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 "vpx_util/vpx_pthread.h"
19 #include "vp8/common/common.h"
20 #include "vp8/common/threading.h"
21 #include "vp8/common/loopfilter.h"
22 #include "vp8/common/extend.h"
23 #include "vpx_ports/vpx_timer.h"
24 #include "decoderthreading.h"
25 #include "detokenize.h"
26 #include "vp8/common/reconintra4x4.h"
27 #include "vp8/common/reconinter.h"
28 #include "vp8/common/reconintra.h"
29 #include "vp8/common/setupintrarecon.h"
30 #if CONFIG_ERROR_CONCEALMENT
31 #include "error_concealment.h"
32 #endif
33
34 #define CALLOC_ARRAY(p, n) \
35 CHECK_MEM_ERROR(&pbi->common.error, (p), vpx_calloc(sizeof(*(p)), (n)))
36 #define CALLOC_ARRAY_ALIGNED(p, n, algn) \
37 do { \
38 CHECK_MEM_ERROR(&pbi->common.error, (p), \
39 vpx_memalign((algn), sizeof(*(p)) * (n))); \
40 memset((p), 0, (n) * sizeof(*(p))); \
41 } while (0)
42
setup_decoding_thread_data(VP8D_COMP * pbi,MACROBLOCKD * xd,MB_ROW_DEC * mbrd,int count)43 static void setup_decoding_thread_data(VP8D_COMP *pbi, MACROBLOCKD *xd,
44 MB_ROW_DEC *mbrd, int count) {
45 VP8_COMMON *const pc = &pbi->common;
46 int i;
47
48 for (i = 0; i < count; ++i) {
49 MACROBLOCKD *mbd = &mbrd[i].mbd;
50 mbd->subpixel_predict = xd->subpixel_predict;
51 mbd->subpixel_predict8x4 = xd->subpixel_predict8x4;
52 mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
53 mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
54
55 mbd->frame_type = pc->frame_type;
56 mbd->pre = xd->pre;
57 mbd->dst = xd->dst;
58
59 mbd->segmentation_enabled = xd->segmentation_enabled;
60 mbd->mb_segment_abs_delta = xd->mb_segment_abs_delta;
61 memcpy(mbd->segment_feature_data, xd->segment_feature_data,
62 sizeof(xd->segment_feature_data));
63
64 /*signed char ref_lf_deltas[MAX_REF_LF_DELTAS];*/
65 memcpy(mbd->ref_lf_deltas, xd->ref_lf_deltas, sizeof(xd->ref_lf_deltas));
66 /*signed char mode_lf_deltas[MAX_MODE_LF_DELTAS];*/
67 memcpy(mbd->mode_lf_deltas, xd->mode_lf_deltas, sizeof(xd->mode_lf_deltas));
68 /*unsigned char mode_ref_lf_delta_enabled;
69 unsigned char mode_ref_lf_delta_update;*/
70 mbd->mode_ref_lf_delta_enabled = xd->mode_ref_lf_delta_enabled;
71 mbd->mode_ref_lf_delta_update = xd->mode_ref_lf_delta_update;
72
73 mbd->current_bc = &pbi->mbc[0];
74
75 memcpy(mbd->dequant_y1_dc, xd->dequant_y1_dc, sizeof(xd->dequant_y1_dc));
76 memcpy(mbd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
77 memcpy(mbd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
78 memcpy(mbd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
79
80 mbd->fullpixel_mask = ~0;
81
82 if (pc->full_pixel) mbd->fullpixel_mask = ~7;
83 }
84
85 for (i = 0; i < pc->mb_rows; ++i)
86 vpx_atomic_store_release(&pbi->mt_current_mb_col[i], -1);
87 }
88
mt_decode_macroblock(VP8D_COMP * pbi,MACROBLOCKD * xd,unsigned int mb_idx)89 static void mt_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd,
90 unsigned int mb_idx) {
91 MB_PREDICTION_MODE mode;
92 int i;
93 #if CONFIG_ERROR_CONCEALMENT
94 int corruption_detected = 0;
95 #else
96 (void)mb_idx;
97 #endif
98
99 if (xd->mode_info_context->mbmi.mb_skip_coeff) {
100 vp8_reset_mb_tokens_context(xd);
101 } else if (!vp8dx_bool_error(xd->current_bc)) {
102 int eobtotal;
103 eobtotal = vp8_decode_mb_tokens(pbi, xd);
104
105 /* Special case: Force the loopfilter to skip when eobtotal is zero */
106 xd->mode_info_context->mbmi.mb_skip_coeff = (eobtotal == 0);
107 }
108
109 mode = xd->mode_info_context->mbmi.mode;
110
111 if (xd->segmentation_enabled) vp8_mb_init_dequantizer(pbi, xd);
112
113 #if CONFIG_ERROR_CONCEALMENT
114
115 if (pbi->ec_active) {
116 int throw_residual;
117 /* When we have independent partitions we can apply residual even
118 * though other partitions within the frame are corrupt.
119 */
120 throw_residual =
121 (!pbi->independent_partitions && pbi->frame_corrupt_residual);
122 throw_residual = (throw_residual || vp8dx_bool_error(xd->current_bc));
123
124 if ((mb_idx >= pbi->mvs_corrupt_from_mb || throw_residual)) {
125 /* MB with corrupt residuals or corrupt mode/motion vectors.
126 * Better to use the predictor as reconstruction.
127 */
128 pbi->frame_corrupt_residual = 1;
129 memset(xd->qcoeff, 0, sizeof(xd->qcoeff));
130
131 corruption_detected = 1;
132
133 /* force idct to be skipped for B_PRED and use the
134 * prediction only for reconstruction
135 * */
136 memset(xd->eobs, 0, 25);
137 }
138 }
139 #endif
140
141 /* do prediction */
142 if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) {
143 vp8_build_intra_predictors_mbuv_s(
144 xd, xd->recon_above[1], xd->recon_above[2], xd->recon_left[1],
145 xd->recon_left[2], xd->recon_left_stride[1], xd->dst.u_buffer,
146 xd->dst.v_buffer, xd->dst.uv_stride);
147
148 if (mode != B_PRED) {
149 vp8_build_intra_predictors_mby_s(
150 xd, xd->recon_above[0], xd->recon_left[0], xd->recon_left_stride[0],
151 xd->dst.y_buffer, xd->dst.y_stride);
152 } else {
153 short *DQC = xd->dequant_y1;
154 int dst_stride = xd->dst.y_stride;
155
156 /* clear out residual eob info */
157 if (xd->mode_info_context->mbmi.mb_skip_coeff) memset(xd->eobs, 0, 25);
158
159 intra_prediction_down_copy(xd, xd->recon_above[0] + 16);
160
161 for (i = 0; i < 16; ++i) {
162 BLOCKD *b = &xd->block[i];
163 unsigned char *dst = xd->dst.y_buffer + b->offset;
164 B_PREDICTION_MODE b_mode = xd->mode_info_context->bmi[i].as_mode;
165 unsigned char *Above;
166 unsigned char *yleft;
167 int left_stride;
168 unsigned char top_left;
169
170 /*Caution: For some b_mode, it needs 8 pixels (4 above + 4
171 * above-right).*/
172 if (i < 4 && pbi->common.filter_level) {
173 Above = xd->recon_above[0] + b->offset;
174 } else {
175 Above = dst - dst_stride;
176 }
177
178 if (i % 4 == 0 && pbi->common.filter_level) {
179 yleft = xd->recon_left[0] + i;
180 left_stride = 1;
181 } else {
182 yleft = dst - 1;
183 left_stride = dst_stride;
184 }
185
186 if ((i == 4 || i == 8 || i == 12) && pbi->common.filter_level) {
187 top_left = *(xd->recon_left[0] + i - 1);
188 } else {
189 top_left = Above[-1];
190 }
191
192 vp8_intra4x4_predict(Above, yleft, left_stride, b_mode, dst, dst_stride,
193 top_left);
194
195 if (xd->eobs[i]) {
196 if (xd->eobs[i] > 1) {
197 vp8_dequant_idct_add(b->qcoeff, DQC, dst, dst_stride);
198 } else {
199 vp8_dc_only_idct_add(b->qcoeff[0] * DQC[0], dst, dst_stride, dst,
200 dst_stride);
201 memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
202 }
203 }
204 }
205 }
206 } else {
207 vp8_build_inter_predictors_mb(xd);
208 }
209
210 #if CONFIG_ERROR_CONCEALMENT
211 if (corruption_detected) {
212 return;
213 }
214 #endif
215
216 if (!xd->mode_info_context->mbmi.mb_skip_coeff) {
217 /* dequantization and idct */
218 if (mode != B_PRED) {
219 short *DQC = xd->dequant_y1;
220
221 if (mode != SPLITMV) {
222 BLOCKD *b = &xd->block[24];
223
224 /* do 2nd order transform on the dc block */
225 if (xd->eobs[24] > 1) {
226 vp8_dequantize_b(b, xd->dequant_y2);
227
228 vp8_short_inv_walsh4x4(&b->dqcoeff[0], xd->qcoeff);
229 memset(b->qcoeff, 0, 16 * sizeof(b->qcoeff[0]));
230 } else {
231 b->dqcoeff[0] = b->qcoeff[0] * xd->dequant_y2[0];
232 vp8_short_inv_walsh4x4_1(&b->dqcoeff[0], xd->qcoeff);
233 memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
234 }
235
236 /* override the dc dequant constant in order to preserve the
237 * dc components
238 */
239 DQC = xd->dequant_y1_dc;
240 }
241
242 vp8_dequant_idct_add_y_block(xd->qcoeff, DQC, xd->dst.y_buffer,
243 xd->dst.y_stride, xd->eobs);
244 }
245
246 vp8_dequant_idct_add_uv_block(xd->qcoeff + 16 * 16, xd->dequant_uv,
247 xd->dst.u_buffer, xd->dst.v_buffer,
248 xd->dst.uv_stride, xd->eobs + 16);
249 }
250 }
251
mt_decode_mb_rows(VP8D_COMP * pbi,MACROBLOCKD * xd,int start_mb_row)252 static void mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd,
253 int start_mb_row) {
254 const vpx_atomic_int *last_row_current_mb_col;
255 vpx_atomic_int *current_mb_col;
256 int mb_row;
257 VP8_COMMON *pc = &pbi->common;
258 const int nsync = pbi->sync_range;
259 const vpx_atomic_int first_row_no_sync_above =
260 VPX_ATOMIC_INIT(pc->mb_cols + nsync);
261 int num_part = 1 << pbi->common.multi_token_partition;
262 int last_mb_row = start_mb_row;
263
264 YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
265 YV12_BUFFER_CONFIG *yv12_fb_lst = pbi->dec_fb_ref[LAST_FRAME];
266
267 int recon_y_stride = yv12_fb_new->y_stride;
268 int recon_uv_stride = yv12_fb_new->uv_stride;
269
270 unsigned char *ref_buffer[MAX_REF_FRAMES][3];
271 unsigned char *dst_buffer[3];
272 int i;
273 int ref_fb_corrupted[MAX_REF_FRAMES];
274
275 ref_fb_corrupted[INTRA_FRAME] = 0;
276
277 for (i = 1; i < MAX_REF_FRAMES; ++i) {
278 YV12_BUFFER_CONFIG *this_fb = pbi->dec_fb_ref[i];
279
280 ref_buffer[i][0] = this_fb->y_buffer;
281 ref_buffer[i][1] = this_fb->u_buffer;
282 ref_buffer[i][2] = this_fb->v_buffer;
283
284 ref_fb_corrupted[i] = this_fb->corrupted;
285 }
286
287 dst_buffer[0] = yv12_fb_new->y_buffer;
288 dst_buffer[1] = yv12_fb_new->u_buffer;
289 dst_buffer[2] = yv12_fb_new->v_buffer;
290
291 xd->up_available = (start_mb_row != 0);
292
293 xd->mode_info_context = pc->mi + pc->mode_info_stride * start_mb_row;
294 xd->mode_info_stride = pc->mode_info_stride;
295
296 for (mb_row = start_mb_row; mb_row < pc->mb_rows;
297 mb_row += (pbi->decoding_thread_count + 1)) {
298 int recon_yoffset, recon_uvoffset;
299 int mb_col;
300 int filter_level;
301 loop_filter_info_n *lfi_n = &pc->lf_info;
302
303 /* save last row processed by this thread */
304 last_mb_row = mb_row;
305 /* select bool coder for current partition */
306 xd->current_bc = &pbi->mbc[mb_row % num_part];
307
308 if (mb_row > 0) {
309 last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row - 1];
310 } else {
311 last_row_current_mb_col = &first_row_no_sync_above;
312 }
313
314 current_mb_col = &pbi->mt_current_mb_col[mb_row];
315
316 recon_yoffset = mb_row * recon_y_stride * 16;
317 recon_uvoffset = mb_row * recon_uv_stride * 8;
318
319 /* reset contexts */
320 xd->above_context = pc->above_context;
321 memset(xd->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));
322
323 xd->left_available = 0;
324
325 xd->mb_to_top_edge = -((mb_row * 16) << 3);
326 xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
327
328 if (pbi->common.filter_level) {
329 xd->recon_above[0] = pbi->mt_yabove_row[mb_row] + 0 * 16 + 32;
330 xd->recon_above[1] = pbi->mt_uabove_row[mb_row] + 0 * 8 + 16;
331 xd->recon_above[2] = pbi->mt_vabove_row[mb_row] + 0 * 8 + 16;
332
333 xd->recon_left[0] = pbi->mt_yleft_col[mb_row];
334 xd->recon_left[1] = pbi->mt_uleft_col[mb_row];
335 xd->recon_left[2] = pbi->mt_vleft_col[mb_row];
336
337 /* TODO: move to outside row loop */
338 xd->recon_left_stride[0] = 1;
339 xd->recon_left_stride[1] = 1;
340 } else {
341 xd->recon_above[0] = dst_buffer[0] + recon_yoffset;
342 xd->recon_above[1] = dst_buffer[1] + recon_uvoffset;
343 xd->recon_above[2] = dst_buffer[2] + recon_uvoffset;
344
345 xd->recon_left[0] = xd->recon_above[0] - 1;
346 xd->recon_left[1] = xd->recon_above[1] - 1;
347 xd->recon_left[2] = xd->recon_above[2] - 1;
348
349 xd->recon_above[0] -= xd->dst.y_stride;
350 xd->recon_above[1] -= xd->dst.uv_stride;
351 xd->recon_above[2] -= xd->dst.uv_stride;
352
353 /* TODO: move to outside row loop */
354 xd->recon_left_stride[0] = xd->dst.y_stride;
355 xd->recon_left_stride[1] = xd->dst.uv_stride;
356
357 setup_intra_recon_left(xd->recon_left[0], xd->recon_left[1],
358 xd->recon_left[2], xd->dst.y_stride,
359 xd->dst.uv_stride);
360 }
361
362 for (mb_col = 0; mb_col < pc->mb_cols; ++mb_col) {
363 if (((mb_col - 1) % nsync) == 0) {
364 vpx_atomic_store_release(current_mb_col, mb_col - 1);
365 }
366
367 if (mb_row && !(mb_col & (nsync - 1))) {
368 vp8_atomic_spin_wait(mb_col, last_row_current_mb_col, nsync);
369 }
370
371 /* Distance of MB to the various image edges.
372 * These are specified to 8th pel as they are always
373 * compared to values that are in 1/8th pel units.
374 */
375 xd->mb_to_left_edge = -((mb_col * 16) << 3);
376 xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
377
378 #if CONFIG_ERROR_CONCEALMENT
379 {
380 int corrupt_residual =
381 (!pbi->independent_partitions && pbi->frame_corrupt_residual) ||
382 vp8dx_bool_error(xd->current_bc);
383 if (pbi->ec_active &&
384 (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) &&
385 corrupt_residual) {
386 /* We have an intra block with corrupt
387 * coefficients, better to conceal with an inter
388 * block.
389 * Interpolate MVs from neighboring MBs
390 *
391 * Note that for the first mb with corrupt
392 * residual in a frame, we might not discover
393 * that before decoding the residual. That
394 * happens after this check, and therefore no
395 * inter concealment will be done.
396 */
397 vp8_interpolate_motion(xd, mb_row, mb_col, pc->mb_rows, pc->mb_cols);
398 }
399 }
400 #endif
401
402 xd->dst.y_buffer = dst_buffer[0] + recon_yoffset;
403 xd->dst.u_buffer = dst_buffer[1] + recon_uvoffset;
404 xd->dst.v_buffer = dst_buffer[2] + recon_uvoffset;
405
406 /* propagate errors from reference frames */
407 xd->corrupted |= ref_fb_corrupted[xd->mode_info_context->mbmi.ref_frame];
408
409 if (xd->corrupted) {
410 // Move current decoding marcoblock to the end of row for all rows
411 // assigned to this thread, such that other threads won't be waiting.
412 for (; mb_row < pc->mb_rows;
413 mb_row += (pbi->decoding_thread_count + 1)) {
414 current_mb_col = &pbi->mt_current_mb_col[mb_row];
415 vpx_atomic_store_release(current_mb_col, pc->mb_cols + nsync);
416 }
417 vpx_internal_error(&xd->error_info, VPX_CODEC_CORRUPT_FRAME,
418 "Corrupted reference frame");
419 }
420
421 if (xd->mode_info_context->mbmi.ref_frame >= LAST_FRAME) {
422 const MV_REFERENCE_FRAME ref = xd->mode_info_context->mbmi.ref_frame;
423 xd->pre.y_buffer = ref_buffer[ref][0] + recon_yoffset;
424 xd->pre.u_buffer = ref_buffer[ref][1] + recon_uvoffset;
425 xd->pre.v_buffer = ref_buffer[ref][2] + recon_uvoffset;
426 } else {
427 // ref_frame is INTRA_FRAME, pre buffer should not be used.
428 xd->pre.y_buffer = 0;
429 xd->pre.u_buffer = 0;
430 xd->pre.v_buffer = 0;
431 }
432 mt_decode_macroblock(pbi, xd, 0);
433
434 xd->left_available = 1;
435
436 /* check if the boolean decoder has suffered an error */
437 xd->corrupted |= vp8dx_bool_error(xd->current_bc);
438
439 xd->recon_above[0] += 16;
440 xd->recon_above[1] += 8;
441 xd->recon_above[2] += 8;
442
443 if (!pbi->common.filter_level) {
444 xd->recon_left[0] += 16;
445 xd->recon_left[1] += 8;
446 xd->recon_left[2] += 8;
447 }
448
449 if (pbi->common.filter_level) {
450 int skip_lf = (xd->mode_info_context->mbmi.mode != B_PRED &&
451 xd->mode_info_context->mbmi.mode != SPLITMV &&
452 xd->mode_info_context->mbmi.mb_skip_coeff);
453
454 const int mode_index =
455 lfi_n->mode_lf_lut[xd->mode_info_context->mbmi.mode];
456 const int seg = xd->mode_info_context->mbmi.segment_id;
457 const int ref_frame = xd->mode_info_context->mbmi.ref_frame;
458
459 filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
460
461 if (mb_row != pc->mb_rows - 1) {
462 /* Save decoded MB last row data for next-row decoding */
463 memcpy((pbi->mt_yabove_row[mb_row + 1] + 32 + mb_col * 16),
464 (xd->dst.y_buffer + 15 * recon_y_stride), 16);
465 memcpy((pbi->mt_uabove_row[mb_row + 1] + 16 + mb_col * 8),
466 (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
467 memcpy((pbi->mt_vabove_row[mb_row + 1] + 16 + mb_col * 8),
468 (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
469 }
470
471 /* save left_col for next MB decoding */
472 if (mb_col != pc->mb_cols - 1) {
473 MODE_INFO *next = xd->mode_info_context + 1;
474
475 if (next->mbmi.ref_frame == INTRA_FRAME) {
476 for (i = 0; i < 16; ++i) {
477 pbi->mt_yleft_col[mb_row][i] =
478 xd->dst.y_buffer[i * recon_y_stride + 15];
479 }
480 for (i = 0; i < 8; ++i) {
481 pbi->mt_uleft_col[mb_row][i] =
482 xd->dst.u_buffer[i * recon_uv_stride + 7];
483 pbi->mt_vleft_col[mb_row][i] =
484 xd->dst.v_buffer[i * recon_uv_stride + 7];
485 }
486 }
487 }
488
489 /* loopfilter on this macroblock. */
490 if (filter_level) {
491 if (pc->filter_type == NORMAL_LOOPFILTER) {
492 loop_filter_info lfi;
493 FRAME_TYPE frame_type = pc->frame_type;
494 const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
495 lfi.mblim = lfi_n->mblim[filter_level];
496 lfi.blim = lfi_n->blim[filter_level];
497 lfi.lim = lfi_n->lim[filter_level];
498 lfi.hev_thr = lfi_n->hev_thr[hev_index];
499
500 if (mb_col > 0)
501 vp8_loop_filter_mbv(xd->dst.y_buffer, xd->dst.u_buffer,
502 xd->dst.v_buffer, recon_y_stride,
503 recon_uv_stride, &lfi);
504
505 if (!skip_lf)
506 vp8_loop_filter_bv(xd->dst.y_buffer, xd->dst.u_buffer,
507 xd->dst.v_buffer, recon_y_stride,
508 recon_uv_stride, &lfi);
509
510 /* don't apply across umv border */
511 if (mb_row > 0)
512 vp8_loop_filter_mbh(xd->dst.y_buffer, xd->dst.u_buffer,
513 xd->dst.v_buffer, recon_y_stride,
514 recon_uv_stride, &lfi);
515
516 if (!skip_lf)
517 vp8_loop_filter_bh(xd->dst.y_buffer, xd->dst.u_buffer,
518 xd->dst.v_buffer, recon_y_stride,
519 recon_uv_stride, &lfi);
520 } else {
521 if (mb_col > 0)
522 vp8_loop_filter_simple_mbv(xd->dst.y_buffer, recon_y_stride,
523 lfi_n->mblim[filter_level]);
524
525 if (!skip_lf)
526 vp8_loop_filter_simple_bv(xd->dst.y_buffer, recon_y_stride,
527 lfi_n->blim[filter_level]);
528
529 /* don't apply across umv border */
530 if (mb_row > 0)
531 vp8_loop_filter_simple_mbh(xd->dst.y_buffer, recon_y_stride,
532 lfi_n->mblim[filter_level]);
533
534 if (!skip_lf)
535 vp8_loop_filter_simple_bh(xd->dst.y_buffer, recon_y_stride,
536 lfi_n->blim[filter_level]);
537 }
538 }
539 }
540
541 recon_yoffset += 16;
542 recon_uvoffset += 8;
543
544 ++xd->mode_info_context; /* next mb */
545
546 xd->above_context++;
547 }
548
549 /* adjust to the next row of mbs */
550 if (pbi->common.filter_level) {
551 if (mb_row != pc->mb_rows - 1) {
552 int lasty = yv12_fb_lst->y_width + VP8BORDERINPIXELS;
553 int lastuv = (yv12_fb_lst->y_width >> 1) + (VP8BORDERINPIXELS >> 1);
554
555 for (i = 0; i < 4; ++i) {
556 pbi->mt_yabove_row[mb_row + 1][lasty + i] =
557 pbi->mt_yabove_row[mb_row + 1][lasty - 1];
558 pbi->mt_uabove_row[mb_row + 1][lastuv + i] =
559 pbi->mt_uabove_row[mb_row + 1][lastuv - 1];
560 pbi->mt_vabove_row[mb_row + 1][lastuv + i] =
561 pbi->mt_vabove_row[mb_row + 1][lastuv - 1];
562 }
563 }
564 } else {
565 vp8_extend_mb_row(yv12_fb_new, xd->dst.y_buffer + 16,
566 xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
567 }
568
569 /* last MB of row is ready just after extension is done */
570 vpx_atomic_store_release(current_mb_col, mb_col + nsync);
571
572 ++xd->mode_info_context; /* skip prediction column */
573 xd->up_available = 1;
574
575 /* since we have multithread */
576 xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
577 }
578
579 /* signal end of decoding of current thread for current frame */
580 if (last_mb_row + (int)pbi->decoding_thread_count + 1 >= pc->mb_rows)
581 vp8_sem_post(&pbi->h_event_end_decoding);
582 }
583
thread_decoding_proc(void * p_data)584 static THREADFN thread_decoding_proc(void *p_data) {
585 int ithread = ((DECODETHREAD_DATA *)p_data)->ithread;
586 VP8D_COMP *pbi = (VP8D_COMP *)(((DECODETHREAD_DATA *)p_data)->ptr1);
587 MB_ROW_DEC *mbrd = (MB_ROW_DEC *)(((DECODETHREAD_DATA *)p_data)->ptr2);
588 ENTROPY_CONTEXT_PLANES mb_row_left_context;
589
590 while (1) {
591 if (vpx_atomic_load_acquire(&pbi->b_multithreaded_rd) == 0) break;
592
593 if (vp8_sem_wait(&pbi->h_event_start_decoding[ithread]) == 0) {
594 if (vpx_atomic_load_acquire(&pbi->b_multithreaded_rd) == 0) {
595 break;
596 } else {
597 MACROBLOCKD *xd = &mbrd->mbd;
598 xd->left_context = &mb_row_left_context;
599 if (setjmp(xd->error_info.jmp)) {
600 xd->error_info.setjmp = 0;
601 // Signal the end of decoding for current thread.
602 vp8_sem_post(&pbi->h_event_end_decoding);
603 continue;
604 }
605 xd->error_info.setjmp = 1;
606 mt_decode_mb_rows(pbi, xd, ithread + 1);
607 xd->error_info.setjmp = 0;
608 }
609 }
610 }
611
612 return THREAD_EXIT_SUCCESS;
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 (vp8_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 (vp8_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 vp8_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 vp8_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 vp8_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 vp8_sem_destroy(&pbi->h_event_start_decoding[i]);
823 }
824
825 if (pbi->allocated_decoding_thread_count) {
826 vp8_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 vp8_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 vp8_sem_wait(&pbi->h_event_end_decoding);
899 return -1;
900 }
901
902 xd->error_info.setjmp = 1;
903 mt_decode_mb_rows(pbi, xd, 0);
904 xd->error_info.setjmp = 0;
905
906 for (i = 0; i < pbi->decoding_thread_count + 1; ++i)
907 vp8_sem_wait(&pbi->h_event_end_decoding); /* add back for each frame */
908
909 return 0;
910 }
911