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    mutex_lock(&lf_sync->mutex_[r - 1]);
44
45    while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
46      pthread_cond_wait(&lf_sync->cond_[r - 1],
47                        &lf_sync->mutex_[r - 1]);
48    }
49    pthread_mutex_unlock(&lf_sync->mutex_[r - 1]);
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, MACROBLOCKD *const xd,
93                                int start, int stop, int y_only,
94                                VP9LfSync *const lf_sync, int num_lf_workers) {
95  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
96  int r, c;  // SB row and col
97  LOOP_FILTER_MASK lfm;
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 **mi_8x8 = 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      int plane;
107
108      sync_read(lf_sync, r, c);
109
110      vp9_setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
111      vp9_setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mi_stride, &lfm);
112
113      for (plane = 0; plane < num_planes; ++plane) {
114        vp9_filter_block_plane(cm, &xd->plane[plane], mi_row, &lfm);
115      }
116
117      sync_write(lf_sync, r, c, sb_cols);
118    }
119  }
120}
121
122// Row-based multi-threaded loopfilter hook
123static int loop_filter_row_worker(void *arg1, void *arg2) {
124  TileWorkerData *const tile_data = (TileWorkerData*)arg1;
125  LFWorkerData *const lf_data = &tile_data->lfdata;
126
127  loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, &lf_data->xd,
128                      lf_data->start, lf_data->stop, lf_data->y_only,
129                      lf_data->lf_sync, lf_data->num_lf_workers);
130  return 1;
131}
132
133// VP9 decoder: Implement multi-threaded loopfilter that uses the tile
134// threads.
135void vp9_loop_filter_frame_mt(VP9Decoder *pbi,
136                              VP9_COMMON *cm,
137                              MACROBLOCKD *xd,
138                              int frame_filter_level,
139                              int y_only, int partial_frame) {
140  // Number of superblock rows and cols
141  const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
142  const int tile_cols = 1 << cm->log2_tile_cols;
143  const int num_workers = MIN(pbi->oxcf.max_threads & ~1, tile_cols);
144  int i;
145
146  // Allocate memory used in thread synchronization.
147  // This always needs to be done even if frame_filter_level is 0.
148  if (!cm->current_video_frame || cm->last_height != cm->height) {
149    VP9LfSync *const lf_sync = &pbi->lf_row_sync;
150
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(pbi->lf_row_sync.cur_sb_col, -1,
170             sizeof(*pbi->lf_row_sync.cur_sb_col) * sb_rows);
171
172  // Set up loopfilter thread data.
173  // The decoder is using num_workers instead of pbi->num_tile_workers
174  // because it has been observed that using more threads on the
175  // loopfilter, than there are tile columns in the frame will hurt
176  // performance on Android. This is because the system will only
177  // schedule the tile decode workers on cores equal to the number
178  // of tile columns. Then if the decoder tries to use more threads for the
179  // loopfilter, it will hurt performance because of contention. If the
180  // multithreading code changes in the future then the number of workers
181  // used by the loopfilter should be revisited.
182  for (i = 0; i < num_workers; ++i) {
183    VP9Worker *const worker = &pbi->tile_workers[i];
184    TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
185    LFWorkerData *const lf_data = &tile_data->lfdata;
186
187    worker->hook = (VP9WorkerHook)loop_filter_row_worker;
188
189    // Loopfilter data
190    lf_data->frame_buffer = get_frame_new_buffer(cm);
191    lf_data->cm = cm;
192    lf_data->xd = pbi->mb;
193    lf_data->start = i;
194    lf_data->stop = sb_rows;
195    lf_data->y_only = y_only;   // always do all planes in decoder
196
197    lf_data->lf_sync = &pbi->lf_row_sync;
198    lf_data->num_lf_workers = num_workers;
199
200    // Start loopfiltering
201    if (i == num_workers - 1) {
202      vp9_worker_execute(worker);
203    } else {
204      vp9_worker_launch(worker);
205    }
206  }
207
208  // Wait till all rows are finished
209  for (i = 0; i < num_workers; ++i) {
210    vp9_worker_sync(&pbi->tile_workers[i]);
211  }
212}
213
214// Set up nsync by width.
215static int get_sync_range(int width) {
216  // nsync numbers are picked by testing. For example, for 4k
217  // video, using 4 gives best performance.
218  if (width < 640)
219    return 1;
220  else if (width <= 1280)
221    return 2;
222  else if (width <= 4096)
223    return 4;
224  else
225    return 8;
226}
227
228// Allocate memory for lf row synchronization
229void vp9_loop_filter_alloc(VP9_COMMON *cm, VP9LfSync *lf_sync, int rows,
230                           int width) {
231#if CONFIG_MULTITHREAD
232  int i;
233
234  CHECK_MEM_ERROR(cm, lf_sync->mutex_,
235                  vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
236  for (i = 0; i < rows; ++i) {
237    pthread_mutex_init(&lf_sync->mutex_[i], NULL);
238  }
239
240  CHECK_MEM_ERROR(cm, lf_sync->cond_,
241                  vpx_malloc(sizeof(*lf_sync->cond_) * rows));
242  for (i = 0; i < rows; ++i) {
243    pthread_cond_init(&lf_sync->cond_[i], NULL);
244  }
245#endif  // CONFIG_MULTITHREAD
246
247  CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
248                  vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
249
250  // Set up nsync.
251  lf_sync->sync_range = get_sync_range(width);
252}
253
254// Deallocate lf synchronization related mutex and data
255void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows) {
256#if CONFIG_MULTITHREAD
257  if (lf_sync != NULL) {
258    int i;
259
260    if (lf_sync->mutex_ != NULL) {
261      for (i = 0; i < rows; ++i) {
262        pthread_mutex_destroy(&lf_sync->mutex_[i]);
263      }
264      vpx_free(lf_sync->mutex_);
265    }
266    if (lf_sync->cond_ != NULL) {
267      for (i = 0; i < rows; ++i) {
268        pthread_cond_destroy(&lf_sync->cond_[i]);
269      }
270      vpx_free(lf_sync->cond_);
271    }
272
273    vpx_free(lf_sync->cur_sb_col);
274    // clear the structure as the source of this call may be a resize in which
275    // case this call will be followed by an _alloc() which may fail.
276    vpx_memset(lf_sync, 0, sizeof(*lf_sync));
277  }
278#else
279  (void)rows;
280  if (lf_sync != NULL) {
281    vpx_free(lf_sync->cur_sb_col);
282    vpx_memset(lf_sync, 0, sizeof(*lf_sync));
283  }
284#endif  // CONFIG_MULTITHREAD
285}
286