ref: dafe064289a917977439ab6f4f002b9946496084
parent: f1d44c1f45bb0828ef4a683ad42544b3c6fc3af1
author: Supradeep T R <supradeep.tr@ittiam.com>
date: Tue Jun 12 09:57:39 EDT 2018
Loopfilter MultiThread Optimization
Adding LPF within the tileworker hook. This means that LPF will be done
immediately after decode, without waiting for all threads to sync.
Performance Improvement -
Platform Resolution 2 Threads 4 Threads
X86 720p 7.24% 22.04%
1080p 5.29% 17.02%
ARM 720p 4.61% 8.75%
1080p 5.55% 12.03%
x86 Improvement measured on Intel Core i7-6700 CPU @ 2.10GHz set
in performance with turbo mode off
ARM Improvement measured on Nexus 6 Snapdragon 805 Quad-core @ 2.65 GHz
Change-Id: Ifa73c71b40db3fa7fa16f54f4e3aa06d1258caae
--- a/vp9/common/vp9_onyxc_int.h
+++ b/vp9/common/vp9_onyxc_int.h
@@ -256,6 +256,8 @@
PARTITION_CONTEXT *above_seg_context;
ENTROPY_CONTEXT *above_context;
int above_context_alloc_cols;
+
+ int lf_row;
} VP9_COMMON;
static INLINE YV12_BUFFER_CONFIG *get_ref_frame(VP9_COMMON *cm, int index) {--- a/vp9/common/vp9_thread_common.c
+++ b/vp9/common/vp9_thread_common.c
@@ -229,6 +229,26 @@
workers, num_workers, lf_sync);
}
+void vp9_lpf_mt_init(VP9LfSync *lf_sync, VP9_COMMON *cm, int frame_filter_level,
+ int num_workers) {+ const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
+
+ if (!frame_filter_level) return;
+
+ if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
+ num_workers > lf_sync->num_workers) {+ vp9_loop_filter_dealloc(lf_sync);
+ vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
+ }
+
+ // Initialize cur_sb_col to -1 for all SB rows.
+ memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
+
+ memset(lf_sync->num_tiles_done, 0,
+ sizeof(*lf_sync->num_tiles_done) * sb_rows);
+ cm->lf_row = 0;
+}
+
// Set up nsync by width.
static INLINE int get_sync_range(int width) {// nsync numbers are picked by testing. For example, for 4k
@@ -266,6 +286,25 @@
pthread_cond_init(&lf_sync->cond_[i], NULL);
}
}
+ pthread_mutex_init(&lf_sync->lf_mutex, NULL);
+
+ CHECK_MEM_ERROR(cm, lf_sync->recon_done_mutex,
+ vpx_malloc(sizeof(*lf_sync->recon_done_mutex) * rows));
+ if (lf_sync->recon_done_mutex) {+ int i;
+ for (i = 0; i < rows; ++i) {+ pthread_mutex_init(&lf_sync->recon_done_mutex[i], NULL);
+ }
+ }
+
+ CHECK_MEM_ERROR(cm, lf_sync->recon_done_cond,
+ vpx_malloc(sizeof(*lf_sync->recon_done_cond) * rows));
+ if (lf_sync->recon_done_cond) {+ int i;
+ for (i = 0; i < rows; ++i) {+ pthread_cond_init(&lf_sync->recon_done_cond[i], NULL);
+ }
+ }
}
#endif // CONFIG_MULTITHREAD
@@ -276,6 +315,11 @@
CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
+ CHECK_MEM_ERROR(cm, lf_sync->num_tiles_done,
+ vpx_malloc(sizeof(*lf_sync->num_tiles_done) *
+ mi_cols_aligned_to_sb(cm->mi_rows) >>
+ MI_BLOCK_SIZE_LOG2));
+
// Set up nsync.
lf_sync->sync_range = get_sync_range(width);
}
@@ -298,13 +342,116 @@
}
vpx_free(lf_sync->cond_);
}
+ if (lf_sync->recon_done_mutex != NULL) {+ int i;
+ for (i = 0; i < lf_sync->rows; ++i) {+ pthread_mutex_destroy(&lf_sync->recon_done_mutex[i]);
+ }
+ vpx_free(lf_sync->recon_done_mutex);
+ }
+
+ pthread_mutex_destroy(&lf_sync->lf_mutex);
+ if (lf_sync->recon_done_cond != NULL) {+ int i;
+ for (i = 0; i < lf_sync->rows; ++i) {+ pthread_cond_destroy(&lf_sync->recon_done_cond[i]);
+ }
+ vpx_free(lf_sync->recon_done_cond);
+ }
#endif // CONFIG_MULTITHREAD
+
vpx_free(lf_sync->lfdata);
vpx_free(lf_sync->cur_sb_col);
+ vpx_free(lf_sync->num_tiles_done);
// clear the structure as the source of this call may be a resize in which
// case this call will be followed by an _alloc() which may fail.
vp9_zero(*lf_sync);
}
+}
+
+static int get_next_row(VP9_COMMON *cm, VP9LfSync *lf_sync) {+ int return_val = -1;
+ int cur_row;
+ const int max_rows = cm->mi_rows;
+
+#if CONFIG_MULTITHREAD
+ const int tile_cols = 1 << cm->log2_tile_cols;
+
+ pthread_mutex_lock(&lf_sync->lf_mutex);
+ if (cm->lf_row < max_rows) {+ cur_row = cm->lf_row >> MI_BLOCK_SIZE_LOG2;
+ return_val = cm->lf_row;
+ cm->lf_row += MI_BLOCK_SIZE;
+ if (cm->lf_row < max_rows) {+ /* If this is not the last row, make sure the next row is also decoded.
+ * This is because the intra predict has to happen before loop filter */
+ cur_row += 1;
+ }
+ }
+ pthread_mutex_unlock(&lf_sync->lf_mutex);
+
+ if (return_val == -1) return return_val;
+
+ pthread_mutex_lock(&lf_sync->recon_done_mutex[cur_row]);
+ if (lf_sync->num_tiles_done[cur_row] < tile_cols) {+ pthread_cond_wait(&lf_sync->recon_done_cond[cur_row],
+ &lf_sync->recon_done_mutex[cur_row]);
+ }
+ pthread_mutex_unlock(&lf_sync->recon_done_mutex[cur_row]);
+#else
+ (void)lf_sync;
+ if (cm->lf_row < max_rows) {+ cur_row = cm->lf_row >> MI_BLOCK_SIZE_LOG2;
+ return_val = cm->lf_row;
+ cm->lf_row += MI_BLOCK_SIZE;
+ if (cm->lf_row < max_rows) {+ /* If this is not the last row, make sure the next row is also decoded.
+ * This is because the intra predict has to happen before loop filter */
+ cur_row += 1;
+ }
+ }
+#endif // CONFIG_MULTITHREAD
+
+ return return_val;
+}
+
+void vp9_loopfilter_rows(LFWorkerData *lf_data, VP9LfSync *lf_sync,
+ MACROBLOCKD *xd) {+ int mi_row;
+ VP9_COMMON *cm = lf_data->cm;
+
+ while (!xd->corrupted && (mi_row = get_next_row(cm, lf_sync)) != -1 &&
+ mi_row < cm->mi_rows) {+ lf_data->start = mi_row;
+ lf_data->stop = mi_row + MI_BLOCK_SIZE;
+
+ thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
+ lf_data->start, lf_data->stop, lf_data->y_only,
+ lf_sync);
+ }
+}
+
+void vp9_set_row(VP9LfSync *lf_sync, int num_tiles, int row, int is_last_row) {+#if CONFIG_MULTITHREAD
+ pthread_mutex_lock(&lf_sync->recon_done_mutex[row]);
+ lf_sync->num_tiles_done[row] += 1;
+ if (num_tiles == lf_sync->num_tiles_done[row]) {+ if (is_last_row) {+ /* The last 2 rows wait on the last row to be done.
+ * So, we have to broadcast the signal in this case.
+ */
+ pthread_cond_broadcast(&lf_sync->recon_done_cond[row]);
+ } else {+ pthread_cond_signal(&lf_sync->recon_done_cond[row]);
+ }
+ }
+ pthread_mutex_unlock(&lf_sync->recon_done_mutex[row]);
+#else
+ (void)lf_sync;
+ (void)num_tiles;
+ (void)row;
+ (void)is_last_row;
+#endif // CONFIG_MULTITHREAD
}
// Accumulate frame counts.
--- a/vp9/common/vp9_thread_common.h
+++ b/vp9/common/vp9_thread_common.h
@@ -37,6 +37,13 @@
// Row-based parallel loopfilter data
LFWorkerData *lfdata;
int num_workers;
+
+#if CONFIG_MULTITHREAD
+ pthread_mutex_t lf_mutex;
+ pthread_mutex_t *recon_done_mutex;
+ pthread_cond_t *recon_done_cond;
+#endif
+ int *num_tiles_done;
} VP9LfSync;
// Allocate memory for loopfilter row synchronization.
@@ -52,6 +59,17 @@
int frame_filter_level, int y_only,
int partial_frame, VPxWorker *workers,
int num_workers, VP9LfSync *lf_sync);
+
+// Multi-threaded loopfilter initialisations
+void vp9_lpf_mt_init(VP9LfSync *lf_sync, struct VP9Common *cm,
+ int frame_filter_level, int num_workers);
+
+void vp9_loopfilter_rows(LFWorkerData *lf_data, VP9LfSync *lf_sync,
+ MACROBLOCKD *xd);
+
+void vp9_set_row(VP9LfSync *lf_sync, int num_tiles, int row, int is_last_row);
+
+void vp9_set_last_decoded_row(struct VP9Common *cm, int tile_col, int mi_row);
void vp9_accumulate_frame_counts(struct FRAME_COUNTS *accum,
const struct FRAME_COUNTS *counts, int is_dec);
--- a/vp9/decoder/vp9_decodeframe.c
+++ b/vp9/decoder/vp9_decodeframe.c
@@ -1487,6 +1487,11 @@
TileInfo *volatile tile = &tile_data->xd.tile;
const int final_col = (1 << pbi->common.log2_tile_cols) - 1;
const uint8_t *volatile bit_reader_end = NULL;
+ VP9_COMMON *cm = &pbi->common;
+
+ LFWorkerData *lf_data = tile_data->lf_data;
+ VP9LfSync *lf_sync = tile_data->lf_sync;
+
volatile int n = tile_data->buf_start;
tile_data->error_info.setjmp = 1;
@@ -1519,6 +1524,13 @@
mi_col += MI_BLOCK_SIZE) {decode_partition(tile_data, pbi, mi_row, mi_col, BLOCK_64X64, 4);
}
+ if (cm->lf.filter_level && !cm->skip_loop_filter) {+ const int aligned_rows = mi_cols_aligned_to_sb(cm->mi_rows);
+ const int sb_rows = (aligned_rows >> MI_BLOCK_SIZE_LOG2);
+ const int is_last_row = (sb_rows - 1 == mi_row >> MI_BLOCK_SIZE_LOG2);
+ vp9_set_row(lf_sync, 1 << cm->log2_tile_cols,
+ mi_row >> MI_BLOCK_SIZE_LOG2, is_last_row);
+ }
}
if (buf->col == final_col) {@@ -1526,6 +1538,11 @@
}
} while (!tile_data->xd.corrupted && ++n <= tile_data->buf_end);
+ if (!tile_data->xd.corrupted && cm->lf.filter_level &&
+ !cm->skip_loop_filter) {+ vp9_loopfilter_rows(lf_data, lf_sync, &tile_data->xd);
+ }
+
tile_data->data_end = bit_reader_end;
return !tile_data->xd.corrupted;
}
@@ -1542,6 +1559,8 @@
VP9_COMMON *const cm = &pbi->common;
const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
const uint8_t *bit_reader_end = NULL;
+ VP9LfSync *lf_row_sync = &pbi->lf_row_sync;
+ YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
const int tile_cols = 1 << cm->log2_tile_cols;
const int tile_rows = 1 << cm->log2_tile_rows;
@@ -1568,6 +1587,12 @@
}
}
+ // Initialize LPF
+ if (cm->lf.filter_level && !cm->skip_loop_filter) {+ vp9_lpf_mt_init(lf_row_sync, cm, cm->lf.filter_level,
+ pbi->num_tile_workers);
+ }
+
// Reset tile decoding hook
for (n = 0; n < num_workers; ++n) {VPxWorker *const worker = &pbi->tile_workers[n];
@@ -1574,6 +1599,14 @@
TileWorkerData *const tile_data =
&pbi->tile_worker_data[n + pbi->total_tiles];
winterface->sync(worker);
+
+ if (cm->lf.filter_level && !cm->skip_loop_filter) {+ tile_data->lf_sync = lf_row_sync;
+ tile_data->lf_data = &tile_data->lf_sync->lfdata[n];
+ vp9_loop_filter_data_reset(tile_data->lf_data, new_fb, cm, pbi->mb.plane);
+ tile_data->lf_data->y_only = 0;
+ }
+
tile_data->xd = pbi->mb;
tile_data->xd.counts =
cm->frame_parallel_decoding_mode ? NULL : &tile_data->counts;
@@ -2094,15 +2127,7 @@
if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1) {// Multi-threaded tile decoder
*p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
- if (!xd->corrupted) {- if (!cm->skip_loop_filter) {- // If multiple threads are used to decode tiles, then we use those
- // threads to do parallel loopfiltering.
- vp9_loop_filter_frame_mt(new_fb, cm, pbi->mb.plane, cm->lf.filter_level,
- 0, 0, pbi->tile_workers, pbi->num_tile_workers,
- &pbi->lf_row_sync);
- }
- } else {+ if (xd->corrupted) {vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
"Decode failed. Frame data is corrupted.");
}
--- a/vp9/decoder/vp9_decoder.h
+++ b/vp9/decoder/vp9_decoder.h
@@ -37,6 +37,8 @@
int buf_start, buf_end; // pbi->tile_buffers to decode, inclusive
vpx_reader bit_reader;
FRAME_COUNTS counts;
+ LFWorkerData *lf_data;
+ VP9LfSync *lf_sync;
DECLARE_ALIGNED(16, MACROBLOCKD, xd);
/* dqcoeff are shared by all the planes. So planes must be decoded serially */
DECLARE_ALIGNED(16, tran_low_t, dqcoeff[32 * 32]);
--- a/vpx_util/vpx_thread.h
+++ b/vpx_util/vpx_thread.h
@@ -159,6 +159,23 @@
return 0;
}
+static INLINE int pthread_cond_broadcast(pthread_cond_t *const condition) {+ int ok = 1;
+#ifdef USE_WINDOWS_CONDITION_VARIABLE
+ WakeAllConditionVariable(condition);
+#else
+ while (WaitForSingleObject(condition->waiting_sem_, 0) == WAIT_OBJECT_0) {+ // a thread is waiting in pthread_cond_wait: allow it to be notified
+ ok &= SetEvent(condition->signal_event_);
+ // wait until the event is consumed so the signaler cannot consume
+ // the event via its own pthread_cond_wait.
+ ok &= (WaitForSingleObject(condition->received_sem_, INFINITE) !=
+ WAIT_OBJECT_0);
+ }
+#endif
+ return !ok;
+}
+
static INLINE int pthread_cond_signal(pthread_cond_t *const condition) {int ok = 1;
#ifdef USE_WINDOWS_CONDITION_VARIABLE