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 #include "ethreading.h"
18 
19 #if CONFIG_MULTITHREAD
20 
21 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x,
22                                     int ok_to_skip);
23 
thread_loopfilter(void * p_data)24 static THREAD_FUNCTION thread_loopfilter(void *p_data) {
25   VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
26   VP8_COMMON *cm = &cpi->common;
27 
28   while (1) {
29     if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) == 0) break;
30 
31     if (sem_wait(&cpi->h_event_start_lpf) == 0) {
32       /* we're shutting down */
33       if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) == 0) break;
34 
35       vp8_loopfilter_frame(cpi, cm);
36 
37       sem_post(&cpi->h_event_end_lpf);
38     }
39   }
40 
41   return 0;
42 }
43 
thread_encoding_proc(void * p_data)44 static THREAD_FUNCTION thread_encoding_proc(void *p_data) {
45   int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
46   VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
47   MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
48   ENTROPY_CONTEXT_PLANES mb_row_left_context;
49 
50   while (1) {
51     if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) == 0) break;
52 
53     if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0) {
54       const int nsync = cpi->mt_sync_range;
55       VP8_COMMON *cm = &cpi->common;
56       int mb_row;
57       MACROBLOCK *x = &mbri->mb;
58       MACROBLOCKD *xd = &x->e_mbd;
59       TOKENEXTRA *tp;
60 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
61       TOKENEXTRA *tp_start = cpi->tok + (1 + ithread) * (16 * 24);
62       const int num_part = (1 << cm->multi_token_partition);
63 #endif
64 
65       int *segment_counts = mbri->segment_counts;
66       int *totalrate = &mbri->totalrate;
67 
68       /* we're shutting down */
69       if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) == 0) break;
70 
71       xd->mode_info_context = cm->mi + cm->mode_info_stride * (ithread + 1);
72       xd->mode_info_stride = cm->mode_info_stride;
73 
74       for (mb_row = ithread + 1; mb_row < cm->mb_rows;
75            mb_row += (cpi->encoding_thread_count + 1)) {
76         int recon_yoffset, recon_uvoffset;
77         int mb_col;
78         int ref_fb_idx = cm->lst_fb_idx;
79         int dst_fb_idx = cm->new_fb_idx;
80         int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
81         int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
82         int map_index = (mb_row * cm->mb_cols);
83         const vpx_atomic_int *last_row_current_mb_col;
84         vpx_atomic_int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
85 
86 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
87         vp8_writer *w = &cpi->bc[1 + (mb_row % num_part)];
88 #else
89         tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
90         cpi->tplist[mb_row].start = tp;
91 #endif
92 
93         last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
94 
95         /* reset above block coeffs */
96         xd->above_context = cm->above_context;
97         xd->left_context = &mb_row_left_context;
98 
99         vp8_zero(mb_row_left_context);
100 
101         xd->up_available = (mb_row != 0);
102         recon_yoffset = (mb_row * recon_y_stride * 16);
103         recon_uvoffset = (mb_row * recon_uv_stride * 8);
104 
105         /* Set the mb activity pointer to the start of the row. */
106         x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
107 
108         /* for each macroblock col in image */
109         for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
110           if (((mb_col - 1) % nsync) == 0) {
111             vpx_atomic_store_release(current_mb_col, mb_col - 1);
112           }
113 
114           if (mb_row && !(mb_col & (nsync - 1))) {
115             vp8_atomic_spin_wait(mb_col, last_row_current_mb_col, nsync);
116           }
117 
118 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
119           tp = tp_start;
120 #endif
121 
122           /* Distance of Mb to the various image edges.
123            * These specified to 8th pel as they are always compared
124            * to values that are in 1/8th pel units
125            */
126           xd->mb_to_left_edge = -((mb_col * 16) << 3);
127           xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
128           xd->mb_to_top_edge = -((mb_row * 16) << 3);
129           xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
130 
131           /* Set up limit values for motion vectors used to prevent
132            * them extending outside the UMV borders
133            */
134           x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
135           x->mv_col_max =
136               ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
137           x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
138           x->mv_row_max =
139               ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
140 
141           xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
142           xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
143           xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
144           xd->left_available = (mb_col != 0);
145 
146           x->rddiv = cpi->RDDIV;
147           x->rdmult = cpi->RDMULT;
148 
149           /* Copy current mb to a buffer */
150           vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
151 
152           if (cpi->oxcf.tuning == VP8_TUNE_SSIM) vp8_activity_masking(cpi, x);
153 
154           /* Is segmentation enabled */
155           /* MB level adjustment to quantizer */
156           if (xd->segmentation_enabled) {
157             /* Code to set segment id in xd->mbmi.segment_id for
158              * current MB (with range checking)
159              */
160             if (cpi->segmentation_map[map_index + mb_col] <= 3) {
161               xd->mode_info_context->mbmi.segment_id =
162                   cpi->segmentation_map[map_index + mb_col];
163             } else {
164               xd->mode_info_context->mbmi.segment_id = 0;
165             }
166 
167             vp8cx_mb_init_quantizer(cpi, x, 1);
168           } else {
169             /* Set to Segment 0 by default */
170             xd->mode_info_context->mbmi.segment_id = 0;
171           }
172 
173           x->active_ptr = cpi->active_map + map_index + mb_col;
174 
175           if (cm->frame_type == KEY_FRAME) {
176             *totalrate += vp8cx_encode_intra_macroblock(cpi, x, &tp);
177 #ifdef MODE_STATS
178             y_modes[xd->mbmi.mode]++;
179 #endif
180           } else {
181             *totalrate += vp8cx_encode_inter_macroblock(
182                 cpi, x, &tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
183 
184 #ifdef MODE_STATS
185             inter_y_modes[xd->mbmi.mode]++;
186 
187             if (xd->mbmi.mode == SPLITMV) {
188               int b;
189 
190               for (b = 0; b < xd->mbmi.partition_count; ++b) {
191                 inter_b_modes[x->partition->bmi[b].mode]++;
192               }
193             }
194 
195 #endif
196             // Keep track of how many (consecutive) times a  block
197             // is coded as ZEROMV_LASTREF, for base layer frames.
198             // Reset to 0 if its coded as anything else.
199             if (cpi->current_layer == 0) {
200               if (xd->mode_info_context->mbmi.mode == ZEROMV &&
201                   xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
202                 // Increment, check for wrap-around.
203                 if (cpi->consec_zero_last[map_index + mb_col] < 255) {
204                   cpi->consec_zero_last[map_index + mb_col] += 1;
205                 }
206                 if (cpi->consec_zero_last_mvbias[map_index + mb_col] < 255) {
207                   cpi->consec_zero_last_mvbias[map_index + mb_col] += 1;
208                 }
209               } else {
210                 cpi->consec_zero_last[map_index + mb_col] = 0;
211                 cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
212               }
213               if (x->zero_last_dot_suppress) {
214                 cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
215               }
216             }
217 
218             /* Special case code for cyclic refresh
219              * If cyclic update enabled then copy
220              * xd->mbmi.segment_id; (which may have been updated
221              * based on mode during
222              * vp8cx_encode_inter_macroblock()) back into the
223              * global segmentation map
224              */
225             if ((cpi->current_layer == 0) &&
226                 (cpi->cyclic_refresh_mode_enabled &&
227                  xd->segmentation_enabled)) {
228               const MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
229               cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
230 
231               /* If the block has been refreshed mark it as clean
232                * (the magnitude of the -ve influences how long it
233                * will be before we consider another refresh):
234                * Else if it was coded (last frame 0,0) and has
235                * not already been refreshed then mark it as a
236                * candidate for cleanup next time (marked 0) else
237                * mark it as dirty (1).
238                */
239               if (mbmi->segment_id) {
240                 cpi->cyclic_refresh_map[map_index + mb_col] = -1;
241               } else if ((mbmi->mode == ZEROMV) &&
242                          (mbmi->ref_frame == LAST_FRAME)) {
243                 if (cpi->cyclic_refresh_map[map_index + mb_col] == 1) {
244                   cpi->cyclic_refresh_map[map_index + mb_col] = 0;
245                 }
246               } else {
247                 cpi->cyclic_refresh_map[map_index + mb_col] = 1;
248               }
249             }
250           }
251 
252 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
253           /* pack tokens for this MB */
254           {
255             int tok_count = tp - tp_start;
256             vp8_pack_tokens(w, tp_start, tok_count);
257           }
258 #else
259           cpi->tplist[mb_row].stop = tp;
260 #endif
261           /* Increment pointer into gf usage flags structure. */
262           x->gf_active_ptr++;
263 
264           /* Increment the activity mask pointers. */
265           x->mb_activity_ptr++;
266 
267           /* adjust to the next column of macroblocks */
268           x->src.y_buffer += 16;
269           x->src.u_buffer += 8;
270           x->src.v_buffer += 8;
271 
272           recon_yoffset += 16;
273           recon_uvoffset += 8;
274 
275           /* Keep track of segment usage */
276           segment_counts[xd->mode_info_context->mbmi.segment_id]++;
277 
278           /* skip to next mb */
279           xd->mode_info_context++;
280           x->partition_info++;
281           xd->above_context++;
282         }
283 
284         vp8_extend_mb_row(&cm->yv12_fb[dst_fb_idx], xd->dst.y_buffer + 16,
285                           xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
286 
287         vpx_atomic_store_release(current_mb_col, mb_col + nsync);
288 
289         /* this is to account for the border */
290         xd->mode_info_context++;
291         x->partition_info++;
292 
293         x->src.y_buffer +=
294             16 * x->src.y_stride * (cpi->encoding_thread_count + 1) -
295             16 * cm->mb_cols;
296         x->src.u_buffer +=
297             8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
298             8 * cm->mb_cols;
299         x->src.v_buffer +=
300             8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
301             8 * cm->mb_cols;
302 
303         xd->mode_info_context +=
304             xd->mode_info_stride * cpi->encoding_thread_count;
305         x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
306         x->gf_active_ptr += cm->mb_cols * cpi->encoding_thread_count;
307       }
308       /* Signal that this thread has completed processing its rows. */
309       sem_post(&cpi->h_event_end_encoding[ithread]);
310     }
311   }
312 
313   /* printf("exit thread %d\n", ithread); */
314   return 0;
315 }
316 
setup_mbby_copy(MACROBLOCK * mbdst,MACROBLOCK * mbsrc)317 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc) {
318   MACROBLOCK *x = mbsrc;
319   MACROBLOCK *z = mbdst;
320   int i;
321 
322   z->ss = x->ss;
323   z->ss_count = x->ss_count;
324   z->searches_per_step = x->searches_per_step;
325   z->errorperbit = x->errorperbit;
326 
327   z->sadperbit16 = x->sadperbit16;
328   z->sadperbit4 = x->sadperbit4;
329 
330   /*
331   z->mv_col_min    = x->mv_col_min;
332   z->mv_col_max    = x->mv_col_max;
333   z->mv_row_min    = x->mv_row_min;
334   z->mv_row_max    = x->mv_row_max;
335   */
336 
337   z->short_fdct4x4 = x->short_fdct4x4;
338   z->short_fdct8x4 = x->short_fdct8x4;
339   z->short_walsh4x4 = x->short_walsh4x4;
340   z->quantize_b = x->quantize_b;
341   z->optimize = x->optimize;
342 
343   /*
344   z->mvc              = x->mvc;
345   z->src.y_buffer      = x->src.y_buffer;
346   z->src.u_buffer      = x->src.u_buffer;
347   z->src.v_buffer      = x->src.v_buffer;
348   */
349 
350   z->mvcost[0] = x->mvcost[0];
351   z->mvcost[1] = x->mvcost[1];
352   z->mvsadcost[0] = x->mvsadcost[0];
353   z->mvsadcost[1] = x->mvsadcost[1];
354 
355   z->token_costs = x->token_costs;
356   z->inter_bmode_costs = x->inter_bmode_costs;
357   z->mbmode_cost = x->mbmode_cost;
358   z->intra_uv_mode_cost = x->intra_uv_mode_cost;
359   z->bmode_costs = x->bmode_costs;
360 
361   for (i = 0; i < 25; ++i) {
362     z->block[i].quant = x->block[i].quant;
363     z->block[i].quant_fast = x->block[i].quant_fast;
364     z->block[i].quant_shift = x->block[i].quant_shift;
365     z->block[i].zbin = x->block[i].zbin;
366     z->block[i].zrun_zbin_boost = x->block[i].zrun_zbin_boost;
367     z->block[i].round = x->block[i].round;
368     z->block[i].src_stride = x->block[i].src_stride;
369   }
370 
371   z->q_index = x->q_index;
372   z->act_zbin_adj = x->act_zbin_adj;
373   z->last_act_zbin_adj = x->last_act_zbin_adj;
374 
375   {
376     MACROBLOCKD *xd = &x->e_mbd;
377     MACROBLOCKD *zd = &z->e_mbd;
378 
379     /*
380     zd->mode_info_context = xd->mode_info_context;
381     zd->mode_info        = xd->mode_info;
382 
383     zd->mode_info_stride  = xd->mode_info_stride;
384     zd->frame_type       = xd->frame_type;
385     zd->up_available     = xd->up_available   ;
386     zd->left_available   = xd->left_available;
387     zd->left_context     = xd->left_context;
388     zd->last_frame_dc     = xd->last_frame_dc;
389     zd->last_frame_dccons = xd->last_frame_dccons;
390     zd->gold_frame_dc     = xd->gold_frame_dc;
391     zd->gold_frame_dccons = xd->gold_frame_dccons;
392     zd->mb_to_left_edge    = xd->mb_to_left_edge;
393     zd->mb_to_right_edge   = xd->mb_to_right_edge;
394     zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
395     zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
396     zd->gf_active_ptr     = xd->gf_active_ptr;
397     zd->frames_since_golden       = xd->frames_since_golden;
398     zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
399     */
400     zd->subpixel_predict = xd->subpixel_predict;
401     zd->subpixel_predict8x4 = xd->subpixel_predict8x4;
402     zd->subpixel_predict8x8 = xd->subpixel_predict8x8;
403     zd->subpixel_predict16x16 = xd->subpixel_predict16x16;
404     zd->segmentation_enabled = xd->segmentation_enabled;
405     zd->mb_segement_abs_delta = xd->mb_segement_abs_delta;
406     memcpy(zd->segment_feature_data, xd->segment_feature_data,
407            sizeof(xd->segment_feature_data));
408 
409     memcpy(zd->dequant_y1_dc, xd->dequant_y1_dc, sizeof(xd->dequant_y1_dc));
410     memcpy(zd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
411     memcpy(zd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
412     memcpy(zd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
413 
414 #if 1
415     /*TODO:  Remove dequant from BLOCKD.  This is a temporary solution until
416      * the quantizer code uses a passed in pointer to the dequant constants.
417      * This will also require modifications to the x86 and neon assembly.
418      * */
419     for (i = 0; i < 16; ++i) zd->block[i].dequant = zd->dequant_y1;
420     for (i = 16; i < 24; ++i) zd->block[i].dequant = zd->dequant_uv;
421     zd->block[24].dequant = zd->dequant_y2;
422 #endif
423 
424     memcpy(z->rd_threshes, x->rd_threshes, sizeof(x->rd_threshes));
425     memcpy(z->rd_thresh_mult, x->rd_thresh_mult, sizeof(x->rd_thresh_mult));
426 
427     z->zbin_over_quant = x->zbin_over_quant;
428     z->zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
429     z->zbin_mode_boost = x->zbin_mode_boost;
430 
431     memset(z->error_bins, 0, sizeof(z->error_bins));
432   }
433 }
434 
vp8cx_init_mbrthread_data(VP8_COMP * cpi,MACROBLOCK * x,MB_ROW_COMP * mbr_ei,int count)435 void vp8cx_init_mbrthread_data(VP8_COMP *cpi, MACROBLOCK *x,
436                                MB_ROW_COMP *mbr_ei, int count) {
437   VP8_COMMON *const cm = &cpi->common;
438   MACROBLOCKD *const xd = &x->e_mbd;
439   int i;
440 
441   for (i = 0; i < count; ++i) {
442     MACROBLOCK *mb = &mbr_ei[i].mb;
443     MACROBLOCKD *mbd = &mb->e_mbd;
444 
445     mbd->subpixel_predict = xd->subpixel_predict;
446     mbd->subpixel_predict8x4 = xd->subpixel_predict8x4;
447     mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
448     mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
449     mb->gf_active_ptr = x->gf_active_ptr;
450 
451     memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
452     mbr_ei[i].totalrate = 0;
453 
454     mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
455 
456     mbd->frame_type = cm->frame_type;
457 
458     mb->src = *cpi->Source;
459     mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
460     mbd->dst = cm->yv12_fb[cm->new_fb_idx];
461 
462     mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
463     mb->src.u_buffer += 8 * x->src.uv_stride * (i + 1);
464     mb->src.v_buffer += 8 * x->src.uv_stride * (i + 1);
465 
466     vp8_build_block_offsets(mb);
467 
468     mbd->left_context = &cm->left_context;
469     mb->mvc = cm->fc.mvc;
470 
471     setup_mbby_copy(&mbr_ei[i].mb, x);
472 
473     mbd->fullpixel_mask = 0xffffffff;
474     if (cm->full_pixel) mbd->fullpixel_mask = 0xfffffff8;
475 
476     vp8_zero(mb->coef_counts);
477     vp8_zero(x->ymode_count);
478     mb->skip_true_count = 0;
479     vp8_zero(mb->MVcount);
480     mb->prediction_error = 0;
481     mb->intra_error = 0;
482     vp8_zero(mb->count_mb_ref_frame_usage);
483     mb->mbs_tested_so_far = 0;
484     mb->mbs_zero_last_dot_suppress = 0;
485   }
486 }
487 
vp8cx_create_encoder_threads(VP8_COMP * cpi)488 int vp8cx_create_encoder_threads(VP8_COMP *cpi) {
489   const VP8_COMMON *cm = &cpi->common;
490 
491   vpx_atomic_init(&cpi->b_multi_threaded, 0);
492   cpi->encoding_thread_count = 0;
493   cpi->b_lpf_running = 0;
494 
495   if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1) {
496     int ithread;
497     int th_count = cpi->oxcf.multi_threaded - 1;
498     int rc = 0;
499 
500     /* don't allocate more threads than cores available */
501     if (cpi->oxcf.multi_threaded > cm->processor_core_count) {
502       th_count = cm->processor_core_count - 1;
503     }
504 
505     /* we have th_count + 1 (main) threads processing one row each */
506     /* no point to have more threads than the sync range allows */
507     if (th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1)) {
508       th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
509     }
510 
511     if (th_count == 0) return 0;
512 
513     CHECK_MEM_ERROR(cpi->h_encoding_thread,
514                     vpx_malloc(sizeof(pthread_t) * th_count));
515     CHECK_MEM_ERROR(cpi->h_event_start_encoding,
516                     vpx_malloc(sizeof(sem_t) * th_count));
517     CHECK_MEM_ERROR(cpi->h_event_end_encoding,
518                     vpx_malloc(sizeof(sem_t) * th_count));
519     CHECK_MEM_ERROR(cpi->mb_row_ei,
520                     vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
521     memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
522     CHECK_MEM_ERROR(cpi->en_thread_data,
523                     vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
524 
525     vpx_atomic_store_release(&cpi->b_multi_threaded, 1);
526     cpi->encoding_thread_count = th_count;
527 
528     /*
529     printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
530            (cpi->encoding_thread_count +1));
531     */
532 
533     for (ithread = 0; ithread < th_count; ++ithread) {
534       ENCODETHREAD_DATA *ethd = &cpi->en_thread_data[ithread];
535 
536       /* Setup block ptrs and offsets */
537       vp8_setup_block_ptrs(&cpi->mb_row_ei[ithread].mb);
538       vp8_setup_block_dptrs(&cpi->mb_row_ei[ithread].mb.e_mbd);
539 
540       sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
541       sem_init(&cpi->h_event_end_encoding[ithread], 0, 0);
542 
543       ethd->ithread = ithread;
544       ethd->ptr1 = (void *)cpi;
545       ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
546 
547       rc = pthread_create(&cpi->h_encoding_thread[ithread], 0,
548                           thread_encoding_proc, ethd);
549       if (rc) break;
550     }
551 
552     if (rc) {
553       /* shutdown other threads */
554       vpx_atomic_store_release(&cpi->b_multi_threaded, 0);
555       for (--ithread; ithread >= 0; ithread--) {
556         pthread_join(cpi->h_encoding_thread[ithread], 0);
557         sem_destroy(&cpi->h_event_start_encoding[ithread]);
558         sem_destroy(&cpi->h_event_end_encoding[ithread]);
559       }
560 
561       /* free thread related resources */
562       vpx_free(cpi->h_event_start_encoding);
563       vpx_free(cpi->h_event_end_encoding);
564       vpx_free(cpi->h_encoding_thread);
565       vpx_free(cpi->mb_row_ei);
566       vpx_free(cpi->en_thread_data);
567 
568       return -1;
569     }
570 
571     {
572       LPFTHREAD_DATA *lpfthd = &cpi->lpf_thread_data;
573 
574       sem_init(&cpi->h_event_start_lpf, 0, 0);
575       sem_init(&cpi->h_event_end_lpf, 0, 0);
576 
577       lpfthd->ptr1 = (void *)cpi;
578       rc = pthread_create(&cpi->h_filter_thread, 0, thread_loopfilter, lpfthd);
579 
580       if (rc) {
581         /* shutdown other threads */
582         vpx_atomic_store_release(&cpi->b_multi_threaded, 0);
583         for (--ithread; ithread >= 0; ithread--) {
584           sem_post(&cpi->h_event_start_encoding[ithread]);
585           sem_post(&cpi->h_event_end_encoding[ithread]);
586           pthread_join(cpi->h_encoding_thread[ithread], 0);
587           sem_destroy(&cpi->h_event_start_encoding[ithread]);
588           sem_destroy(&cpi->h_event_end_encoding[ithread]);
589         }
590         sem_destroy(&cpi->h_event_end_lpf);
591         sem_destroy(&cpi->h_event_start_lpf);
592 
593         /* free thread related resources */
594         vpx_free(cpi->h_event_start_encoding);
595         vpx_free(cpi->h_event_end_encoding);
596         vpx_free(cpi->h_encoding_thread);
597         vpx_free(cpi->mb_row_ei);
598         vpx_free(cpi->en_thread_data);
599 
600         return -2;
601       }
602     }
603   }
604   return 0;
605 }
606 
vp8cx_remove_encoder_threads(VP8_COMP * cpi)607 void vp8cx_remove_encoder_threads(VP8_COMP *cpi) {
608   if (vpx_atomic_load_acquire(&cpi->b_multi_threaded)) {
609     /* shutdown other threads */
610     vpx_atomic_store_release(&cpi->b_multi_threaded, 0);
611     {
612       int i;
613 
614       for (i = 0; i < cpi->encoding_thread_count; ++i) {
615         sem_post(&cpi->h_event_start_encoding[i]);
616         sem_post(&cpi->h_event_end_encoding[i]);
617 
618         pthread_join(cpi->h_encoding_thread[i], 0);
619 
620         sem_destroy(&cpi->h_event_start_encoding[i]);
621         sem_destroy(&cpi->h_event_end_encoding[i]);
622       }
623 
624       sem_post(&cpi->h_event_start_lpf);
625       pthread_join(cpi->h_filter_thread, 0);
626     }
627 
628     sem_destroy(&cpi->h_event_end_lpf);
629     sem_destroy(&cpi->h_event_start_lpf);
630 
631     /* free thread related resources */
632     vpx_free(cpi->h_event_start_encoding);
633     vpx_free(cpi->h_event_end_encoding);
634     vpx_free(cpi->h_encoding_thread);
635     vpx_free(cpi->mb_row_ei);
636     vpx_free(cpi->en_thread_data);
637   }
638 }
639 #endif
640