Merge "Skip computation of distortion in vp8_pick_inter_mode if active_map is used"
[libvpx.git] / vp8 / encoder / ethreading.c
blob1a37f03b9acea94ff8898d6c569a634b8cb1b7ad
1 /*
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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 */
11 #include "onyx_int.h"
12 #include "vp8/common/threading.h"
13 #include "vp8/common/common.h"
14 #include "vp8/common/extend.h"
16 #if CONFIG_MULTITHREAD
18 extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
19 TOKENEXTRA **t, int recon_yoffset,
20 int recon_uvoffset);
21 extern int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x,
22 TOKENEXTRA **t);
23 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x);
24 extern void vp8_build_block_offsets(MACROBLOCK *x);
25 extern void vp8_setup_block_ptrs(MACROBLOCK *x);
27 extern void loopfilter_frame(VP8_COMP *cpi, VP8_COMMON *cm);
29 static THREAD_FUNCTION loopfilter_thread(void *p_data)
31 VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
32 VP8_COMMON *cm = &cpi->common;
34 while (1)
36 if (cpi->b_multi_threaded == 0)
37 break;
39 if (sem_wait(&cpi->h_event_start_lpf) == 0)
41 if (cpi->b_multi_threaded == FALSE) // we're shutting down
42 break;
44 loopfilter_frame(cpi, cm);
46 sem_post(&cpi->h_event_end_lpf);
50 return 0;
53 static
54 THREAD_FUNCTION thread_encoding_proc(void *p_data)
56 int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
57 VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
58 MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
59 ENTROPY_CONTEXT_PLANES mb_row_left_context;
61 const int nsync = cpi->mt_sync_range;
62 //printf("Started thread %d\n", ithread);
64 while (1)
66 if (cpi->b_multi_threaded == 0)
67 break;
69 //if(WaitForSingleObject(cpi->h_event_mbrencoding[ithread], INFINITE) == WAIT_OBJECT_0)
70 if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0)
72 VP8_COMMON *cm = &cpi->common;
73 int mb_row;
74 MACROBLOCK *x = &mbri->mb;
75 MACROBLOCKD *xd = &x->e_mbd;
76 TOKENEXTRA *tp ;
78 int *segment_counts = mbri->segment_counts;
79 int *totalrate = &mbri->totalrate;
81 if (cpi->b_multi_threaded == FALSE) // we're shutting down
82 break;
84 for (mb_row = ithread + 1; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
87 int i;
88 int recon_yoffset, recon_uvoffset;
89 int mb_col;
90 int ref_fb_idx = cm->lst_fb_idx;
91 int dst_fb_idx = cm->new_fb_idx;
92 int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
93 int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
94 int map_index = (mb_row * cm->mb_cols);
95 volatile int *last_row_current_mb_col;
97 tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
99 last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
101 // reset above block coeffs
102 xd->above_context = cm->above_context;
103 xd->left_context = &mb_row_left_context;
105 vp8_zero(mb_row_left_context);
107 xd->up_available = (mb_row != 0);
108 recon_yoffset = (mb_row * recon_y_stride * 16);
109 recon_uvoffset = (mb_row * recon_uv_stride * 8);
111 cpi->tplist[mb_row].start = tp;
113 //printf("Thread mb_row = %d\n", mb_row);
115 // Set the mb activity pointer to the start of the row.
116 x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
118 // for each macroblock col in image
119 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
121 if ((mb_col & (nsync - 1)) == 0)
123 while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != cm->mb_cols - 1)
125 x86_pause_hint();
126 thread_sleep(0);
130 // Distance of Mb to the various image edges.
131 // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
132 xd->mb_to_left_edge = -((mb_col * 16) << 3);
133 xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
134 xd->mb_to_top_edge = -((mb_row * 16) << 3);
135 xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
137 // Set up limit values for motion vectors used to prevent them extending outside the UMV borders
138 x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
139 x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
140 x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
141 x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
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);
148 x->rddiv = cpi->RDDIV;
149 x->rdmult = cpi->RDMULT;
151 //Copy current mb to a buffer
152 RECON_INVOKE(&xd->rtcd->recon, copy16x16)(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
154 if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
155 vp8_activity_masking(cpi, x);
157 // Is segmentation enabled
158 // MB level adjutment to quantizer
159 if (xd->segmentation_enabled)
161 // Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
162 if (cpi->segmentation_map[map_index + mb_col] <= 3)
163 xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index + mb_col];
164 else
165 xd->mode_info_context->mbmi.segment_id = 0;
167 vp8cx_mb_init_quantizer(cpi, x);
169 else
170 xd->mode_info_context->mbmi.segment_id = 0; // Set to Segment 0 by default
172 x->active_ptr = cpi->active_map + map_index + mb_col;
174 if (cm->frame_type == KEY_FRAME)
176 *totalrate += vp8cx_encode_intra_macro_block(cpi, x, &tp);
177 #ifdef MODE_STATS
178 y_modes[xd->mbmi.mode] ++;
179 #endif
181 else
183 *totalrate += vp8cx_encode_inter_macroblock(cpi, x, &tp, recon_yoffset, recon_uvoffset);
185 #ifdef MODE_STATS
186 inter_y_modes[xd->mbmi.mode] ++;
188 if (xd->mbmi.mode == SPLITMV)
190 int b;
192 for (b = 0; b < xd->mbmi.partition_count; b++)
194 inter_b_modes[x->partition->bmi[b].mode] ++;
198 #endif
200 // Count of last ref frame 0,0 useage
201 if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
202 cpi->inter_zz_count++;
204 // Special case code for cyclic refresh
205 // If cyclic update enabled then copy xd->mbmi.segment_id; (which may have been updated based on mode
206 // during vp8cx_encode_inter_macroblock()) back into the global sgmentation map
207 if (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)
209 const MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
210 cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
212 // If the block has been refreshed mark it as clean (the magnitude of the -ve influences how long it will be before we consider another refresh):
213 // Else if it was coded (last frame 0,0) and has not already been refreshed then mark it as a candidate for cleanup next time (marked 0)
214 // else mark it as dirty (1).
215 if (mbmi->segment_id)
216 cpi->cyclic_refresh_map[map_index + mb_col] = -1;
217 else if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
219 if (cpi->cyclic_refresh_map[map_index + mb_col] == 1)
220 cpi->cyclic_refresh_map[map_index + mb_col] = 0;
222 else
223 cpi->cyclic_refresh_map[map_index + mb_col] = 1;
227 cpi->tplist[mb_row].stop = tp;
229 // Increment pointer into gf useage flags structure.
230 x->gf_active_ptr++;
232 // Increment the activity mask pointers.
233 x->mb_activity_ptr++;
235 /* save the block info */
236 for (i = 0; i < 16; i++)
237 xd->mode_info_context->bmi[i] = xd->block[i].bmi;
239 // adjust to the next column of macroblocks
240 x->src.y_buffer += 16;
241 x->src.u_buffer += 8;
242 x->src.v_buffer += 8;
244 recon_yoffset += 16;
245 recon_uvoffset += 8;
247 // Keep track of segment useage
248 segment_counts[xd->mode_info_context->mbmi.segment_id]++;
250 // skip to next mb
251 xd->mode_info_context++;
252 x->partition_info++;
253 xd->above_context++;
255 cpi->mt_current_mb_col[mb_row] = mb_col;
258 //extend the recon for intra prediction
259 vp8_extend_mb_row(
260 &cm->yv12_fb[dst_fb_idx],
261 xd->dst.y_buffer + 16,
262 xd->dst.u_buffer + 8,
263 xd->dst.v_buffer + 8);
265 // this is to account for the border
266 xd->mode_info_context++;
267 x->partition_info++;
269 x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
270 x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
271 x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
273 xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
274 x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
275 x->gf_active_ptr += cm->mb_cols * cpi->encoding_thread_count;
277 if (mb_row == cm->mb_rows - 1)
279 //SetEvent(cpi->h_event_main);
280 sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
286 //printf("exit thread %d\n", ithread);
287 return 0;
290 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
293 MACROBLOCK *x = mbsrc;
294 MACROBLOCK *z = mbdst;
295 int i;
297 z->ss = x->ss;
298 z->ss_count = x->ss_count;
299 z->searches_per_step = x->searches_per_step;
300 z->errorperbit = x->errorperbit;
302 z->sadperbit16 = x->sadperbit16;
303 z->sadperbit4 = x->sadperbit4;
306 z->mv_col_min = x->mv_col_min;
307 z->mv_col_max = x->mv_col_max;
308 z->mv_row_min = x->mv_row_min;
309 z->mv_row_max = x->mv_row_max;
310 z->vector_range = x->vector_range ;
313 z->vp8_short_fdct4x4 = x->vp8_short_fdct4x4;
314 z->vp8_short_fdct8x4 = x->vp8_short_fdct8x4;
315 z->short_walsh4x4 = x->short_walsh4x4;
316 z->quantize_b = x->quantize_b;
317 z->quantize_b_pair = x->quantize_b_pair;
318 z->optimize = x->optimize;
321 z->mvc = x->mvc;
322 z->src.y_buffer = x->src.y_buffer;
323 z->src.u_buffer = x->src.u_buffer;
324 z->src.v_buffer = x->src.v_buffer;
328 vpx_memcpy(z->mvcosts, x->mvcosts, sizeof(x->mvcosts));
329 z->mvcost[0] = &z->mvcosts[0][mv_max+1];
330 z->mvcost[1] = &z->mvcosts[1][mv_max+1];
331 z->mvsadcost[0] = &z->mvsadcosts[0][mvfp_max+1];
332 z->mvsadcost[1] = &z->mvsadcosts[1][mvfp_max+1];
335 vpx_memcpy(z->token_costs, x->token_costs, sizeof(x->token_costs));
336 vpx_memcpy(z->inter_bmode_costs, x->inter_bmode_costs, sizeof(x->inter_bmode_costs));
337 //memcpy(z->mvcosts, x->mvcosts, sizeof(x->mvcosts));
338 //memcpy(z->mvcost, x->mvcost, sizeof(x->mvcost));
339 vpx_memcpy(z->mbmode_cost, x->mbmode_cost, sizeof(x->mbmode_cost));
340 vpx_memcpy(z->intra_uv_mode_cost, x->intra_uv_mode_cost, sizeof(x->intra_uv_mode_cost));
341 vpx_memcpy(z->bmode_costs, x->bmode_costs, sizeof(x->bmode_costs));
343 for (i = 0; i < 25; i++)
345 z->block[i].quant = x->block[i].quant;
346 z->block[i].quant_fast = x->block[i].quant_fast;
347 z->block[i].quant_shift = x->block[i].quant_shift;
348 z->block[i].zbin = x->block[i].zbin;
349 z->block[i].zrun_zbin_boost = x->block[i].zrun_zbin_boost;
350 z->block[i].round = x->block[i].round;
352 z->block[i].src = x->block[i].src;
354 z->block[i].src_stride = x->block[i].src_stride;
355 z->block[i].force_empty = x->block[i].force_empty;
360 MACROBLOCKD *xd = &x->e_mbd;
361 MACROBLOCKD *zd = &z->e_mbd;
364 zd->mode_info_context = xd->mode_info_context;
365 zd->mode_info = xd->mode_info;
367 zd->mode_info_stride = xd->mode_info_stride;
368 zd->frame_type = xd->frame_type;
369 zd->up_available = xd->up_available ;
370 zd->left_available = xd->left_available;
371 zd->left_context = xd->left_context;
372 zd->last_frame_dc = xd->last_frame_dc;
373 zd->last_frame_dccons = xd->last_frame_dccons;
374 zd->gold_frame_dc = xd->gold_frame_dc;
375 zd->gold_frame_dccons = xd->gold_frame_dccons;
376 zd->mb_to_left_edge = xd->mb_to_left_edge;
377 zd->mb_to_right_edge = xd->mb_to_right_edge;
378 zd->mb_to_top_edge = xd->mb_to_top_edge ;
379 zd->mb_to_bottom_edge = xd->mb_to_bottom_edge;
380 zd->gf_active_ptr = xd->gf_active_ptr;
381 zd->frames_since_golden = xd->frames_since_golden;
382 zd->frames_till_alt_ref_frame = xd->frames_till_alt_ref_frame;
384 zd->subpixel_predict = xd->subpixel_predict;
385 zd->subpixel_predict8x4 = xd->subpixel_predict8x4;
386 zd->subpixel_predict8x8 = xd->subpixel_predict8x8;
387 zd->subpixel_predict16x16 = xd->subpixel_predict16x16;
388 zd->segmentation_enabled = xd->segmentation_enabled;
389 zd->mb_segement_abs_delta = xd->mb_segement_abs_delta;
390 vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
392 for (i = 0; i < 25; i++)
394 zd->block[i].dequant = xd->block[i].dequant;
399 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
400 MACROBLOCK *x,
401 MB_ROW_COMP *mbr_ei,
402 int mb_row,
403 int count
407 VP8_COMMON *const cm = & cpi->common;
408 MACROBLOCKD *const xd = & x->e_mbd;
409 int i;
410 (void) mb_row;
412 for (i = 0; i < count; i++)
414 MACROBLOCK *mb = & mbr_ei[i].mb;
415 MACROBLOCKD *mbd = &mb->e_mbd;
417 mbd->subpixel_predict = xd->subpixel_predict;
418 mbd->subpixel_predict8x4 = xd->subpixel_predict8x4;
419 mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
420 mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
421 #if CONFIG_RUNTIME_CPU_DETECT
422 mbd->rtcd = xd->rtcd;
423 #endif
424 mb->gf_active_ptr = x->gf_active_ptr;
426 mb->vector_range = 32;
428 vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
429 mbr_ei[i].totalrate = 0;
431 mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
433 mbd->mode_info_context = cm->mi + x->e_mbd.mode_info_stride * (i + 1);
434 mbd->mode_info_stride = cm->mode_info_stride;
436 mbd->frame_type = cm->frame_type;
438 mbd->frames_since_golden = cm->frames_since_golden;
439 mbd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
441 mb->src = * cpi->Source;
442 mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
443 mbd->dst = cm->yv12_fb[cm->new_fb_idx];
445 mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
446 mb->src.u_buffer += 8 * x->src.uv_stride * (i + 1);
447 mb->src.v_buffer += 8 * x->src.uv_stride * (i + 1);
449 vp8_build_block_offsets(mb);
451 vp8_setup_block_dptrs(mbd);
453 vp8_setup_block_ptrs(mb);
455 mbd->left_context = &cm->left_context;
456 mb->mvc = cm->fc.mvc;
458 setup_mbby_copy(&mbr_ei[i].mb, x);
463 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
465 const VP8_COMMON * cm = &cpi->common;
467 cpi->b_multi_threaded = 0;
468 cpi->encoding_thread_count = 0;
470 if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
472 int ithread;
473 int th_count = cpi->oxcf.multi_threaded - 1;
475 /* don't allocate more threads than cores available */
476 if (cpi->oxcf.multi_threaded > cm->processor_core_count)
477 th_count = cm->processor_core_count - 1;
479 /* we have th_count + 1 (main) threads processing one row each */
480 /* no point to have more threads than the sync range allows */
481 if(th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1))
483 th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
486 if(th_count == 0)
487 return;
489 CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * th_count));
490 CHECK_MEM_ERROR(cpi->h_event_start_encoding, vpx_malloc(sizeof(sem_t) * th_count));
491 CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
492 vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
493 CHECK_MEM_ERROR(cpi->en_thread_data,
494 vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
495 CHECK_MEM_ERROR(cpi->mt_current_mb_col,
496 vpx_malloc(sizeof(*cpi->mt_current_mb_col) * cm->mb_rows));
498 sem_init(&cpi->h_event_end_encoding, 0, 0);
500 cpi->b_multi_threaded = 1;
501 cpi->encoding_thread_count = th_count;
504 printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
505 (cpi->encoding_thread_count +1));
508 for (ithread = 0; ithread < th_count; ithread++)
510 ENCODETHREAD_DATA * ethd = &cpi->en_thread_data[ithread];
512 sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
513 ethd->ithread = ithread;
514 ethd->ptr1 = (void *)cpi;
515 ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
517 pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, ethd);
521 LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
523 sem_init(&cpi->h_event_start_lpf, 0, 0);
524 sem_init(&cpi->h_event_end_lpf, 0, 0);
526 lpfthd->ptr1 = (void *)cpi;
527 pthread_create(&cpi->h_filter_thread, 0, loopfilter_thread, lpfthd);
533 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
535 if (cpi->b_multi_threaded)
537 //shutdown other threads
538 cpi->b_multi_threaded = 0;
540 int i;
542 for (i = 0; i < cpi->encoding_thread_count; i++)
544 //SetEvent(cpi->h_event_mbrencoding[i]);
545 sem_post(&cpi->h_event_start_encoding[i]);
546 pthread_join(cpi->h_encoding_thread[i], 0);
548 sem_destroy(&cpi->h_event_start_encoding[i]);
551 sem_post(&cpi->h_event_start_lpf);
552 pthread_join(cpi->h_filter_thread, 0);
555 sem_destroy(&cpi->h_event_end_encoding);
556 sem_destroy(&cpi->h_event_end_lpf);
557 sem_destroy(&cpi->h_event_start_lpf);
559 //free thread related resources
560 vpx_free(cpi->h_event_start_encoding);
561 vpx_free(cpi->h_encoding_thread);
562 vpx_free(cpi->mb_row_ei);
563 vpx_free(cpi->en_thread_data);
564 vpx_free(cpi->mt_current_mb_col);
567 #endif