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 #include "onyx_int.h"
12 #include "vp8/common/threading.h"
13 #include "vp8/common/common.h"
14 #include "vp8/common/extend.h"
15 #include "bitstream.h"
16 #include "encodeframe.h"
17 
18 #if CONFIG_MULTITHREAD
19 
20 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x, int ok_to_skip);
21 
22 extern void vp8_loopfilter_frame(VP8_COMP *cpi, VP8_COMMON *cm);
23 
thread_loopfilter(void * p_data)24 static THREAD_FUNCTION thread_loopfilter(void *p_data)
25 {
26     VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
27     VP8_COMMON *cm = &cpi->common;
28 
29     while (1)
30     {
31         if (cpi->b_multi_threaded == 0)
32             break;
33 
34         if (sem_wait(&cpi->h_event_start_lpf) == 0)
35         {
36             if (cpi->b_multi_threaded == 0) /* we're shutting down */
37                 break;
38 
39             vp8_loopfilter_frame(cpi, cm);
40 
41             sem_post(&cpi->h_event_end_lpf);
42         }
43     }
44 
45     return 0;
46 }
47 
48 static
thread_encoding_proc(void * p_data)49 THREAD_FUNCTION thread_encoding_proc(void *p_data)
50 {
51     int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
52     VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
53     MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
54     ENTROPY_CONTEXT_PLANES mb_row_left_context;
55 
56     while (1)
57     {
58         if (cpi->b_multi_threaded == 0)
59             break;
60 
61         if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0)
62         {
63             const int nsync = cpi->mt_sync_range;
64             VP8_COMMON *cm = &cpi->common;
65             int mb_row;
66             MACROBLOCK *x = &mbri->mb;
67             MACROBLOCKD *xd = &x->e_mbd;
68             TOKENEXTRA *tp ;
69 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
70             TOKENEXTRA *tp_start = cpi->tok + (1 + ithread) * (16 * 24);
71             const int num_part = (1 << cm->multi_token_partition);
72 #endif
73 
74             int *segment_counts = mbri->segment_counts;
75             int *totalrate = &mbri->totalrate;
76 
77             if (cpi->b_multi_threaded == 0) /* we're shutting down */
78                 break;
79 
80             for (mb_row = ithread + 1; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
81             {
82 
83                 int recon_yoffset, recon_uvoffset;
84                 int mb_col;
85                 int ref_fb_idx = cm->lst_fb_idx;
86                 int dst_fb_idx = cm->new_fb_idx;
87                 int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
88                 int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
89                 int map_index = (mb_row * cm->mb_cols);
90                 volatile const int *last_row_current_mb_col;
91                 volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
92 
93 #if  (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
94                 vp8_writer *w = &cpi->bc[1 + (mb_row % num_part)];
95 #else
96                 tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
97                 cpi->tplist[mb_row].start = tp;
98 #endif
99 
100                 last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
101 
102                 /* reset above block coeffs */
103                 xd->above_context = cm->above_context;
104                 xd->left_context = &mb_row_left_context;
105 
106                 vp8_zero(mb_row_left_context);
107 
108                 xd->up_available = (mb_row != 0);
109                 recon_yoffset = (mb_row * recon_y_stride * 16);
110                 recon_uvoffset = (mb_row * recon_uv_stride * 8);
111 
112                 /* Set the mb activity pointer to the start of the row. */
113                 x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
114 
115                 /* for each macroblock col in image */
116                 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
117                 {
118                     *current_mb_col = mb_col - 1;
119 
120                     if ((mb_col & (nsync - 1)) == 0)
121                     {
122                         while (mb_col > (*last_row_current_mb_col - nsync))
123                         {
124                             x86_pause_hint();
125                             thread_sleep(0);
126                         }
127                     }
128 
129 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
130                     tp = tp_start;
131 #endif
132 
133                     /* Distance of Mb to the various image edges.
134                      * These specified to 8th pel as they are always compared
135                      * to values that are in 1/8th pel units
136                      */
137                     xd->mb_to_left_edge = -((mb_col * 16) << 3);
138                     xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
139                     xd->mb_to_top_edge = -((mb_row * 16) << 3);
140                     xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
141 
142                     /* Set up limit values for motion vectors used to prevent
143                      * them extending outside the UMV borders
144                      */
145                     x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
146                     x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
147                     x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
148                     x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
149 
150                     xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
151                     xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
152                     xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
153                     xd->left_available = (mb_col != 0);
154 
155                     x->rddiv = cpi->RDDIV;
156                     x->rdmult = cpi->RDMULT;
157 
158                     /* Copy current mb to a buffer */
159                     vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
160 
161                     if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
162                         vp8_activity_masking(cpi, x);
163 
164                     /* Is segmentation enabled */
165                     /* MB level adjustment to quantizer */
166                     if (xd->segmentation_enabled)
167                     {
168                         /* Code to set segment id in xd->mbmi.segment_id for
169                          * current MB (with range checking)
170                          */
171                         if (cpi->segmentation_map[map_index + mb_col] <= 3)
172                             xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index + mb_col];
173                         else
174                             xd->mode_info_context->mbmi.segment_id = 0;
175 
176                         vp8cx_mb_init_quantizer(cpi, x, 1);
177                     }
178                     else
179                         /* Set to Segment 0 by default */
180                         xd->mode_info_context->mbmi.segment_id = 0;
181 
182                     x->active_ptr = cpi->active_map + map_index + mb_col;
183 
184                     if (cm->frame_type == KEY_FRAME)
185                     {
186                         *totalrate += vp8cx_encode_intra_macroblock(cpi, x, &tp);
187 #ifdef MODE_STATS
188                         y_modes[xd->mbmi.mode] ++;
189 #endif
190                     }
191                     else
192                     {
193                         *totalrate += vp8cx_encode_inter_macroblock(cpi, x, &tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
194 
195 #ifdef MODE_STATS
196                         inter_y_modes[xd->mbmi.mode] ++;
197 
198                         if (xd->mbmi.mode == SPLITMV)
199                         {
200                             int b;
201 
202                             for (b = 0; b < xd->mbmi.partition_count; b++)
203                             {
204                                 inter_b_modes[x->partition->bmi[b].mode] ++;
205                             }
206                         }
207 
208 #endif
209                         // Keep track of how many (consecutive) times a  block
210                         // is coded as ZEROMV_LASTREF, for base layer frames.
211                         // Reset to 0 if its coded as anything else.
212                         if (cpi->current_layer == 0) {
213                           if (xd->mode_info_context->mbmi.mode == ZEROMV &&
214                               xd->mode_info_context->mbmi.ref_frame ==
215                                   LAST_FRAME) {
216                             // Increment, check for wrap-around.
217                             if (cpi->consec_zero_last[map_index+mb_col] < 255)
218                               cpi->consec_zero_last[map_index+mb_col] +=
219                                   1;
220                           } else {
221                             cpi->consec_zero_last[map_index+mb_col] = 0;
222                           }
223                         }
224 
225                         /* Special case code for cyclic refresh
226                          * If cyclic update enabled then copy
227                          * xd->mbmi.segment_id; (which may have been updated
228                          * based on mode during
229                          * vp8cx_encode_inter_macroblock()) back into the
230                          * global segmentation map
231                          */
232                         if ((cpi->current_layer == 0) &&
233                             (cpi->cyclic_refresh_mode_enabled &&
234                              xd->segmentation_enabled))
235                         {
236                             const MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
237                             cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
238 
239                             /* If the block has been refreshed mark it as clean
240                              * (the magnitude of the -ve influences how long it
241                              * will be before we consider another refresh):
242                              * Else if it was coded (last frame 0,0) and has
243                              * not already been refreshed then mark it as a
244                              * candidate for cleanup next time (marked 0) else
245                              * mark it as dirty (1).
246                              */
247                             if (mbmi->segment_id)
248                                 cpi->cyclic_refresh_map[map_index + mb_col] = -1;
249                             else if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
250                             {
251                                 if (cpi->cyclic_refresh_map[map_index + mb_col] == 1)
252                                     cpi->cyclic_refresh_map[map_index + mb_col] = 0;
253                             }
254                             else
255                                 cpi->cyclic_refresh_map[map_index + mb_col] = 1;
256 
257                         }
258                     }
259 
260 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
261                     /* pack tokens for this MB */
262                     {
263                         int tok_count = tp - tp_start;
264                         pack_tokens(w, tp_start, tok_count);
265                     }
266 #else
267                     cpi->tplist[mb_row].stop = tp;
268 #endif
269                     /* Increment pointer into gf usage flags structure. */
270                     x->gf_active_ptr++;
271 
272                     /* Increment the activity mask pointers. */
273                     x->mb_activity_ptr++;
274 
275                     /* adjust to the next column of macroblocks */
276                     x->src.y_buffer += 16;
277                     x->src.u_buffer += 8;
278                     x->src.v_buffer += 8;
279 
280                     recon_yoffset += 16;
281                     recon_uvoffset += 8;
282 
283                     /* Keep track of segment usage */
284                     segment_counts[xd->mode_info_context->mbmi.segment_id]++;
285 
286                     /* skip to next mb */
287                     xd->mode_info_context++;
288                     x->partition_info++;
289                     xd->above_context++;
290                 }
291 
292                 vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
293                                     xd->dst.y_buffer + 16,
294                                     xd->dst.u_buffer + 8,
295                                     xd->dst.v_buffer + 8);
296 
297                 *current_mb_col = mb_col + nsync;
298 
299                 /* this is to account for the border */
300                 xd->mode_info_context++;
301                 x->partition_info++;
302 
303                 x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
304                 x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
305                 x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
306 
307                 xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
308                 x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
309                 x->gf_active_ptr   += cm->mb_cols * cpi->encoding_thread_count;
310 
311                 if (mb_row == cm->mb_rows - 1)
312                 {
313                     sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
314                 }
315             }
316         }
317     }
318 
319     /* printf("exit thread %d\n", ithread); */
320     return 0;
321 }
322 
setup_mbby_copy(MACROBLOCK * mbdst,MACROBLOCK * mbsrc)323 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
324 {
325 
326     MACROBLOCK *x = mbsrc;
327     MACROBLOCK *z = mbdst;
328     int i;
329 
330     z->ss               = x->ss;
331     z->ss_count          = x->ss_count;
332     z->searches_per_step  = x->searches_per_step;
333     z->errorperbit      = x->errorperbit;
334 
335     z->sadperbit16      = x->sadperbit16;
336     z->sadperbit4       = x->sadperbit4;
337 
338     /*
339     z->mv_col_min    = x->mv_col_min;
340     z->mv_col_max    = x->mv_col_max;
341     z->mv_row_min    = x->mv_row_min;
342     z->mv_row_max    = x->mv_row_max;
343     */
344 
345     z->short_fdct4x4     = x->short_fdct4x4;
346     z->short_fdct8x4     = x->short_fdct8x4;
347     z->short_walsh4x4    = x->short_walsh4x4;
348     z->quantize_b        = x->quantize_b;
349     z->quantize_b_pair   = x->quantize_b_pair;
350     z->optimize          = x->optimize;
351 
352     /*
353     z->mvc              = x->mvc;
354     z->src.y_buffer      = x->src.y_buffer;
355     z->src.u_buffer      = x->src.u_buffer;
356     z->src.v_buffer      = x->src.v_buffer;
357     */
358 
359     z->mvcost[0] =  x->mvcost[0];
360     z->mvcost[1] =  x->mvcost[1];
361     z->mvsadcost[0] =  x->mvsadcost[0];
362     z->mvsadcost[1] =  x->mvsadcost[1];
363 
364     z->token_costs = x->token_costs;
365     z->inter_bmode_costs = x->inter_bmode_costs;
366     z->mbmode_cost = x->mbmode_cost;
367     z->intra_uv_mode_cost = x->intra_uv_mode_cost;
368     z->bmode_costs = x->bmode_costs;
369 
370     for (i = 0; i < 25; i++)
371     {
372         z->block[i].quant           = x->block[i].quant;
373         z->block[i].quant_fast      = x->block[i].quant_fast;
374         z->block[i].quant_shift     = x->block[i].quant_shift;
375         z->block[i].zbin            = x->block[i].zbin;
376         z->block[i].zrun_zbin_boost = x->block[i].zrun_zbin_boost;
377         z->block[i].round           = x->block[i].round;
378         z->block[i].src_stride      = x->block[i].src_stride;
379     }
380 
381     z->q_index           = x->q_index;
382     z->act_zbin_adj      = x->act_zbin_adj;
383     z->last_act_zbin_adj = x->last_act_zbin_adj;
384 
385     {
386         MACROBLOCKD *xd = &x->e_mbd;
387         MACROBLOCKD *zd = &z->e_mbd;
388 
389         /*
390         zd->mode_info_context = xd->mode_info_context;
391         zd->mode_info        = xd->mode_info;
392 
393         zd->mode_info_stride  = xd->mode_info_stride;
394         zd->frame_type       = xd->frame_type;
395         zd->up_available     = xd->up_available   ;
396         zd->left_available   = xd->left_available;
397         zd->left_context     = xd->left_context;
398         zd->last_frame_dc     = xd->last_frame_dc;
399         zd->last_frame_dccons = xd->last_frame_dccons;
400         zd->gold_frame_dc     = xd->gold_frame_dc;
401         zd->gold_frame_dccons = xd->gold_frame_dccons;
402         zd->mb_to_left_edge    = xd->mb_to_left_edge;
403         zd->mb_to_right_edge   = xd->mb_to_right_edge;
404         zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
405         zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
406         zd->gf_active_ptr     = xd->gf_active_ptr;
407         zd->frames_since_golden       = xd->frames_since_golden;
408         zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
409         */
410         zd->subpixel_predict         = xd->subpixel_predict;
411         zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
412         zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
413         zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
414         zd->segmentation_enabled     = xd->segmentation_enabled;
415         zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
416         vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data,
417                    sizeof(xd->segment_feature_data));
418 
419         vpx_memcpy(zd->dequant_y1_dc, xd->dequant_y1_dc,
420                    sizeof(xd->dequant_y1_dc));
421         vpx_memcpy(zd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
422         vpx_memcpy(zd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
423         vpx_memcpy(zd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
424 
425 #if 1
426         /*TODO:  Remove dequant from BLOCKD.  This is a temporary solution until
427          * the quantizer code uses a passed in pointer to the dequant constants.
428          * This will also require modifications to the x86 and neon assembly.
429          * */
430         for (i = 0; i < 16; i++)
431             zd->block[i].dequant = zd->dequant_y1;
432         for (i = 16; i < 24; i++)
433             zd->block[i].dequant = zd->dequant_uv;
434         zd->block[24].dequant = zd->dequant_y2;
435 #endif
436 
437 
438         vpx_memcpy(z->rd_threshes, x->rd_threshes, sizeof(x->rd_threshes));
439         vpx_memcpy(z->rd_thresh_mult, x->rd_thresh_mult,
440                    sizeof(x->rd_thresh_mult));
441 
442         z->zbin_over_quant = x->zbin_over_quant;
443         z->zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
444         z->zbin_mode_boost = x->zbin_mode_boost;
445 
446         vpx_memset(z->error_bins, 0, sizeof(z->error_bins));
447     }
448 }
449 
vp8cx_init_mbrthread_data(VP8_COMP * cpi,MACROBLOCK * x,MB_ROW_COMP * mbr_ei,int count)450 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
451                                MACROBLOCK *x,
452                                MB_ROW_COMP *mbr_ei,
453                                int count
454                               )
455 {
456 
457     VP8_COMMON *const cm = & cpi->common;
458     MACROBLOCKD *const xd = & x->e_mbd;
459     int i;
460 
461     for (i = 0; i < count; i++)
462     {
463         MACROBLOCK *mb = & mbr_ei[i].mb;
464         MACROBLOCKD *mbd = &mb->e_mbd;
465 
466         mbd->subpixel_predict        = xd->subpixel_predict;
467         mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
468         mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
469         mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
470         mb->gf_active_ptr            = x->gf_active_ptr;
471 
472         vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
473         mbr_ei[i].totalrate = 0;
474 
475         mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
476 
477         mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
478         mbd->mode_info_stride  = cm->mode_info_stride;
479 
480         mbd->frame_type = cm->frame_type;
481 
482         mb->src = * cpi->Source;
483         mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
484         mbd->dst = cm->yv12_fb[cm->new_fb_idx];
485 
486         mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
487         mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
488         mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
489 
490         vp8_build_block_offsets(mb);
491 
492         mbd->left_context = &cm->left_context;
493         mb->mvc = cm->fc.mvc;
494 
495         setup_mbby_copy(&mbr_ei[i].mb, x);
496 
497         mbd->fullpixel_mask = 0xffffffff;
498         if(cm->full_pixel)
499             mbd->fullpixel_mask = 0xfffffff8;
500 
501         vp8_zero(mb->coef_counts);
502         vp8_zero(x->ymode_count);
503         mb->skip_true_count = 0;
504         vp8_zero(mb->MVcount);
505         mb->prediction_error = 0;
506         mb->intra_error = 0;
507         vp8_zero(mb->count_mb_ref_frame_usage);
508         mb->mbs_tested_so_far = 0;
509     }
510 }
511 
vp8cx_create_encoder_threads(VP8_COMP * cpi)512 int vp8cx_create_encoder_threads(VP8_COMP *cpi)
513 {
514     const VP8_COMMON * cm = &cpi->common;
515 
516     cpi->b_multi_threaded = 0;
517     cpi->encoding_thread_count = 0;
518     cpi->b_lpf_running = 0;
519 
520     if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
521     {
522         int ithread;
523         int th_count = cpi->oxcf.multi_threaded - 1;
524         int rc = 0;
525 
526         /* don't allocate more threads than cores available */
527         if (cpi->oxcf.multi_threaded > cm->processor_core_count)
528             th_count = cm->processor_core_count - 1;
529 
530         /* we have th_count + 1 (main) threads processing one row each */
531         /* no point to have more threads than the sync range allows */
532         if(th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1))
533         {
534             th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
535         }
536 
537         if(th_count == 0)
538             return 0;
539 
540         CHECK_MEM_ERROR(cpi->h_encoding_thread,
541                         vpx_malloc(sizeof(pthread_t) * th_count));
542         CHECK_MEM_ERROR(cpi->h_event_start_encoding,
543                         vpx_malloc(sizeof(sem_t) * th_count));
544         CHECK_MEM_ERROR(cpi->mb_row_ei,
545                         vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
546         vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
547         CHECK_MEM_ERROR(cpi->en_thread_data,
548                         vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
549 
550         sem_init(&cpi->h_event_end_encoding, 0, 0);
551 
552         cpi->b_multi_threaded = 1;
553         cpi->encoding_thread_count = th_count;
554 
555         /*
556         printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
557                (cpi->encoding_thread_count +1));
558         */
559 
560         for (ithread = 0; ithread < th_count; ithread++)
561         {
562             ENCODETHREAD_DATA *ethd = &cpi->en_thread_data[ithread];
563 
564             /* Setup block ptrs and offsets */
565             vp8_setup_block_ptrs(&cpi->mb_row_ei[ithread].mb);
566             vp8_setup_block_dptrs(&cpi->mb_row_ei[ithread].mb.e_mbd);
567 
568             sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
569 
570             ethd->ithread = ithread;
571             ethd->ptr1 = (void *)cpi;
572             ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
573 
574             rc = pthread_create(&cpi->h_encoding_thread[ithread], 0,
575                                 thread_encoding_proc, ethd);
576             if(rc)
577                 break;
578         }
579 
580         if(rc)
581         {
582             /* shutdown other threads */
583             cpi->b_multi_threaded = 0;
584             for(--ithread; ithread >= 0; ithread--)
585             {
586                 pthread_join(cpi->h_encoding_thread[ithread], 0);
587                 sem_destroy(&cpi->h_event_start_encoding[ithread]);
588             }
589             sem_destroy(&cpi->h_event_end_encoding);
590 
591             /* free thread related resources */
592             vpx_free(cpi->h_event_start_encoding);
593             vpx_free(cpi->h_encoding_thread);
594             vpx_free(cpi->mb_row_ei);
595             vpx_free(cpi->en_thread_data);
596 
597             return -1;
598         }
599 
600 
601         {
602             LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
603 
604             sem_init(&cpi->h_event_start_lpf, 0, 0);
605             sem_init(&cpi->h_event_end_lpf, 0, 0);
606 
607             lpfthd->ptr1 = (void *)cpi;
608             rc = pthread_create(&cpi->h_filter_thread, 0, thread_loopfilter,
609                                 lpfthd);
610 
611             if(rc)
612             {
613                 /* shutdown other threads */
614                 cpi->b_multi_threaded = 0;
615                 for(--ithread; ithread >= 0; ithread--)
616                 {
617                     sem_post(&cpi->h_event_start_encoding[ithread]);
618                     pthread_join(cpi->h_encoding_thread[ithread], 0);
619                     sem_destroy(&cpi->h_event_start_encoding[ithread]);
620                 }
621                 sem_destroy(&cpi->h_event_end_encoding);
622                 sem_destroy(&cpi->h_event_end_lpf);
623                 sem_destroy(&cpi->h_event_start_lpf);
624 
625                 /* free thread related resources */
626                 vpx_free(cpi->h_event_start_encoding);
627                 vpx_free(cpi->h_encoding_thread);
628                 vpx_free(cpi->mb_row_ei);
629                 vpx_free(cpi->en_thread_data);
630 
631                 return -2;
632             }
633         }
634     }
635     return 0;
636 }
637 
vp8cx_remove_encoder_threads(VP8_COMP * cpi)638 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
639 {
640     if (cpi->b_multi_threaded)
641     {
642         /* shutdown other threads */
643         cpi->b_multi_threaded = 0;
644         {
645             int i;
646 
647             for (i = 0; i < cpi->encoding_thread_count; i++)
648             {
649                 sem_post(&cpi->h_event_start_encoding[i]);
650                 pthread_join(cpi->h_encoding_thread[i], 0);
651 
652                 sem_destroy(&cpi->h_event_start_encoding[i]);
653             }
654 
655             sem_post(&cpi->h_event_start_lpf);
656             pthread_join(cpi->h_filter_thread, 0);
657         }
658 
659         sem_destroy(&cpi->h_event_end_encoding);
660         sem_destroy(&cpi->h_event_end_lpf);
661         sem_destroy(&cpi->h_event_start_lpf);
662 
663         /* free thread related resources */
664         vpx_free(cpi->h_event_start_encoding);
665         vpx_free(cpi->h_encoding_thread);
666         vpx_free(cpi->mb_row_ei);
667         vpx_free(cpi->en_thread_data);
668     }
669 }
670 #endif
671