Line data Source code
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 :
18 : #if CONFIG_MULTITHREAD
19 :
20 : extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x,
21 : int ok_to_skip);
22 :
23 0 : static THREAD_FUNCTION thread_loopfilter(void *p_data) {
24 0 : VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
25 0 : VP8_COMMON *cm = &cpi->common;
26 :
27 : while (1) {
28 0 : if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0) break;
29 :
30 0 : if (sem_wait(&cpi->h_event_start_lpf) == 0) {
31 : /* we're shutting down */
32 0 : if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0) break;
33 :
34 0 : vp8_loopfilter_frame(cpi, cm);
35 :
36 0 : sem_post(&cpi->h_event_end_lpf);
37 : }
38 : }
39 :
40 0 : return 0;
41 : }
42 :
43 0 : static THREAD_FUNCTION thread_encoding_proc(void *p_data) {
44 0 : int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
45 0 : VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
46 0 : MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
47 : ENTROPY_CONTEXT_PLANES mb_row_left_context;
48 :
49 : while (1) {
50 0 : if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0) break;
51 :
52 0 : if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0) {
53 0 : const int nsync = cpi->mt_sync_range;
54 0 : VP8_COMMON *cm = &cpi->common;
55 : int mb_row;
56 0 : MACROBLOCK *x = &mbri->mb;
57 0 : MACROBLOCKD *xd = &x->e_mbd;
58 : TOKENEXTRA *tp;
59 : #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
60 : TOKENEXTRA *tp_start = cpi->tok + (1 + ithread) * (16 * 24);
61 : const int num_part = (1 << cm->multi_token_partition);
62 : #endif
63 :
64 0 : int *segment_counts = mbri->segment_counts;
65 0 : int *totalrate = &mbri->totalrate;
66 :
67 : /* we're shutting down */
68 0 : if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0) break;
69 :
70 0 : xd->mode_info_context = cm->mi + cm->mode_info_stride * (ithread + 1);
71 0 : xd->mode_info_stride = cm->mode_info_stride;
72 :
73 0 : for (mb_row = ithread + 1; mb_row < cm->mb_rows;
74 0 : mb_row += (cpi->encoding_thread_count + 1)) {
75 : int recon_yoffset, recon_uvoffset;
76 : int mb_col;
77 0 : int ref_fb_idx = cm->lst_fb_idx;
78 0 : int dst_fb_idx = cm->new_fb_idx;
79 0 : int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
80 0 : int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
81 0 : int map_index = (mb_row * cm->mb_cols);
82 : const int *last_row_current_mb_col;
83 0 : int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
84 :
85 : #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
86 : vp8_writer *w = &cpi->bc[1 + (mb_row % num_part)];
87 : #else
88 0 : tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
89 0 : cpi->tplist[mb_row].start = tp;
90 : #endif
91 :
92 0 : last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
93 :
94 : /* reset above block coeffs */
95 0 : xd->above_context = cm->above_context;
96 0 : xd->left_context = &mb_row_left_context;
97 :
98 0 : vp8_zero(mb_row_left_context);
99 :
100 0 : xd->up_available = (mb_row != 0);
101 0 : recon_yoffset = (mb_row * recon_y_stride * 16);
102 0 : recon_uvoffset = (mb_row * recon_uv_stride * 8);
103 :
104 : /* Set the mb activity pointer to the start of the row. */
105 0 : x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
106 :
107 : /* for each macroblock col in image */
108 0 : for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
109 0 : if (((mb_col - 1) % nsync) == 0) {
110 0 : pthread_mutex_t *mutex = &cpi->pmutex[mb_row];
111 0 : protected_write(mutex, current_mb_col, mb_col - 1);
112 : }
113 :
114 0 : if (mb_row && !(mb_col & (nsync - 1))) {
115 0 : pthread_mutex_t *mutex = &cpi->pmutex[mb_row - 1];
116 0 : sync_read(mutex, mb_col, last_row_current_mb_col, nsync);
117 : }
118 :
119 : #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
120 : tp = tp_start;
121 : #endif
122 :
123 : /* Distance of Mb to the various image edges.
124 : * These specified to 8th pel as they are always compared
125 : * to values that are in 1/8th pel units
126 : */
127 0 : xd->mb_to_left_edge = -((mb_col * 16) << 3);
128 0 : xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
129 0 : xd->mb_to_top_edge = -((mb_row * 16) << 3);
130 0 : xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
131 :
132 : /* Set up limit values for motion vectors used to prevent
133 : * them extending outside the UMV borders
134 : */
135 0 : x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
136 0 : x->mv_col_max =
137 0 : ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
138 0 : x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
139 0 : x->mv_row_max =
140 0 : ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
141 :
142 0 : xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
143 0 : xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
144 0 : xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
145 0 : xd->left_available = (mb_col != 0);
146 :
147 0 : x->rddiv = cpi->RDDIV;
148 0 : x->rdmult = cpi->RDMULT;
149 :
150 : /* Copy current mb to a buffer */
151 0 : vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
152 :
153 0 : if (cpi->oxcf.tuning == VP8_TUNE_SSIM) vp8_activity_masking(cpi, x);
154 :
155 : /* Is segmentation enabled */
156 : /* MB level adjustment to quantizer */
157 0 : if (xd->segmentation_enabled) {
158 : /* Code to set segment id in xd->mbmi.segment_id for
159 : * current MB (with range checking)
160 : */
161 0 : if (cpi->segmentation_map[map_index + mb_col] <= 3) {
162 0 : xd->mode_info_context->mbmi.segment_id =
163 0 : cpi->segmentation_map[map_index + mb_col];
164 : } else {
165 0 : xd->mode_info_context->mbmi.segment_id = 0;
166 : }
167 :
168 0 : vp8cx_mb_init_quantizer(cpi, x, 1);
169 : } else {
170 : /* Set to Segment 0 by default */
171 0 : xd->mode_info_context->mbmi.segment_id = 0;
172 : }
173 :
174 0 : x->active_ptr = cpi->active_map + map_index + mb_col;
175 :
176 0 : if (cm->frame_type == KEY_FRAME) {
177 0 : *totalrate += vp8cx_encode_intra_macroblock(cpi, x, &tp);
178 : #ifdef MODE_STATS
179 : y_modes[xd->mbmi.mode]++;
180 : #endif
181 : } else {
182 0 : *totalrate += vp8cx_encode_inter_macroblock(
183 : cpi, x, &tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
184 :
185 : #ifdef MODE_STATS
186 : inter_y_modes[xd->mbmi.mode]++;
187 :
188 : if (xd->mbmi.mode == SPLITMV) {
189 : int b;
190 :
191 : for (b = 0; b < xd->mbmi.partition_count; ++b) {
192 : inter_b_modes[x->partition->bmi[b].mode]++;
193 : }
194 : }
195 :
196 : #endif
197 : // Keep track of how many (consecutive) times a block
198 : // is coded as ZEROMV_LASTREF, for base layer frames.
199 : // Reset to 0 if its coded as anything else.
200 0 : if (cpi->current_layer == 0) {
201 0 : if (xd->mode_info_context->mbmi.mode == ZEROMV &&
202 0 : xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
203 : // Increment, check for wrap-around.
204 0 : if (cpi->consec_zero_last[map_index + mb_col] < 255) {
205 0 : cpi->consec_zero_last[map_index + mb_col] += 1;
206 : }
207 0 : if (cpi->consec_zero_last_mvbias[map_index + mb_col] < 255) {
208 0 : cpi->consec_zero_last_mvbias[map_index + mb_col] += 1;
209 : }
210 : } else {
211 0 : cpi->consec_zero_last[map_index + mb_col] = 0;
212 0 : cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
213 : }
214 0 : if (x->zero_last_dot_suppress) {
215 0 : cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
216 : }
217 : }
218 :
219 : /* Special case code for cyclic refresh
220 : * If cyclic update enabled then copy
221 : * xd->mbmi.segment_id; (which may have been updated
222 : * based on mode during
223 : * vp8cx_encode_inter_macroblock()) back into the
224 : * global segmentation map
225 : */
226 0 : if ((cpi->current_layer == 0) &&
227 0 : (cpi->cyclic_refresh_mode_enabled &&
228 0 : xd->segmentation_enabled)) {
229 0 : const MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
230 0 : cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
231 :
232 : /* If the block has been refreshed mark it as clean
233 : * (the magnitude of the -ve influences how long it
234 : * will be before we consider another refresh):
235 : * Else if it was coded (last frame 0,0) and has
236 : * not already been refreshed then mark it as a
237 : * candidate for cleanup next time (marked 0) else
238 : * mark it as dirty (1).
239 : */
240 0 : if (mbmi->segment_id) {
241 0 : cpi->cyclic_refresh_map[map_index + mb_col] = -1;
242 0 : } else if ((mbmi->mode == ZEROMV) &&
243 0 : (mbmi->ref_frame == LAST_FRAME)) {
244 0 : if (cpi->cyclic_refresh_map[map_index + mb_col] == 1) {
245 0 : cpi->cyclic_refresh_map[map_index + mb_col] = 0;
246 : }
247 : } else {
248 0 : cpi->cyclic_refresh_map[map_index + mb_col] = 1;
249 : }
250 : }
251 : }
252 :
253 : #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
254 : /* pack tokens for this MB */
255 : {
256 : int tok_count = tp - tp_start;
257 : vp8_pack_tokens(w, tp_start, tok_count);
258 : }
259 : #else
260 0 : cpi->tplist[mb_row].stop = tp;
261 : #endif
262 : /* Increment pointer into gf usage flags structure. */
263 0 : x->gf_active_ptr++;
264 :
265 : /* Increment the activity mask pointers. */
266 0 : x->mb_activity_ptr++;
267 :
268 : /* adjust to the next column of macroblocks */
269 0 : x->src.y_buffer += 16;
270 0 : x->src.u_buffer += 8;
271 0 : x->src.v_buffer += 8;
272 :
273 0 : recon_yoffset += 16;
274 0 : recon_uvoffset += 8;
275 :
276 : /* Keep track of segment usage */
277 0 : segment_counts[xd->mode_info_context->mbmi.segment_id]++;
278 :
279 : /* skip to next mb */
280 0 : xd->mode_info_context++;
281 0 : x->partition_info++;
282 0 : xd->above_context++;
283 : }
284 :
285 0 : vp8_extend_mb_row(&cm->yv12_fb[dst_fb_idx], xd->dst.y_buffer + 16,
286 0 : xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
287 :
288 0 : protected_write(&cpi->pmutex[mb_row], current_mb_col, mb_col + nsync);
289 :
290 : /* this is to account for the border */
291 0 : xd->mode_info_context++;
292 0 : x->partition_info++;
293 :
294 0 : x->src.y_buffer +=
295 0 : 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) -
296 0 : 16 * cm->mb_cols;
297 0 : x->src.u_buffer +=
298 0 : 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
299 0 : 8 * cm->mb_cols;
300 0 : x->src.v_buffer +=
301 0 : 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
302 0 : 8 * cm->mb_cols;
303 :
304 0 : xd->mode_info_context +=
305 0 : xd->mode_info_stride * cpi->encoding_thread_count;
306 0 : x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
307 0 : x->gf_active_ptr += cm->mb_cols * cpi->encoding_thread_count;
308 : }
309 : /* Signal that this thread has completed processing its rows. */
310 0 : sem_post(&cpi->h_event_end_encoding[ithread]);
311 : }
312 : }
313 :
314 : /* printf("exit thread %d\n", ithread); */
315 0 : return 0;
316 : }
317 :
318 0 : static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc) {
319 0 : MACROBLOCK *x = mbsrc;
320 0 : MACROBLOCK *z = mbdst;
321 : int i;
322 :
323 0 : z->ss = x->ss;
324 0 : z->ss_count = x->ss_count;
325 0 : z->searches_per_step = x->searches_per_step;
326 0 : z->errorperbit = x->errorperbit;
327 :
328 0 : z->sadperbit16 = x->sadperbit16;
329 0 : z->sadperbit4 = x->sadperbit4;
330 :
331 : /*
332 : z->mv_col_min = x->mv_col_min;
333 : z->mv_col_max = x->mv_col_max;
334 : z->mv_row_min = x->mv_row_min;
335 : z->mv_row_max = x->mv_row_max;
336 : */
337 :
338 0 : z->short_fdct4x4 = x->short_fdct4x4;
339 0 : z->short_fdct8x4 = x->short_fdct8x4;
340 0 : z->short_walsh4x4 = x->short_walsh4x4;
341 0 : z->quantize_b = x->quantize_b;
342 0 : z->optimize = x->optimize;
343 :
344 : /*
345 : z->mvc = x->mvc;
346 : z->src.y_buffer = x->src.y_buffer;
347 : z->src.u_buffer = x->src.u_buffer;
348 : z->src.v_buffer = x->src.v_buffer;
349 : */
350 :
351 0 : z->mvcost[0] = x->mvcost[0];
352 0 : z->mvcost[1] = x->mvcost[1];
353 0 : z->mvsadcost[0] = x->mvsadcost[0];
354 0 : z->mvsadcost[1] = x->mvsadcost[1];
355 :
356 0 : z->token_costs = x->token_costs;
357 0 : z->inter_bmode_costs = x->inter_bmode_costs;
358 0 : z->mbmode_cost = x->mbmode_cost;
359 0 : z->intra_uv_mode_cost = x->intra_uv_mode_cost;
360 0 : z->bmode_costs = x->bmode_costs;
361 :
362 0 : for (i = 0; i < 25; ++i) {
363 0 : z->block[i].quant = x->block[i].quant;
364 0 : z->block[i].quant_fast = x->block[i].quant_fast;
365 0 : z->block[i].quant_shift = x->block[i].quant_shift;
366 0 : z->block[i].zbin = x->block[i].zbin;
367 0 : z->block[i].zrun_zbin_boost = x->block[i].zrun_zbin_boost;
368 0 : z->block[i].round = x->block[i].round;
369 0 : z->block[i].src_stride = x->block[i].src_stride;
370 : }
371 :
372 0 : z->q_index = x->q_index;
373 0 : z->act_zbin_adj = x->act_zbin_adj;
374 0 : z->last_act_zbin_adj = x->last_act_zbin_adj;
375 :
376 : {
377 0 : MACROBLOCKD *xd = &x->e_mbd;
378 0 : MACROBLOCKD *zd = &z->e_mbd;
379 :
380 : /*
381 : zd->mode_info_context = xd->mode_info_context;
382 : zd->mode_info = xd->mode_info;
383 :
384 : zd->mode_info_stride = xd->mode_info_stride;
385 : zd->frame_type = xd->frame_type;
386 : zd->up_available = xd->up_available ;
387 : zd->left_available = xd->left_available;
388 : zd->left_context = xd->left_context;
389 : zd->last_frame_dc = xd->last_frame_dc;
390 : zd->last_frame_dccons = xd->last_frame_dccons;
391 : zd->gold_frame_dc = xd->gold_frame_dc;
392 : zd->gold_frame_dccons = xd->gold_frame_dccons;
393 : zd->mb_to_left_edge = xd->mb_to_left_edge;
394 : zd->mb_to_right_edge = xd->mb_to_right_edge;
395 : zd->mb_to_top_edge = xd->mb_to_top_edge ;
396 : zd->mb_to_bottom_edge = xd->mb_to_bottom_edge;
397 : zd->gf_active_ptr = xd->gf_active_ptr;
398 : zd->frames_since_golden = xd->frames_since_golden;
399 : zd->frames_till_alt_ref_frame = xd->frames_till_alt_ref_frame;
400 : */
401 0 : zd->subpixel_predict = xd->subpixel_predict;
402 0 : zd->subpixel_predict8x4 = xd->subpixel_predict8x4;
403 0 : zd->subpixel_predict8x8 = xd->subpixel_predict8x8;
404 0 : zd->subpixel_predict16x16 = xd->subpixel_predict16x16;
405 0 : zd->segmentation_enabled = xd->segmentation_enabled;
406 0 : zd->mb_segement_abs_delta = xd->mb_segement_abs_delta;
407 0 : memcpy(zd->segment_feature_data, xd->segment_feature_data,
408 : sizeof(xd->segment_feature_data));
409 :
410 0 : memcpy(zd->dequant_y1_dc, xd->dequant_y1_dc, sizeof(xd->dequant_y1_dc));
411 0 : memcpy(zd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
412 0 : memcpy(zd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
413 0 : memcpy(zd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
414 :
415 : #if 1
416 : /*TODO: Remove dequant from BLOCKD. This is a temporary solution until
417 : * the quantizer code uses a passed in pointer to the dequant constants.
418 : * This will also require modifications to the x86 and neon assembly.
419 : * */
420 0 : for (i = 0; i < 16; ++i) zd->block[i].dequant = zd->dequant_y1;
421 0 : for (i = 16; i < 24; ++i) zd->block[i].dequant = zd->dequant_uv;
422 0 : zd->block[24].dequant = zd->dequant_y2;
423 : #endif
424 :
425 0 : memcpy(z->rd_threshes, x->rd_threshes, sizeof(x->rd_threshes));
426 0 : memcpy(z->rd_thresh_mult, x->rd_thresh_mult, sizeof(x->rd_thresh_mult));
427 :
428 0 : z->zbin_over_quant = x->zbin_over_quant;
429 0 : z->zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
430 0 : z->zbin_mode_boost = x->zbin_mode_boost;
431 :
432 0 : memset(z->error_bins, 0, sizeof(z->error_bins));
433 : }
434 0 : }
435 :
436 0 : void vp8cx_init_mbrthread_data(VP8_COMP *cpi, MACROBLOCK *x,
437 : MB_ROW_COMP *mbr_ei, int count) {
438 0 : VP8_COMMON *const cm = &cpi->common;
439 0 : MACROBLOCKD *const xd = &x->e_mbd;
440 : int i;
441 :
442 0 : for (i = 0; i < count; ++i) {
443 0 : MACROBLOCK *mb = &mbr_ei[i].mb;
444 0 : MACROBLOCKD *mbd = &mb->e_mbd;
445 :
446 0 : mbd->subpixel_predict = xd->subpixel_predict;
447 0 : mbd->subpixel_predict8x4 = xd->subpixel_predict8x4;
448 0 : mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
449 0 : mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
450 0 : mb->gf_active_ptr = x->gf_active_ptr;
451 :
452 0 : memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
453 0 : mbr_ei[i].totalrate = 0;
454 :
455 0 : mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
456 :
457 0 : mbd->frame_type = cm->frame_type;
458 :
459 0 : mb->src = *cpi->Source;
460 0 : mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
461 0 : mbd->dst = cm->yv12_fb[cm->new_fb_idx];
462 :
463 0 : mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
464 0 : mb->src.u_buffer += 8 * x->src.uv_stride * (i + 1);
465 0 : mb->src.v_buffer += 8 * x->src.uv_stride * (i + 1);
466 :
467 0 : vp8_build_block_offsets(mb);
468 :
469 0 : mbd->left_context = &cm->left_context;
470 0 : mb->mvc = cm->fc.mvc;
471 :
472 0 : setup_mbby_copy(&mbr_ei[i].mb, x);
473 :
474 0 : mbd->fullpixel_mask = 0xffffffff;
475 0 : if (cm->full_pixel) mbd->fullpixel_mask = 0xfffffff8;
476 :
477 0 : vp8_zero(mb->coef_counts);
478 0 : vp8_zero(x->ymode_count);
479 0 : mb->skip_true_count = 0;
480 0 : vp8_zero(mb->MVcount);
481 0 : mb->prediction_error = 0;
482 0 : mb->intra_error = 0;
483 0 : vp8_zero(mb->count_mb_ref_frame_usage);
484 0 : mb->mbs_tested_so_far = 0;
485 0 : mb->mbs_zero_last_dot_suppress = 0;
486 : }
487 0 : }
488 :
489 0 : int vp8cx_create_encoder_threads(VP8_COMP *cpi) {
490 0 : const VP8_COMMON *cm = &cpi->common;
491 :
492 0 : cpi->b_multi_threaded = 0;
493 0 : cpi->encoding_thread_count = 0;
494 0 : cpi->b_lpf_running = 0;
495 :
496 0 : pthread_mutex_init(&cpi->mt_mutex, NULL);
497 :
498 0 : if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1) {
499 : int ithread;
500 0 : int th_count = cpi->oxcf.multi_threaded - 1;
501 0 : int rc = 0;
502 :
503 : /* don't allocate more threads than cores available */
504 0 : if (cpi->oxcf.multi_threaded > cm->processor_core_count) {
505 0 : th_count = cm->processor_core_count - 1;
506 : }
507 :
508 : /* we have th_count + 1 (main) threads processing one row each */
509 : /* no point to have more threads than the sync range allows */
510 0 : if (th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1)) {
511 0 : th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
512 : }
513 :
514 0 : if (th_count == 0) return 0;
515 :
516 0 : CHECK_MEM_ERROR(cpi->h_encoding_thread,
517 : vpx_malloc(sizeof(pthread_t) * th_count));
518 0 : CHECK_MEM_ERROR(cpi->h_event_start_encoding,
519 : vpx_malloc(sizeof(sem_t) * th_count));
520 0 : CHECK_MEM_ERROR(cpi->h_event_end_encoding,
521 : vpx_malloc(sizeof(sem_t) * th_count));
522 0 : CHECK_MEM_ERROR(cpi->mb_row_ei,
523 : vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
524 0 : memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
525 0 : CHECK_MEM_ERROR(cpi->en_thread_data,
526 : vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
527 :
528 0 : cpi->b_multi_threaded = 1;
529 0 : cpi->encoding_thread_count = th_count;
530 :
531 : /*
532 : printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
533 : (cpi->encoding_thread_count +1));
534 : */
535 :
536 0 : for (ithread = 0; ithread < th_count; ++ithread) {
537 0 : ENCODETHREAD_DATA *ethd = &cpi->en_thread_data[ithread];
538 :
539 : /* Setup block ptrs and offsets */
540 0 : vp8_setup_block_ptrs(&cpi->mb_row_ei[ithread].mb);
541 0 : vp8_setup_block_dptrs(&cpi->mb_row_ei[ithread].mb.e_mbd);
542 :
543 0 : sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
544 0 : sem_init(&cpi->h_event_end_encoding[ithread], 0, 0);
545 :
546 0 : ethd->ithread = ithread;
547 0 : ethd->ptr1 = (void *)cpi;
548 0 : ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
549 :
550 0 : rc = pthread_create(&cpi->h_encoding_thread[ithread], 0,
551 : thread_encoding_proc, ethd);
552 0 : if (rc) break;
553 : }
554 :
555 0 : if (rc) {
556 : /* shutdown other threads */
557 0 : protected_write(&cpi->mt_mutex, &cpi->b_multi_threaded, 0);
558 0 : for (--ithread; ithread >= 0; ithread--) {
559 0 : pthread_join(cpi->h_encoding_thread[ithread], 0);
560 0 : sem_destroy(&cpi->h_event_start_encoding[ithread]);
561 0 : sem_destroy(&cpi->h_event_end_encoding[ithread]);
562 : }
563 :
564 : /* free thread related resources */
565 0 : vpx_free(cpi->h_event_start_encoding);
566 0 : vpx_free(cpi->h_event_end_encoding);
567 0 : vpx_free(cpi->h_encoding_thread);
568 0 : vpx_free(cpi->mb_row_ei);
569 0 : vpx_free(cpi->en_thread_data);
570 :
571 0 : pthread_mutex_destroy(&cpi->mt_mutex);
572 :
573 0 : return -1;
574 : }
575 :
576 : {
577 0 : LPFTHREAD_DATA *lpfthd = &cpi->lpf_thread_data;
578 :
579 0 : sem_init(&cpi->h_event_start_lpf, 0, 0);
580 0 : sem_init(&cpi->h_event_end_lpf, 0, 0);
581 :
582 0 : lpfthd->ptr1 = (void *)cpi;
583 0 : rc = pthread_create(&cpi->h_filter_thread, 0, thread_loopfilter, lpfthd);
584 :
585 0 : if (rc) {
586 : /* shutdown other threads */
587 0 : protected_write(&cpi->mt_mutex, &cpi->b_multi_threaded, 0);
588 0 : for (--ithread; ithread >= 0; ithread--) {
589 0 : sem_post(&cpi->h_event_start_encoding[ithread]);
590 0 : sem_post(&cpi->h_event_end_encoding[ithread]);
591 0 : pthread_join(cpi->h_encoding_thread[ithread], 0);
592 0 : sem_destroy(&cpi->h_event_start_encoding[ithread]);
593 0 : sem_destroy(&cpi->h_event_end_encoding[ithread]);
594 : }
595 0 : sem_destroy(&cpi->h_event_end_lpf);
596 0 : sem_destroy(&cpi->h_event_start_lpf);
597 :
598 : /* free thread related resources */
599 0 : vpx_free(cpi->h_event_start_encoding);
600 0 : vpx_free(cpi->h_event_end_encoding);
601 0 : vpx_free(cpi->h_encoding_thread);
602 0 : vpx_free(cpi->mb_row_ei);
603 0 : vpx_free(cpi->en_thread_data);
604 :
605 0 : pthread_mutex_destroy(&cpi->mt_mutex);
606 :
607 0 : return -2;
608 : }
609 : }
610 : }
611 0 : return 0;
612 : }
613 :
614 0 : void vp8cx_remove_encoder_threads(VP8_COMP *cpi) {
615 0 : if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded)) {
616 : /* shutdown other threads */
617 0 : protected_write(&cpi->mt_mutex, &cpi->b_multi_threaded, 0);
618 : {
619 : int i;
620 :
621 0 : for (i = 0; i < cpi->encoding_thread_count; ++i) {
622 0 : sem_post(&cpi->h_event_start_encoding[i]);
623 0 : sem_post(&cpi->h_event_end_encoding[i]);
624 :
625 0 : pthread_join(cpi->h_encoding_thread[i], 0);
626 :
627 0 : sem_destroy(&cpi->h_event_start_encoding[i]);
628 0 : sem_destroy(&cpi->h_event_end_encoding[i]);
629 : }
630 :
631 0 : sem_post(&cpi->h_event_start_lpf);
632 0 : pthread_join(cpi->h_filter_thread, 0);
633 : }
634 :
635 0 : sem_destroy(&cpi->h_event_end_lpf);
636 0 : sem_destroy(&cpi->h_event_start_lpf);
637 :
638 : /* free thread related resources */
639 0 : vpx_free(cpi->h_event_start_encoding);
640 0 : vpx_free(cpi->h_event_end_encoding);
641 0 : vpx_free(cpi->h_encoding_thread);
642 0 : vpx_free(cpi->mb_row_ei);
643 0 : vpx_free(cpi->en_thread_data);
644 : }
645 0 : pthread_mutex_destroy(&cpi->mt_mutex);
646 0 : }
647 : #endif
|
