = x->searches_per_step;
 245     z->errorperbit      = x->errorperbit;
 246
 247     z->sadperbit16      = x->sadperbit16;
 248     z->sadperbit4       = x->sadperbit4;
 249     z->errthresh        = x->errthresh;
 250
 251     /*
 252     z->mv_col_min    = x->mv_col_min;
 253     z->mv_col_max    = x->mv_col_max;
 254     z->mv_row_min    = x->mv_row_min;
 255     z->mv_row_max    = x->mv_row_max;
 256     z->vector_range = x->vector_range ;
 257     */
 258
 259     z->vp8_short_fdct4x4     = x->vp8_short_fdct4x4;
 260     z->vp8_short_fdct8x4     = x->vp8_short_fdct8x4;
 261     z->short_walsh4x4    = x->short_walsh4x4;
 262     z->quantize_b        = x->quantize_b;
 263
 264     /*
 265     z->mvc              = x->mvc;
 266     z->src.y_buffer      = x->src.y_buffer;
 267     z->src.u_buffer      = x->src.u_buffer;
 268     z->src.v_buffer      = x->src.v_buffer;
 269     */
 270
 271
 272     vpx_memcpy(z->mvcosts,          x->mvcosts,         sizeof(x->mvcosts));
 273     z->mvcost[0] = &z->mvcosts[0][mv_max+1];
 274     z->mvcost[1] = &z->mvcosts[1][mv_max+1];
 275     z->mvsadcost[0] = &z->mvsadcosts[0][mv_max+1];
 276     z->mvsadcost[1] = &z->mvsadcosts[1][mv_max+1];
 277
 278
 279     vpx_memcpy(z->token_costs,       x->token_costs,      sizeof(x->token_costs));
 280     vpx_memcpy(z->inter_bmode_costs,  x->inter_bmode_costs, sizeof(x->inter_bmode_costs));
 281     //memcpy(z->mvcosts,            x->mvcosts,         sizeof(x->mvcosts));
 282     //memcpy(z->mvcost,         x->mvcost,          sizeof(x->mvcost));
 283     vpx_memcpy(z->mbmode_cost,       x->mbmode_cost,      sizeof(x->mbmode_cost));
 284     vpx_memcpy(z->intra_uv_mode_cost,  x->intra_uv_mode_cost, sizeof(x->intra_uv_mode_cost));
 285     vpx_memcpy(z->bmode_costs,       x->bmode_costs,      sizeof(x->bmode_costs));
 286
 287     for (i = 0; i < 25; i++)
 288     {
 289         z->block[i].quant           = x->block[i].quant;
 290         z->block[i].quant_shift     = x->block[i].quant_shift;
 291         z->block[i].zbin            = x->block[i].zbin;
 292         z->block[i].zrun_zbin_boost   = x->block[i].zrun_zbin_boost;
 293         z->block[i].round           = x->block[i].round;
 294         /*
 295         z->block[i].src             = x->block[i].src;
 296         */
 297         z->block[i].src_stride       = x->block[i].src_stride;
 298         z->block[i].force_empty      = x->block[i].force_empty;
 299
 300     }
 301
 302     {
 303         MACROBLOCKD *xd = &x->e_mbd;
 304         MACROBLOCKD *zd = &z->e_mbd;
 305
 306         /*
 307         zd->mode_info_context = xd->mode_info_context;
 308         zd->mode_info        = xd->mode_info;
 309
 310         zd->mode_info_stride  = xd->mode_info_stride;
 311         zd->frame_type       = xd->frame_type;
 312         zd->up_available     = xd->up_available   ;
 313         zd->left_available   = xd->left_available;
 314         zd->left_context     = xd->left_context;
 315         zd->last_frame_dc     = xd->last_frame_dc;
 316         zd->last_frame_dccons = xd->last_frame_dccons;
 317         zd->gold_frame_dc     = xd->gold_frame_dc;
 318         zd->gold_frame_dccons = xd->gold_frame_dccons;
 319         zd->mb_to_left_edge    = xd->mb_to_left_edge;
 320         zd->mb_to_right_edge   = xd->mb_to_right_edge;
 321         zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
 322         zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
 323         zd->gf_active_ptr     = xd->gf_active_ptr;
 324         zd->frames_since_golden       = xd->frames_since_golden;
 325         zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
 326         */
 327         zd->subpixel_predict         = xd->subpixel_predict;
 328         zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
 329         zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
 330         zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
 331         zd->segmentation_enabled     = xd->segmentation_enabled;
 332         zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
 333         vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
 334
 335         for (i = 0; i < 25; i++)
 336         {
 337             zd->block[i].dequant = xd->block[i].dequant;
 338         }
 339     }
 340 }
 341
 342
 343 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
 344                                MACROBLOCK *x,
 345                                MB_ROW_COMP *mbr_ei,
 346                                int mb_row,
 347                                int count
 348                               )
 349 {
 350
 351     VP8_COMMON *const cm = & cpi->common;
 352     MACROBLOCKD *const xd = & x->e_mbd;
 353     int i;
 354     (void) mb_row;
 355
 356     for (i = 0; i < count; i++)
 357     {
 358         MACROBLOCK *mb = & mbr_ei[i].mb;
 359         MACROBLOCKD *mbd = &mb->e_mbd;
 360
 361         mbd->subpixel_predict        = xd->subpixel_predict;
 362         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
 363         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
 364         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
 365 #if CONFIG_RUNTIME_CPU_DETECT
 366         mbd->rtcd                   = xd->rtcd;
 367 #endif
 368         mb->gf_active_ptr            = x->gf_active_ptr;
 369
 370         mb->vector_range             = 32;
 371
 372         vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
 373         mbr_ei[i].totalrate = 0;
 374
 375         mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
 376
 377         mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
 378         mbd->mode_info_stride  = cm->mode_info_stride;
 379
 380         mbd->frame_type = cm->frame_type;
 381
 382         mbd->frames_since_golden = cm->frames_since_golden;
 383         mbd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
 384
 385         mb->src = * cpi->Source;
 386         mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
 387         mbd->dst = cm->yv12_fb[cm->new_fb_idx];
 388
 389         mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
 390         mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
 391         mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
 392
 393
 394         vp8_build_block_offsets(mb);
 395
 396         vp8_setup_block_dptrs(mbd);
 397
 398         vp8_setup_block_ptrs(mb);
 399
 400         mb->activity_sum = 0;
 401
 402         mbd->left_context = &cm->left_context;
 403         mb->mvc = cm->fc.mvc;
 404
 405         setup_mbby_copy(&mbr_ei[i].mb, x);
 406
 407     }
 408 }
 409
 410
 411 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
 412 {
 413     cpi->b_multi_threaded = 0;
 414
 415     cpi->processor_core_count = 32; //vp8_get_proc_core_count();
 416
 417     CHECK_MEM_ERROR(cpi->tplist, vpx_malloc(sizeof(TOKENLIST) * cpi->common.mb_rows));
 418
 419 #if CONFIG_MULTITHREAD
 420
 421     if (cpi->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
 422     {
 423         int ithread;
 424
 425         if (cpi->oxcf.multi_threaded > cpi->processor_core_count)
 426             cpi->encoding_thread_count = cpi->processor_core_count - 1;
 427         else
 428             cpi->encoding_thread_count = cpi->oxcf.multi_threaded - 1;
 429
 430
 431         CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * cpi->encoding_thread_count));
 432         CHECK_MEM_ERROR(cpi->h_event_mbrencoding, vpx_malloc(sizeof(sem_t) * cpi->encoding_thread_count));
 433         CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count));
 434         vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count);
 435         CHECK_MEM_ERROR(cpi->en_thread_data, vpx_malloc(sizeof(ENCODETHREAD_DATA) * cpi->encoding_thread_count));
 436         //cpi->h_event_main = CreateEvent(NULL, FALSE, FALSE, NULL);
 437         sem_init(&cpi->h_event_main, 0, 0);
 438
 439         cpi->b_multi_threaded = 1;
 440
 441         //printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n", (cpi->encoding_thread_count +1));
 442
 443         for (ithread = 0; ithread < cpi->encoding_thread_count; ithread++)
 444         {
 445             //cpi->h_event_mbrencoding[ithread] = CreateEvent(NULL, FALSE, FALSE, NULL);
 446             sem_init(&cpi->h_event_mbrencoding[ithread], 0, 0);
 447             cpi->en_thread_data[ithread].ithread = ithread;
 448             cpi->en_thread_data[ithread].ptr1 = (void *)cpi;
 449             cpi->en_thread_data[ithread].ptr2 = (void *)&cpi->mb_row_ei[ithread];
 450
 451             //printf(" call begin thread %d \n", ithread);
 452
 453             //cpi->h_encoding_thread[ithread] =   (HANDLE)_beginthreadex(
 454             //  NULL,           // security
 455             //  0,              // stksize
 456             //  thread_encoding_proc,
 457             //  (&cpi->en_thread_data[ithread]),          // Thread data
 458             //  0,
 459             //  NULL);
 460
 461             pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, (&cpi->en_thread_data[ithread]));
 462
 463         }
 464
 465     }
 466
 467 #endif
 468 }
 469
 470 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
 471 {
 472 #if CONFIG_MULTITHREAD
 473
 474     if (cpi->b_multi_threaded)
 475     {
 476         //shutdown other threads
 477         cpi->b_multi_threaded = 0;
 478         {
 479             int i;
 480
 481             for (i = 0; i < cpi->encoding_thread_count; i++)
 482             {
 483                 //SetEvent(cpi->h_event_mbrencoding[i]);
 484                 sem_post(&cpi->h_event_mbrencoding[i]);
 485                 pthread_join(cpi->h_encoding_thread[i], 0);
 486             }
 487
 488             for (i = 0; i < cpi->encoding_thread_count; i++)
 489                 sem_destroy(&cpi->h_event_mbrencoding[i]);
 490         }
 491         //free thread related resources
 492         vpx_free(cpi->h_event_mbrencoding);
 493         vpx_free(cpi->h_encoding_thread);
 494         vpx_free(cpi->mb_row_ei);
 495         vpx_free(cpi->en_thread_data);
 496     }
 497
 498 #endif
 499     vpx_free(cpi->tplist);
 500 }