1/*
2 *  Copyright (c) 2014 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
13#include "vpx_mem/vpx_mem.h"
14
15#include "vp9/common/vp9_reconinter.h"
16
17#include "vp9/decoder/vp9_dthread.h"
18#include "vp9/decoder/vp9_decoder.h"
19
20#if CONFIG_MULTITHREAD
21static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
22  const int kMaxTryLocks = 4000;
23  int locked = 0;
24  int i;
25
26  for (i = 0; i < kMaxTryLocks; ++i) {
27    if (!pthread_mutex_trylock(mutex)) {
28      locked = 1;
29      break;
30    }
31  }
32
33  if (!locked)
34    pthread_mutex_lock(mutex);
35}
36#endif  // CONFIG_MULTITHREAD
37
38static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
39#if CONFIG_MULTITHREAD
40  const int nsync = lf_sync->sync_range;
41
42  if (r && !(c & (nsync - 1))) {
43    pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
44    mutex_lock(mutex);
45
46    while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
47      pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
48    }
49    pthread_mutex_unlock(mutex);
50  }
51#else
52  (void)lf_sync;
53  (void)r;
54  (void)c;
55#endif  // CONFIG_MULTITHREAD
56}
57
58static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
59                              const int sb_cols) {
60#if CONFIG_MULTITHREAD
61  const int nsync = lf_sync->sync_range;
62  int cur;
63  // Only signal when there are enough filtered SB for next row to run.
64  int sig = 1;
65
66  if (c < sb_cols - 1) {
67    cur = c;
68    if (c % nsync)
69      sig = 0;
70  } else {
71    cur = sb_cols + nsync;
72  }
73
74  if (sig) {
75    mutex_lock(&lf_sync->mutex_[r]);
76
77    lf_sync->cur_sb_col[r] = cur;
78
79    pthread_cond_signal(&lf_sync->cond_[r]);
80    pthread_mutex_unlock(&lf_sync->mutex_[r]);
81  }
82#else
83  (void)lf_sync;
84  (void)r;
85  (void)c;
86  (void)sb_cols;
87#endif  // CONFIG_MULTITHREAD
88}
89
90// Implement row loopfiltering for each thread.
91static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer,
92                                VP9_COMMON *const cm,
93                                struct macroblockd_plane planes[MAX_MB_PLANE],
94                                int start, int stop, int y_only,
95                                VP9LfSync *const lf_sync, int num_lf_workers) {
96  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
97  int r, c;  // SB row and col
98  const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
99
100  for (r = start; r < stop; r += num_lf_workers) {
101    const int mi_row = r << MI_BLOCK_SIZE_LOG2;
102    MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
103
104    for (c = 0; c < sb_cols; ++c) {
105      const int mi_col = c << MI_BLOCK_SIZE_LOG2;
106      LOOP_FILTER_MASK lfm;
107      int plane;
108
109      sync_read(lf_sync, r, c);
110
111      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
112      vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm);
113
114      for (plane = 0; plane < num_planes; ++plane) {
115        vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
116      }
117
118      sync_write(lf_sync, r, c, sb_cols);
119    }
120  }
121}
122
123// Row-based multi-threaded loopfilter hook
124static int loop_filter_row_worker(void *arg1, void *arg2) {
125  TileWorkerData *const tile_data = (TileWorkerData*)arg1;
126  LFWorkerData *const lf_data = &tile_data->lfdata;
127  (void) arg2;
128  loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
129                      lf_data->start, lf_data->stop, lf_data->y_only,
130                      lf_data->lf_sync, lf_data->num_lf_workers);
131  return 1;
132}
133
134// VP9 decoder: Implement multi-threaded loopfilter that uses the tile
135// threads.
136void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
137                              VP9Decoder *pbi, VP9_COMMON *cm,
138                              int frame_filter_level,
139                              int y_only) {
140  VP9LfSync *const lf_sync = &pbi->lf_row_sync;
141  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
142  // Number of superblock rows and cols
143  const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
144  const int tile_cols = 1 << cm->log2_tile_cols;
145  const int num_workers = MIN(pbi->max_threads & ~1, tile_cols);
146  int i;
147
148  // Allocate memory used in thread synchronization.
149  // This always needs to be done even if frame_filter_level is 0.
150  if (!cm->current_video_frame || cm->last_height != cm->height) {
151    if (cm->last_height != cm->height) {
152      const int aligned_last_height =
153          ALIGN_POWER_OF_TWO(cm->last_height, MI_SIZE_LOG2);
154      const int last_sb_rows =
155          mi_cols_aligned_to_sb(aligned_last_height >> MI_SIZE_LOG2) >>
156          MI_BLOCK_SIZE_LOG2;
157
158      vp9_loop_filter_dealloc(lf_sync, last_sb_rows);
159    }
160
161    vp9_loop_filter_alloc(cm, lf_sync, sb_rows, cm->width);
162  }
163
164  if (!frame_filter_level) return;
165
166  vp9_loop_filter_frame_init(cm, frame_filter_level);
167
168  // Initialize cur_sb_col to -1 for all SB rows.
169  vpx_memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
170
171  // Set up loopfilter thread data.
172  // The decoder is using num_workers instead of pbi->num_tile_workers
173  // because it has been observed that using more threads on the
174  // loopfilter, than there are tile columns in the frame will hurt
175  // performance on Android. This is because the system will only
176  // schedule the tile decode workers on cores equal to the number
177  // of tile columns. Then if the decoder tries to use more threads for the
178  // loopfilter, it will hurt performance because of contention. If the
179  // multithreading code changes in the future then the number of workers
180  // used by the loopfilter should be revisited.
181  for (i = 0; i < num_workers; ++i) {
182    VP9Worker *const worker = &pbi->tile_workers[i];
183    TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
184    LFWorkerData *const lf_data = &tile_data->lfdata;
185
186    worker->hook = (VP9WorkerHook)loop_filter_row_worker;
187
188    // Loopfilter data
189    lf_data->frame_buffer = frame;
190    lf_data->cm = cm;
191    vp9_copy(lf_data->planes, pbi->mb.plane);
192    lf_data->start = i;
193    lf_data->stop = sb_rows;
194    lf_data->y_only = y_only;   // always do all planes in decoder
195
196    lf_data->lf_sync = lf_sync;
197    lf_data->num_lf_workers = num_workers;
198
199    // Start loopfiltering
200    if (i == num_workers - 1) {
201      winterface->execute(worker);
202    } else {
203      winterface->launch(worker);
204    }
205  }
206
207  // Wait till all rows are finished
208  for (i = 0; i < num_workers; ++i) {
209    winterface->sync(&pbi->tile_workers[i]);
210  }
211}
212
213// Set up nsync by width.
214static int get_sync_range(int width) {
215  // nsync numbers are picked by testing. For example, for 4k
216  // video, using 4 gives best performance.
217  if (width < 640)
218    return 1;
219  else if (width <= 1280)
220    return 2;
221  else if (width <= 4096)
222    return 4;
223  else
224    return 8;
225}
226
227// Allocate memory for lf row synchronization
228void vp9_loop_filter_alloc(VP9_COMMON *cm, VP9LfSync *lf_sync, int rows,
229                           int width) {
230#if CONFIG_MULTITHREAD
231  int i;
232
233  CHECK_MEM_ERROR(cm, lf_sync->mutex_,
234                  vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
235  for (i = 0; i < rows; ++i) {
236    pthread_mutex_init(&lf_sync->mutex_[i], NULL);
237  }
238
239  CHECK_MEM_ERROR(cm, lf_sync->cond_,
240                  vpx_malloc(sizeof(*lf_sync->cond_) * rows));
241  for (i = 0; i < rows; ++i) {
242    pthread_cond_init(&lf_sync->cond_[i], NULL);
243  }
244#endif  // CONFIG_MULTITHREAD
245
246  CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
247                  vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
248
249  // Set up nsync.
250  lf_sync->sync_range = get_sync_range(width);
251}
252
253// Deallocate lf synchronization related mutex and data
254void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows) {
255#if !CONFIG_MULTITHREAD
256  (void)rows;
257#endif  // !CONFIG_MULTITHREAD
258
259  if (lf_sync != NULL) {
260#if CONFIG_MULTITHREAD
261    int i;
262
263    if (lf_sync->mutex_ != NULL) {
264      for (i = 0; i < rows; ++i) {
265        pthread_mutex_destroy(&lf_sync->mutex_[i]);
266      }
267      vpx_free(lf_sync->mutex_);
268    }
269    if (lf_sync->cond_ != NULL) {
270      for (i = 0; i < rows; ++i) {
271        pthread_cond_destroy(&lf_sync->cond_[i]);
272      }
273      vpx_free(lf_sync->cond_);
274    }
275#endif  // CONFIG_MULTITHREAD
276    vpx_free(lf_sync->cur_sb_col);
277    // clear the structure as the source of this call may be a resize in which
278    // case this call will be followed by an _alloc() which may fail.
279    vp9_zero(*lf_sync);
280  }
281}
282