ref: e76eaf05b1a20b906d166b7e8db402a9883c6025
parent: 1be4267ec21e3235027f83e7442c25e7eefbd836
author: Yunqing Wang <yunqingwang@google.com>
date: Tue Jan 6 09:14:26 EST 2015
vp9_ethread: add parallel loopfilter 1. Added row-based loopfilter in encoder; 2. Moved common multi-threaded loopfilter functions from decoder to common; 3. Merged multi-threaded loopfilter code, and made encoder/ decoder call same function to reduce code duplication. Encoder tests showed that 1% - 2% speedup was seen for good-quality 2-pass mode(at speed 3); 1% - 3% speedup using 2 threads and 4% - 6% speedup using 4 threads were seen for real-time mode(at speed 7). Change-Id: I8a4ac51c2ad9bab9fa7b864e90743931c53ec1c4
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -1149,10 +1149,10 @@
}
#endif // CONFIG_VP9_HIGHBITDEPTH
-static void filter_block_plane_non420(VP9_COMMON *cm,
- struct macroblockd_plane *plane,
- MODE_INFO *mi_8x8,
- int mi_row, int mi_col) {+void vp9_filter_block_plane_non420(VP9_COMMON *cm,
+ struct macroblockd_plane *plane,
+ MODE_INFO *mi_8x8,
+ int mi_row, int mi_col) {const int ss_x = plane->subsampling_x;
const int ss_y = plane->subsampling_y;
const int row_step = 1 << ss_y;
@@ -1598,8 +1598,8 @@
if (use_420)
vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
else
- filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
- mi_row, mi_col);
+ vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
+ mi_row, mi_col);
}
}
}
--- a/vp9/common/vp9_loopfilter.h
+++ b/vp9/common/vp9_loopfilter.h
@@ -97,6 +97,11 @@
int mi_row,
LOOP_FILTER_MASK *lfm);
+void vp9_filter_block_plane_non420(struct VP9Common *cm,
+ struct macroblockd_plane *plane,
+ MODE_INFO *mi_8x8,
+ int mi_row, int mi_col);
+
void vp9_loop_filter_init(struct VP9Common *cm);
// Update the loop filter for the current frame.
--- /dev/null
+++ b/vp9/common/vp9_loopfilter_thread.c
@@ -1,0 +1,291 @@
+/*
+ * Copyright (c) 2014 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 "./vpx_config.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp9/common/vp9_loopfilter_thread.h"
+#include "vp9/common/vp9_reconinter.h"
+
+#if CONFIG_MULTITHREAD
+static INLINE void mutex_lock(pthread_mutex_t *const mutex) {+ const int kMaxTryLocks = 4000;
+ int locked = 0;
+ int i;
+
+ for (i = 0; i < kMaxTryLocks; ++i) {+ if (!pthread_mutex_trylock(mutex)) {+ locked = 1;
+ break;
+ }
+ }
+
+ if (!locked)
+ pthread_mutex_lock(mutex);
+}
+#endif // CONFIG_MULTITHREAD
+
+static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {+#if CONFIG_MULTITHREAD
+ const int nsync = lf_sync->sync_range;
+
+ if (r && !(c & (nsync - 1))) {+ pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
+ mutex_lock(mutex);
+
+ while (c > lf_sync->cur_sb_col[r - 1] - nsync) {+ pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
+ }
+ pthread_mutex_unlock(mutex);
+ }
+#else
+ (void)lf_sync;
+ (void)r;
+ (void)c;
+#endif // CONFIG_MULTITHREAD
+}
+
+static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
+ const int sb_cols) {+#if CONFIG_MULTITHREAD
+ const int nsync = lf_sync->sync_range;
+ int cur;
+ // Only signal when there are enough filtered SB for next row to run.
+ int sig = 1;
+
+ if (c < sb_cols - 1) {+ cur = c;
+ if (c % nsync)
+ sig = 0;
+ } else {+ cur = sb_cols + nsync;
+ }
+
+ if (sig) {+ mutex_lock(&lf_sync->mutex_[r]);
+
+ lf_sync->cur_sb_col[r] = cur;
+
+ pthread_cond_signal(&lf_sync->cond_[r]);
+ pthread_mutex_unlock(&lf_sync->mutex_[r]);
+ }
+#else
+ (void)lf_sync;
+ (void)r;
+ (void)c;
+ (void)sb_cols;
+#endif // CONFIG_MULTITHREAD
+}
+
+// Implement row loopfiltering for each thread.
+static INLINE
+void thread_loop_filter_rows(const YV12_BUFFER_CONFIG *const frame_buffer,
+ VP9_COMMON *const cm,
+ struct macroblockd_plane planes[MAX_MB_PLANE],
+ int start, int stop, int y_only,
+ VP9LfSync *const lf_sync) {+ const int num_planes = y_only ? 1 : MAX_MB_PLANE;
+ const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
+ int mi_row, mi_col;
+
+ for (mi_row = start; mi_row < stop;
+ mi_row += lf_sync->num_workers * MI_BLOCK_SIZE) {+ MODE_INFO *const mi = cm->mi + mi_row * cm->mi_stride;
+
+ for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {+ const int r = mi_row >> MI_BLOCK_SIZE_LOG2;
+ const int c = mi_col >> MI_BLOCK_SIZE_LOG2;
+ LOOP_FILTER_MASK lfm;
+ int plane;
+
+ sync_read(lf_sync, r, c);
+
+ vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
+ vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm);
+
+ for (plane = 0; plane < num_planes; ++plane) {+ vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
+ }
+
+ sync_write(lf_sync, r, c, sb_cols);
+ }
+ }
+}
+
+// Row-based multi-threaded loopfilter hook
+static int loop_filter_row_worker(VP9LfSync *const lf_sync,
+ LFWorkerData *const lf_data) {+ 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);
+ return 1;
+}
+
+static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame,
+ VP9_COMMON *cm,
+ struct macroblockd_plane planes[MAX_MB_PLANE],
+ int start, int stop, int y_only,
+ VP9Worker *workers, int nworkers,
+ VP9LfSync *lf_sync) {+ const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+ // Number of superblock rows and cols
+ const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
+ // Decoder may allocate more threads than number of tiles based on user's
+ // input.
+ const int tile_cols = 1 << cm->log2_tile_cols;
+ const int num_workers = MIN(nworkers, tile_cols);
+ int i;
+
+ if (!lf_sync->sync_range || cm->last_height != cm->height ||
+ 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.
+ vpx_memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
+
+ // Set up loopfilter thread data.
+ // The decoder is capping num_workers because it has been observed that using
+ // more threads on the loopfilter than there are cores will hurt performance
+ // on Android. This is because the system will only schedule the tile decode
+ // workers on cores equal to the number of tile columns. Then if the decoder
+ // tries to use more threads for the loopfilter, it will hurt performance
+ // because of contention. If the multithreading code changes in the future
+ // then the number of workers used by the loopfilter should be revisited.
+ for (i = 0; i < num_workers; ++i) {+ VP9Worker *const worker = &workers[i];
+ LFWorkerData *const lf_data = &lf_sync->lfdata[i];
+
+ worker->hook = (VP9WorkerHook)loop_filter_row_worker;
+ worker->data1 = lf_sync;
+ worker->data2 = lf_data;
+
+ // Loopfilter data
+ vp9_loop_filter_data_reset(lf_data, frame, cm, planes);
+ lf_data->start = start + i * MI_BLOCK_SIZE;
+ lf_data->stop = stop;
+ lf_data->y_only = y_only;
+
+ // Start loopfiltering
+ if (i == num_workers - 1) {+ winterface->execute(worker);
+ } else {+ winterface->launch(worker);
+ }
+ }
+
+ // Wait till all rows are finished
+ for (i = 0; i < num_workers; ++i) {+ winterface->sync(&workers[i]);
+ }
+}
+
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+ VP9_COMMON *cm,
+ struct macroblockd_plane planes[MAX_MB_PLANE],
+ int frame_filter_level,
+ int y_only, int partial_frame,
+ VP9Worker *workers, int num_workers,
+ VP9LfSync *lf_sync) {+ int start_mi_row, end_mi_row, mi_rows_to_filter;
+
+ if (!frame_filter_level) return;
+
+ start_mi_row = 0;
+ mi_rows_to_filter = cm->mi_rows;
+ if (partial_frame && cm->mi_rows > 8) {+ start_mi_row = cm->mi_rows >> 1;
+ start_mi_row &= 0xfffffff8;
+ mi_rows_to_filter = MAX(cm->mi_rows / 8, 8);
+ }
+ end_mi_row = start_mi_row + mi_rows_to_filter;
+ vp9_loop_filter_frame_init(cm, frame_filter_level);
+
+ loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row,
+ y_only, workers, num_workers, lf_sync);
+}
+
+// Set up nsync by width.
+static INLINE int get_sync_range(int width) {+ // nsync numbers are picked by testing. For example, for 4k
+ // video, using 4 gives best performance.
+ if (width < 640)
+ return 1;
+ else if (width <= 1280)
+ return 2;
+ else if (width <= 4096)
+ return 4;
+ else
+ return 8;
+}
+
+// Allocate memory for lf row synchronization
+void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
+ int width, int num_workers) {+ lf_sync->rows = rows;
+#if CONFIG_MULTITHREAD
+ {+ int i;
+
+ CHECK_MEM_ERROR(cm, lf_sync->mutex_,
+ vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
+ if (lf_sync->mutex_) {+ for (i = 0; i < rows; ++i) {+ pthread_mutex_init(&lf_sync->mutex_[i], NULL);
+ }
+ }
+
+ CHECK_MEM_ERROR(cm, lf_sync->cond_,
+ vpx_malloc(sizeof(*lf_sync->cond_) * rows));
+ if (lf_sync->cond_) {+ for (i = 0; i < rows; ++i) {+ pthread_cond_init(&lf_sync->cond_[i], NULL);
+ }
+ }
+ }
+#endif // CONFIG_MULTITHREAD
+
+ CHECK_MEM_ERROR(cm, lf_sync->lfdata,
+ vpx_malloc(num_workers * sizeof(*lf_sync->lfdata)));
+ lf_sync->num_workers = num_workers;
+
+ CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
+ vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
+
+ // Set up nsync.
+ lf_sync->sync_range = get_sync_range(width);
+}
+
+// Deallocate lf synchronization related mutex and data
+void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) {+ if (lf_sync != NULL) {+#if CONFIG_MULTITHREAD
+ int i;
+
+ if (lf_sync->mutex_ != NULL) {+ for (i = 0; i < lf_sync->rows; ++i) {+ pthread_mutex_destroy(&lf_sync->mutex_[i]);
+ }
+ vpx_free(lf_sync->mutex_);
+ }
+ if (lf_sync->cond_ != NULL) {+ for (i = 0; i < lf_sync->rows; ++i) {+ pthread_cond_destroy(&lf_sync->cond_[i]);
+ }
+ vpx_free(lf_sync->cond_);
+ }
+#endif // CONFIG_MULTITHREAD
+ vpx_free(lf_sync->lfdata);
+ vpx_free(lf_sync->cur_sb_col);
+ // 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);
+ }
+}
--- /dev/null
+++ b/vp9/common/vp9_loopfilter_thread.h
@@ -1,0 +1,53 @@
+/*
+ * Copyright (c) 2014 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.
+ */
+
+#ifndef VP9_COMMON_VP9_LOOPFILTER_THREAD_H_
+#define VP9_COMMON_VP9_LOOPFILTER_THREAD_H_
+#include "./vpx_config.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_thread.h"
+
+struct VP9Common;
+
+// Loopfilter row synchronization
+typedef struct VP9LfSyncData {+#if CONFIG_MULTITHREAD
+ pthread_mutex_t *mutex_;
+ pthread_cond_t *cond_;
+#endif
+ // Allocate memory to store the loop-filtered superblock index in each row.
+ int *cur_sb_col;
+ // The optimal sync_range for different resolution and platform should be
+ // determined by testing. Currently, it is chosen to be a power-of-2 number.
+ int sync_range;
+ int rows;
+
+ // Row-based parallel loopfilter data
+ LFWorkerData *lfdata;
+ int num_workers;
+} VP9LfSync;
+
+// Allocate memory for loopfilter row synchronization.
+void vp9_loop_filter_alloc(VP9LfSync *lf_sync, struct VP9Common *cm, int rows,
+ int width, int num_workers);
+
+// Deallocate loopfilter synchronization related mutex and data.
+void vp9_loop_filter_dealloc(VP9LfSync *lf_sync);
+
+// Multi-threaded loopfilter that uses the tile threads.
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+ struct VP9Common *cm,
+ struct macroblockd_plane planes[MAX_MB_PLANE],
+ int frame_filter_level,
+ int y_only, int partial_frame,
+ VP9Worker *workers, int num_workers,
+ VP9LfSync *lf_sync);
+
+#endif // VP9_COMMON_VP9_LOOPFILTER_THREAD_H_
--- a/vp9/decoder/vp9_decodeframe.c
+++ b/vp9/decoder/vp9_decodeframe.c
@@ -36,7 +36,6 @@
#include "vp9/decoder/vp9_decodemv.h"
#include "vp9/decoder/vp9_decoder.h"
#include "vp9/decoder/vp9_dsubexp.h"
-#include "vp9/decoder/vp9_dthread.h"
#include "vp9/decoder/vp9_read_bit_buffer.h"
#include "vp9/decoder/vp9_reader.h"
@@ -1591,9 +1590,9 @@
if (!xd->corrupted) {// If multiple threads are used to decode tiles, then we use those threads
// to do parallel loopfiltering.
- vp9_loop_filter_frame_mt(&pbi->lf_row_sync, new_fb, pbi->mb.plane, cm,
- pbi->tile_workers, pbi->num_tile_workers,
- cm->lf.filter_level, 0);
+ 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 {vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
"Decode failed. Frame data is corrupted.");
--- a/vp9/decoder/vp9_decoder.c
+++ b/vp9/decoder/vp9_decoder.c
@@ -32,7 +32,6 @@
#include "vp9/decoder/vp9_decodeframe.h"
#include "vp9/decoder/vp9_decoder.h"
#include "vp9/decoder/vp9_detokenize.h"
-#include "vp9/decoder/vp9_dthread.h"
static void initialize_dec(void) {static volatile int init_done = 0;
--- a/vp9/decoder/vp9_decoder.h
+++ b/vp9/decoder/vp9_decoder.h
@@ -15,13 +15,12 @@
#include "vpx/vpx_codec.h"
#include "vpx_scale/yv12config.h"
-
+#include "vp9/common/vp9_loopfilter_thread.h"
#include "vp9/common/vp9_onyxc_int.h"
#include "vp9/common/vp9_ppflags.h"
#include "vp9/common/vp9_thread.h"
+#include "vp9/decoder/vp9_reader.h"
-#include "vp9/decoder/vp9_dthread.h"
-
#ifdef __cplusplus
extern "C" {#endif
@@ -32,6 +31,13 @@
vp9_reader bit_reader;
DECLARE_ALIGNED(16, MACROBLOCKD, xd);
} TileData;
+
+typedef struct TileWorkerData {+ VP9_COMMON *cm;
+ vp9_reader bit_reader;
+ DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+ struct vpx_internal_error_info error_info;
+} TileWorkerData;
typedef struct VP9Decoder {DECLARE_ALIGNED(16, MACROBLOCKD, mb);
--- a/vp9/decoder/vp9_dthread.c
+++ /dev/null
@@ -1,272 +1,0 @@
-/*
- * Copyright (c) 2014 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 "./vpx_config.h"
-
-#include "vpx_mem/vpx_mem.h"
-
-#include "vp9/common/vp9_reconinter.h"
-
-#include "vp9/decoder/vp9_dthread.h"
-#include "vp9/decoder/vp9_decoder.h"
-
-#if CONFIG_MULTITHREAD
-static INLINE void mutex_lock(pthread_mutex_t *const mutex) {- const int kMaxTryLocks = 4000;
- int locked = 0;
- int i;
-
- for (i = 0; i < kMaxTryLocks; ++i) {- if (!pthread_mutex_trylock(mutex)) {- locked = 1;
- break;
- }
- }
-
- if (!locked)
- pthread_mutex_lock(mutex);
-}
-#endif // CONFIG_MULTITHREAD
-
-static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {-#if CONFIG_MULTITHREAD
- const int nsync = lf_sync->sync_range;
-
- if (r && !(c & (nsync - 1))) {- pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
- mutex_lock(mutex);
-
- while (c > lf_sync->cur_sb_col[r - 1] - nsync) {- pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
- }
- pthread_mutex_unlock(mutex);
- }
-#else
- (void)lf_sync;
- (void)r;
- (void)c;
-#endif // CONFIG_MULTITHREAD
-}
-
-static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
- const int sb_cols) {-#if CONFIG_MULTITHREAD
- const int nsync = lf_sync->sync_range;
- int cur;
- // Only signal when there are enough filtered SB for next row to run.
- int sig = 1;
-
- if (c < sb_cols - 1) {- cur = c;
- if (c % nsync)
- sig = 0;
- } else {- cur = sb_cols + nsync;
- }
-
- if (sig) {- mutex_lock(&lf_sync->mutex_[r]);
-
- lf_sync->cur_sb_col[r] = cur;
-
- pthread_cond_signal(&lf_sync->cond_[r]);
- pthread_mutex_unlock(&lf_sync->mutex_[r]);
- }
-#else
- (void)lf_sync;
- (void)r;
- (void)c;
- (void)sb_cols;
-#endif // CONFIG_MULTITHREAD
-}
-
-// Implement row loopfiltering for each thread.
-static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer,
- VP9_COMMON *const cm,
- struct macroblockd_plane planes[MAX_MB_PLANE],
- int start, int stop, int y_only,
- VP9LfSync *const lf_sync) {- const int num_planes = y_only ? 1 : MAX_MB_PLANE;
- int r, c; // SB row and col
- const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
-
- for (r = start; r < stop; r += lf_sync->num_workers) {- const int mi_row = r << MI_BLOCK_SIZE_LOG2;
- MODE_INFO *const mi = cm->mi + mi_row * cm->mi_stride;
-
- for (c = 0; c < sb_cols; ++c) {- const int mi_col = c << MI_BLOCK_SIZE_LOG2;
- LOOP_FILTER_MASK lfm;
- int plane;
-
- sync_read(lf_sync, r, c);
-
- vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
- vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm);
-
- for (plane = 0; plane < num_planes; ++plane) {- vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
- }
-
- sync_write(lf_sync, r, c, sb_cols);
- }
- }
-}
-
-// Row-based multi-threaded loopfilter hook
-static int loop_filter_row_worker(VP9LfSync *const lf_sync,
- LFWorkerData *const lf_data) {- loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
- lf_data->start, lf_data->stop, lf_data->y_only, lf_sync);
- return 1;
-}
-
-// VP9 decoder: Implement multi-threaded loopfilter that uses the tile
-// threads.
-void vp9_loop_filter_frame_mt(VP9LfSync *lf_sync,
- YV12_BUFFER_CONFIG *frame,
- struct macroblockd_plane planes[MAX_MB_PLANE],
- VP9_COMMON *cm,
- VP9Worker *workers, int nworkers,
- int frame_filter_level,
- int y_only) {- const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
- // Number of superblock rows and cols
- const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
- const int tile_cols = 1 << cm->log2_tile_cols;
- const int num_workers = MIN(nworkers, tile_cols);
- int i;
-
- if (!frame_filter_level) return;
-
- if (!lf_sync->sync_range || cm->last_height != cm->height ||
- 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);
- }
-
- vp9_loop_filter_frame_init(cm, frame_filter_level);
-
- // Initialize cur_sb_col to -1 for all SB rows.
- vpx_memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
-
- // Set up loopfilter thread data.
- // The decoder is capping num_workers because it has been observed that using
- // more threads on the loopfilter than there are cores will hurt performance
- // on Android. This is because the system will only schedule the tile decode
- // workers on cores equal to the number of tile columns. Then if the decoder
- // tries to use more threads for the loopfilter, it will hurt performance
- // because of contention. If the multithreading code changes in the future
- // then the number of workers used by the loopfilter should be revisited.
- for (i = 0; i < num_workers; ++i) {- VP9Worker *const worker = &workers[i];
- LFWorkerData *const lf_data = &lf_sync->lfdata[i];
-
- worker->hook = (VP9WorkerHook)loop_filter_row_worker;
- worker->data1 = lf_sync;
- worker->data2 = lf_data;
-
- // Loopfilter data
- vp9_loop_filter_data_reset(lf_data, frame, cm, planes);
- lf_data->start = i;
- lf_data->stop = sb_rows;
- lf_data->y_only = y_only;
-
- // Start loopfiltering
- if (i == num_workers - 1) {- winterface->execute(worker);
- } else {- winterface->launch(worker);
- }
- }
-
- // Wait till all rows are finished
- for (i = 0; i < num_workers; ++i) {- winterface->sync(&workers[i]);
- }
-}
-
-// Set up nsync by width.
-static int get_sync_range(int width) {- // nsync numbers are picked by testing. For example, for 4k
- // video, using 4 gives best performance.
- if (width < 640)
- return 1;
- else if (width <= 1280)
- return 2;
- else if (width <= 4096)
- return 4;
- else
- return 8;
-}
-
-// Allocate memory for lf row synchronization
-void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
- int width, int num_workers) {- lf_sync->rows = rows;
-#if CONFIG_MULTITHREAD
- {- int i;
-
- CHECK_MEM_ERROR(cm, lf_sync->mutex_,
- vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
- if (lf_sync->mutex_) {- for (i = 0; i < rows; ++i) {- pthread_mutex_init(&lf_sync->mutex_[i], NULL);
- }
- }
-
- CHECK_MEM_ERROR(cm, lf_sync->cond_,
- vpx_malloc(sizeof(*lf_sync->cond_) * rows));
- if (lf_sync->cond_) {- for (i = 0; i < rows; ++i) {- pthread_cond_init(&lf_sync->cond_[i], NULL);
- }
- }
- }
-#endif // CONFIG_MULTITHREAD
-
- CHECK_MEM_ERROR(cm, lf_sync->lfdata,
- vpx_malloc(num_workers * sizeof(*lf_sync->lfdata)));
- lf_sync->num_workers = num_workers;
-
- CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
- vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
-
- // Set up nsync.
- lf_sync->sync_range = get_sync_range(width);
-}
-
-// Deallocate lf synchronization related mutex and data
-void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) {- if (lf_sync != NULL) {-#if CONFIG_MULTITHREAD
- int i;
-
- if (lf_sync->mutex_ != NULL) {- for (i = 0; i < lf_sync->rows; ++i) {- pthread_mutex_destroy(&lf_sync->mutex_[i]);
- }
- vpx_free(lf_sync->mutex_);
- }
- if (lf_sync->cond_ != NULL) {- for (i = 0; i < lf_sync->rows; ++i) {- pthread_cond_destroy(&lf_sync->cond_[i]);
- }
- vpx_free(lf_sync->cond_);
- }
-#endif // CONFIG_MULTITHREAD
- vpx_free(lf_sync->lfdata);
- vpx_free(lf_sync->cur_sb_col);
- // 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);
- }
-}
--- a/vp9/decoder/vp9_dthread.h
+++ /dev/null
@@ -1,63 +1,0 @@
-/*
- * Copyright (c) 2014 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.
- */
-
-#ifndef VP9_DECODER_VP9_DTHREAD_H_
-#define VP9_DECODER_VP9_DTHREAD_H_
-
-#include "./vpx_config.h"
-#include "vp9/common/vp9_thread.h"
-#include "vp9/decoder/vp9_reader.h"
-#include "vpx/internal/vpx_codec_internal.h"
-
-struct VP9Common;
-struct VP9Decoder;
-
-typedef struct TileWorkerData {- struct VP9Common *cm;
- vp9_reader bit_reader;
- DECLARE_ALIGNED(16, struct macroblockd, xd);
- struct vpx_internal_error_info error_info;
-} TileWorkerData;
-
-// Loopfilter row synchronization
-typedef struct VP9LfSyncData {-#if CONFIG_MULTITHREAD
- pthread_mutex_t *mutex_;
- pthread_cond_t *cond_;
-#endif
- // Allocate memory to store the loop-filtered superblock index in each row.
- int *cur_sb_col;
- // The optimal sync_range for different resolution and platform should be
- // determined by testing. Currently, it is chosen to be a power-of-2 number.
- int sync_range;
- int rows;
-
- // Row-based parallel loopfilter data
- LFWorkerData *lfdata;
- int num_workers;
-} VP9LfSync;
-
-// Allocate memory for loopfilter row synchronization.
-void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
- int width, int num_workers);
-
-// Deallocate loopfilter synchronization related mutex and data.
-void vp9_loop_filter_dealloc(VP9LfSync *lf_sync);
-
-// Multi-threaded loopfilter that uses the tile threads.
-void vp9_loop_filter_frame_mt(VP9LfSync *lf_sync,
- YV12_BUFFER_CONFIG *frame,
- struct macroblockd_plane planes[MAX_MB_PLANE],
- struct VP9Common *cm,
- VP9Worker *workers, int num_workers,
- int frame_filter_level,
- int y_only);
-
-#endif // VP9_DECODER_VP9_DTHREAD_H_
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -1786,7 +1786,7 @@
for (t = 0; t < cpi->num_workers; ++t) {VP9Worker *const worker = &cpi->workers[t];
- EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
+ EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
// Deallocate allocated threads.
vp9_get_worker_interface()->end(worker);
@@ -1797,11 +1797,13 @@
vp9_free_pc_tree(thread_data->td);
vpx_free(thread_data->td);
}
-
- vpx_free(worker->data1);
}
+ vpx_free(cpi->tile_thr_data);
vpx_free(cpi->workers);
+ if (cpi->num_workers > 1)
+ vp9_loop_filter_dealloc(&cpi->lf_row_sync);
+
dealloc_compressor_data(cpi);
for (i = 0; i < sizeof(cpi->mbgraph_stats) /
@@ -2437,7 +2439,13 @@
}
if (lf->filter_level > 0) {- vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
+ if (cpi->num_workers > 1)
+ vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
+ lf->filter_level, 0, 0,
+ cpi->workers, cpi->num_workers,
+ &cpi->lf_row_sync);
+ else
+ vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
}
vp9_extend_frame_inner_borders(cm->frame_to_show);
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -19,6 +19,7 @@
#include "vp9/common/vp9_ppflags.h"
#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_loopfilter_thread.h"
#include "vp9/common/vp9_onyxc_int.h"
#include "vp9/common/vp9_thread.h"
@@ -36,6 +37,7 @@
#include "vp9/encoder/vp9_svc_layercontext.h"
#include "vp9/encoder/vp9_tokenize.h"
#include "vp9/encoder/vp9_variance.h"
+
#if CONFIG_VP9_TEMPORAL_DENOISING
#include "vp9/encoder/vp9_denoiser.h"
#endif
@@ -262,6 +264,8 @@
PC_TREE *pc_root;
} ThreadData;
+struct EncWorkerData;
+
typedef struct VP9_COMP {QUANTS quants;
ThreadData td;
@@ -447,6 +451,8 @@
// Multi-threading
int num_workers;
VP9Worker *workers;
+ struct EncWorkerData *tile_thr_data;
+ VP9LfSync lf_row_sync;
} VP9_COMP;
void vp9_initialize_enc(void);
--- a/vp9/encoder/vp9_ethread.c
+++ b/vp9/encoder/vp9_ethread.c
@@ -167,16 +167,15 @@
CHECK_MEM_ERROR(cm, cpi->workers,
vpx_malloc(num_workers * sizeof(*cpi->workers)));
+ CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
+ vpx_calloc(num_workers, sizeof(*cpi->tile_thr_data)));
+
for (i = 0; i < num_workers; i++) {VP9Worker *const worker = &cpi->workers[i];
- EncWorkerData *thread_data;
+ EncWorkerData *thread_data = &cpi->tile_thr_data[i];
++cpi->num_workers;
-
winterface->init(worker);
- CHECK_MEM_ERROR(cm, worker->data1,
- (EncWorkerData*)vpx_calloc(1, sizeof(EncWorkerData)));
- thread_data = (EncWorkerData*)worker->data1;
if (i < num_workers - 1) {thread_data->cpi = cpi;
@@ -205,17 +204,18 @@
thread_data->td = &cpi->td;
}
- // data2 is unused.
- worker->data2 = NULL;
-
winterface->sync(worker);
- worker->hook = (VP9WorkerHook)enc_worker_hook;
}
}
for (i = 0; i < num_workers; i++) {VP9Worker *const worker = &cpi->workers[i];
- EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
+ EncWorkerData *thread_data;
+
+ worker->hook = (VP9WorkerHook)enc_worker_hook;
+ worker->data1 = &cpi->tile_thr_data[i];
+ worker->data2 = NULL;
+ thread_data = (EncWorkerData*)worker->data1;
// Before encoding a frame, copy the thread data from cpi.
thread_data->td->mb = cpi->td.mb;
--- a/vp9/encoder/vp9_picklpf.c
+++ b/vp9/encoder/vp9_picklpf.c
@@ -39,8 +39,14 @@
VP9_COMMON *const cm = &cpi->common;
int64_t filt_err;
- vp9_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level, 1,
- partial_frame);
+ if (cpi->num_workers > 1)
+ vp9_loop_filter_frame_mt(cm->frame_to_show, cm, cpi->td.mb.e_mbd.plane,
+ filt_level, 1, partial_frame,
+ cpi->workers, cpi->num_workers, &cpi->lf_row_sync);
+ else
+ vp9_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level,
+ 1, partial_frame);
+
#if CONFIG_VP9_HIGHBITDEPTH
if (cm->use_highbitdepth) {filt_err = vp9_highbd_get_y_sse(sd, cm->frame_to_show);
--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -33,6 +33,7 @@
VP9_COMMON_SRCS-yes += common/vp9_enums.h
VP9_COMMON_SRCS-yes += common/vp9_idct.h
VP9_COMMON_SRCS-yes += common/vp9_loopfilter.h
+VP9_COMMON_SRCS-yes += common/vp9_loopfilter_thread.h
VP9_COMMON_SRCS-yes += common/vp9_mv.h
VP9_COMMON_SRCS-yes += common/vp9_onyxc_int.h
VP9_COMMON_SRCS-yes += common/vp9_pred_common.h
@@ -56,6 +57,7 @@
VP9_COMMON_SRCS-yes += common/vp9_tile_common.c
VP9_COMMON_SRCS-yes += common/vp9_loopfilter.c
VP9_COMMON_SRCS-yes += common/vp9_loopfilter_filters.c
+VP9_COMMON_SRCS-yes += common/vp9_loopfilter_thread.c
VP9_COMMON_SRCS-yes += common/vp9_mvref_common.c
VP9_COMMON_SRCS-yes += common/vp9_mvref_common.h
VP9_COMMON_SRCS-yes += common/vp9_quant_common.c
--- a/vp9/vp9dx.mk
+++ b/vp9/vp9dx.mk
@@ -21,8 +21,6 @@
VP9_DX_SRCS-yes += decoder/vp9_decodeframe.c
VP9_DX_SRCS-yes += decoder/vp9_decodeframe.h
VP9_DX_SRCS-yes += decoder/vp9_detokenize.c
-VP9_DX_SRCS-yes += decoder/vp9_dthread.c
-VP9_DX_SRCS-yes += decoder/vp9_dthread.h
VP9_DX_SRCS-yes += decoder/vp9_reader.h
VP9_DX_SRCS-yes += decoder/vp9_reader.c
VP9_DX_SRCS-yes += decoder/vp9_read_bit_buffer.c