vpx/vp9/encoder/vp9_multi_thread.c
Ranjit Kumar Tulabandu 71061e9332 Row based multi-threading of encoding stage
(Yunqing Wang)
This patch implements the row-based multi-threading within tiles in
the encoding pass, and substantially speeds up the multi-threaded
encoder in VP9.

Speed tests at speed 1 on STDHD(using 4 tiles) set show that the
average speedups of the encoding pass(second pass in the 2-pass
encoding) is 7% while using 2 threads, 16% while using 4 threads,
85% while using 8 threads, and 116% while using 16 threads.

Change-Id: I12e41dbc171951958af9e6d098efd6e2c82827de
2017-02-15 00:49:34 +00:00

323 lines
11 KiB
C

/*
* Copyright (c) 2017 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include <assert.h>
#include "vp9/encoder/vp9_encoder.h"
#include "vp9/encoder/vp9_ethread.h"
#include "vp9/encoder/vp9_multi_thread.h"
void *vp9_enc_grp_get_next_job(MultiThreadHandle *multi_thread_ctxt,
int tile_id) {
RowMTInfo *row_mt_info;
JobQueueHandle *job_queue_hdl = NULL;
void *next = NULL;
JobNode *job_info = NULL;
#if CONFIG_MULTITHREAD
pthread_mutex_t *mutex_handle = NULL;
#endif
row_mt_info = (RowMTInfo *)(&multi_thread_ctxt->row_mt_info[tile_id]);
job_queue_hdl = (JobQueueHandle *)&row_mt_info->job_queue_hdl;
#if CONFIG_MULTITHREAD
mutex_handle = &row_mt_info->job_mutex;
#endif
// lock the mutex for queue access
#if CONFIG_MULTITHREAD
pthread_mutex_lock(mutex_handle);
#endif
next = job_queue_hdl->next;
if (NULL != next) {
JobQueue *job_queue = (JobQueue *)next;
job_info = &job_queue->job_info;
// Update the next job in the queue
job_queue_hdl->next = job_queue->next;
job_queue_hdl->num_jobs_acquired++;
}
#if CONFIG_MULTITHREAD
pthread_mutex_unlock(mutex_handle);
#endif
return job_info;
}
void vp9_row_mt_mem_alloc(VP9_COMP *cpi) {
struct VP9Common *cm = &cpi->common;
MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
int tile_row, tile_col;
const int tile_cols = 1 << cm->log2_tile_cols;
const int tile_rows = 1 << cm->log2_tile_rows;
const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
int jobs_per_tile_col, total_jobs;
jobs_per_tile_col = VPXMAX(cm->mb_rows, sb_rows);
// Calculate the total number of jobs
total_jobs = jobs_per_tile_col * tile_cols;
multi_thread_ctxt->allocated_tile_cols = tile_cols;
multi_thread_ctxt->allocated_tile_rows = tile_rows;
multi_thread_ctxt->allocated_vert_unit_rows = jobs_per_tile_col;
multi_thread_ctxt->job_queue =
(JobQueue *)vpx_memalign(32, total_jobs * sizeof(JobQueue));
#if CONFIG_MULTITHREAD
// Create mutex for each tile
for (tile_col = 0; tile_col < tile_cols; tile_col++) {
RowMTInfo *row_mt_info = &multi_thread_ctxt->row_mt_info[tile_col];
pthread_mutex_init(&row_mt_info->job_mutex, NULL);
}
#endif
// Allocate memory for row based multi-threading
for (tile_col = 0; tile_col < tile_cols; tile_col++) {
TileDataEnc *this_tile = &cpi->tile_data[tile_col];
vp9_row_mt_sync_mem_alloc(&this_tile->row_mt_sync, cm, jobs_per_tile_col);
}
// Assign the sync pointer of tile row zero for every tile row > 0
for (tile_row = 1; tile_row < tile_rows; tile_row++) {
for (tile_col = 0; tile_col < tile_cols; tile_col++) {
TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
TileDataEnc *this_col_tile = &cpi->tile_data[tile_col];
this_tile->row_mt_sync = this_col_tile->row_mt_sync;
}
}
// Calculate the number of vertical units in the given tile row
for (tile_row = 0; tile_row < tile_rows; tile_row++) {
TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols];
TileInfo *tile_info = &this_tile->tile_info;
multi_thread_ctxt->num_tile_vert_sbs[tile_row] =
get_num_vert_units(*tile_info, MI_BLOCK_SIZE_LOG2);
}
#if CONFIG_MULTITHREAD
for (tile_row = 0; tile_row < tile_rows; tile_row++) {
for (tile_col = 0; tile_col < tile_cols; tile_col++) {
TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
CHECK_MEM_ERROR(cm, this_tile->search_count_mutex,
vpx_malloc(sizeof(*this_tile->search_count_mutex)));
pthread_mutex_init(this_tile->search_count_mutex, NULL);
CHECK_MEM_ERROR(cm, this_tile->enc_row_mt_mutex,
vpx_malloc(sizeof(*this_tile->enc_row_mt_mutex)));
pthread_mutex_init(this_tile->enc_row_mt_mutex, NULL);
}
}
#endif
}
void vp9_row_mt_mem_dealloc(VP9_COMP *cpi) {
MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
int tile_col;
#if CONFIG_MULTITHREAD
int tile_row;
#endif
// Deallocate memory for job queue
if (multi_thread_ctxt->job_queue) vpx_free(multi_thread_ctxt->job_queue);
#if CONFIG_MULTITHREAD
// Destroy mutex for each tile
for (tile_col = 0; tile_col < multi_thread_ctxt->allocated_tile_cols;
tile_col++) {
RowMTInfo *row_mt_info = &multi_thread_ctxt->row_mt_info[tile_col];
if (row_mt_info) pthread_mutex_destroy(&row_mt_info->job_mutex);
}
#endif
// Free row based multi-threading sync memory
for (tile_col = 0; tile_col < multi_thread_ctxt->allocated_tile_cols;
tile_col++) {
TileDataEnc *this_tile = &cpi->tile_data[tile_col];
vp9_row_mt_sync_mem_dealloc(&this_tile->row_mt_sync);
}
#if CONFIG_MULTITHREAD
for (tile_row = 0; tile_row < multi_thread_ctxt->allocated_tile_rows;
tile_row++) {
for (tile_col = 0; tile_col < multi_thread_ctxt->allocated_tile_cols;
tile_col++) {
TileDataEnc *this_tile =
&cpi->tile_data[tile_row * multi_thread_ctxt->allocated_tile_cols +
tile_col];
pthread_mutex_destroy(this_tile->search_count_mutex);
vpx_free(this_tile->search_count_mutex);
this_tile->search_count_mutex = NULL;
pthread_mutex_destroy(this_tile->enc_row_mt_mutex);
vpx_free(this_tile->enc_row_mt_mutex);
this_tile->enc_row_mt_mutex = NULL;
}
}
#endif
}
void vp9_multi_thread_tile_init(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
const int tile_cols = 1 << cm->log2_tile_cols;
const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
int i;
for (i = 0; i < tile_cols; i++) {
TileDataEnc *this_tile = &cpi->tile_data[i];
int jobs_per_tile_col = cpi->oxcf.pass == 1 ? cm->mb_rows : sb_rows;
// Initialize cur_col to -1 for all rows.
memset(this_tile->row_mt_sync.cur_col, -1,
sizeof(*this_tile->row_mt_sync.cur_col) * jobs_per_tile_col);
vp9_zero(this_tile->fp_data);
this_tile->fp_data.image_data_start_row = INVALID_ROW;
}
}
void vp9_assign_tile_to_thread(MultiThreadHandle *multi_thread_ctxt,
int tile_cols, int num_workers) {
int tile_id = 0;
int i;
// Allocating the threads for the tiles
for (i = 0; i < num_workers; i++) {
multi_thread_ctxt->thread_id_to_tile_id[i] = tile_id++;
if (tile_id == tile_cols) tile_id = 0;
}
}
int vp9_get_job_queue_status(MultiThreadHandle *multi_thread_ctxt,
int cur_tile_id) {
RowMTInfo *row_mt_info;
JobQueueHandle *job_queue_hndl;
#if CONFIG_MULTITHREAD
pthread_mutex_t *mutex;
#endif
int num_jobs_remaining;
row_mt_info = &multi_thread_ctxt->row_mt_info[cur_tile_id];
job_queue_hndl = &row_mt_info->job_queue_hdl;
#if CONFIG_MULTITHREAD
mutex = &row_mt_info->job_mutex;
#endif
#if CONFIG_MULTITHREAD
pthread_mutex_lock(mutex);
#endif
num_jobs_remaining =
multi_thread_ctxt->jobs_per_tile_col - job_queue_hndl->num_jobs_acquired;
#if CONFIG_MULTITHREAD
pthread_mutex_unlock(mutex);
#endif
return (num_jobs_remaining);
}
void vp9_prepare_job_queue(VP9_COMP *cpi, JOB_TYPE job_type) {
VP9_COMMON *const cm = &cpi->common;
MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
JobQueue *job_queue = multi_thread_ctxt->job_queue;
const int tile_cols = 1 << cm->log2_tile_cols;
int job_row_num, jobs_per_tile, jobs_per_tile_col, total_jobs;
const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
int tile_col, i;
jobs_per_tile_col = (job_type != ENCODE_JOB) ? cm->mb_rows : sb_rows;
total_jobs = jobs_per_tile_col * tile_cols;
multi_thread_ctxt->jobs_per_tile_col = jobs_per_tile_col;
// memset the entire job queue buffer to zero
memset(job_queue, 0, total_jobs * sizeof(JobQueue));
// Job queue preparation
for (tile_col = 0; tile_col < tile_cols; tile_col++) {
RowMTInfo *tile_ctxt = &multi_thread_ctxt->row_mt_info[tile_col];
JobQueue *job_queue_curr, *job_queue_temp;
int tile_row = 0;
tile_ctxt->job_queue_hdl.next = (void *)job_queue;
tile_ctxt->job_queue_hdl.num_jobs_acquired = 0;
job_queue_curr = job_queue;
job_queue_temp = job_queue;
// loop over all the vertical rows
for (job_row_num = 0, jobs_per_tile = 0; job_row_num < jobs_per_tile_col;
job_row_num++, jobs_per_tile++) {
job_queue_curr->job_info.vert_unit_row_num = job_row_num;
job_queue_curr->job_info.tile_col_id = tile_col;
job_queue_curr->job_info.tile_row_id = tile_row;
job_queue_curr->next = (void *)(job_queue_temp + 1);
job_queue_curr = ++job_queue_temp;
if (ENCODE_JOB == job_type) {
if (jobs_per_tile >=
multi_thread_ctxt->num_tile_vert_sbs[tile_row] - 1) {
tile_row++;
jobs_per_tile = -1;
}
}
}
// Set the last pointer to NULL
job_queue_curr += -1;
job_queue_curr->next = (void *)NULL;
// Move to the next tile
job_queue += jobs_per_tile_col;
}
for (i = 0; i < cpi->num_workers; i++) {
EncWorkerData *thread_data;
thread_data = &cpi->tile_thr_data[i];
thread_data->thread_id = i;
for (tile_col = 0; tile_col < tile_cols; tile_col++)
thread_data->tile_completion_status[tile_col] = 0;
}
}
int vp9_get_tiles_proc_status(MultiThreadHandle *multi_thread_ctxt,
int *tile_completion_status, int *cur_tile_id,
int tile_cols) {
int tile_col;
int tile_id = -1; // Stores the tile ID with minimum proc done
int max_num_jobs_remaining = 0;
int num_jobs_remaining;
// Mark the completion to avoid check in the loop
tile_completion_status[*cur_tile_id] = 1;
// Check for the status of all the tiles
for (tile_col = 0; tile_col < tile_cols; tile_col++) {
if (tile_completion_status[tile_col] == 0) {
num_jobs_remaining =
vp9_get_job_queue_status(multi_thread_ctxt, tile_col);
// Mark the completion to avoid checks during future switches across tiles
if (num_jobs_remaining == 0) tile_completion_status[tile_col] = 1;
if (num_jobs_remaining > max_num_jobs_remaining) {
max_num_jobs_remaining = num_jobs_remaining;
tile_id = tile_col;
}
}
}
if (-1 == tile_id) {
return 1;
} else {
// Update the cur ID to the next tile ID that will be processed,
// which will be the least processed tile
*cur_tile_id = tile_id;
return 0;
}
}