• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Copyright (c) 2017 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 <assert.h>
12 
13 #include "vp9/encoder/vp9_encoder.h"
14 #include "vp9/encoder/vp9_ethread.h"
15 #include "vp9/encoder/vp9_multi_thread.h"
16 
vp9_enc_grp_get_next_job(MultiThreadHandle * multi_thread_ctxt,int tile_id)17 void *vp9_enc_grp_get_next_job(MultiThreadHandle *multi_thread_ctxt,
18                                int tile_id) {
19   RowMTInfo *row_mt_info;
20   JobQueueHandle *job_queue_hdl = NULL;
21   void *next = NULL;
22   JobNode *job_info = NULL;
23 #if CONFIG_MULTITHREAD
24   pthread_mutex_t *mutex_handle = NULL;
25 #endif
26 
27   row_mt_info = (RowMTInfo *)(&multi_thread_ctxt->row_mt_info[tile_id]);
28   job_queue_hdl = (JobQueueHandle *)&row_mt_info->job_queue_hdl;
29 #if CONFIG_MULTITHREAD
30   mutex_handle = &row_mt_info->job_mutex;
31 #endif
32 
33 // lock the mutex for queue access
34 #if CONFIG_MULTITHREAD
35   pthread_mutex_lock(mutex_handle);
36 #endif
37   next = job_queue_hdl->next;
38   if (NULL != next) {
39     JobQueue *job_queue = (JobQueue *)next;
40     job_info = &job_queue->job_info;
41     // Update the next job in the queue
42     job_queue_hdl->next = job_queue->next;
43     job_queue_hdl->num_jobs_acquired++;
44   }
45 
46 #if CONFIG_MULTITHREAD
47   pthread_mutex_unlock(mutex_handle);
48 #endif
49 
50   return job_info;
51 }
52 
vp9_row_mt_mem_alloc(VP9_COMP * cpi)53 void vp9_row_mt_mem_alloc(VP9_COMP *cpi) {
54   struct VP9Common *cm = &cpi->common;
55   MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
56   int tile_row, tile_col;
57   const int tile_cols = 1 << cm->log2_tile_cols;
58   const int tile_rows = 1 << cm->log2_tile_rows;
59   const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
60   int jobs_per_tile_col, total_jobs;
61 
62   jobs_per_tile_col = VPXMAX(cm->mb_rows, sb_rows);
63   // Calculate the total number of jobs
64   total_jobs = jobs_per_tile_col * tile_cols;
65 
66   multi_thread_ctxt->allocated_tile_cols = tile_cols;
67   multi_thread_ctxt->allocated_tile_rows = tile_rows;
68   multi_thread_ctxt->allocated_vert_unit_rows = jobs_per_tile_col;
69 
70   multi_thread_ctxt->job_queue =
71       (JobQueue *)vpx_memalign(32, total_jobs * sizeof(JobQueue));
72 
73 #if CONFIG_MULTITHREAD
74   // Create mutex for each tile
75   for (tile_col = 0; tile_col < tile_cols; tile_col++) {
76     RowMTInfo *row_mt_info = &multi_thread_ctxt->row_mt_info[tile_col];
77     pthread_mutex_init(&row_mt_info->job_mutex, NULL);
78   }
79 #endif
80 
81   // Allocate memory for row based multi-threading
82   for (tile_col = 0; tile_col < tile_cols; tile_col++) {
83     TileDataEnc *this_tile = &cpi->tile_data[tile_col];
84     vp9_row_mt_sync_mem_alloc(&this_tile->row_mt_sync, cm, jobs_per_tile_col);
85     if (cpi->sf.adaptive_rd_thresh_row_mt) {
86       const int sb_rows =
87           (mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2) + 1;
88       int i;
89       this_tile->row_base_thresh_freq_fact =
90           (int *)vpx_calloc(sb_rows * BLOCK_SIZES * MAX_MODES,
91                             sizeof(*(this_tile->row_base_thresh_freq_fact)));
92       for (i = 0; i < sb_rows * BLOCK_SIZES * MAX_MODES; i++)
93         this_tile->row_base_thresh_freq_fact[i] = RD_THRESH_INIT_FACT;
94     }
95   }
96 
97   // Assign the sync pointer of tile row zero for every tile row > 0
98   for (tile_row = 1; tile_row < tile_rows; tile_row++) {
99     for (tile_col = 0; tile_col < tile_cols; tile_col++) {
100       TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
101       TileDataEnc *this_col_tile = &cpi->tile_data[tile_col];
102       this_tile->row_mt_sync = this_col_tile->row_mt_sync;
103     }
104   }
105 
106   // Calculate the number of vertical units in the given tile row
107   for (tile_row = 0; tile_row < tile_rows; tile_row++) {
108     TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols];
109     TileInfo *tile_info = &this_tile->tile_info;
110     multi_thread_ctxt->num_tile_vert_sbs[tile_row] =
111         get_num_vert_units(*tile_info, MI_BLOCK_SIZE_LOG2);
112   }
113 }
114 
vp9_row_mt_mem_dealloc(VP9_COMP * cpi)115 void vp9_row_mt_mem_dealloc(VP9_COMP *cpi) {
116   MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
117   int tile_col;
118 #if CONFIG_MULTITHREAD
119   int tile_row;
120 #endif
121 
122   // Deallocate memory for job queue
123   if (multi_thread_ctxt->job_queue) vpx_free(multi_thread_ctxt->job_queue);
124 
125 #if CONFIG_MULTITHREAD
126   // Destroy mutex for each tile
127   for (tile_col = 0; tile_col < multi_thread_ctxt->allocated_tile_cols;
128        tile_col++) {
129     RowMTInfo *row_mt_info = &multi_thread_ctxt->row_mt_info[tile_col];
130     if (row_mt_info) pthread_mutex_destroy(&row_mt_info->job_mutex);
131   }
132 #endif
133 
134   // Free row based multi-threading sync memory
135   for (tile_col = 0; tile_col < multi_thread_ctxt->allocated_tile_cols;
136        tile_col++) {
137     TileDataEnc *this_tile = &cpi->tile_data[tile_col];
138     vp9_row_mt_sync_mem_dealloc(&this_tile->row_mt_sync);
139   }
140 
141 #if CONFIG_MULTITHREAD
142   for (tile_row = 0; tile_row < multi_thread_ctxt->allocated_tile_rows;
143        tile_row++) {
144     for (tile_col = 0; tile_col < multi_thread_ctxt->allocated_tile_cols;
145          tile_col++) {
146       TileDataEnc *this_tile =
147           &cpi->tile_data[tile_row * multi_thread_ctxt->allocated_tile_cols +
148                           tile_col];
149       if (cpi->sf.adaptive_rd_thresh_row_mt) {
150         if (this_tile->row_base_thresh_freq_fact != NULL) {
151           vpx_free(this_tile->row_base_thresh_freq_fact);
152           this_tile->row_base_thresh_freq_fact = NULL;
153         }
154       }
155     }
156   }
157 #endif
158 }
159 
vp9_multi_thread_tile_init(VP9_COMP * cpi)160 void vp9_multi_thread_tile_init(VP9_COMP *cpi) {
161   VP9_COMMON *const cm = &cpi->common;
162   const int tile_cols = 1 << cm->log2_tile_cols;
163   const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
164   int i;
165 
166   for (i = 0; i < tile_cols; i++) {
167     TileDataEnc *this_tile = &cpi->tile_data[i];
168     int jobs_per_tile_col = cpi->oxcf.pass == 1 ? cm->mb_rows : sb_rows;
169 
170     // Initialize cur_col to -1 for all rows.
171     memset(this_tile->row_mt_sync.cur_col, -1,
172            sizeof(*this_tile->row_mt_sync.cur_col) * jobs_per_tile_col);
173     vp9_zero(this_tile->fp_data);
174     this_tile->fp_data.image_data_start_row = INVALID_ROW;
175   }
176 }
177 
vp9_assign_tile_to_thread(MultiThreadHandle * multi_thread_ctxt,int tile_cols,int num_workers)178 void vp9_assign_tile_to_thread(MultiThreadHandle *multi_thread_ctxt,
179                                int tile_cols, int num_workers) {
180   int tile_id = 0;
181   int i;
182 
183   // Allocating the threads for the tiles
184   for (i = 0; i < num_workers; i++) {
185     multi_thread_ctxt->thread_id_to_tile_id[i] = tile_id++;
186     if (tile_id == tile_cols) tile_id = 0;
187   }
188 }
189 
vp9_get_job_queue_status(MultiThreadHandle * multi_thread_ctxt,int cur_tile_id)190 int vp9_get_job_queue_status(MultiThreadHandle *multi_thread_ctxt,
191                              int cur_tile_id) {
192   RowMTInfo *row_mt_info;
193   JobQueueHandle *job_queue_hndl;
194 #if CONFIG_MULTITHREAD
195   pthread_mutex_t *mutex;
196 #endif
197   int num_jobs_remaining;
198 
199   row_mt_info = &multi_thread_ctxt->row_mt_info[cur_tile_id];
200   job_queue_hndl = &row_mt_info->job_queue_hdl;
201 #if CONFIG_MULTITHREAD
202   mutex = &row_mt_info->job_mutex;
203 #endif
204 
205 #if CONFIG_MULTITHREAD
206   pthread_mutex_lock(mutex);
207 #endif
208   num_jobs_remaining =
209       multi_thread_ctxt->jobs_per_tile_col - job_queue_hndl->num_jobs_acquired;
210 #if CONFIG_MULTITHREAD
211   pthread_mutex_unlock(mutex);
212 #endif
213 
214   return (num_jobs_remaining);
215 }
216 
vp9_prepare_job_queue(VP9_COMP * cpi,JOB_TYPE job_type)217 void vp9_prepare_job_queue(VP9_COMP *cpi, JOB_TYPE job_type) {
218   VP9_COMMON *const cm = &cpi->common;
219   MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
220   JobQueue *job_queue = multi_thread_ctxt->job_queue;
221   const int tile_cols = 1 << cm->log2_tile_cols;
222   int job_row_num, jobs_per_tile, jobs_per_tile_col, total_jobs;
223   const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
224   int tile_col, i;
225 
226   jobs_per_tile_col = (job_type != ENCODE_JOB) ? cm->mb_rows : sb_rows;
227   total_jobs = jobs_per_tile_col * tile_cols;
228 
229   multi_thread_ctxt->jobs_per_tile_col = jobs_per_tile_col;
230   // memset the entire job queue buffer to zero
231   memset(job_queue, 0, total_jobs * sizeof(JobQueue));
232 
233   // Job queue preparation
234   for (tile_col = 0; tile_col < tile_cols; tile_col++) {
235     RowMTInfo *tile_ctxt = &multi_thread_ctxt->row_mt_info[tile_col];
236     JobQueue *job_queue_curr, *job_queue_temp;
237     int tile_row = 0;
238 
239     tile_ctxt->job_queue_hdl.next = (void *)job_queue;
240     tile_ctxt->job_queue_hdl.num_jobs_acquired = 0;
241 
242     job_queue_curr = job_queue;
243     job_queue_temp = job_queue;
244 
245     // loop over all the vertical rows
246     for (job_row_num = 0, jobs_per_tile = 0; job_row_num < jobs_per_tile_col;
247          job_row_num++, jobs_per_tile++) {
248       job_queue_curr->job_info.vert_unit_row_num = job_row_num;
249       job_queue_curr->job_info.tile_col_id = tile_col;
250       job_queue_curr->job_info.tile_row_id = tile_row;
251       job_queue_curr->next = (void *)(job_queue_temp + 1);
252       job_queue_curr = ++job_queue_temp;
253 
254       if (ENCODE_JOB == job_type) {
255         if (jobs_per_tile >=
256             multi_thread_ctxt->num_tile_vert_sbs[tile_row] - 1) {
257           tile_row++;
258           jobs_per_tile = -1;
259         }
260       }
261     }
262 
263     // Set the last pointer to NULL
264     job_queue_curr += -1;
265     job_queue_curr->next = (void *)NULL;
266 
267     // Move to the next tile
268     job_queue += jobs_per_tile_col;
269   }
270 
271   for (i = 0; i < cpi->num_workers; i++) {
272     EncWorkerData *thread_data;
273     thread_data = &cpi->tile_thr_data[i];
274     thread_data->thread_id = i;
275 
276     for (tile_col = 0; tile_col < tile_cols; tile_col++)
277       thread_data->tile_completion_status[tile_col] = 0;
278   }
279 }
280 
vp9_get_tiles_proc_status(MultiThreadHandle * multi_thread_ctxt,int * tile_completion_status,int * cur_tile_id,int tile_cols)281 int vp9_get_tiles_proc_status(MultiThreadHandle *multi_thread_ctxt,
282                               int *tile_completion_status, int *cur_tile_id,
283                               int tile_cols) {
284   int tile_col;
285   int tile_id = -1;  // Stores the tile ID with minimum proc done
286   int max_num_jobs_remaining = 0;
287   int num_jobs_remaining;
288 
289   // Mark the completion to avoid check in the loop
290   tile_completion_status[*cur_tile_id] = 1;
291   // Check for the status of all the tiles
292   for (tile_col = 0; tile_col < tile_cols; tile_col++) {
293     if (tile_completion_status[tile_col] == 0) {
294       num_jobs_remaining =
295           vp9_get_job_queue_status(multi_thread_ctxt, tile_col);
296       // Mark the completion to avoid checks during future switches across tiles
297       if (num_jobs_remaining == 0) tile_completion_status[tile_col] = 1;
298       if (num_jobs_remaining > max_num_jobs_remaining) {
299         max_num_jobs_remaining = num_jobs_remaining;
300         tile_id = tile_col;
301       }
302     }
303   }
304 
305   if (-1 == tile_id) {
306     return 1;
307   } else {
308     // Update the cur ID to the next tile ID that will be processed,
309     // which will be the least processed tile
310     *cur_tile_id = tile_id;
311     return 0;
312   }
313 }
314