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.
13 #include "threading.h"
18 extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t, int recon_yoffset, int recon_uvoffset);
19 extern int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t);
20 extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x);
21 extern void vp8_build_block_offsets(MACROBLOCK *x);
22 extern void vp8_setup_block_ptrs(MACROBLOCK *x);
25 THREAD_FUNCTION thread_encoding_proc(void *p_data)
27 #if CONFIG_MULTITHREAD
28 int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
29 VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
30 MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
31 ENTROPY_CONTEXT_PLANES mb_row_left_context;
33 //printf("Started thread %d\n", ithread);
37 if (cpi->b_multi_threaded == 0)
40 //if(WaitForSingleObject(cpi->h_event_mbrencoding[ithread], INFINITE) == WAIT_OBJECT_0)
41 if (sem_wait(&cpi->h_event_mbrencoding[ithread]) == 0)
43 if (cpi->b_multi_threaded == FALSE) // we're shutting down
47 VP8_COMMON *cm = &cpi->common;
48 int mb_row = mbri->mb_row;
49 MACROBLOCK *x = &mbri->mb;
50 MACROBLOCKD *xd = &x->e_mbd;
51 TOKENEXTRA **tp = &mbri->tp;
52 int *segment_counts = mbri->segment_counts;
53 int *totalrate = &mbri->totalrate;
57 int recon_yoffset, recon_uvoffset;
59 int ref_fb_idx = cm->lst_fb_idx;
60 int dst_fb_idx = cm->new_fb_idx;
61 int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
62 int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
63 volatile int *last_row_current_mb_col;
66 last_row_current_mb_col = &cpi->mb_row_ei[ithread-1].current_mb_col;
68 last_row_current_mb_col = &cpi->current_mb_col_main;
70 // reset above block coeffs
71 xd->above_context = cm->above_context;
72 xd->left_context = &mb_row_left_context;
74 vp8_zero(mb_row_left_context);
76 xd->up_available = (mb_row != 0);
77 recon_yoffset = (mb_row * recon_y_stride * 16);
78 recon_uvoffset = (mb_row * recon_uv_stride * 8);
81 cpi->tplist[mb_row].start = *tp;
83 //printf("Thread mb_row = %d\n", mb_row);
85 // for each macroblock col in image
86 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
88 int seg_map_index = (mb_row * cm->mb_cols);
90 while (mb_col > (*last_row_current_mb_col - 1) && *last_row_current_mb_col != cm->mb_cols - 1)
96 // Distance of Mb to the various image edges.
97 // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
98 xd->mb_to_left_edge = -((mb_col * 16) << 3);
99 xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
100 xd->mb_to_top_edge = -((mb_row * 16) << 3);
101 xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
103 // Set up limit values for motion vectors used to prevent them extending outside the UMV borders
104 x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
105 x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
106 x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
107 x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
109 xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
110 xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
111 xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
112 xd->left_available = (mb_col != 0);
114 // Is segmentation enabled
115 // MB level adjutment to quantizer
116 if (xd->segmentation_enabled)
118 // Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
119 if (cpi->segmentation_map[seg_map_index+mb_col] <= 3)
120 xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[seg_map_index+mb_col];
122 xd->mode_info_context->mbmi.segment_id = 0;
124 vp8cx_mb_init_quantizer(cpi, x);
127 xd->mode_info_context->mbmi.segment_id = 0; // Set to Segment 0 by default
130 if (cm->frame_type == KEY_FRAME)
132 *totalrate += vp8cx_encode_intra_macro_block(cpi, x, tp);
134 y_modes[xd->mbmi.mode] ++;
139 *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset);
142 inter_y_modes[xd->mbmi.mode] ++;
144 if (xd->mbmi.mode == SPLITMV)
148 for (b = 0; b < xd->mbmi.partition_count; b++)
150 inter_b_modes[xd->mbmi.partition_bmi[b].mode] ++;
156 // Count of last ref frame 0,0 useage
157 if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
158 cpi->inter_zz_count ++;
162 cpi->tplist[mb_row].stop = *tp;
164 x->gf_active_ptr++; // Increment pointer into gf useage flags structure for next mb
166 for (i = 0; i < 16; i++)
167 vpx_memcpy(&xd->mode_info_context->bmi[i], &xd->block[i].bmi, sizeof(xd->block[i].bmi));
169 // adjust to the next column of macroblocks
170 x->src.y_buffer += 16;
171 x->src.u_buffer += 8;
172 x->src.v_buffer += 8;
177 // Keep track of segment useage
178 segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
181 xd->mode_info_context++;
185 cpi->mb_row_ei[ithread].current_mb_col = mb_col;
189 //extend the recon for intra prediction
191 &cm->yv12_fb[dst_fb_idx],
192 xd->dst.y_buffer + 16,
193 xd->dst.u_buffer + 8,
194 xd->dst.v_buffer + 8);
196 // this is to account for the border
197 xd->mode_info_context++;
199 x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
200 x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
201 x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
203 xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
205 if (ithread == (cpi->encoding_thread_count - 1) || mb_row == cm->mb_rows - 1)
207 //SetEvent(cpi->h_event_main);
208 sem_post(&cpi->h_event_main);
221 //printf("exit thread %d\n", ithread);
225 static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
228 MACROBLOCK *x = mbsrc;
229 MACROBLOCK *z = mbdst;
233 z->ss_count = x->ss_count;
234 z->searches_per_step = x->searches_per_step;
235 z->errorperbit = x->errorperbit;
237 z->sadperbit16 = x->sadperbit16;
238 z->sadperbit4 = x->sadperbit4;
239 z->errthresh = x->errthresh;
241 z->rdmult = x->rdmult;
244 z->mv_col_min = x->mv_col_min;
245 z->mv_col_max = x->mv_col_max;
246 z->mv_row_min = x->mv_row_min;
247 z->mv_row_max = x->mv_row_max;
248 z->vector_range = x->vector_range ;
251 z->vp8_short_fdct4x4 = x->vp8_short_fdct4x4;
252 z->vp8_short_fdct8x4 = x->vp8_short_fdct8x4;
253 z->short_walsh4x4 = x->short_walsh4x4;
254 z->quantize_b = x->quantize_b;
258 z->src.y_buffer = x->src.y_buffer;
259 z->src.u_buffer = x->src.u_buffer;
260 z->src.v_buffer = x->src.v_buffer;
264 vpx_memcpy(z->mvcosts, x->mvcosts, sizeof(x->mvcosts));
265 z->mvcost[0] = &z->mvcosts[0][mv_max+1];
266 z->mvcost[1] = &z->mvcosts[1][mv_max+1];
267 z->mvsadcost[0] = &z->mvsadcosts[0][mv_max+1];
268 z->mvsadcost[1] = &z->mvsadcosts[1][mv_max+1];
271 vpx_memcpy(z->token_costs, x->token_costs, sizeof(x->token_costs));
272 vpx_memcpy(z->inter_bmode_costs, x->inter_bmode_costs, sizeof(x->inter_bmode_costs));
273 //memcpy(z->mvcosts, x->mvcosts, sizeof(x->mvcosts));
274 //memcpy(z->mvcost, x->mvcost, sizeof(x->mvcost));
275 vpx_memcpy(z->mbmode_cost, x->mbmode_cost, sizeof(x->mbmode_cost));
276 vpx_memcpy(z->intra_uv_mode_cost, x->intra_uv_mode_cost, sizeof(x->intra_uv_mode_cost));
277 vpx_memcpy(z->bmode_costs, x->bmode_costs, sizeof(x->bmode_costs));
279 for (i = 0; i < 25; i++)
281 z->block[i].quant = x->block[i].quant;
282 z->block[i].quant_shift = x->block[i].quant_shift;
283 z->block[i].zbin = x->block[i].zbin;
284 z->block[i].zrun_zbin_boost = x->block[i].zrun_zbin_boost;
285 z->block[i].round = x->block[i].round;
287 z->block[i].src = x->block[i].src;
289 z->block[i].src_stride = x->block[i].src_stride;
290 z->block[i].force_empty = x->block[i].force_empty;
295 MACROBLOCKD *xd = &x->e_mbd;
296 MACROBLOCKD *zd = &z->e_mbd;
299 zd->mode_info_context = xd->mode_info_context;
300 zd->mode_info = xd->mode_info;
302 zd->mode_info_stride = xd->mode_info_stride;
303 zd->frame_type = xd->frame_type;
304 zd->up_available = xd->up_available ;
305 zd->left_available = xd->left_available;
306 zd->left_context = xd->left_context;
307 zd->last_frame_dc = xd->last_frame_dc;
308 zd->last_frame_dccons = xd->last_frame_dccons;
309 zd->gold_frame_dc = xd->gold_frame_dc;
310 zd->gold_frame_dccons = xd->gold_frame_dccons;
311 zd->mb_to_left_edge = xd->mb_to_left_edge;
312 zd->mb_to_right_edge = xd->mb_to_right_edge;
313 zd->mb_to_top_edge = xd->mb_to_top_edge ;
314 zd->mb_to_bottom_edge = xd->mb_to_bottom_edge;
315 zd->gf_active_ptr = xd->gf_active_ptr;
316 zd->frames_since_golden = xd->frames_since_golden;
317 zd->frames_till_alt_ref_frame = xd->frames_till_alt_ref_frame;
319 zd->subpixel_predict = xd->subpixel_predict;
320 zd->subpixel_predict8x4 = xd->subpixel_predict8x4;
321 zd->subpixel_predict8x8 = xd->subpixel_predict8x8;
322 zd->subpixel_predict16x16 = xd->subpixel_predict16x16;
323 zd->segmentation_enabled = xd->segmentation_enabled;
324 zd->mb_segement_abs_delta = xd->mb_segement_abs_delta;
325 vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
327 for (i = 0; i < 25; i++)
329 zd->block[i].dequant = xd->block[i].dequant;
335 void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
343 VP8_COMMON *const cm = & cpi->common;
344 MACROBLOCKD *const xd = & x->e_mbd;
348 for (i = 0; i < count; i++)
350 MACROBLOCK *mb = & mbr_ei[i].mb;
351 MACROBLOCKD *mbd = &mb->e_mbd;
353 mbd->subpixel_predict = xd->subpixel_predict;
354 mbd->subpixel_predict8x4 = xd->subpixel_predict8x4;
355 mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
356 mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
357 #if CONFIG_RUNTIME_CPU_DETECT
358 mbd->rtcd = xd->rtcd;
360 mb->gf_active_ptr = x->gf_active_ptr;
362 mb->vector_range = 32;
364 vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
365 mbr_ei[i].totalrate = 0;
367 mbd->mode_info = cm->mi - 1;
368 mbd->mode_info_context = cm->mi + x->e_mbd.mode_info_stride * (i + 1);
369 mbd->mode_info_stride = cm->mode_info_stride;
371 mbd->frame_type = cm->frame_type;
373 mbd->frames_since_golden = cm->frames_since_golden;
374 mbd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
376 mb->src = * cpi->Source;
377 mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
378 mbd->dst = cm->yv12_fb[cm->new_fb_idx];
380 mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
381 mb->src.u_buffer += 8 * x->src.uv_stride * (i + 1);
382 mb->src.v_buffer += 8 * x->src.uv_stride * (i + 1);
385 vp8_build_block_offsets(mb);
387 vp8_setup_block_dptrs(mbd);
389 vp8_setup_block_ptrs(mb);
391 mb->rddiv = cpi->RDDIV;
392 mb->rdmult = cpi->RDMULT;
394 mbd->left_context = &cm->left_context;
395 mb->mvc = cm->fc.mvc;
397 setup_mbby_copy(&mbr_ei[i].mb, x);
403 void vp8cx_create_encoder_threads(VP8_COMP *cpi)
405 cpi->b_multi_threaded = 0;
407 cpi->processor_core_count = 32; //vp8_get_proc_core_count();
409 CHECK_MEM_ERROR(cpi->tplist, vpx_malloc(sizeof(TOKENLIST) * cpi->common.mb_rows));
411 #if CONFIG_MULTITHREAD
413 if (cpi->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
417 if (cpi->oxcf.multi_threaded > cpi->processor_core_count)
418 cpi->encoding_thread_count = cpi->processor_core_count - 1;
420 cpi->encoding_thread_count = cpi->oxcf.multi_threaded - 1;
423 CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * cpi->encoding_thread_count));
424 CHECK_MEM_ERROR(cpi->h_event_mbrencoding, vpx_malloc(sizeof(sem_t) * cpi->encoding_thread_count));
425 CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count));
426 vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count);
427 CHECK_MEM_ERROR(cpi->en_thread_data, vpx_malloc(sizeof(ENCODETHREAD_DATA) * cpi->encoding_thread_count));
428 //cpi->h_event_main = CreateEvent(NULL, FALSE, FALSE, NULL);
429 sem_init(&cpi->h_event_main, 0, 0);
431 cpi->b_multi_threaded = 1;
433 //printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n", (cpi->encoding_thread_count +1));
435 for (ithread = 0; ithread < cpi->encoding_thread_count; ithread++)
437 //cpi->h_event_mbrencoding[ithread] = CreateEvent(NULL, FALSE, FALSE, NULL);
438 sem_init(&cpi->h_event_mbrencoding[ithread], 0, 0);
439 cpi->en_thread_data[ithread].ithread = ithread;
440 cpi->en_thread_data[ithread].ptr1 = (void *)cpi;
441 cpi->en_thread_data[ithread].ptr2 = (void *)&cpi->mb_row_ei[ithread];
443 //printf(" call begin thread %d \n", ithread);
445 //cpi->h_encoding_thread[ithread] = (HANDLE)_beginthreadex(
448 // thread_encoding_proc,
449 // (&cpi->en_thread_data[ithread]), // Thread data
453 pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, (&cpi->en_thread_data[ithread]));
462 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
464 #if CONFIG_MULTITHREAD
466 if (cpi->b_multi_threaded)
468 //shutdown other threads
469 cpi->b_multi_threaded = 0;
473 for (i = 0; i < cpi->encoding_thread_count; i++)
475 //SetEvent(cpi->h_event_mbrencoding[i]);
476 sem_post(&cpi->h_event_mbrencoding[i]);
477 pthread_join(cpi->h_encoding_thread[i], 0);
480 for (i = 0; i < cpi->encoding_thread_count; i++)
481 sem_destroy(&cpi->h_event_mbrencoding[i]);
483 //free thread related resources
484 vpx_free(cpi->h_event_mbrencoding);
485 vpx_free(cpi->h_encoding_thread);
486 vpx_free(cpi->mb_row_ei);
487 vpx_free(cpi->en_thread_data);
491 vpx_free(cpi->tplist);