xref: /aosp_15_r20/external/libvpx/vp8/encoder/ethreading.c (revision fb1b10ab9aebc7c7068eedab379b749d7e3900be)
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 #include <stddef.h>
11 
12 #include "onyx_int.h"
13 #include "vpx_util/vpx_pthread.h"
14 #include "vp8/common/threading.h"
15 #include "vp8/common/common.h"
16 #include "vp8/common/extend.h"
17 #include "bitstream.h"
18 #include "encodeframe.h"
19 #include "ethreading.h"
20 
21 #if CONFIG_MULTITHREAD
22 
23 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x,
24                                     int ok_to_skip);
25 
thread_loopfilter(void * p_data)26 static THREADFN thread_loopfilter(void *p_data) {
27   VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
28   VP8_COMMON *cm = &cpi->common;
29 
30   while (1) {
31     if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) == 0) break;
32 
33     if (vp8_sem_wait(&cpi->h_event_start_lpf) == 0) {
34       /* we're shutting down */
35       if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) == 0) break;
36 
37       vp8_loopfilter_frame(cpi, cm);
38 
39       vp8_sem_post(&cpi->h_event_end_lpf);
40     }
41   }
42 
43   return THREAD_EXIT_SUCCESS;
44 }
45 
thread_encoding_proc(void * p_data)46 static THREADFN thread_encoding_proc(void *p_data) {
47   int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
48   VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
49   MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
50   ENTROPY_CONTEXT_PLANES mb_row_left_context;
51 
52   while (1) {
53     if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) == 0) break;
54 
55     if (vp8_sem_wait(&cpi->h_event_start_encoding[ithread]) == 0) {
56       const int nsync = cpi->mt_sync_range;
57       VP8_COMMON *cm = &cpi->common;
58       int mb_row;
59       MACROBLOCK *x = &mbri->mb;
60       MACROBLOCKD *xd = &x->e_mbd;
61       TOKENEXTRA *tp;
62 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
63       TOKENEXTRA *tp_start = cpi->tok + (1 + ithread) * (16 * 24);
64       const int num_part = (1 << cm->multi_token_partition);
65 #endif
66 
67       int *segment_counts = mbri->segment_counts;
68       int *totalrate = &mbri->totalrate;
69 
70       /* we're shutting down */
71       if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) == 0) break;
72 
73       xd->mode_info_context = cm->mi + cm->mode_info_stride * (ithread + 1);
74       xd->mode_info_stride = cm->mode_info_stride;
75 
76       for (mb_row = ithread + 1; mb_row < cm->mb_rows;
77            mb_row += (cpi->encoding_thread_count + 1)) {
78         int recon_yoffset, recon_uvoffset;
79         int mb_col;
80         int ref_fb_idx = cm->lst_fb_idx;
81         int dst_fb_idx = cm->new_fb_idx;
82         int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
83         int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
84         int map_index = (mb_row * cm->mb_cols);
85         const vpx_atomic_int *last_row_current_mb_col;
86         vpx_atomic_int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
87 
88 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
89         vp8_writer *w = &cpi->bc[1 + (mb_row % num_part)];
90 #else
91         tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
92         cpi->tplist[mb_row].start = tp;
93 #endif
94 
95         last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
96 
97         /* reset above block coeffs */
98         xd->above_context = cm->above_context;
99         xd->left_context = &mb_row_left_context;
100 
101         vp8_zero(mb_row_left_context);
102 
103         xd->up_available = (mb_row != 0);
104         recon_yoffset = (mb_row * recon_y_stride * 16);
105         recon_uvoffset = (mb_row * recon_uv_stride * 8);
106 
107         /* Set the mb activity pointer to the start of the row. */
108         x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
109 
110         /* for each macroblock col in image */
111         for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
112           if (((mb_col - 1) % nsync) == 0) {
113             vpx_atomic_store_release(current_mb_col, mb_col - 1);
114           }
115 
116           if (mb_row && !(mb_col & (nsync - 1))) {
117             vp8_atomic_spin_wait(mb_col, last_row_current_mb_col, nsync);
118           }
119 
120 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
121           tp = tp_start;
122 #endif
123 
124           /* Distance of Mb to the various image edges.
125            * These specified to 8th pel as they are always compared
126            * to values that are in 1/8th pel units
127            */
128           xd->mb_to_left_edge = -((mb_col * 16) << 3);
129           xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
130           xd->mb_to_top_edge = -((mb_row * 16) << 3);
131           xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
132 
133           /* Set up limit values for motion vectors used to prevent
134            * them extending outside the UMV borders
135            */
136           x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
137           x->mv_col_max =
138               ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
139           x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
140           x->mv_row_max =
141               ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
142 
143           xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
144           xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
145           xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
146           xd->left_available = (mb_col != 0);
147 
148           x->rddiv = cpi->RDDIV;
149           x->rdmult = cpi->RDMULT;
150 
151           /* Copy current mb to a buffer */
152           vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
153 
154           if (cpi->oxcf.tuning == VP8_TUNE_SSIM) vp8_activity_masking(cpi, x);
155 
156           /* Is segmentation enabled */
157           /* MB level adjustment to quantizer */
158           if (xd->segmentation_enabled) {
159             /* Code to set segment id in xd->mbmi.segment_id for
160              * current MB (with range checking)
161              */
162             if (cpi->segmentation_map[map_index + mb_col] <= 3) {
163               xd->mode_info_context->mbmi.segment_id =
164                   cpi->segmentation_map[map_index + mb_col];
165             } else {
166               xd->mode_info_context->mbmi.segment_id = 0;
167             }
168 
169             vp8cx_mb_init_quantizer(cpi, x, 1);
170           } else {
171             /* Set to Segment 0 by default */
172             xd->mode_info_context->mbmi.segment_id = 0;
173           }
174 
175           x->active_ptr = cpi->active_map + map_index + mb_col;
176 
177           if (cm->frame_type == KEY_FRAME) {
178             *totalrate += vp8cx_encode_intra_macroblock(cpi, x, &tp);
179 #ifdef MODE_STATS
180             y_modes[xd->mbmi.mode]++;
181 #endif
182           } else {
183             *totalrate += vp8cx_encode_inter_macroblock(
184                 cpi, x, &tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
185 
186 #ifdef MODE_STATS
187             inter_y_modes[xd->mbmi.mode]++;
188 
189             if (xd->mbmi.mode == SPLITMV) {
190               int b;
191 
192               for (b = 0; b < xd->mbmi.partition_count; ++b) {
193                 inter_b_modes[x->partition->bmi[b].mode]++;
194               }
195             }
196 
197 #endif
198             // Keep track of how many (consecutive) times a  block
199             // is coded as ZEROMV_LASTREF, for base layer frames.
200             // Reset to 0 if its coded as anything else.
201             if (cpi->current_layer == 0) {
202               if (xd->mode_info_context->mbmi.mode == ZEROMV &&
203                   xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
204                 // Increment, check for wrap-around.
205                 if (cpi->consec_zero_last[map_index + mb_col] < 255) {
206                   cpi->consec_zero_last[map_index + mb_col] += 1;
207                 }
208                 if (cpi->consec_zero_last_mvbias[map_index + mb_col] < 255) {
209                   cpi->consec_zero_last_mvbias[map_index + mb_col] += 1;
210                 }
211               } else {
212                 cpi->consec_zero_last[map_index + mb_col] = 0;
213                 cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
214               }
215               if (x->zero_last_dot_suppress) {
216                 cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
217               }
218             }
219 
220             /* Special case code for cyclic refresh
221              * If cyclic update enabled then copy
222              * xd->mbmi.segment_id; (which may have been updated
223              * based on mode during
224              * vp8cx_encode_inter_macroblock()) back into the
225              * global segmentation map
226              */
227             if ((cpi->current_layer == 0) &&
228                 (cpi->cyclic_refresh_mode_enabled &&
229                  xd->segmentation_enabled)) {
230               const MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
231               cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
232 
233               /* If the block has been refreshed mark it as clean
234                * (the magnitude of the -ve influences how long it
235                * will be before we consider another refresh):
236                * Else if it was coded (last frame 0,0) and has
237                * not already been refreshed then mark it as a
238                * candidate for cleanup next time (marked 0) else
239                * mark it as dirty (1).
240                */
241               if (mbmi->segment_id) {
242                 cpi->cyclic_refresh_map[map_index + mb_col] = -1;
243               } else if ((mbmi->mode == ZEROMV) &&
244                          (mbmi->ref_frame == LAST_FRAME)) {
245                 if (cpi->cyclic_refresh_map[map_index + mb_col] == 1) {
246                   cpi->cyclic_refresh_map[map_index + mb_col] = 0;
247                 }
248               } else {
249                 cpi->cyclic_refresh_map[map_index + mb_col] = 1;
250               }
251             }
252           }
253 
254 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
255           /* pack tokens for this MB */
256           {
257             int tok_count = tp - tp_start;
258             vp8_pack_tokens(w, tp_start, tok_count);
259           }
260 #else
261           cpi->tplist[mb_row].stop = tp;
262 #endif
263           /* Increment pointer into gf usage flags structure. */
264           x->gf_active_ptr++;
265 
266           /* Increment the activity mask pointers. */
267           x->mb_activity_ptr++;
268 
269           /* adjust to the next column of macroblocks */
270           x->src.y_buffer += 16;
271           x->src.u_buffer += 8;
272           x->src.v_buffer += 8;
273 
274           recon_yoffset += 16;
275           recon_uvoffset += 8;
276 
277           /* Keep track of segment usage */
278           segment_counts[xd->mode_info_context->mbmi.segment_id]++;
279 
280           /* skip to next mb */
281           xd->mode_info_context++;
282           x->partition_info++;
283           xd->above_context++;
284         }
285 
286         vp8_extend_mb_row(&cm->yv12_fb[dst_fb_idx], xd->dst.y_buffer + 16,
287                           xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
288 
289         vpx_atomic_store_release(current_mb_col, mb_col + nsync);
290 
291         /* this is to account for the border */
292         xd->mode_info_context++;
293         x->partition_info++;
294 
295         x->src.y_buffer +=
296             16 * x->src.y_stride * (cpi->encoding_thread_count + 1) -
297             16 * cm->mb_cols;
298         x->src.u_buffer +=
299             8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
300             8 * cm->mb_cols;
301         x->src.v_buffer +=
302             8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
303             8 * cm->mb_cols;
304 
305         xd->mode_info_context +=
306             xd->mode_info_stride * cpi->encoding_thread_count;
307         x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
308         x->gf_active_ptr += cm->mb_cols * cpi->encoding_thread_count;
309       }
310       /* Signal that this thread has completed processing its rows. */
311       vp8_sem_post(&cpi->h_event_end_encoding[ithread]);
312     }
313   }
314 
315   /* printf("exit thread %d\n", ithread); */
316   return THREAD_EXIT_SUCCESS;
317 }
318 
setup_mbby_copy(MACROBLOCK * mbdst,MACROBLOCK * mbsrc)319 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc) {
320   MACROBLOCK *x = mbsrc;
321   MACROBLOCK *z = mbdst;
322   int i;
323 
324   z->ss = x->ss;
325   z->ss_count = x->ss_count;
326   z->searches_per_step = x->searches_per_step;
327   z->errorperbit = x->errorperbit;
328 
329   z->sadperbit16 = x->sadperbit16;
330   z->sadperbit4 = x->sadperbit4;
331 
332   /*
333   z->mv_col_min    = x->mv_col_min;
334   z->mv_col_max    = x->mv_col_max;
335   z->mv_row_min    = x->mv_row_min;
336   z->mv_row_max    = x->mv_row_max;
337   */
338 
339   z->short_fdct4x4 = x->short_fdct4x4;
340   z->short_fdct8x4 = x->short_fdct8x4;
341   z->short_walsh4x4 = x->short_walsh4x4;
342   z->quantize_b = x->quantize_b;
343   z->optimize = x->optimize;
344 
345   /*
346   z->mvc              = x->mvc;
347   z->src.y_buffer      = x->src.y_buffer;
348   z->src.u_buffer      = x->src.u_buffer;
349   z->src.v_buffer      = x->src.v_buffer;
350   */
351 
352   z->mvcost[0] = x->mvcost[0];
353   z->mvcost[1] = x->mvcost[1];
354   z->mvsadcost[0] = x->mvsadcost[0];
355   z->mvsadcost[1] = x->mvsadcost[1];
356 
357   z->token_costs = x->token_costs;
358   z->inter_bmode_costs = x->inter_bmode_costs;
359   z->mbmode_cost = x->mbmode_cost;
360   z->intra_uv_mode_cost = x->intra_uv_mode_cost;
361   z->bmode_costs = x->bmode_costs;
362 
363   for (i = 0; i < 25; ++i) {
364     z->block[i].quant = x->block[i].quant;
365     z->block[i].quant_fast = x->block[i].quant_fast;
366     z->block[i].quant_shift = x->block[i].quant_shift;
367     z->block[i].zbin = x->block[i].zbin;
368     z->block[i].zrun_zbin_boost = x->block[i].zrun_zbin_boost;
369     z->block[i].round = x->block[i].round;
370     z->block[i].src_stride = x->block[i].src_stride;
371   }
372 
373   z->q_index = x->q_index;
374   z->act_zbin_adj = x->act_zbin_adj;
375   z->last_act_zbin_adj = x->last_act_zbin_adj;
376 
377   {
378     MACROBLOCKD *xd = &x->e_mbd;
379     MACROBLOCKD *zd = &z->e_mbd;
380 
381     /*
382     zd->mode_info_context = xd->mode_info_context;
383     zd->mode_info        = xd->mode_info;
384 
385     zd->mode_info_stride  = xd->mode_info_stride;
386     zd->frame_type       = xd->frame_type;
387     zd->up_available     = xd->up_available   ;
388     zd->left_available   = xd->left_available;
389     zd->left_context     = xd->left_context;
390     zd->last_frame_dc     = xd->last_frame_dc;
391     zd->last_frame_dccons = xd->last_frame_dccons;
392     zd->gold_frame_dc     = xd->gold_frame_dc;
393     zd->gold_frame_dccons = xd->gold_frame_dccons;
394     zd->mb_to_left_edge    = xd->mb_to_left_edge;
395     zd->mb_to_right_edge   = xd->mb_to_right_edge;
396     zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
397     zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
398     zd->gf_active_ptr     = xd->gf_active_ptr;
399     zd->frames_since_golden       = xd->frames_since_golden;
400     zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
401     */
402     zd->subpixel_predict = xd->subpixel_predict;
403     zd->subpixel_predict8x4 = xd->subpixel_predict8x4;
404     zd->subpixel_predict8x8 = xd->subpixel_predict8x8;
405     zd->subpixel_predict16x16 = xd->subpixel_predict16x16;
406     zd->segmentation_enabled = xd->segmentation_enabled;
407     zd->mb_segment_abs_delta = xd->mb_segment_abs_delta;
408     memcpy(zd->segment_feature_data, xd->segment_feature_data,
409            sizeof(xd->segment_feature_data));
410 
411     memcpy(zd->dequant_y1_dc, xd->dequant_y1_dc, sizeof(xd->dequant_y1_dc));
412     memcpy(zd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
413     memcpy(zd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
414     memcpy(zd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
415 
416 #if 1
417     /*TODO:  Remove dequant from BLOCKD.  This is a temporary solution until
418      * the quantizer code uses a passed in pointer to the dequant constants.
419      * This will also require modifications to the x86 and neon assembly.
420      * */
421     for (i = 0; i < 16; ++i) zd->block[i].dequant = zd->dequant_y1;
422     for (i = 16; i < 24; ++i) zd->block[i].dequant = zd->dequant_uv;
423     zd->block[24].dequant = zd->dequant_y2;
424 #endif
425 
426     memcpy(z->rd_threshes, x->rd_threshes, sizeof(x->rd_threshes));
427     memcpy(z->rd_thresh_mult, x->rd_thresh_mult, sizeof(x->rd_thresh_mult));
428 
429     z->zbin_over_quant = x->zbin_over_quant;
430     z->zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
431     z->zbin_mode_boost = x->zbin_mode_boost;
432 
433     memset(z->error_bins, 0, sizeof(z->error_bins));
434   }
435 }
436 
vp8cx_init_mbrthread_data(VP8_COMP * cpi,MACROBLOCK * x,MB_ROW_COMP * mbr_ei,int count)437 void vp8cx_init_mbrthread_data(VP8_COMP *cpi, MACROBLOCK *x,
438                                MB_ROW_COMP *mbr_ei, int count) {
439   VP8_COMMON *const cm = &cpi->common;
440   MACROBLOCKD *const xd = &x->e_mbd;
441   int i;
442 
443   for (i = 0; i < count; ++i) {
444     MACROBLOCK *mb = &mbr_ei[i].mb;
445     MACROBLOCKD *mbd = &mb->e_mbd;
446 
447     mbd->subpixel_predict = xd->subpixel_predict;
448     mbd->subpixel_predict8x4 = xd->subpixel_predict8x4;
449     mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
450     mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
451     mb->gf_active_ptr = x->gf_active_ptr;
452 
453     memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
454     mbr_ei[i].totalrate = 0;
455 
456     mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
457 
458     mbd->frame_type = cm->frame_type;
459 
460     mb->src = *cpi->Source;
461     mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
462     mbd->dst = cm->yv12_fb[cm->new_fb_idx];
463 
464     mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
465     mb->src.u_buffer += 8 * x->src.uv_stride * (i + 1);
466     mb->src.v_buffer += 8 * x->src.uv_stride * (i + 1);
467 
468     vp8_build_block_offsets(mb);
469 
470     mbd->left_context = &cm->left_context;
471     mb->mvc = cm->fc.mvc;
472 
473     setup_mbby_copy(&mbr_ei[i].mb, x);
474 
475     mbd->fullpixel_mask = ~0;
476     if (cm->full_pixel) mbd->fullpixel_mask = ~7;
477 
478     vp8_zero(mb->coef_counts);
479     vp8_zero(x->ymode_count);
480     mb->skip_true_count = 0;
481     vp8_zero(mb->MVcount);
482     mb->prediction_error = 0;
483     mb->intra_error = 0;
484     vp8_zero(mb->count_mb_ref_frame_usage);
485     mb->mbs_tested_so_far = 0;
486     mb->mbs_zero_last_dot_suppress = 0;
487   }
488 }
489 
vp8cx_create_encoder_threads(VP8_COMP * cpi)490 int vp8cx_create_encoder_threads(VP8_COMP *cpi) {
491   const VP8_COMMON *cm = &cpi->common;
492   int th_count = 0;
493 
494   if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1) {
495     th_count = cpi->oxcf.multi_threaded - 1;
496 
497     /* don't allocate more threads than cores available */
498     if (cpi->oxcf.multi_threaded > cm->processor_core_count) {
499       th_count = cm->processor_core_count - 1;
500     }
501 
502     /* we have th_count + 1 (main) threads processing one row each */
503     /* no point to have more threads than the sync range allows */
504     if (th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1)) {
505       th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
506     }
507   }
508   if (th_count == cpi->encoding_thread_count) return 0;
509 
510   vp8cx_remove_encoder_threads(cpi);
511   if (th_count != 0) {
512     int ithread;
513     int rc = 0;
514 
515     CHECK_MEM_ERROR(&cpi->common.error, cpi->h_encoding_thread,
516                     vpx_malloc(sizeof(pthread_t) * th_count));
517     CHECK_MEM_ERROR(&cpi->common.error, cpi->h_event_start_encoding,
518                     vpx_malloc(sizeof(vp8_sem_t) * th_count));
519     CHECK_MEM_ERROR(&cpi->common.error, cpi->h_event_end_encoding,
520                     vpx_malloc(sizeof(vp8_sem_t) * th_count));
521     CHECK_MEM_ERROR(&cpi->common.error, cpi->mb_row_ei,
522                     vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
523     memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
524     CHECK_MEM_ERROR(&cpi->common.error, cpi->en_thread_data,
525                     vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
526 
527     vpx_atomic_store_release(&cpi->b_multi_threaded, 1);
528     cpi->encoding_thread_count = th_count;
529 
530     /*
531     printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
532            (cpi->encoding_thread_count +1));
533     */
534 
535     for (ithread = 0; ithread < th_count; ++ithread) {
536       ENCODETHREAD_DATA *ethd = &cpi->en_thread_data[ithread];
537 
538       /* Setup block ptrs and offsets */
539       vp8_setup_block_ptrs(&cpi->mb_row_ei[ithread].mb);
540       vp8_setup_block_dptrs(&cpi->mb_row_ei[ithread].mb.e_mbd);
541 
542       vp8_sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
543       vp8_sem_init(&cpi->h_event_end_encoding[ithread], 0, 0);
544 
545       ethd->ithread = ithread;
546       ethd->ptr1 = (void *)cpi;
547       ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
548 
549       rc = pthread_create(&cpi->h_encoding_thread[ithread], 0,
550                           thread_encoding_proc, ethd);
551       if (rc) break;
552     }
553 
554     if (rc) {
555       /* shutdown other threads */
556       vpx_atomic_store_release(&cpi->b_multi_threaded, 0);
557       for (--ithread; ithread >= 0; ithread--) {
558         vp8_sem_post(&cpi->h_event_start_encoding[ithread]);
559         vp8_sem_post(&cpi->h_event_end_encoding[ithread]);
560         pthread_join(cpi->h_encoding_thread[ithread], 0);
561         vp8_sem_destroy(&cpi->h_event_start_encoding[ithread]);
562         vp8_sem_destroy(&cpi->h_event_end_encoding[ithread]);
563       }
564 
565       /* free thread related resources */
566       vpx_free(cpi->h_event_start_encoding);
567       cpi->h_event_start_encoding = NULL;
568       vpx_free(cpi->h_event_end_encoding);
569       cpi->h_event_end_encoding = NULL;
570       vpx_free(cpi->h_encoding_thread);
571       cpi->h_encoding_thread = NULL;
572       vpx_free(cpi->mb_row_ei);
573       cpi->mb_row_ei = NULL;
574       vpx_free(cpi->en_thread_data);
575       cpi->en_thread_data = NULL;
576       cpi->encoding_thread_count = 0;
577 
578       return -1;
579     }
580 
581     {
582       LPFTHREAD_DATA *lpfthd = &cpi->lpf_thread_data;
583 
584       vp8_sem_init(&cpi->h_event_start_lpf, 0, 0);
585       vp8_sem_init(&cpi->h_event_end_lpf, 0, 0);
586 
587       lpfthd->ptr1 = (void *)cpi;
588       rc = pthread_create(&cpi->h_filter_thread, 0, thread_loopfilter, lpfthd);
589 
590       if (rc) {
591         /* shutdown other threads */
592         vpx_atomic_store_release(&cpi->b_multi_threaded, 0);
593         for (--ithread; ithread >= 0; ithread--) {
594           vp8_sem_post(&cpi->h_event_start_encoding[ithread]);
595           vp8_sem_post(&cpi->h_event_end_encoding[ithread]);
596           pthread_join(cpi->h_encoding_thread[ithread], 0);
597           vp8_sem_destroy(&cpi->h_event_start_encoding[ithread]);
598           vp8_sem_destroy(&cpi->h_event_end_encoding[ithread]);
599         }
600         vp8_sem_destroy(&cpi->h_event_end_lpf);
601         vp8_sem_destroy(&cpi->h_event_start_lpf);
602 
603         /* free thread related resources */
604         vpx_free(cpi->h_event_start_encoding);
605         cpi->h_event_start_encoding = NULL;
606         vpx_free(cpi->h_event_end_encoding);
607         cpi->h_event_end_encoding = NULL;
608         vpx_free(cpi->h_encoding_thread);
609         cpi->h_encoding_thread = NULL;
610         vpx_free(cpi->mb_row_ei);
611         cpi->mb_row_ei = NULL;
612         vpx_free(cpi->en_thread_data);
613         cpi->en_thread_data = NULL;
614         cpi->encoding_thread_count = 0;
615 
616         return -2;
617       }
618     }
619   }
620   return 0;
621 }
622 
vp8cx_remove_encoder_threads(VP8_COMP * cpi)623 void vp8cx_remove_encoder_threads(VP8_COMP *cpi) {
624   if (vpx_atomic_load_acquire(&cpi->b_multi_threaded)) {
625     /* shutdown other threads */
626     vpx_atomic_store_release(&cpi->b_multi_threaded, 0);
627     {
628       int i;
629 
630       for (i = 0; i < cpi->encoding_thread_count; ++i) {
631         vp8_sem_post(&cpi->h_event_start_encoding[i]);
632         vp8_sem_post(&cpi->h_event_end_encoding[i]);
633 
634         pthread_join(cpi->h_encoding_thread[i], 0);
635 
636         vp8_sem_destroy(&cpi->h_event_start_encoding[i]);
637         vp8_sem_destroy(&cpi->h_event_end_encoding[i]);
638       }
639 
640       vp8_sem_post(&cpi->h_event_start_lpf);
641       pthread_join(cpi->h_filter_thread, 0);
642     }
643 
644     vp8_sem_destroy(&cpi->h_event_end_lpf);
645     vp8_sem_destroy(&cpi->h_event_start_lpf);
646     cpi->b_lpf_running = 0;
647 
648     /* free thread related resources */
649     vpx_free(cpi->mt_current_mb_col);
650     cpi->mt_current_mb_col = NULL;
651     cpi->mt_current_mb_col_size = 0;
652     vpx_free(cpi->h_event_start_encoding);
653     cpi->h_event_start_encoding = NULL;
654     vpx_free(cpi->h_event_end_encoding);
655     cpi->h_event_end_encoding = NULL;
656     vpx_free(cpi->h_encoding_thread);
657     cpi->h_encoding_thread = NULL;
658     vpx_free(cpi->mb_row_ei);
659     cpi->mb_row_ei = NULL;
660     vpx_free(cpi->en_thread_data);
661     cpi->en_thread_data = NULL;
662     cpi->encoding_thread_count = 0;
663   }
664 }
665 #endif
666