2 * Copyright (c) 2010 The VP8 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.
15 #include "onyxd_int.h"
16 #include "vpx_mem/vpx_mem.h"
17 #include "threading.h"
19 #include "loopfilter.h"
21 #include "vpx_ports/vpx_timer.h"
23 extern void vp8_decode_mb_row(VP8D_COMP *pbi,
28 extern void vp8_build_uvmvs(MACROBLOCKD *x, int fullpixel);
29 extern void vp8_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd);
31 void vp8_setup_decoding_thread_data(VP8D_COMP *pbi, MACROBLOCKD *xd, MB_ROW_DEC *mbrd, int count)
36 #if CONFIG_MULTITHREAD
37 VP8_COMMON *const pc = & pbi->common;
40 for (i = 0; i < count; i++)
42 MACROBLOCKD *mbd = &mbrd[i].mbd;
43 #if CONFIG_RUNTIME_CPU_DETECT
48 mbd->subpixel_predict = xd->subpixel_predict;
49 mbd->subpixel_predict8x4 = xd->subpixel_predict8x4;
50 mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
51 mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
52 mbd->gf_active_ptr = xd->gf_active_ptr;
54 mbd->mode_info = pc->mi - 1;
55 mbd->mode_info_context = pc->mi + pc->mode_info_stride * (i + 1);
56 mbd->mode_info_stride = pc->mode_info_stride;
58 mbd->frame_type = pc->frame_type;
59 mbd->frames_since_golden = pc->frames_since_golden;
60 mbd->frames_till_alt_ref_frame = pc->frames_till_alt_ref_frame;
62 mbd->pre = pc->yv12_fb[pc->lst_fb_idx];
63 mbd->dst = pc->yv12_fb[pc->new_fb_idx];
65 vp8_setup_block_dptrs(mbd);
66 vp8_build_block_doffsets(mbd);
67 mbd->segmentation_enabled = xd->segmentation_enabled;
68 mbd->mb_segement_abs_delta = xd->mb_segement_abs_delta;
69 vpx_memcpy(mbd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
71 mbd->mbmi.mode = DC_PRED;
72 mbd->mbmi.uv_mode = DC_PRED;
74 mbd->current_bc = &pbi->bc2;
76 for (j = 0; j < 25; j++)
78 mbd->block[j].dequant = xd->block[j].dequant;
91 THREAD_FUNCTION vp8_thread_decoding_proc(void *p_data)
93 #if CONFIG_MULTITHREAD
94 int ithread = ((DECODETHREAD_DATA *)p_data)->ithread;
95 VP8D_COMP *pbi = (VP8D_COMP *)(((DECODETHREAD_DATA *)p_data)->ptr1);
96 MB_ROW_DEC *mbrd = (MB_ROW_DEC *)(((DECODETHREAD_DATA *)p_data)->ptr2);
97 ENTROPY_CONTEXT mb_row_left_context[4][4];
101 if (pbi->b_multithreaded_rd == 0)
104 //if(WaitForSingleObject(pbi->h_event_mbrdecoding[ithread], INFINITE) == WAIT_OBJECT_0)
105 if (sem_wait(&pbi->h_event_mbrdecoding[ithread]) == 0)
107 if (pbi->b_multithreaded_rd == 0)
111 VP8_COMMON *pc = &pbi->common;
112 int mb_row = mbrd->mb_row;
113 MACROBLOCKD *xd = &mbrd->mbd;
115 //printf("ithread:%d mb_row %d\n", ithread, mb_row);
117 int recon_yoffset, recon_uvoffset;
119 int ref_fb_idx = pc->lst_fb_idx;
120 int dst_fb_idx = pc->new_fb_idx;
121 int recon_y_stride = pc->yv12_fb[ref_fb_idx].y_stride;
122 int recon_uv_stride = pc->yv12_fb[ref_fb_idx].uv_stride;
124 volatile int *last_row_current_mb_col;
127 last_row_current_mb_col = &pbi->mb_row_di[ithread-1].current_mb_col;
129 last_row_current_mb_col = &pbi->current_mb_col_main;
131 recon_yoffset = mb_row * recon_y_stride * 16;
132 recon_uvoffset = mb_row * recon_uv_stride * 8;
133 // reset above block coeffs
135 xd->above_context[Y1CONTEXT] = pc->above_context[Y1CONTEXT];
136 xd->above_context[UCONTEXT ] = pc->above_context[UCONTEXT];
137 xd->above_context[VCONTEXT ] = pc->above_context[VCONTEXT];
138 xd->above_context[Y2CONTEXT] = pc->above_context[Y2CONTEXT];
139 xd->left_context = mb_row_left_context;
140 vpx_memset(mb_row_left_context, 0, sizeof(mb_row_left_context));
141 xd->up_available = (mb_row != 0);
143 xd->mb_to_top_edge = -((mb_row * 16)) << 3;
144 xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
146 for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
149 while (mb_col > (*last_row_current_mb_col - 1) && *last_row_current_mb_col != pc->mb_cols - 1)
155 // Take a copy of the mode and Mv information for this macroblock into the xd->mbmi
156 // the partition_bmi array is unused in the decoder, so don't copy it.
157 vpx_memcpy(&xd->mbmi, &xd->mode_info_context->mbmi,
158 sizeof(MB_MODE_INFO) - sizeof(xd->mbmi.partition_bmi));
160 if (xd->mbmi.mode == SPLITMV || xd->mbmi.mode == B_PRED)
162 for (i = 0; i < 16; i++)
164 BLOCKD *d = &xd->block[i];
165 vpx_memcpy(&d->bmi, &xd->mode_info_context->bmi[i], sizeof(B_MODE_INFO));
169 // Distance of Mb to the various image edges.
170 // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
171 xd->mb_to_left_edge = -((mb_col * 16) << 3);
172 xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
174 xd->dst.y_buffer = pc->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
175 xd->dst.u_buffer = pc->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
176 xd->dst.v_buffer = pc->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
178 xd->left_available = (mb_col != 0);
180 // Select the appropriate reference frame for this MB
181 if (xd->mbmi.ref_frame == LAST_FRAME)
182 ref_fb_idx = pc->lst_fb_idx;
183 else if (xd->mbmi.ref_frame == GOLDEN_FRAME)
184 ref_fb_idx = pc->gld_fb_idx;
186 ref_fb_idx = pc->alt_fb_idx;
188 xd->pre.y_buffer = pc->yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
189 xd->pre.u_buffer = pc->yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
190 xd->pre.v_buffer = pc->yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
192 vp8_build_uvmvs(xd, pc->full_pixel);
194 vp8_decode_macroblock(pbi, xd);
200 ++xd->mode_info_context; /* next mb */
202 xd->gf_active_ptr++; // GF useage flag for next MB
204 xd->above_context[Y1CONTEXT] += 4;
205 xd->above_context[UCONTEXT ] += 2;
206 xd->above_context[VCONTEXT ] += 2;
207 xd->above_context[Y2CONTEXT] ++;
208 pbi->mb_row_di[ithread].current_mb_col = mb_col;
212 // adjust to the next row of mbs
214 &pc->yv12_fb[dst_fb_idx],
215 xd->dst.y_buffer + 16, xd->dst.u_buffer + 8, xd->dst.v_buffer + 8
218 ++xd->mode_info_context; /* skip prediction column */
220 // since we have multithread
221 xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
223 //memcpy(&pbi->lpfmb, &pbi->mb, sizeof(pbi->mb));
224 if ((mb_row & 1) == 1)
226 pbi->last_mb_row_decoded = mb_row;
227 //printf("S%d", pbi->last_mb_row_decoded);
230 if (ithread == (pbi->decoding_thread_count - 1) || mb_row == pc->mb_rows - 1)
232 //SetEvent(pbi->h_event_main);
233 sem_post(&pbi->h_event_main);
247 THREAD_FUNCTION vp8_thread_loop_filter(void *p_data)
249 #if CONFIG_MULTITHREAD
250 VP8D_COMP *pbi = (VP8D_COMP *)p_data;
254 if (pbi->b_multithreaded_lf == 0)
257 //printf("before waiting for start_lpf\n");
259 //if(WaitForSingleObject(pbi->h_event_start_lpf, INFINITE) == WAIT_OBJECT_0)
260 if (sem_wait(&pbi->h_event_start_lpf) == 0)
262 if (pbi->b_multithreaded_lf == 0) // we're shutting down
267 VP8_COMMON *cm = &pbi->common;
268 MACROBLOCKD *mbd = &pbi->lpfmb;
269 int default_filt_lvl = pbi->common.filter_level;
271 YV12_BUFFER_CONFIG *post = &cm->yv12_fb[cm->new_fb_idx];
272 loop_filter_info *lfi = cm->lf_info;
273 int frame_type = cm->frame_type;
278 int baseline_filter_level[MAX_MB_SEGMENTS];
280 int alt_flt_enabled = mbd->segmentation_enabled;
283 unsigned char *y_ptr, *u_ptr, *v_ptr;
285 volatile int *last_mb_row_decoded = &pbi->last_mb_row_decoded;
287 //MODE_INFO * this_mb_mode_info = cm->mi;
288 mbd->mode_info_context = cm->mi; // Point at base of Mb MODE_INFO list
290 // Note the baseline filter values for each segment
293 for (i = 0; i < MAX_MB_SEGMENTS; i++)
295 if (mbd->mb_segement_abs_delta == SEGMENT_ABSDATA)
296 baseline_filter_level[i] = mbd->segment_feature_data[MB_LVL_ALT_LF][i];
299 baseline_filter_level[i] = default_filt_lvl + mbd->segment_feature_data[MB_LVL_ALT_LF][i];
300 baseline_filter_level[i] = (baseline_filter_level[i] >= 0) ? ((baseline_filter_level[i] <= MAX_LOOP_FILTER) ? baseline_filter_level[i] : MAX_LOOP_FILTER) : 0; // Clamp to valid range
306 for (i = 0; i < MAX_MB_SEGMENTS; i++)
307 baseline_filter_level[i] = default_filt_lvl;
310 // Initialize the loop filter for this frame.
311 if ((cm->last_filter_type != cm->filter_type) || (cm->last_sharpness_level != cm->sharpness_level))
312 vp8_init_loop_filter(cm);
313 else if (frame_type != cm->last_frame_type)
314 vp8_frame_init_loop_filter(lfi, frame_type);
316 // Set up the buffer pointers
317 y_ptr = post->y_buffer;
318 u_ptr = post->u_buffer;
319 v_ptr = post->v_buffer;
321 // vp8_filter each macro block
322 for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
325 while (mb_row >= *last_mb_row_decoded)
331 //printf("R%d", mb_row);
332 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
334 int Segment = (alt_flt_enabled) ? mbd->mode_info_context->mbmi.segment_id : 0;
336 filter_level = baseline_filter_level[Segment];
338 // Apply any context driven MB level adjustment
339 vp8_adjust_mb_lf_value(mbd, &filter_level);
344 cm->lf_mbv(y_ptr, u_ptr, v_ptr, post->y_stride, post->uv_stride, &lfi[filter_level], cm->simpler_lpf);
346 if (mbd->mode_info_context->mbmi.dc_diff > 0)
347 cm->lf_bv(y_ptr, u_ptr, v_ptr, post->y_stride, post->uv_stride, &lfi[filter_level], cm->simpler_lpf);
349 // don't apply across umv border
351 cm->lf_mbh(y_ptr, u_ptr, v_ptr, post->y_stride, post->uv_stride, &lfi[filter_level], cm->simpler_lpf);
353 if (mbd->mode_info_context->mbmi.dc_diff > 0)
354 cm->lf_bh(y_ptr, u_ptr, v_ptr, post->y_stride, post->uv_stride, &lfi[filter_level], cm->simpler_lpf);
361 mbd->mode_info_context++; // step to next MB
365 y_ptr += post->y_stride * 16 - post->y_width;
366 u_ptr += post->uv_stride * 8 - post->uv_width;
367 v_ptr += post->uv_stride * 8 - post->uv_width;
369 mbd->mode_info_context++; // Skip border mb
372 //printf("R%d\n", mb_row);
373 // When done, signal main thread that ME is finished
374 //SetEvent(pbi->h_event_lpf);
375 sem_post(&pbi->h_event_lpf);
387 void vp8_decoder_create_threads(VP8D_COMP *pbi)
389 #if CONFIG_MULTITHREAD
393 pbi->b_multithreaded_rd = 0;
394 pbi->b_multithreaded_lf = 0;
395 pbi->allocated_decoding_thread_count = 0;
396 core_count = (pbi->max_threads > 16) ? 16 : pbi->max_threads; //vp8_get_proc_core_count();
399 sem_init(&pbi->h_event_lpf, 0, 0);
400 sem_init(&pbi->h_event_start_lpf, 0, 0);
401 pbi->b_multithreaded_lf = 1;
402 pthread_create(&pbi->h_thread_lpf, 0, vp8_thread_loop_filter, (pbi));
407 pbi->b_multithreaded_rd = 1;
408 pbi->decoding_thread_count = core_count - 1;
410 CHECK_MEM_ERROR(pbi->h_decoding_thread, vpx_malloc(sizeof(pthread_t) * pbi->decoding_thread_count));
411 CHECK_MEM_ERROR(pbi->h_event_mbrdecoding, vpx_malloc(sizeof(sem_t) * pbi->decoding_thread_count));
412 CHECK_MEM_ERROR(pbi->mb_row_di, vpx_memalign(32, sizeof(MB_ROW_DEC) * pbi->decoding_thread_count));
413 vpx_memset(pbi->mb_row_di, 0, sizeof(MB_ROW_DEC) * pbi->decoding_thread_count);
414 CHECK_MEM_ERROR(pbi->de_thread_data, vpx_malloc(sizeof(DECODETHREAD_DATA) * pbi->decoding_thread_count));
416 for (ithread = 0; ithread < pbi->decoding_thread_count; ithread++)
418 sem_init(&pbi->h_event_mbrdecoding[ithread], 0, 0);
420 pbi->de_thread_data[ithread].ithread = ithread;
421 pbi->de_thread_data[ithread].ptr1 = (void *)pbi;
422 pbi->de_thread_data[ithread].ptr2 = (void *) &pbi->mb_row_di[ithread];
424 pthread_create(&pbi->h_decoding_thread[ithread], 0, vp8_thread_decoding_proc, (&pbi->de_thread_data[ithread]));
428 sem_init(&pbi->h_event_main, 0, 0);
429 pbi->allocated_decoding_thread_count = pbi->decoding_thread_count;
437 void vp8_decoder_remove_threads(VP8D_COMP *pbi)
439 #if CONFIG_MULTITHREAD
441 if (pbi->b_multithreaded_lf)
443 pbi->b_multithreaded_lf = 0;
444 sem_post(&pbi->h_event_start_lpf);
445 pthread_join(pbi->h_thread_lpf, 0);
446 sem_destroy(&pbi->h_event_start_lpf);
449 //shutdown MB Decoding thread;
450 if (pbi->b_multithreaded_rd)
452 pbi->b_multithreaded_rd = 0;
453 // allow all threads to exit
457 for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
460 sem_post(&pbi->h_event_mbrdecoding[i]);
461 pthread_join(pbi->h_decoding_thread[i], NULL);
467 for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
469 sem_destroy(&pbi->h_event_mbrdecoding[i]);
475 sem_destroy(&pbi->h_event_main);
477 if (pbi->h_decoding_thread)
479 vpx_free(pbi->h_decoding_thread);
480 pbi->h_decoding_thread = NULL;
483 if (pbi->h_event_mbrdecoding)
485 vpx_free(pbi->h_event_mbrdecoding);
486 pbi->h_event_mbrdecoding = NULL;
491 vpx_free(pbi->mb_row_di);
492 pbi->mb_row_di = NULL ;
495 if (pbi->de_thread_data)
497 vpx_free(pbi->de_thread_data);
498 pbi->de_thread_data = NULL;
508 void vp8_start_lfthread(VP8D_COMP *pbi)
510 #if CONFIG_MULTITHREAD
511 memcpy(&pbi->lpfmb, &pbi->mb, sizeof(pbi->mb));
512 pbi->last_mb_row_decoded = 0;
513 sem_post(&pbi->h_event_start_lpf);
519 void vp8_stop_lfthread(VP8D_COMP *pbi)
521 #if CONFIG_MULTITHREAD
522 struct vpx_usec_timer timer;
524 vpx_usec_timer_start(&timer);
526 sem_wait(&pbi->h_event_lpf);
528 vpx_usec_timer_mark(&timer);
529 pbi->time_loop_filtering += vpx_usec_timer_elapsed(&timer);
536 void vp8_mtdecode_mb_rows(VP8D_COMP *pbi,
539 #if CONFIG_MULTITHREAD
541 VP8_COMMON *pc = &pbi->common;
544 int num_part = 1 << pbi->common.multi_token_partition;
546 vp8_setup_decoding_thread_data(pbi, xd, pbi->mb_row_di, pbi->decoding_thread_count);
548 for (mb_row = 0; mb_row < pc->mb_rows; mb_row += (pbi->decoding_thread_count + 1))
551 pbi->current_mb_col_main = -1;
553 xd->current_bc = &pbi->mbc[ibc];
559 for (i = 0; i < pbi->decoding_thread_count; i++)
561 if ((mb_row + i + 1) >= pc->mb_rows)
564 pbi->mb_row_di[i].mb_row = mb_row + i + 1;
565 pbi->mb_row_di[i].mbd.current_bc = &pbi->mbc[ibc];
571 pbi->mb_row_di[i].current_mb_col = -1;
572 sem_post(&pbi->h_event_mbrdecoding[i]);
575 vp8_decode_mb_row(pbi, pc, mb_row, xd);
577 xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
579 if (mb_row < pc->mb_rows - 1)
581 sem_wait(&pbi->h_event_main);
585 pbi->last_mb_row_decoded = mb_row;