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 #include "vp9/encoder/vp9_temporal_filter.h"
17
vp9_enc_grp_get_next_job(MultiThreadHandle * multi_thread_ctxt,int tile_id)18 void *vp9_enc_grp_get_next_job(MultiThreadHandle *multi_thread_ctxt,
19 int tile_id) {
20 RowMTInfo *row_mt_info;
21 JobQueueHandle *job_queue_hdl = NULL;
22 void *next = NULL;
23 JobNode *job_info = NULL;
24 #if CONFIG_MULTITHREAD
25 pthread_mutex_t *mutex_handle = NULL;
26 #endif
27
28 row_mt_info = (RowMTInfo *)(&multi_thread_ctxt->row_mt_info[tile_id]);
29 job_queue_hdl = (JobQueueHandle *)&row_mt_info->job_queue_hdl;
30 #if CONFIG_MULTITHREAD
31 mutex_handle = &row_mt_info->job_mutex;
32 #endif
33
34 // lock the mutex for queue access
35 #if CONFIG_MULTITHREAD
36 pthread_mutex_lock(mutex_handle);
37 #endif
38 next = job_queue_hdl->next;
39 if (NULL != next) {
40 JobQueue *job_queue = (JobQueue *)next;
41 job_info = &job_queue->job_info;
42 // Update the next job in the queue
43 job_queue_hdl->next = job_queue->next;
44 job_queue_hdl->num_jobs_acquired++;
45 }
46
47 #if CONFIG_MULTITHREAD
48 pthread_mutex_unlock(mutex_handle);
49 #endif
50
51 return job_info;
52 }
53
vp9_row_mt_alloc_rd_thresh(VP9_COMP * const cpi,TileDataEnc * const this_tile)54 void vp9_row_mt_alloc_rd_thresh(VP9_COMP *const cpi,
55 TileDataEnc *const this_tile) {
56 VP9_COMMON *const cm = &cpi->common;
57 const int sb_rows =
58 (mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2) + 1;
59 int i;
60
61 this_tile->row_base_thresh_freq_fact =
62 (int *)vpx_calloc(sb_rows * BLOCK_SIZES * MAX_MODES,
63 sizeof(*(this_tile->row_base_thresh_freq_fact)));
64 for (i = 0; i < sb_rows * BLOCK_SIZES * MAX_MODES; i++)
65 this_tile->row_base_thresh_freq_fact[i] = RD_THRESH_INIT_FACT;
66 }
67
vp9_row_mt_mem_alloc(VP9_COMP * cpi)68 void vp9_row_mt_mem_alloc(VP9_COMP *cpi) {
69 struct VP9Common *cm = &cpi->common;
70 MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
71 int tile_row, tile_col;
72 const int tile_cols = 1 << cm->log2_tile_cols;
73 const int tile_rows = 1 << cm->log2_tile_rows;
74 const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
75 int jobs_per_tile_col, total_jobs;
76
77 // Allocate memory that is large enough for all row_mt stages. First pass
78 // uses 16x16 block size.
79 jobs_per_tile_col = VPXMAX(cm->mb_rows, sb_rows);
80 // Calculate the total number of jobs
81 total_jobs = jobs_per_tile_col * tile_cols;
82
83 multi_thread_ctxt->allocated_tile_cols = tile_cols;
84 multi_thread_ctxt->allocated_tile_rows = tile_rows;
85 multi_thread_ctxt->allocated_vert_unit_rows = jobs_per_tile_col;
86
87 multi_thread_ctxt->job_queue =
88 (JobQueue *)vpx_memalign(32, total_jobs * sizeof(JobQueue));
89
90 #if CONFIG_MULTITHREAD
91 // Create mutex for each tile
92 for (tile_col = 0; tile_col < tile_cols; tile_col++) {
93 RowMTInfo *row_mt_info = &multi_thread_ctxt->row_mt_info[tile_col];
94 pthread_mutex_init(&row_mt_info->job_mutex, NULL);
95 }
96 #endif
97
98 // Allocate memory for row based multi-threading
99 for (tile_col = 0; tile_col < tile_cols; tile_col++) {
100 TileDataEnc *this_tile = &cpi->tile_data[tile_col];
101 vp9_row_mt_sync_mem_alloc(&this_tile->row_mt_sync, cm, jobs_per_tile_col);
102 if (cpi->sf.adaptive_rd_thresh_row_mt) {
103 if (this_tile->row_base_thresh_freq_fact != NULL) {
104 vpx_free(this_tile->row_base_thresh_freq_fact);
105 this_tile->row_base_thresh_freq_fact = NULL;
106 }
107 vp9_row_mt_alloc_rd_thresh(cpi, this_tile);
108 }
109 }
110
111 // Assign the sync pointer of tile row zero for every tile row > 0
112 for (tile_row = 1; tile_row < tile_rows; tile_row++) {
113 for (tile_col = 0; tile_col < tile_cols; tile_col++) {
114 TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
115 TileDataEnc *this_col_tile = &cpi->tile_data[tile_col];
116 this_tile->row_mt_sync = this_col_tile->row_mt_sync;
117 }
118 }
119
120 // Calculate the number of vertical units in the given tile row
121 for (tile_row = 0; tile_row < tile_rows; tile_row++) {
122 TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols];
123 TileInfo *tile_info = &this_tile->tile_info;
124 multi_thread_ctxt->num_tile_vert_sbs[tile_row] =
125 get_num_vert_units(*tile_info, MI_BLOCK_SIZE_LOG2);
126 }
127 }
128
vp9_row_mt_mem_dealloc(VP9_COMP * cpi)129 void vp9_row_mt_mem_dealloc(VP9_COMP *cpi) {
130 MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
131 int tile_col;
132 #if CONFIG_MULTITHREAD
133 int tile_row;
134 #endif
135
136 // Deallocate memory for job queue
137 if (multi_thread_ctxt->job_queue) vpx_free(multi_thread_ctxt->job_queue);
138
139 #if CONFIG_MULTITHREAD
140 // Destroy mutex for each tile
141 for (tile_col = 0; tile_col < multi_thread_ctxt->allocated_tile_cols;
142 tile_col++) {
143 RowMTInfo *row_mt_info = &multi_thread_ctxt->row_mt_info[tile_col];
144 if (row_mt_info) pthread_mutex_destroy(&row_mt_info->job_mutex);
145 }
146 #endif
147
148 // Free row based multi-threading sync memory
149 for (tile_col = 0; tile_col < multi_thread_ctxt->allocated_tile_cols;
150 tile_col++) {
151 TileDataEnc *this_tile = &cpi->tile_data[tile_col];
152 vp9_row_mt_sync_mem_dealloc(&this_tile->row_mt_sync);
153 }
154
155 #if CONFIG_MULTITHREAD
156 for (tile_row = 0; tile_row < multi_thread_ctxt->allocated_tile_rows;
157 tile_row++) {
158 for (tile_col = 0; tile_col < multi_thread_ctxt->allocated_tile_cols;
159 tile_col++) {
160 TileDataEnc *this_tile =
161 &cpi->tile_data[tile_row * multi_thread_ctxt->allocated_tile_cols +
162 tile_col];
163 if (this_tile->row_base_thresh_freq_fact != NULL) {
164 vpx_free(this_tile->row_base_thresh_freq_fact);
165 this_tile->row_base_thresh_freq_fact = NULL;
166 }
167 }
168 }
169 #endif
170 }
171
vp9_multi_thread_tile_init(VP9_COMP * cpi)172 void vp9_multi_thread_tile_init(VP9_COMP *cpi) {
173 VP9_COMMON *const cm = &cpi->common;
174 const int tile_cols = 1 << cm->log2_tile_cols;
175 const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
176 int i;
177
178 for (i = 0; i < tile_cols; i++) {
179 TileDataEnc *this_tile = &cpi->tile_data[i];
180 int jobs_per_tile_col = cpi->oxcf.pass == 1 ? cm->mb_rows : sb_rows;
181
182 // Initialize cur_col to -1 for all rows.
183 memset(this_tile->row_mt_sync.cur_col, -1,
184 sizeof(*this_tile->row_mt_sync.cur_col) * jobs_per_tile_col);
185 vp9_zero(this_tile->fp_data);
186 this_tile->fp_data.image_data_start_row = INVALID_ROW;
187 }
188 }
189
vp9_assign_tile_to_thread(MultiThreadHandle * multi_thread_ctxt,int tile_cols,int num_workers)190 void vp9_assign_tile_to_thread(MultiThreadHandle *multi_thread_ctxt,
191 int tile_cols, int num_workers) {
192 int tile_id = 0;
193 int i;
194
195 // Allocating the threads for the tiles
196 for (i = 0; i < num_workers; i++) {
197 multi_thread_ctxt->thread_id_to_tile_id[i] = tile_id++;
198 if (tile_id == tile_cols) tile_id = 0;
199 }
200 }
201
vp9_get_job_queue_status(MultiThreadHandle * multi_thread_ctxt,int cur_tile_id)202 int vp9_get_job_queue_status(MultiThreadHandle *multi_thread_ctxt,
203 int cur_tile_id) {
204 RowMTInfo *row_mt_info;
205 JobQueueHandle *job_queue_hndl;
206 #if CONFIG_MULTITHREAD
207 pthread_mutex_t *mutex;
208 #endif
209 int num_jobs_remaining;
210
211 row_mt_info = &multi_thread_ctxt->row_mt_info[cur_tile_id];
212 job_queue_hndl = &row_mt_info->job_queue_hdl;
213 #if CONFIG_MULTITHREAD
214 mutex = &row_mt_info->job_mutex;
215 #endif
216
217 #if CONFIG_MULTITHREAD
218 pthread_mutex_lock(mutex);
219 #endif
220 num_jobs_remaining =
221 multi_thread_ctxt->jobs_per_tile_col - job_queue_hndl->num_jobs_acquired;
222 #if CONFIG_MULTITHREAD
223 pthread_mutex_unlock(mutex);
224 #endif
225
226 return (num_jobs_remaining);
227 }
228
vp9_prepare_job_queue(VP9_COMP * cpi,JOB_TYPE job_type)229 void vp9_prepare_job_queue(VP9_COMP *cpi, JOB_TYPE job_type) {
230 VP9_COMMON *const cm = &cpi->common;
231 MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
232 JobQueue *job_queue = multi_thread_ctxt->job_queue;
233 const int tile_cols = 1 << cm->log2_tile_cols;
234 int job_row_num, jobs_per_tile, jobs_per_tile_col = 0, total_jobs;
235 const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
236 int tile_col, i;
237
238 switch (job_type) {
239 case ENCODE_JOB: jobs_per_tile_col = sb_rows; break;
240 case FIRST_PASS_JOB: jobs_per_tile_col = cm->mb_rows; break;
241 case ARNR_JOB:
242 jobs_per_tile_col = ((cm->mi_rows + TF_ROUND) >> TF_SHIFT);
243 break;
244 default: assert(0);
245 }
246
247 total_jobs = jobs_per_tile_col * tile_cols;
248
249 multi_thread_ctxt->jobs_per_tile_col = jobs_per_tile_col;
250 // memset the entire job queue buffer to zero
251 memset(job_queue, 0, total_jobs * sizeof(JobQueue));
252
253 // Job queue preparation
254 for (tile_col = 0; tile_col < tile_cols; tile_col++) {
255 RowMTInfo *tile_ctxt = &multi_thread_ctxt->row_mt_info[tile_col];
256 JobQueue *job_queue_curr, *job_queue_temp;
257 int tile_row = 0;
258
259 tile_ctxt->job_queue_hdl.next = (void *)job_queue;
260 tile_ctxt->job_queue_hdl.num_jobs_acquired = 0;
261
262 job_queue_curr = job_queue;
263 job_queue_temp = job_queue;
264
265 // loop over all the vertical rows
266 for (job_row_num = 0, jobs_per_tile = 0; job_row_num < jobs_per_tile_col;
267 job_row_num++, jobs_per_tile++) {
268 job_queue_curr->job_info.vert_unit_row_num = job_row_num;
269 job_queue_curr->job_info.tile_col_id = tile_col;
270 job_queue_curr->job_info.tile_row_id = tile_row;
271 job_queue_curr->next = (void *)(job_queue_temp + 1);
272 job_queue_curr = ++job_queue_temp;
273
274 if (ENCODE_JOB == job_type) {
275 if (jobs_per_tile >=
276 multi_thread_ctxt->num_tile_vert_sbs[tile_row] - 1) {
277 tile_row++;
278 jobs_per_tile = -1;
279 }
280 }
281 }
282
283 // Set the last pointer to NULL
284 job_queue_curr += -1;
285 job_queue_curr->next = (void *)NULL;
286
287 // Move to the next tile
288 job_queue += jobs_per_tile_col;
289 }
290
291 for (i = 0; i < cpi->num_workers; i++) {
292 EncWorkerData *thread_data;
293 thread_data = &cpi->tile_thr_data[i];
294 thread_data->thread_id = i;
295
296 for (tile_col = 0; tile_col < tile_cols; tile_col++)
297 thread_data->tile_completion_status[tile_col] = 0;
298 }
299 }
300
vp9_get_tiles_proc_status(MultiThreadHandle * multi_thread_ctxt,int * tile_completion_status,int * cur_tile_id,int tile_cols)301 int vp9_get_tiles_proc_status(MultiThreadHandle *multi_thread_ctxt,
302 int *tile_completion_status, int *cur_tile_id,
303 int tile_cols) {
304 int tile_col;
305 int tile_id = -1; // Stores the tile ID with minimum proc done
306 int max_num_jobs_remaining = 0;
307 int num_jobs_remaining;
308
309 // Mark the completion to avoid check in the loop
310 tile_completion_status[*cur_tile_id] = 1;
311 // Check for the status of all the tiles
312 for (tile_col = 0; tile_col < tile_cols; tile_col++) {
313 if (tile_completion_status[tile_col] == 0) {
314 num_jobs_remaining =
315 vp9_get_job_queue_status(multi_thread_ctxt, tile_col);
316 // Mark the completion to avoid checks during future switches across tiles
317 if (num_jobs_remaining == 0) tile_completion_status[tile_col] = 1;
318 if (num_jobs_remaining > max_num_jobs_remaining) {
319 max_num_jobs_remaining = num_jobs_remaining;
320 tile_id = tile_col;
321 }
322 }
323 }
324
325 if (-1 == tile_id) {
326 return 1;
327 } else {
328 // Update the cur ID to the next tile ID that will be processed,
329 // which will be the least processed tile
330 *cur_tile_id = tile_id;
331 return 0;
332 }
333 }
334