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 <errno.h>
12
13 #include "vpx_config.h"
14 #include "vp8_rtcd.h"
15 #if !defined(_WIN32) && CONFIG_OS_SUPPORT == 1
16 #include <unistd.h>
17 #endif
18 #include "onyxd_int.h"
19 #include "vpx_mem/vpx_mem.h"
20 #include "vp8/common/common.h"
21 #include "vp8/common/threading.h"
22 #include "vp8/common/loopfilter.h"
23 #include "vp8/common/extend.h"
24 #include "vpx_ports/vpx_timer.h"
25 #include "decoderthreading.h"
26 #include "detokenize.h"
27 #include "vp8/common/reconintra4x4.h"
28 #include "vp8/common/reconinter.h"
29 #include "vp8/common/reconintra.h"
30 #include "vp8/common/setupintrarecon.h"
31 #if CONFIG_ERROR_CONCEALMENT
32 #include "error_concealment.h"
33 #endif
34
35 #define CALLOC_ARRAY(p, n) \
36 CHECK_MEM_ERROR(&pbi->common.error, (p), vpx_calloc(sizeof(*(p)), (n)))
37 #define CALLOC_ARRAY_ALIGNED(p, n, algn) \
38 do { \
39 CHECK_MEM_ERROR(&pbi->common.error, (p), \
40 vpx_memalign((algn), sizeof(*(p)) * (n))); \
41 memset((p), 0, (n) * sizeof(*(p))); \
42 } while (0)
43
setup_decoding_thread_data(VP8D_COMP * pbi,MACROBLOCKD * xd,MB_ROW_DEC * mbrd,int count)44 static void setup_decoding_thread_data(VP8D_COMP *pbi, MACROBLOCKD *xd,
45 MB_ROW_DEC *mbrd, int count) {
46 VP8_COMMON *const pc = &pbi->common;
47 int i;
48
49 for (i = 0; i < count; ++i) {
50 MACROBLOCKD *mbd = &mbrd[i].mbd;
51 mbd->subpixel_predict = xd->subpixel_predict;
52 mbd->subpixel_predict8x4 = xd->subpixel_predict8x4;
53 mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
54 mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
55
56 mbd->frame_type = pc->frame_type;
57 mbd->pre = xd->pre;
58 mbd->dst = xd->dst;
59
60 mbd->segmentation_enabled = xd->segmentation_enabled;
61 mbd->mb_segment_abs_delta = xd->mb_segment_abs_delta;
62 memcpy(mbd->segment_feature_data, xd->segment_feature_data,
63 sizeof(xd->segment_feature_data));
64
65 /*signed char ref_lf_deltas[MAX_REF_LF_DELTAS];*/
66 memcpy(mbd->ref_lf_deltas, xd->ref_lf_deltas, sizeof(xd->ref_lf_deltas));
67 /*signed char mode_lf_deltas[MAX_MODE_LF_DELTAS];*/
68 memcpy(mbd->mode_lf_deltas, xd->mode_lf_deltas, sizeof(xd->mode_lf_deltas));
69 /*unsigned char mode_ref_lf_delta_enabled;
70 unsigned char mode_ref_lf_delta_update;*/
71 mbd->mode_ref_lf_delta_enabled = xd->mode_ref_lf_delta_enabled;
72 mbd->mode_ref_lf_delta_update = xd->mode_ref_lf_delta_update;
73
74 mbd->current_bc = &pbi->mbc[0];
75
76 memcpy(mbd->dequant_y1_dc, xd->dequant_y1_dc, sizeof(xd->dequant_y1_dc));
77 memcpy(mbd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
78 memcpy(mbd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
79 memcpy(mbd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
80
81 mbd->fullpixel_mask = ~0;
82
83 if (pc->full_pixel) mbd->fullpixel_mask = ~7;
84 }
85
86 for (i = 0; i < pc->mb_rows; ++i)
87 vpx_atomic_store_release(&pbi->mt_current_mb_col[i], -1);
88 }
89
mt_decode_macroblock(VP8D_COMP * pbi,MACROBLOCKD * xd,unsigned int mb_idx)90 static void mt_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd,
91 unsigned int mb_idx) {
92 MB_PREDICTION_MODE mode;
93 int i;
94 #if CONFIG_ERROR_CONCEALMENT
95 int corruption_detected = 0;
96 #else
97 (void)mb_idx;
98 #endif
99
100 if (xd->mode_info_context->mbmi.mb_skip_coeff) {
101 vp8_reset_mb_tokens_context(xd);
102 } else if (!vp8dx_bool_error(xd->current_bc)) {
103 int eobtotal;
104 eobtotal = vp8_decode_mb_tokens(pbi, xd);
105
106 /* Special case: Force the loopfilter to skip when eobtotal is zero */
107 xd->mode_info_context->mbmi.mb_skip_coeff = (eobtotal == 0);
108 }
109
110 mode = xd->mode_info_context->mbmi.mode;
111
112 if (xd->segmentation_enabled) vp8_mb_init_dequantizer(pbi, xd);
113
114 #if CONFIG_ERROR_CONCEALMENT
115
116 if (pbi->ec_active) {
117 int throw_residual;
118 /* When we have independent partitions we can apply residual even
119 * though other partitions within the frame are corrupt.
120 */
121 throw_residual =
122 (!pbi->independent_partitions && pbi->frame_corrupt_residual);
123 throw_residual = (throw_residual || vp8dx_bool_error(xd->current_bc));
124
125 if ((mb_idx >= pbi->mvs_corrupt_from_mb || throw_residual)) {
126 /* MB with corrupt residuals or corrupt mode/motion vectors.
127 * Better to use the predictor as reconstruction.
128 */
129 pbi->frame_corrupt_residual = 1;
130 memset(xd->qcoeff, 0, sizeof(xd->qcoeff));
131
132 corruption_detected = 1;
133
134 /* force idct to be skipped for B_PRED and use the
135 * prediction only for reconstruction
136 * */
137 memset(xd->eobs, 0, 25);
138 }
139 }
140 #endif
141
142 /* do prediction */
143 if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) {
144 vp8_build_intra_predictors_mbuv_s(
145 xd, xd->recon_above[1], xd->recon_above[2], xd->recon_left[1],
146 xd->recon_left[2], xd->recon_left_stride[1], xd->dst.u_buffer,
147 xd->dst.v_buffer, xd->dst.uv_stride);
148
149 if (mode != B_PRED) {
150 vp8_build_intra_predictors_mby_s(
151 xd, xd->recon_above[0], xd->recon_left[0], xd->recon_left_stride[0],
152 xd->dst.y_buffer, xd->dst.y_stride);
153 } else {
154 short *DQC = xd->dequant_y1;
155 int dst_stride = xd->dst.y_stride;
156
157 /* clear out residual eob info */
158 if (xd->mode_info_context->mbmi.mb_skip_coeff) memset(xd->eobs, 0, 25);
159
160 intra_prediction_down_copy(xd, xd->recon_above[0] + 16);
161
162 for (i = 0; i < 16; ++i) {
163 BLOCKD *b = &xd->block[i];
164 unsigned char *dst = xd->dst.y_buffer + b->offset;
165 B_PREDICTION_MODE b_mode = xd->mode_info_context->bmi[i].as_mode;
166 unsigned char *Above;
167 unsigned char *yleft;
168 int left_stride;
169 unsigned char top_left;
170
171 /*Caution: For some b_mode, it needs 8 pixels (4 above + 4
172 * above-right).*/
173 if (i < 4 && pbi->common.filter_level) {
174 Above = xd->recon_above[0] + b->offset;
175 } else {
176 Above = dst - dst_stride;
177 }
178
179 if (i % 4 == 0 && pbi->common.filter_level) {
180 yleft = xd->recon_left[0] + i;
181 left_stride = 1;
182 } else {
183 yleft = dst - 1;
184 left_stride = dst_stride;
185 }
186
187 if ((i == 4 || i == 8 || i == 12) && pbi->common.filter_level) {
188 top_left = *(xd->recon_left[0] + i - 1);
189 } else {
190 top_left = Above[-1];
191 }
192
193 vp8_intra4x4_predict(Above, yleft, left_stride, b_mode, dst, dst_stride,
194 top_left);
195
196 if (xd->eobs[i]) {
197 if (xd->eobs[i] > 1) {
198 vp8_dequant_idct_add(b->qcoeff, DQC, dst, dst_stride);
199 } else {
200 vp8_dc_only_idct_add(b->qcoeff[0] * DQC[0], dst, dst_stride, dst,
201 dst_stride);
202 memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
203 }
204 }
205 }
206 }
207 } else {
208 vp8_build_inter_predictors_mb(xd);
209 }
210
211 #if CONFIG_ERROR_CONCEALMENT
212 if (corruption_detected) {
213 return;
214 }
215 #endif
216
217 if (!xd->mode_info_context->mbmi.mb_skip_coeff) {
218 /* dequantization and idct */
219 if (mode != B_PRED) {
220 short *DQC = xd->dequant_y1;
221
222 if (mode != SPLITMV) {
223 BLOCKD *b = &xd->block[24];
224
225 /* do 2nd order transform on the dc block */
226 if (xd->eobs[24] > 1) {
227 vp8_dequantize_b(b, xd->dequant_y2);
228
229 vp8_short_inv_walsh4x4(&b->dqcoeff[0], xd->qcoeff);
230 memset(b->qcoeff, 0, 16 * sizeof(b->qcoeff[0]));
231 } else {
232 b->dqcoeff[0] = b->qcoeff[0] * xd->dequant_y2[0];
233 vp8_short_inv_walsh4x4_1(&b->dqcoeff[0], xd->qcoeff);
234 memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
235 }
236
237 /* override the dc dequant constant in order to preserve the
238 * dc components
239 */
240 DQC = xd->dequant_y1_dc;
241 }
242
243 vp8_dequant_idct_add_y_block(xd->qcoeff, DQC, xd->dst.y_buffer,
244 xd->dst.y_stride, xd->eobs);
245 }
246
247 vp8_dequant_idct_add_uv_block(xd->qcoeff + 16 * 16, xd->dequant_uv,
248 xd->dst.u_buffer, xd->dst.v_buffer,
249 xd->dst.uv_stride, xd->eobs + 16);
250 }
251 }
252
mt_decode_mb_rows(VP8D_COMP * pbi,MACROBLOCKD * xd,int start_mb_row)253 static void mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd,
254 int start_mb_row) {
255 const vpx_atomic_int *last_row_current_mb_col;
256 vpx_atomic_int *current_mb_col;
257 int mb_row;
258 VP8_COMMON *pc = &pbi->common;
259 const int nsync = pbi->sync_range;
260 const vpx_atomic_int first_row_no_sync_above =
261 VPX_ATOMIC_INIT(pc->mb_cols + nsync);
262 int num_part = 1 << pbi->common.multi_token_partition;
263 int last_mb_row = start_mb_row;
264
265 YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
266 YV12_BUFFER_CONFIG *yv12_fb_lst = pbi->dec_fb_ref[LAST_FRAME];
267
268 int recon_y_stride = yv12_fb_new->y_stride;
269 int recon_uv_stride = yv12_fb_new->uv_stride;
270
271 unsigned char *ref_buffer[MAX_REF_FRAMES][3];
272 unsigned char *dst_buffer[3];
273 int i;
274 int ref_fb_corrupted[MAX_REF_FRAMES];
275
276 ref_fb_corrupted[INTRA_FRAME] = 0;
277
278 for (i = 1; i < MAX_REF_FRAMES; ++i) {
279 YV12_BUFFER_CONFIG *this_fb = pbi->dec_fb_ref[i];
280
281 ref_buffer[i][0] = this_fb->y_buffer;
282 ref_buffer[i][1] = this_fb->u_buffer;
283 ref_buffer[i][2] = this_fb->v_buffer;
284
285 ref_fb_corrupted[i] = this_fb->corrupted;
286 }
287
288 dst_buffer[0] = yv12_fb_new->y_buffer;
289 dst_buffer[1] = yv12_fb_new->u_buffer;
290 dst_buffer[2] = yv12_fb_new->v_buffer;
291
292 xd->up_available = (start_mb_row != 0);
293
294 xd->mode_info_context = pc->mi + pc->mode_info_stride * start_mb_row;
295 xd->mode_info_stride = pc->mode_info_stride;
296
297 for (mb_row = start_mb_row; mb_row < pc->mb_rows;
298 mb_row += (pbi->decoding_thread_count + 1)) {
299 int recon_yoffset, recon_uvoffset;
300 int mb_col;
301 int filter_level;
302 loop_filter_info_n *lfi_n = &pc->lf_info;
303
304 /* save last row processed by this thread */
305 last_mb_row = mb_row;
306 /* select bool coder for current partition */
307 xd->current_bc = &pbi->mbc[mb_row % num_part];
308
309 if (mb_row > 0) {
310 last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row - 1];
311 } else {
312 last_row_current_mb_col = &first_row_no_sync_above;
313 }
314
315 current_mb_col = &pbi->mt_current_mb_col[mb_row];
316
317 recon_yoffset = mb_row * recon_y_stride * 16;
318 recon_uvoffset = mb_row * recon_uv_stride * 8;
319
320 /* reset contexts */
321 xd->above_context = pc->above_context;
322 memset(xd->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));
323
324 xd->left_available = 0;
325
326 xd->mb_to_top_edge = -((mb_row * 16) << 3);
327 xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
328
329 if (pbi->common.filter_level) {
330 xd->recon_above[0] = pbi->mt_yabove_row[mb_row] + 0 * 16 + 32;
331 xd->recon_above[1] = pbi->mt_uabove_row[mb_row] + 0 * 8 + 16;
332 xd->recon_above[2] = pbi->mt_vabove_row[mb_row] + 0 * 8 + 16;
333
334 xd->recon_left[0] = pbi->mt_yleft_col[mb_row];
335 xd->recon_left[1] = pbi->mt_uleft_col[mb_row];
336 xd->recon_left[2] = pbi->mt_vleft_col[mb_row];
337
338 /* TODO: move to outside row loop */
339 xd->recon_left_stride[0] = 1;
340 xd->recon_left_stride[1] = 1;
341 } else {
342 xd->recon_above[0] = dst_buffer[0] + recon_yoffset;
343 xd->recon_above[1] = dst_buffer[1] + recon_uvoffset;
344 xd->recon_above[2] = dst_buffer[2] + recon_uvoffset;
345
346 xd->recon_left[0] = xd->recon_above[0] - 1;
347 xd->recon_left[1] = xd->recon_above[1] - 1;
348 xd->recon_left[2] = xd->recon_above[2] - 1;
349
350 xd->recon_above[0] -= xd->dst.y_stride;
351 xd->recon_above[1] -= xd->dst.uv_stride;
352 xd->recon_above[2] -= xd->dst.uv_stride;
353
354 /* TODO: move to outside row loop */
355 xd->recon_left_stride[0] = xd->dst.y_stride;
356 xd->recon_left_stride[1] = xd->dst.uv_stride;
357
358 setup_intra_recon_left(xd->recon_left[0], xd->recon_left[1],
359 xd->recon_left[2], xd->dst.y_stride,
360 xd->dst.uv_stride);
361 }
362
363 for (mb_col = 0; mb_col < pc->mb_cols; ++mb_col) {
364 if (((mb_col - 1) % nsync) == 0) {
365 vpx_atomic_store_release(current_mb_col, mb_col - 1);
366 }
367
368 if (mb_row && !(mb_col & (nsync - 1))) {
369 vp8_atomic_spin_wait(mb_col, last_row_current_mb_col, nsync);
370 }
371
372 /* Distance of MB to the various image edges.
373 * These are specified to 8th pel as they are always
374 * compared to values that are in 1/8th pel units.
375 */
376 xd->mb_to_left_edge = -((mb_col * 16) << 3);
377 xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
378
379 #if CONFIG_ERROR_CONCEALMENT
380 {
381 int corrupt_residual =
382 (!pbi->independent_partitions && pbi->frame_corrupt_residual) ||
383 vp8dx_bool_error(xd->current_bc);
384 if (pbi->ec_active &&
385 (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) &&
386 corrupt_residual) {
387 /* We have an intra block with corrupt
388 * coefficients, better to conceal with an inter
389 * block.
390 * Interpolate MVs from neighboring MBs
391 *
392 * Note that for the first mb with corrupt
393 * residual in a frame, we might not discover
394 * that before decoding the residual. That
395 * happens after this check, and therefore no
396 * inter concealment will be done.
397 */
398 vp8_interpolate_motion(xd, mb_row, mb_col, pc->mb_rows, pc->mb_cols);
399 }
400 }
401 #endif
402
403 xd->dst.y_buffer = dst_buffer[0] + recon_yoffset;
404 xd->dst.u_buffer = dst_buffer[1] + recon_uvoffset;
405 xd->dst.v_buffer = dst_buffer[2] + recon_uvoffset;
406
407 /* propagate errors from reference frames */
408 xd->corrupted |= ref_fb_corrupted[xd->mode_info_context->mbmi.ref_frame];
409
410 if (xd->corrupted) {
411 // Move current decoding marcoblock to the end of row for all rows
412 // assigned to this thread, such that other threads won't be waiting.
413 for (; mb_row < pc->mb_rows;
414 mb_row += (pbi->decoding_thread_count + 1)) {
415 current_mb_col = &pbi->mt_current_mb_col[mb_row];
416 vpx_atomic_store_release(current_mb_col, pc->mb_cols + nsync);
417 }
418 vpx_internal_error(&xd->error_info, VPX_CODEC_CORRUPT_FRAME,
419 "Corrupted reference frame");
420 }
421
422 if (xd->mode_info_context->mbmi.ref_frame >= LAST_FRAME) {
423 const MV_REFERENCE_FRAME ref = xd->mode_info_context->mbmi.ref_frame;
424 xd->pre.y_buffer = ref_buffer[ref][0] + recon_yoffset;
425 xd->pre.u_buffer = ref_buffer[ref][1] + recon_uvoffset;
426 xd->pre.v_buffer = ref_buffer[ref][2] + recon_uvoffset;
427 } else {
428 // ref_frame is INTRA_FRAME, pre buffer should not be used.
429 xd->pre.y_buffer = 0;
430 xd->pre.u_buffer = 0;
431 xd->pre.v_buffer = 0;
432 }
433 mt_decode_macroblock(pbi, xd, 0);
434
435 xd->left_available = 1;
436
437 /* check if the boolean decoder has suffered an error */
438 xd->corrupted |= vp8dx_bool_error(xd->current_bc);
439
440 xd->recon_above[0] += 16;
441 xd->recon_above[1] += 8;
442 xd->recon_above[2] += 8;
443
444 if (!pbi->common.filter_level) {
445 xd->recon_left[0] += 16;
446 xd->recon_left[1] += 8;
447 xd->recon_left[2] += 8;
448 }
449
450 if (pbi->common.filter_level) {
451 int skip_lf = (xd->mode_info_context->mbmi.mode != B_PRED &&
452 xd->mode_info_context->mbmi.mode != SPLITMV &&
453 xd->mode_info_context->mbmi.mb_skip_coeff);
454
455 const int mode_index =
456 lfi_n->mode_lf_lut[xd->mode_info_context->mbmi.mode];
457 const int seg = xd->mode_info_context->mbmi.segment_id;
458 const int ref_frame = xd->mode_info_context->mbmi.ref_frame;
459
460 filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
461
462 if (mb_row != pc->mb_rows - 1) {
463 /* Save decoded MB last row data for next-row decoding */
464 memcpy((pbi->mt_yabove_row[mb_row + 1] + 32 + mb_col * 16),
465 (xd->dst.y_buffer + 15 * recon_y_stride), 16);
466 memcpy((pbi->mt_uabove_row[mb_row + 1] + 16 + mb_col * 8),
467 (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
468 memcpy((pbi->mt_vabove_row[mb_row + 1] + 16 + mb_col * 8),
469 (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
470 }
471
472 /* save left_col for next MB decoding */
473 if (mb_col != pc->mb_cols - 1) {
474 MODE_INFO *next = xd->mode_info_context + 1;
475
476 if (next->mbmi.ref_frame == INTRA_FRAME) {
477 for (i = 0; i < 16; ++i) {
478 pbi->mt_yleft_col[mb_row][i] =
479 xd->dst.y_buffer[i * recon_y_stride + 15];
480 }
481 for (i = 0; i < 8; ++i) {
482 pbi->mt_uleft_col[mb_row][i] =
483 xd->dst.u_buffer[i * recon_uv_stride + 7];
484 pbi->mt_vleft_col[mb_row][i] =
485 xd->dst.v_buffer[i * recon_uv_stride + 7];
486 }
487 }
488 }
489
490 /* loopfilter on this macroblock. */
491 if (filter_level) {
492 if (pc->filter_type == NORMAL_LOOPFILTER) {
493 loop_filter_info lfi;
494 FRAME_TYPE frame_type = pc->frame_type;
495 const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
496 lfi.mblim = lfi_n->mblim[filter_level];
497 lfi.blim = lfi_n->blim[filter_level];
498 lfi.lim = lfi_n->lim[filter_level];
499 lfi.hev_thr = lfi_n->hev_thr[hev_index];
500
501 if (mb_col > 0)
502 vp8_loop_filter_mbv(xd->dst.y_buffer, xd->dst.u_buffer,
503 xd->dst.v_buffer, recon_y_stride,
504 recon_uv_stride, &lfi);
505
506 if (!skip_lf)
507 vp8_loop_filter_bv(xd->dst.y_buffer, xd->dst.u_buffer,
508 xd->dst.v_buffer, recon_y_stride,
509 recon_uv_stride, &lfi);
510
511 /* don't apply across umv border */
512 if (mb_row > 0)
513 vp8_loop_filter_mbh(xd->dst.y_buffer, xd->dst.u_buffer,
514 xd->dst.v_buffer, recon_y_stride,
515 recon_uv_stride, &lfi);
516
517 if (!skip_lf)
518 vp8_loop_filter_bh(xd->dst.y_buffer, xd->dst.u_buffer,
519 xd->dst.v_buffer, recon_y_stride,
520 recon_uv_stride, &lfi);
521 } else {
522 if (mb_col > 0)
523 vp8_loop_filter_simple_mbv(xd->dst.y_buffer, recon_y_stride,
524 lfi_n->mblim[filter_level]);
525
526 if (!skip_lf)
527 vp8_loop_filter_simple_bv(xd->dst.y_buffer, recon_y_stride,
528 lfi_n->blim[filter_level]);
529
530 /* don't apply across umv border */
531 if (mb_row > 0)
532 vp8_loop_filter_simple_mbh(xd->dst.y_buffer, recon_y_stride,
533 lfi_n->mblim[filter_level]);
534
535 if (!skip_lf)
536 vp8_loop_filter_simple_bh(xd->dst.y_buffer, recon_y_stride,
537 lfi_n->blim[filter_level]);
538 }
539 }
540 }
541
542 recon_yoffset += 16;
543 recon_uvoffset += 8;
544
545 ++xd->mode_info_context; /* next mb */
546
547 xd->above_context++;
548 }
549
550 /* adjust to the next row of mbs */
551 if (pbi->common.filter_level) {
552 if (mb_row != pc->mb_rows - 1) {
553 int lasty = yv12_fb_lst->y_width + VP8BORDERINPIXELS;
554 int lastuv = (yv12_fb_lst->y_width >> 1) + (VP8BORDERINPIXELS >> 1);
555
556 for (i = 0; i < 4; ++i) {
557 pbi->mt_yabove_row[mb_row + 1][lasty + i] =
558 pbi->mt_yabove_row[mb_row + 1][lasty - 1];
559 pbi->mt_uabove_row[mb_row + 1][lastuv + i] =
560 pbi->mt_uabove_row[mb_row + 1][lastuv - 1];
561 pbi->mt_vabove_row[mb_row + 1][lastuv + i] =
562 pbi->mt_vabove_row[mb_row + 1][lastuv - 1];
563 }
564 }
565 } else {
566 vp8_extend_mb_row(yv12_fb_new, xd->dst.y_buffer + 16,
567 xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
568 }
569
570 /* last MB of row is ready just after extension is done */
571 vpx_atomic_store_release(current_mb_col, mb_col + nsync);
572
573 ++xd->mode_info_context; /* skip prediction column */
574 xd->up_available = 1;
575
576 /* since we have multithread */
577 xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
578 }
579
580 /* signal end of decoding of current thread for current frame */
581 if (last_mb_row + (int)pbi->decoding_thread_count + 1 >= pc->mb_rows)
582 sem_post(&pbi->h_event_end_decoding);
583 }
584
thread_decoding_proc(void * p_data)585 static THREAD_FUNCTION thread_decoding_proc(void *p_data) {
586 int ithread = ((DECODETHREAD_DATA *)p_data)->ithread;
587 VP8D_COMP *pbi = (VP8D_COMP *)(((DECODETHREAD_DATA *)p_data)->ptr1);
588 MB_ROW_DEC *mbrd = (MB_ROW_DEC *)(((DECODETHREAD_DATA *)p_data)->ptr2);
589 ENTROPY_CONTEXT_PLANES mb_row_left_context;
590
591 while (1) {
592 if (vpx_atomic_load_acquire(&pbi->b_multithreaded_rd) == 0) break;
593
594 if (sem_wait(&pbi->h_event_start_decoding[ithread]) == 0) {
595 if (vpx_atomic_load_acquire(&pbi->b_multithreaded_rd) == 0) {
596 break;
597 } else {
598 MACROBLOCKD *xd = &mbrd->mbd;
599 xd->left_context = &mb_row_left_context;
600 if (setjmp(xd->error_info.jmp)) {
601 xd->error_info.setjmp = 0;
602 // Signal the end of decoding for current thread.
603 sem_post(&pbi->h_event_end_decoding);
604 continue;
605 }
606 xd->error_info.setjmp = 1;
607 mt_decode_mb_rows(pbi, xd, ithread + 1);
608 }
609 }
610 }
611
612 return 0;
613 }
614
vp8_decoder_create_threads(VP8D_COMP * pbi)615 void vp8_decoder_create_threads(VP8D_COMP *pbi) {
616 int core_count = 0;
617 unsigned int ithread;
618
619 vpx_atomic_init(&pbi->b_multithreaded_rd, 0);
620 pbi->allocated_decoding_thread_count = 0;
621
622 /* limit decoding threads to the max number of token partitions */
623 core_count = (pbi->max_threads > 8) ? 8 : pbi->max_threads;
624
625 /* limit decoding threads to the available cores */
626 if (core_count > pbi->common.processor_core_count) {
627 core_count = pbi->common.processor_core_count;
628 }
629
630 if (core_count > 1) {
631 vpx_atomic_init(&pbi->b_multithreaded_rd, 1);
632 pbi->decoding_thread_count = core_count - 1;
633
634 CALLOC_ARRAY(pbi->h_decoding_thread, pbi->decoding_thread_count);
635 CALLOC_ARRAY(pbi->h_event_start_decoding, pbi->decoding_thread_count);
636 CALLOC_ARRAY_ALIGNED(pbi->mb_row_di, pbi->decoding_thread_count, 32);
637 CALLOC_ARRAY(pbi->de_thread_data, pbi->decoding_thread_count);
638
639 if (sem_init(&pbi->h_event_end_decoding, 0, 0)) {
640 vpx_internal_error(&pbi->common.error, VPX_CODEC_MEM_ERROR,
641 "Failed to initialize semaphore");
642 }
643
644 for (ithread = 0; ithread < pbi->decoding_thread_count; ++ithread) {
645 if (sem_init(&pbi->h_event_start_decoding[ithread], 0, 0)) break;
646
647 vp8_setup_block_dptrs(&pbi->mb_row_di[ithread].mbd);
648
649 pbi->de_thread_data[ithread].ithread = ithread;
650 pbi->de_thread_data[ithread].ptr1 = (void *)pbi;
651 pbi->de_thread_data[ithread].ptr2 = (void *)&pbi->mb_row_di[ithread];
652
653 if (pthread_create(&pbi->h_decoding_thread[ithread], 0,
654 thread_decoding_proc, &pbi->de_thread_data[ithread])) {
655 sem_destroy(&pbi->h_event_start_decoding[ithread]);
656 break;
657 }
658 }
659
660 pbi->allocated_decoding_thread_count = ithread;
661 if (pbi->allocated_decoding_thread_count !=
662 (int)pbi->decoding_thread_count) {
663 /* the remainder of cleanup cases will be handled in
664 * vp8_decoder_remove_threads(). */
665 if (pbi->allocated_decoding_thread_count == 0) {
666 sem_destroy(&pbi->h_event_end_decoding);
667 }
668 vpx_internal_error(&pbi->common.error, VPX_CODEC_MEM_ERROR,
669 "Failed to create threads");
670 }
671 }
672 }
673
vp8mt_de_alloc_temp_buffers(VP8D_COMP * pbi,int mb_rows)674 void vp8mt_de_alloc_temp_buffers(VP8D_COMP *pbi, int mb_rows) {
675 int i;
676
677 vpx_free(pbi->mt_current_mb_col);
678 pbi->mt_current_mb_col = NULL;
679
680 /* Free above_row buffers. */
681 if (pbi->mt_yabove_row) {
682 for (i = 0; i < mb_rows; ++i) {
683 vpx_free(pbi->mt_yabove_row[i]);
684 pbi->mt_yabove_row[i] = NULL;
685 }
686 vpx_free(pbi->mt_yabove_row);
687 pbi->mt_yabove_row = NULL;
688 }
689
690 if (pbi->mt_uabove_row) {
691 for (i = 0; i < mb_rows; ++i) {
692 vpx_free(pbi->mt_uabove_row[i]);
693 pbi->mt_uabove_row[i] = NULL;
694 }
695 vpx_free(pbi->mt_uabove_row);
696 pbi->mt_uabove_row = NULL;
697 }
698
699 if (pbi->mt_vabove_row) {
700 for (i = 0; i < mb_rows; ++i) {
701 vpx_free(pbi->mt_vabove_row[i]);
702 pbi->mt_vabove_row[i] = NULL;
703 }
704 vpx_free(pbi->mt_vabove_row);
705 pbi->mt_vabove_row = NULL;
706 }
707
708 /* Free left_col buffers. */
709 if (pbi->mt_yleft_col) {
710 for (i = 0; i < mb_rows; ++i) {
711 vpx_free(pbi->mt_yleft_col[i]);
712 pbi->mt_yleft_col[i] = NULL;
713 }
714 vpx_free(pbi->mt_yleft_col);
715 pbi->mt_yleft_col = NULL;
716 }
717
718 if (pbi->mt_uleft_col) {
719 for (i = 0; i < mb_rows; ++i) {
720 vpx_free(pbi->mt_uleft_col[i]);
721 pbi->mt_uleft_col[i] = NULL;
722 }
723 vpx_free(pbi->mt_uleft_col);
724 pbi->mt_uleft_col = NULL;
725 }
726
727 if (pbi->mt_vleft_col) {
728 for (i = 0; i < mb_rows; ++i) {
729 vpx_free(pbi->mt_vleft_col[i]);
730 pbi->mt_vleft_col[i] = NULL;
731 }
732 vpx_free(pbi->mt_vleft_col);
733 pbi->mt_vleft_col = NULL;
734 }
735 }
736
vp8mt_alloc_temp_buffers(VP8D_COMP * pbi,int width,int prev_mb_rows)737 void vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows) {
738 VP8_COMMON *const pc = &pbi->common;
739 int i;
740 int uv_width;
741
742 if (vpx_atomic_load_acquire(&pbi->b_multithreaded_rd)) {
743 vp8mt_de_alloc_temp_buffers(pbi, prev_mb_rows);
744
745 /* our internal buffers are always multiples of 16 */
746 if ((width & 0xf) != 0) width += 16 - (width & 0xf);
747
748 if (width < 640) {
749 pbi->sync_range = 1;
750 } else if (width <= 1280) {
751 pbi->sync_range = 8;
752 } else if (width <= 2560) {
753 pbi->sync_range = 16;
754 } else {
755 pbi->sync_range = 32;
756 }
757
758 uv_width = width >> 1;
759
760 /* Allocate a vpx_atomic_int for each mb row. */
761 CHECK_MEM_ERROR(&pc->error, pbi->mt_current_mb_col,
762 vpx_malloc(sizeof(*pbi->mt_current_mb_col) * pc->mb_rows));
763 for (i = 0; i < pc->mb_rows; ++i)
764 vpx_atomic_init(&pbi->mt_current_mb_col[i], 0);
765
766 /* Allocate memory for above_row buffers. */
767 CALLOC_ARRAY(pbi->mt_yabove_row, pc->mb_rows);
768 for (i = 0; i < pc->mb_rows; ++i) {
769 CHECK_MEM_ERROR(&pc->error, pbi->mt_yabove_row[i],
770 vpx_memalign(16, sizeof(unsigned char) *
771 (width + (VP8BORDERINPIXELS << 1))));
772 vp8_zero_array(pbi->mt_yabove_row[i], width + (VP8BORDERINPIXELS << 1));
773 }
774
775 CALLOC_ARRAY(pbi->mt_uabove_row, pc->mb_rows);
776 for (i = 0; i < pc->mb_rows; ++i) {
777 CHECK_MEM_ERROR(&pc->error, pbi->mt_uabove_row[i],
778 vpx_memalign(16, sizeof(unsigned char) *
779 (uv_width + VP8BORDERINPIXELS)));
780 vp8_zero_array(pbi->mt_uabove_row[i], uv_width + VP8BORDERINPIXELS);
781 }
782
783 CALLOC_ARRAY(pbi->mt_vabove_row, pc->mb_rows);
784 for (i = 0; i < pc->mb_rows; ++i) {
785 CHECK_MEM_ERROR(&pc->error, pbi->mt_vabove_row[i],
786 vpx_memalign(16, sizeof(unsigned char) *
787 (uv_width + VP8BORDERINPIXELS)));
788 vp8_zero_array(pbi->mt_vabove_row[i], uv_width + VP8BORDERINPIXELS);
789 }
790
791 /* Allocate memory for left_col buffers. */
792 CALLOC_ARRAY(pbi->mt_yleft_col, pc->mb_rows);
793 for (i = 0; i < pc->mb_rows; ++i)
794 CHECK_MEM_ERROR(&pc->error, pbi->mt_yleft_col[i],
795 vpx_calloc(sizeof(unsigned char) * 16, 1));
796
797 CALLOC_ARRAY(pbi->mt_uleft_col, pc->mb_rows);
798 for (i = 0; i < pc->mb_rows; ++i)
799 CHECK_MEM_ERROR(&pc->error, pbi->mt_uleft_col[i],
800 vpx_calloc(sizeof(unsigned char) * 8, 1));
801
802 CALLOC_ARRAY(pbi->mt_vleft_col, pc->mb_rows);
803 for (i = 0; i < pc->mb_rows; ++i)
804 CHECK_MEM_ERROR(&pc->error, pbi->mt_vleft_col[i],
805 vpx_calloc(sizeof(unsigned char) * 8, 1));
806 }
807 }
808
vp8_decoder_remove_threads(VP8D_COMP * pbi)809 void vp8_decoder_remove_threads(VP8D_COMP *pbi) {
810 /* shutdown MB Decoding thread; */
811 if (vpx_atomic_load_acquire(&pbi->b_multithreaded_rd)) {
812 int i;
813 vpx_atomic_store_release(&pbi->b_multithreaded_rd, 0);
814
815 /* allow all threads to exit */
816 for (i = 0; i < pbi->allocated_decoding_thread_count; ++i) {
817 sem_post(&pbi->h_event_start_decoding[i]);
818 pthread_join(pbi->h_decoding_thread[i], NULL);
819 }
820
821 for (i = 0; i < pbi->allocated_decoding_thread_count; ++i) {
822 sem_destroy(&pbi->h_event_start_decoding[i]);
823 }
824
825 if (pbi->allocated_decoding_thread_count) {
826 sem_destroy(&pbi->h_event_end_decoding);
827 }
828
829 vpx_free(pbi->h_decoding_thread);
830 pbi->h_decoding_thread = NULL;
831
832 vpx_free(pbi->h_event_start_decoding);
833 pbi->h_event_start_decoding = NULL;
834
835 vpx_free(pbi->mb_row_di);
836 pbi->mb_row_di = NULL;
837
838 vpx_free(pbi->de_thread_data);
839 pbi->de_thread_data = NULL;
840
841 vp8mt_de_alloc_temp_buffers(pbi, pbi->common.mb_rows);
842 }
843 }
844
vp8mt_decode_mb_rows(VP8D_COMP * pbi,MACROBLOCKD * xd)845 int vp8mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd) {
846 VP8_COMMON *pc = &pbi->common;
847 unsigned int i;
848 int j;
849
850 int filter_level = pc->filter_level;
851 YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
852
853 if (filter_level) {
854 /* Set above_row buffer to 127 for decoding first MB row */
855 memset(pbi->mt_yabove_row[0] + VP8BORDERINPIXELS - 1, 127,
856 yv12_fb_new->y_width + 5);
857 memset(pbi->mt_uabove_row[0] + (VP8BORDERINPIXELS >> 1) - 1, 127,
858 (yv12_fb_new->y_width >> 1) + 5);
859 memset(pbi->mt_vabove_row[0] + (VP8BORDERINPIXELS >> 1) - 1, 127,
860 (yv12_fb_new->y_width >> 1) + 5);
861
862 for (j = 1; j < pc->mb_rows; ++j) {
863 memset(pbi->mt_yabove_row[j] + VP8BORDERINPIXELS - 1, (unsigned char)129,
864 1);
865 memset(pbi->mt_uabove_row[j] + (VP8BORDERINPIXELS >> 1) - 1,
866 (unsigned char)129, 1);
867 memset(pbi->mt_vabove_row[j] + (VP8BORDERINPIXELS >> 1) - 1,
868 (unsigned char)129, 1);
869 }
870
871 /* Set left_col to 129 initially */
872 for (j = 0; j < pc->mb_rows; ++j) {
873 memset(pbi->mt_yleft_col[j], (unsigned char)129, 16);
874 memset(pbi->mt_uleft_col[j], (unsigned char)129, 8);
875 memset(pbi->mt_vleft_col[j], (unsigned char)129, 8);
876 }
877
878 /* Initialize the loop filter for this frame. */
879 vp8_loop_filter_frame_init(pc, &pbi->mb, filter_level);
880 } else {
881 vp8_setup_intra_recon_top_line(yv12_fb_new);
882 }
883
884 setup_decoding_thread_data(pbi, xd, pbi->mb_row_di,
885 pbi->decoding_thread_count);
886
887 for (i = 0; i < pbi->decoding_thread_count; ++i) {
888 sem_post(&pbi->h_event_start_decoding[i]);
889 }
890
891 if (setjmp(xd->error_info.jmp)) {
892 xd->error_info.setjmp = 0;
893 xd->corrupted = 1;
894 // Wait for other threads to finish. This prevents other threads decoding
895 // the current frame while the main thread starts decoding the next frame,
896 // which causes a data race.
897 for (i = 0; i < pbi->decoding_thread_count; ++i) {
898 errno = 0;
899 while (sem_wait(&pbi->h_event_end_decoding) != 0 && errno == EINTR) {
900 }
901 }
902 return -1;
903 }
904
905 xd->error_info.setjmp = 1;
906 mt_decode_mb_rows(pbi, xd, 0);
907
908 for (i = 0; i < pbi->decoding_thread_count + 1; ++i) {
909 /* add back for each frame */
910 errno = 0;
911 while (sem_wait(&pbi->h_event_end_decoding) != 0 && errno == EINTR) {
912 }
913 }
914
915 return 0;
916 }
917