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 "vpx_config.h"
12 #include "vp8_rtcd.h"
13 #if !defined(WIN32) && CONFIG_OS_SUPPORT == 1
14 #include <unistd.h>
15 #endif
16 #include "onyxd_int.h"
17 #include "vpx_mem/vpx_mem.h"
18 #include "vp8/common/threading.h"
19
20 #include "vp8/common/loopfilter.h"
21 #include "vp8/common/extend.h"
22 #include "vpx_ports/vpx_timer.h"
23 #include "detokenize.h"
24 #include "vp8/common/reconintra4x4.h"
25 #include "vp8/common/reconinter.h"
26 #include "vp8/common/reconintra.h"
27 #include "vp8/common/setupintrarecon.h"
28 #if CONFIG_ERROR_CONCEALMENT
29 #include "error_concealment.h"
30 #endif
31
32 #define CALLOC_ARRAY(p, n) CHECK_MEM_ERROR((p), vpx_calloc(sizeof(*(p)), (n)))
33 #define CALLOC_ARRAY_ALIGNED(p, n, algn) \
34 do { \
35 CHECK_MEM_ERROR((p), vpx_memalign((algn), sizeof(*(p)) * (n))); \
36 memset((p), 0, (n) * sizeof(*(p))); \
37 } while (0)
38
39 void vp8_mb_init_dequantizer(VP8D_COMP *pbi, MACROBLOCKD *xd);
40
setup_decoding_thread_data(VP8D_COMP * pbi,MACROBLOCKD * xd,MB_ROW_DEC * mbrd,int count)41 static void setup_decoding_thread_data(VP8D_COMP *pbi, MACROBLOCKD *xd,
42 MB_ROW_DEC *mbrd, int count) {
43 VP8_COMMON *const pc = &pbi->common;
44 int i;
45
46 for (i = 0; i < count; ++i) {
47 MACROBLOCKD *mbd = &mbrd[i].mbd;
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
53 mbd->frame_type = pc->frame_type;
54 mbd->pre = xd->pre;
55 mbd->dst = xd->dst;
56
57 mbd->segmentation_enabled = xd->segmentation_enabled;
58 mbd->mb_segement_abs_delta = xd->mb_segement_abs_delta;
59 memcpy(mbd->segment_feature_data, xd->segment_feature_data,
60 sizeof(xd->segment_feature_data));
61
62 /*signed char ref_lf_deltas[MAX_REF_LF_DELTAS];*/
63 memcpy(mbd->ref_lf_deltas, xd->ref_lf_deltas, sizeof(xd->ref_lf_deltas));
64 /*signed char mode_lf_deltas[MAX_MODE_LF_DELTAS];*/
65 memcpy(mbd->mode_lf_deltas, xd->mode_lf_deltas, sizeof(xd->mode_lf_deltas));
66 /*unsigned char mode_ref_lf_delta_enabled;
67 unsigned char mode_ref_lf_delta_update;*/
68 mbd->mode_ref_lf_delta_enabled = xd->mode_ref_lf_delta_enabled;
69 mbd->mode_ref_lf_delta_update = xd->mode_ref_lf_delta_update;
70
71 mbd->current_bc = &pbi->mbc[0];
72
73 memcpy(mbd->dequant_y1_dc, xd->dequant_y1_dc, sizeof(xd->dequant_y1_dc));
74 memcpy(mbd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
75 memcpy(mbd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
76 memcpy(mbd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
77
78 mbd->fullpixel_mask = 0xffffffff;
79
80 if (pc->full_pixel) mbd->fullpixel_mask = 0xfffffff8;
81 }
82
83 for (i = 0; i < pc->mb_rows; ++i) pbi->mt_current_mb_col[i] = -1;
84 }
85
mt_decode_macroblock(VP8D_COMP * pbi,MACROBLOCKD * xd,unsigned int mb_idx)86 static void mt_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd,
87 unsigned int mb_idx) {
88 MB_PREDICTION_MODE mode;
89 int i;
90 #if CONFIG_ERROR_CONCEALMENT
91 int corruption_detected = 0;
92 #else
93 (void)mb_idx;
94 #endif
95
96 if (xd->mode_info_context->mbmi.mb_skip_coeff) {
97 vp8_reset_mb_tokens_context(xd);
98 } else if (!vp8dx_bool_error(xd->current_bc)) {
99 int eobtotal;
100 eobtotal = vp8_decode_mb_tokens(pbi, xd);
101
102 /* Special case: Force the loopfilter to skip when eobtotal is zero */
103 xd->mode_info_context->mbmi.mb_skip_coeff = (eobtotal == 0);
104 }
105
106 mode = xd->mode_info_context->mbmi.mode;
107
108 if (xd->segmentation_enabled) vp8_mb_init_dequantizer(pbi, xd);
109
110 #if CONFIG_ERROR_CONCEALMENT
111
112 if (pbi->ec_active) {
113 int throw_residual;
114 /* When we have independent partitions we can apply residual even
115 * though other partitions within the frame are corrupt.
116 */
117 throw_residual =
118 (!pbi->independent_partitions && pbi->frame_corrupt_residual);
119 throw_residual = (throw_residual || vp8dx_bool_error(xd->current_bc));
120
121 if ((mb_idx >= pbi->mvs_corrupt_from_mb || throw_residual)) {
122 /* MB with corrupt residuals or corrupt mode/motion vectors.
123 * Better to use the predictor as reconstruction.
124 */
125 pbi->frame_corrupt_residual = 1;
126 memset(xd->qcoeff, 0, sizeof(xd->qcoeff));
127
128 corruption_detected = 1;
129
130 /* force idct to be skipped for B_PRED and use the
131 * prediction only for reconstruction
132 * */
133 memset(xd->eobs, 0, 25);
134 }
135 }
136 #endif
137
138 /* do prediction */
139 if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) {
140 vp8_build_intra_predictors_mbuv_s(
141 xd, xd->recon_above[1], xd->recon_above[2], xd->recon_left[1],
142 xd->recon_left[2], xd->recon_left_stride[1], xd->dst.u_buffer,
143 xd->dst.v_buffer, xd->dst.uv_stride);
144
145 if (mode != B_PRED) {
146 vp8_build_intra_predictors_mby_s(
147 xd, xd->recon_above[0], xd->recon_left[0], xd->recon_left_stride[0],
148 xd->dst.y_buffer, xd->dst.y_stride);
149 } else {
150 short *DQC = xd->dequant_y1;
151 int dst_stride = xd->dst.y_stride;
152
153 /* clear out residual eob info */
154 if (xd->mode_info_context->mbmi.mb_skip_coeff) memset(xd->eobs, 0, 25);
155
156 intra_prediction_down_copy(xd, xd->recon_above[0] + 16);
157
158 for (i = 0; i < 16; ++i) {
159 BLOCKD *b = &xd->block[i];
160 unsigned char *dst = xd->dst.y_buffer + b->offset;
161 B_PREDICTION_MODE b_mode = xd->mode_info_context->bmi[i].as_mode;
162 unsigned char *Above;
163 unsigned char *yleft;
164 int left_stride;
165 unsigned char top_left;
166
167 /*Caution: For some b_mode, it needs 8 pixels (4 above + 4
168 * above-right).*/
169 if (i < 4 && pbi->common.filter_level) {
170 Above = xd->recon_above[0] + b->offset;
171 } else {
172 Above = dst - dst_stride;
173 }
174
175 if (i % 4 == 0 && pbi->common.filter_level) {
176 yleft = xd->recon_left[0] + i;
177 left_stride = 1;
178 } else {
179 yleft = dst - 1;
180 left_stride = dst_stride;
181 }
182
183 if ((i == 4 || i == 8 || i == 12) && pbi->common.filter_level) {
184 top_left = *(xd->recon_left[0] + i - 1);
185 } else {
186 top_left = Above[-1];
187 }
188
189 vp8_intra4x4_predict(Above, yleft, left_stride, b_mode, dst, dst_stride,
190 top_left);
191
192 if (xd->eobs[i]) {
193 if (xd->eobs[i] > 1) {
194 vp8_dequant_idct_add(b->qcoeff, DQC, dst, dst_stride);
195 } else {
196 vp8_dc_only_idct_add(b->qcoeff[0] * DQC[0], dst, dst_stride, dst,
197 dst_stride);
198 memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
199 }
200 }
201 }
202 }
203 } else {
204 vp8_build_inter_predictors_mb(xd);
205 }
206
207 #if CONFIG_ERROR_CONCEALMENT
208 if (corruption_detected) {
209 return;
210 }
211 #endif
212
213 if (!xd->mode_info_context->mbmi.mb_skip_coeff) {
214 /* dequantization and idct */
215 if (mode != B_PRED) {
216 short *DQC = xd->dequant_y1;
217
218 if (mode != SPLITMV) {
219 BLOCKD *b = &xd->block[24];
220
221 /* do 2nd order transform on the dc block */
222 if (xd->eobs[24] > 1) {
223 vp8_dequantize_b(b, xd->dequant_y2);
224
225 vp8_short_inv_walsh4x4(&b->dqcoeff[0], xd->qcoeff);
226 memset(b->qcoeff, 0, 16 * sizeof(b->qcoeff[0]));
227 } else {
228 b->dqcoeff[0] = b->qcoeff[0] * xd->dequant_y2[0];
229 vp8_short_inv_walsh4x4_1(&b->dqcoeff[0], xd->qcoeff);
230 memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
231 }
232
233 /* override the dc dequant constant in order to preserve the
234 * dc components
235 */
236 DQC = xd->dequant_y1_dc;
237 }
238
239 vp8_dequant_idct_add_y_block(xd->qcoeff, DQC, xd->dst.y_buffer,
240 xd->dst.y_stride, xd->eobs);
241 }
242
243 vp8_dequant_idct_add_uv_block(xd->qcoeff + 16 * 16, xd->dequant_uv,
244 xd->dst.u_buffer, xd->dst.v_buffer,
245 xd->dst.uv_stride, xd->eobs + 16);
246 }
247 }
248
mt_decode_mb_rows(VP8D_COMP * pbi,MACROBLOCKD * xd,int start_mb_row)249 static void mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd,
250 int start_mb_row) {
251 const int *last_row_current_mb_col;
252 int *current_mb_col;
253 int mb_row;
254 VP8_COMMON *pc = &pbi->common;
255 const int nsync = pbi->sync_range;
256 const int first_row_no_sync_above = pc->mb_cols + nsync;
257 int num_part = 1 << pbi->common.multi_token_partition;
258 int last_mb_row = start_mb_row;
259
260 YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
261 YV12_BUFFER_CONFIG *yv12_fb_lst = pbi->dec_fb_ref[LAST_FRAME];
262
263 int recon_y_stride = yv12_fb_new->y_stride;
264 int recon_uv_stride = yv12_fb_new->uv_stride;
265
266 unsigned char *ref_buffer[MAX_REF_FRAMES][3];
267 unsigned char *dst_buffer[3];
268 int i;
269 int ref_fb_corrupted[MAX_REF_FRAMES];
270
271 ref_fb_corrupted[INTRA_FRAME] = 0;
272
273 for (i = 1; i < MAX_REF_FRAMES; ++i) {
274 YV12_BUFFER_CONFIG *this_fb = pbi->dec_fb_ref[i];
275
276 ref_buffer[i][0] = this_fb->y_buffer;
277 ref_buffer[i][1] = this_fb->u_buffer;
278 ref_buffer[i][2] = this_fb->v_buffer;
279
280 ref_fb_corrupted[i] = this_fb->corrupted;
281 }
282
283 dst_buffer[0] = yv12_fb_new->y_buffer;
284 dst_buffer[1] = yv12_fb_new->u_buffer;
285 dst_buffer[2] = yv12_fb_new->v_buffer;
286
287 xd->up_available = (start_mb_row != 0);
288
289 xd->mode_info_context = pc->mi + pc->mode_info_stride * start_mb_row;
290 xd->mode_info_stride = pc->mode_info_stride;
291
292 for (mb_row = start_mb_row; mb_row < pc->mb_rows;
293 mb_row += (pbi->decoding_thread_count + 1)) {
294 int recon_yoffset, recon_uvoffset;
295 int mb_col;
296 int filter_level;
297 loop_filter_info_n *lfi_n = &pc->lf_info;
298
299 /* save last row processed by this thread */
300 last_mb_row = mb_row;
301 /* select bool coder for current partition */
302 xd->current_bc = &pbi->mbc[mb_row % num_part];
303
304 if (mb_row > 0) {
305 last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row - 1];
306 } else {
307 last_row_current_mb_col = &first_row_no_sync_above;
308 }
309
310 current_mb_col = &pbi->mt_current_mb_col[mb_row];
311
312 recon_yoffset = mb_row * recon_y_stride * 16;
313 recon_uvoffset = mb_row * recon_uv_stride * 8;
314
315 /* reset contexts */
316 xd->above_context = pc->above_context;
317 memset(xd->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));
318
319 xd->left_available = 0;
320
321 xd->mb_to_top_edge = -((mb_row * 16) << 3);
322 xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
323
324 if (pbi->common.filter_level) {
325 xd->recon_above[0] = pbi->mt_yabove_row[mb_row] + 0 * 16 + 32;
326 xd->recon_above[1] = pbi->mt_uabove_row[mb_row] + 0 * 8 + 16;
327 xd->recon_above[2] = pbi->mt_vabove_row[mb_row] + 0 * 8 + 16;
328
329 xd->recon_left[0] = pbi->mt_yleft_col[mb_row];
330 xd->recon_left[1] = pbi->mt_uleft_col[mb_row];
331 xd->recon_left[2] = pbi->mt_vleft_col[mb_row];
332
333 /* TODO: move to outside row loop */
334 xd->recon_left_stride[0] = 1;
335 xd->recon_left_stride[1] = 1;
336 } else {
337 xd->recon_above[0] = dst_buffer[0] + recon_yoffset;
338 xd->recon_above[1] = dst_buffer[1] + recon_uvoffset;
339 xd->recon_above[2] = dst_buffer[2] + recon_uvoffset;
340
341 xd->recon_left[0] = xd->recon_above[0] - 1;
342 xd->recon_left[1] = xd->recon_above[1] - 1;
343 xd->recon_left[2] = xd->recon_above[2] - 1;
344
345 xd->recon_above[0] -= xd->dst.y_stride;
346 xd->recon_above[1] -= xd->dst.uv_stride;
347 xd->recon_above[2] -= xd->dst.uv_stride;
348
349 /* TODO: move to outside row loop */
350 xd->recon_left_stride[0] = xd->dst.y_stride;
351 xd->recon_left_stride[1] = xd->dst.uv_stride;
352
353 setup_intra_recon_left(xd->recon_left[0], xd->recon_left[1],
354 xd->recon_left[2], xd->dst.y_stride,
355 xd->dst.uv_stride);
356 }
357
358 for (mb_col = 0; mb_col < pc->mb_cols; ++mb_col) {
359 if (((mb_col - 1) % nsync) == 0) {
360 pthread_mutex_t *mutex = &pbi->pmutex[mb_row];
361 protected_write(mutex, current_mb_col, mb_col - 1);
362 }
363
364 if (mb_row && !(mb_col & (nsync - 1))) {
365 pthread_mutex_t *mutex = &pbi->pmutex[mb_row - 1];
366 sync_read(mutex, mb_col, last_row_current_mb_col, nsync);
367 }
368
369 /* Distance of MB to the various image edges.
370 * These are specified to 8th pel as they are always
371 * compared to values that are in 1/8th pel units.
372 */
373 xd->mb_to_left_edge = -((mb_col * 16) << 3);
374 xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
375
376 #if CONFIG_ERROR_CONCEALMENT
377 {
378 int corrupt_residual =
379 (!pbi->independent_partitions && pbi->frame_corrupt_residual) ||
380 vp8dx_bool_error(xd->current_bc);
381 if (pbi->ec_active &&
382 (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) &&
383 corrupt_residual) {
384 /* We have an intra block with corrupt
385 * coefficients, better to conceal with an inter
386 * block.
387 * Interpolate MVs from neighboring MBs
388 *
389 * Note that for the first mb with corrupt
390 * residual in a frame, we might not discover
391 * that before decoding the residual. That
392 * happens after this check, and therefore no
393 * inter concealment will be done.
394 */
395 vp8_interpolate_motion(xd, mb_row, mb_col, pc->mb_rows, pc->mb_cols);
396 }
397 }
398 #endif
399
400 xd->dst.y_buffer = dst_buffer[0] + recon_yoffset;
401 xd->dst.u_buffer = dst_buffer[1] + recon_uvoffset;
402 xd->dst.v_buffer = dst_buffer[2] + recon_uvoffset;
403
404 xd->pre.y_buffer =
405 ref_buffer[xd->mode_info_context->mbmi.ref_frame][0] + recon_yoffset;
406 xd->pre.u_buffer =
407 ref_buffer[xd->mode_info_context->mbmi.ref_frame][1] + recon_uvoffset;
408 xd->pre.v_buffer =
409 ref_buffer[xd->mode_info_context->mbmi.ref_frame][2] + recon_uvoffset;
410
411 /* propagate errors from reference frames */
412 xd->corrupted |= ref_fb_corrupted[xd->mode_info_context->mbmi.ref_frame];
413
414 mt_decode_macroblock(pbi, xd, 0);
415
416 xd->left_available = 1;
417
418 /* check if the boolean decoder has suffered an error */
419 xd->corrupted |= vp8dx_bool_error(xd->current_bc);
420
421 xd->recon_above[0] += 16;
422 xd->recon_above[1] += 8;
423 xd->recon_above[2] += 8;
424
425 if (!pbi->common.filter_level) {
426 xd->recon_left[0] += 16;
427 xd->recon_left[1] += 8;
428 xd->recon_left[2] += 8;
429 }
430
431 if (pbi->common.filter_level) {
432 int skip_lf = (xd->mode_info_context->mbmi.mode != B_PRED &&
433 xd->mode_info_context->mbmi.mode != SPLITMV &&
434 xd->mode_info_context->mbmi.mb_skip_coeff);
435
436 const int mode_index =
437 lfi_n->mode_lf_lut[xd->mode_info_context->mbmi.mode];
438 const int seg = xd->mode_info_context->mbmi.segment_id;
439 const int ref_frame = xd->mode_info_context->mbmi.ref_frame;
440
441 filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
442
443 if (mb_row != pc->mb_rows - 1) {
444 /* Save decoded MB last row data for next-row decoding */
445 memcpy((pbi->mt_yabove_row[mb_row + 1] + 32 + mb_col * 16),
446 (xd->dst.y_buffer + 15 * recon_y_stride), 16);
447 memcpy((pbi->mt_uabove_row[mb_row + 1] + 16 + mb_col * 8),
448 (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
449 memcpy((pbi->mt_vabove_row[mb_row + 1] + 16 + mb_col * 8),
450 (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
451 }
452
453 /* save left_col for next MB decoding */
454 if (mb_col != pc->mb_cols - 1) {
455 MODE_INFO *next = xd->mode_info_context + 1;
456
457 if (next->mbmi.ref_frame == INTRA_FRAME) {
458 for (i = 0; i < 16; ++i) {
459 pbi->mt_yleft_col[mb_row][i] =
460 xd->dst.y_buffer[i * recon_y_stride + 15];
461 }
462 for (i = 0; i < 8; ++i) {
463 pbi->mt_uleft_col[mb_row][i] =
464 xd->dst.u_buffer[i * recon_uv_stride + 7];
465 pbi->mt_vleft_col[mb_row][i] =
466 xd->dst.v_buffer[i * recon_uv_stride + 7];
467 }
468 }
469 }
470
471 /* loopfilter on this macroblock. */
472 if (filter_level) {
473 if (pc->filter_type == NORMAL_LOOPFILTER) {
474 loop_filter_info lfi;
475 FRAME_TYPE frame_type = pc->frame_type;
476 const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
477 lfi.mblim = lfi_n->mblim[filter_level];
478 lfi.blim = lfi_n->blim[filter_level];
479 lfi.lim = lfi_n->lim[filter_level];
480 lfi.hev_thr = lfi_n->hev_thr[hev_index];
481
482 if (mb_col > 0)
483 vp8_loop_filter_mbv(xd->dst.y_buffer, xd->dst.u_buffer,
484 xd->dst.v_buffer, recon_y_stride,
485 recon_uv_stride, &lfi);
486
487 if (!skip_lf)
488 vp8_loop_filter_bv(xd->dst.y_buffer, xd->dst.u_buffer,
489 xd->dst.v_buffer, recon_y_stride,
490 recon_uv_stride, &lfi);
491
492 /* don't apply across umv border */
493 if (mb_row > 0)
494 vp8_loop_filter_mbh(xd->dst.y_buffer, xd->dst.u_buffer,
495 xd->dst.v_buffer, recon_y_stride,
496 recon_uv_stride, &lfi);
497
498 if (!skip_lf)
499 vp8_loop_filter_bh(xd->dst.y_buffer, xd->dst.u_buffer,
500 xd->dst.v_buffer, recon_y_stride,
501 recon_uv_stride, &lfi);
502 } else {
503 if (mb_col > 0)
504 vp8_loop_filter_simple_mbv(xd->dst.y_buffer, recon_y_stride,
505 lfi_n->mblim[filter_level]);
506
507 if (!skip_lf)
508 vp8_loop_filter_simple_bv(xd->dst.y_buffer, recon_y_stride,
509 lfi_n->blim[filter_level]);
510
511 /* don't apply across umv border */
512 if (mb_row > 0)
513 vp8_loop_filter_simple_mbh(xd->dst.y_buffer, recon_y_stride,
514 lfi_n->mblim[filter_level]);
515
516 if (!skip_lf)
517 vp8_loop_filter_simple_bh(xd->dst.y_buffer, recon_y_stride,
518 lfi_n->blim[filter_level]);
519 }
520 }
521 }
522
523 recon_yoffset += 16;
524 recon_uvoffset += 8;
525
526 ++xd->mode_info_context; /* next mb */
527
528 xd->above_context++;
529 }
530
531 /* adjust to the next row of mbs */
532 if (pbi->common.filter_level) {
533 if (mb_row != pc->mb_rows - 1) {
534 int lasty = yv12_fb_lst->y_width + VP8BORDERINPIXELS;
535 int lastuv = (yv12_fb_lst->y_width >> 1) + (VP8BORDERINPIXELS >> 1);
536
537 for (i = 0; i < 4; ++i) {
538 pbi->mt_yabove_row[mb_row + 1][lasty + i] =
539 pbi->mt_yabove_row[mb_row + 1][lasty - 1];
540 pbi->mt_uabove_row[mb_row + 1][lastuv + i] =
541 pbi->mt_uabove_row[mb_row + 1][lastuv - 1];
542 pbi->mt_vabove_row[mb_row + 1][lastuv + i] =
543 pbi->mt_vabove_row[mb_row + 1][lastuv - 1];
544 }
545 }
546 } else {
547 vp8_extend_mb_row(yv12_fb_new, xd->dst.y_buffer + 16,
548 xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
549 }
550
551 /* last MB of row is ready just after extension is done */
552 protected_write(&pbi->pmutex[mb_row], current_mb_col, mb_col + nsync);
553
554 ++xd->mode_info_context; /* skip prediction column */
555 xd->up_available = 1;
556
557 /* since we have multithread */
558 xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
559 }
560
561 /* signal end of frame decoding if this thread processed the last mb_row */
562 if (last_mb_row == (pc->mb_rows - 1)) sem_post(&pbi->h_event_end_decoding);
563 }
564
thread_decoding_proc(void * p_data)565 static THREAD_FUNCTION thread_decoding_proc(void *p_data) {
566 int ithread = ((DECODETHREAD_DATA *)p_data)->ithread;
567 VP8D_COMP *pbi = (VP8D_COMP *)(((DECODETHREAD_DATA *)p_data)->ptr1);
568 MB_ROW_DEC *mbrd = (MB_ROW_DEC *)(((DECODETHREAD_DATA *)p_data)->ptr2);
569 ENTROPY_CONTEXT_PLANES mb_row_left_context;
570
571 while (1) {
572 if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd) == 0) break;
573
574 if (sem_wait(&pbi->h_event_start_decoding[ithread]) == 0) {
575 if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd) == 0) {
576 break;
577 } else {
578 MACROBLOCKD *xd = &mbrd->mbd;
579 xd->left_context = &mb_row_left_context;
580
581 mt_decode_mb_rows(pbi, xd, ithread + 1);
582 }
583 }
584 }
585
586 return 0;
587 }
588
vp8_decoder_create_threads(VP8D_COMP * pbi)589 void vp8_decoder_create_threads(VP8D_COMP *pbi) {
590 int core_count = 0;
591 unsigned int ithread;
592
593 pbi->b_multithreaded_rd = 0;
594 pbi->allocated_decoding_thread_count = 0;
595 pthread_mutex_init(&pbi->mt_mutex, NULL);
596
597 /* limit decoding threads to the max number of token partitions */
598 core_count = (pbi->max_threads > 8) ? 8 : pbi->max_threads;
599
600 /* limit decoding threads to the available cores */
601 if (core_count > pbi->common.processor_core_count) {
602 core_count = pbi->common.processor_core_count;
603 }
604
605 if (core_count > 1) {
606 pbi->b_multithreaded_rd = 1;
607 pbi->decoding_thread_count = core_count - 1;
608
609 CALLOC_ARRAY(pbi->h_decoding_thread, pbi->decoding_thread_count);
610 CALLOC_ARRAY(pbi->h_event_start_decoding, pbi->decoding_thread_count);
611 CALLOC_ARRAY_ALIGNED(pbi->mb_row_di, pbi->decoding_thread_count, 32);
612 CALLOC_ARRAY(pbi->de_thread_data, pbi->decoding_thread_count);
613
614 if (sem_init(&pbi->h_event_end_decoding, 0, 0)) {
615 vpx_internal_error(&pbi->common.error, VPX_CODEC_MEM_ERROR,
616 "Failed to initialize semaphore");
617 }
618
619 for (ithread = 0; ithread < pbi->decoding_thread_count; ++ithread) {
620 if (sem_init(&pbi->h_event_start_decoding[ithread], 0, 0)) break;
621
622 vp8_setup_block_dptrs(&pbi->mb_row_di[ithread].mbd);
623
624 pbi->de_thread_data[ithread].ithread = ithread;
625 pbi->de_thread_data[ithread].ptr1 = (void *)pbi;
626 pbi->de_thread_data[ithread].ptr2 = (void *)&pbi->mb_row_di[ithread];
627
628 if (pthread_create(&pbi->h_decoding_thread[ithread], 0,
629 thread_decoding_proc, &pbi->de_thread_data[ithread])) {
630 sem_destroy(&pbi->h_event_start_decoding[ithread]);
631 break;
632 }
633 }
634
635 pbi->allocated_decoding_thread_count = ithread;
636 if (pbi->allocated_decoding_thread_count !=
637 (int)pbi->decoding_thread_count) {
638 /* the remainder of cleanup cases will be handled in
639 * vp8_decoder_remove_threads(). */
640 if (pbi->allocated_decoding_thread_count == 0) {
641 sem_destroy(&pbi->h_event_end_decoding);
642 }
643 vpx_internal_error(&pbi->common.error, VPX_CODEC_MEM_ERROR,
644 "Failed to create threads");
645 }
646 }
647 }
648
vp8mt_de_alloc_temp_buffers(VP8D_COMP * pbi,int mb_rows)649 void vp8mt_de_alloc_temp_buffers(VP8D_COMP *pbi, int mb_rows) {
650 int i;
651
652 /* De-allocate mutex */
653 if (pbi->pmutex != NULL) {
654 for (i = 0; i < mb_rows; ++i) {
655 pthread_mutex_destroy(&pbi->pmutex[i]);
656 }
657
658 vpx_free(pbi->pmutex);
659 pbi->pmutex = NULL;
660 }
661
662 vpx_free(pbi->mt_current_mb_col);
663 pbi->mt_current_mb_col = NULL;
664
665 /* Free above_row buffers. */
666 if (pbi->mt_yabove_row) {
667 for (i = 0; i < mb_rows; ++i) {
668 vpx_free(pbi->mt_yabove_row[i]);
669 pbi->mt_yabove_row[i] = NULL;
670 }
671 vpx_free(pbi->mt_yabove_row);
672 pbi->mt_yabove_row = NULL;
673 }
674
675 if (pbi->mt_uabove_row) {
676 for (i = 0; i < mb_rows; ++i) {
677 vpx_free(pbi->mt_uabove_row[i]);
678 pbi->mt_uabove_row[i] = NULL;
679 }
680 vpx_free(pbi->mt_uabove_row);
681 pbi->mt_uabove_row = NULL;
682 }
683
684 if (pbi->mt_vabove_row) {
685 for (i = 0; i < mb_rows; ++i) {
686 vpx_free(pbi->mt_vabove_row[i]);
687 pbi->mt_vabove_row[i] = NULL;
688 }
689 vpx_free(pbi->mt_vabove_row);
690 pbi->mt_vabove_row = NULL;
691 }
692
693 /* Free left_col buffers. */
694 if (pbi->mt_yleft_col) {
695 for (i = 0; i < mb_rows; ++i) {
696 vpx_free(pbi->mt_yleft_col[i]);
697 pbi->mt_yleft_col[i] = NULL;
698 }
699 vpx_free(pbi->mt_yleft_col);
700 pbi->mt_yleft_col = NULL;
701 }
702
703 if (pbi->mt_uleft_col) {
704 for (i = 0; i < mb_rows; ++i) {
705 vpx_free(pbi->mt_uleft_col[i]);
706 pbi->mt_uleft_col[i] = NULL;
707 }
708 vpx_free(pbi->mt_uleft_col);
709 pbi->mt_uleft_col = NULL;
710 }
711
712 if (pbi->mt_vleft_col) {
713 for (i = 0; i < mb_rows; ++i) {
714 vpx_free(pbi->mt_vleft_col[i]);
715 pbi->mt_vleft_col[i] = NULL;
716 }
717 vpx_free(pbi->mt_vleft_col);
718 pbi->mt_vleft_col = NULL;
719 }
720 }
721
vp8mt_alloc_temp_buffers(VP8D_COMP * pbi,int width,int prev_mb_rows)722 void vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows) {
723 VP8_COMMON *const pc = &pbi->common;
724 int i;
725 int uv_width;
726
727 if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd)) {
728 vp8mt_de_alloc_temp_buffers(pbi, prev_mb_rows);
729
730 /* our internal buffers are always multiples of 16 */
731 if ((width & 0xf) != 0) width += 16 - (width & 0xf);
732
733 if (width < 640) {
734 pbi->sync_range = 1;
735 } else if (width <= 1280) {
736 pbi->sync_range = 8;
737 } else if (width <= 2560) {
738 pbi->sync_range = 16;
739 } else {
740 pbi->sync_range = 32;
741 }
742
743 uv_width = width >> 1;
744
745 /* Allocate mutex */
746 CHECK_MEM_ERROR(pbi->pmutex,
747 vpx_malloc(sizeof(*pbi->pmutex) * pc->mb_rows));
748 if (pbi->pmutex) {
749 for (i = 0; i < pc->mb_rows; ++i) {
750 pthread_mutex_init(&pbi->pmutex[i], NULL);
751 }
752 }
753
754 /* Allocate an int for each mb row. */
755 CALLOC_ARRAY(pbi->mt_current_mb_col, pc->mb_rows);
756
757 /* Allocate memory for above_row buffers. */
758 CALLOC_ARRAY(pbi->mt_yabove_row, pc->mb_rows);
759 for (i = 0; i < pc->mb_rows; ++i)
760 CHECK_MEM_ERROR(pbi->mt_yabove_row[i],
761 vpx_memalign(16, sizeof(unsigned char) *
762 (width + (VP8BORDERINPIXELS << 1))));
763
764 CALLOC_ARRAY(pbi->mt_uabove_row, pc->mb_rows);
765 for (i = 0; i < pc->mb_rows; ++i)
766 CHECK_MEM_ERROR(pbi->mt_uabove_row[i],
767 vpx_memalign(16, sizeof(unsigned char) *
768 (uv_width + VP8BORDERINPIXELS)));
769
770 CALLOC_ARRAY(pbi->mt_vabove_row, pc->mb_rows);
771 for (i = 0; i < pc->mb_rows; ++i)
772 CHECK_MEM_ERROR(pbi->mt_vabove_row[i],
773 vpx_memalign(16, sizeof(unsigned char) *
774 (uv_width + VP8BORDERINPIXELS)));
775
776 /* Allocate memory for left_col buffers. */
777 CALLOC_ARRAY(pbi->mt_yleft_col, pc->mb_rows);
778 for (i = 0; i < pc->mb_rows; ++i)
779 CHECK_MEM_ERROR(pbi->mt_yleft_col[i],
780 vpx_calloc(sizeof(unsigned char) * 16, 1));
781
782 CALLOC_ARRAY(pbi->mt_uleft_col, pc->mb_rows);
783 for (i = 0; i < pc->mb_rows; ++i)
784 CHECK_MEM_ERROR(pbi->mt_uleft_col[i],
785 vpx_calloc(sizeof(unsigned char) * 8, 1));
786
787 CALLOC_ARRAY(pbi->mt_vleft_col, pc->mb_rows);
788 for (i = 0; i < pc->mb_rows; ++i)
789 CHECK_MEM_ERROR(pbi->mt_vleft_col[i],
790 vpx_calloc(sizeof(unsigned char) * 8, 1));
791 }
792 }
793
vp8_decoder_remove_threads(VP8D_COMP * pbi)794 void vp8_decoder_remove_threads(VP8D_COMP *pbi) {
795 /* shutdown MB Decoding thread; */
796 if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd)) {
797 int i;
798 protected_write(&pbi->mt_mutex, &pbi->b_multithreaded_rd, 0);
799
800 /* allow all threads to exit */
801 for (i = 0; i < pbi->allocated_decoding_thread_count; ++i) {
802 sem_post(&pbi->h_event_start_decoding[i]);
803 pthread_join(pbi->h_decoding_thread[i], NULL);
804 }
805
806 for (i = 0; i < pbi->allocated_decoding_thread_count; ++i) {
807 sem_destroy(&pbi->h_event_start_decoding[i]);
808 }
809
810 if (pbi->allocated_decoding_thread_count) {
811 sem_destroy(&pbi->h_event_end_decoding);
812 }
813
814 vpx_free(pbi->h_decoding_thread);
815 pbi->h_decoding_thread = NULL;
816
817 vpx_free(pbi->h_event_start_decoding);
818 pbi->h_event_start_decoding = NULL;
819
820 vpx_free(pbi->mb_row_di);
821 pbi->mb_row_di = NULL;
822
823 vpx_free(pbi->de_thread_data);
824 pbi->de_thread_data = NULL;
825
826 vp8mt_de_alloc_temp_buffers(pbi, pbi->common.mb_rows);
827 }
828 pthread_mutex_destroy(&pbi->mt_mutex);
829 }
830
vp8mt_decode_mb_rows(VP8D_COMP * pbi,MACROBLOCKD * xd)831 void vp8mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd) {
832 VP8_COMMON *pc = &pbi->common;
833 unsigned int i;
834 int j;
835
836 int filter_level = pc->filter_level;
837 YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
838
839 if (filter_level) {
840 /* Set above_row buffer to 127 for decoding first MB row */
841 memset(pbi->mt_yabove_row[0] + VP8BORDERINPIXELS - 1, 127,
842 yv12_fb_new->y_width + 5);
843 memset(pbi->mt_uabove_row[0] + (VP8BORDERINPIXELS >> 1) - 1, 127,
844 (yv12_fb_new->y_width >> 1) + 5);
845 memset(pbi->mt_vabove_row[0] + (VP8BORDERINPIXELS >> 1) - 1, 127,
846 (yv12_fb_new->y_width >> 1) + 5);
847
848 for (j = 1; j < pc->mb_rows; ++j) {
849 memset(pbi->mt_yabove_row[j] + VP8BORDERINPIXELS - 1, (unsigned char)129,
850 1);
851 memset(pbi->mt_uabove_row[j] + (VP8BORDERINPIXELS >> 1) - 1,
852 (unsigned char)129, 1);
853 memset(pbi->mt_vabove_row[j] + (VP8BORDERINPIXELS >> 1) - 1,
854 (unsigned char)129, 1);
855 }
856
857 /* Set left_col to 129 initially */
858 for (j = 0; j < pc->mb_rows; ++j) {
859 memset(pbi->mt_yleft_col[j], (unsigned char)129, 16);
860 memset(pbi->mt_uleft_col[j], (unsigned char)129, 8);
861 memset(pbi->mt_vleft_col[j], (unsigned char)129, 8);
862 }
863
864 /* Initialize the loop filter for this frame. */
865 vp8_loop_filter_frame_init(pc, &pbi->mb, filter_level);
866 } else {
867 vp8_setup_intra_recon_top_line(yv12_fb_new);
868 }
869
870 setup_decoding_thread_data(pbi, xd, pbi->mb_row_di,
871 pbi->decoding_thread_count);
872
873 for (i = 0; i < pbi->decoding_thread_count; ++i) {
874 sem_post(&pbi->h_event_start_decoding[i]);
875 }
876
877 mt_decode_mb_rows(pbi, xd, 0);
878
879 sem_wait(&pbi->h_event_end_decoding); /* add back for each frame */
880 }
881