diff options
Diffstat (limited to 'vp9/decoder')
-rw-r--r-- | vp9/decoder/vp9_decodeframe.c | 7 | ||||
-rw-r--r-- | vp9/decoder/vp9_decoder.c | 1 | ||||
-rw-r--r-- | vp9/decoder/vp9_decoder.h | 12 | ||||
-rw-r--r-- | vp9/decoder/vp9_dthread.c | 272 | ||||
-rw-r--r-- | vp9/decoder/vp9_dthread.h | 63 |
5 files changed, 12 insertions, 343 deletions
diff --git a/vp9/decoder/vp9_decodeframe.c b/vp9/decoder/vp9_decodeframe.c index d135199c1..ea4edbffe 100644 --- 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 @@ void vp9_decode_frame(VP9Decoder *pbi, 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."); diff --git a/vp9/decoder/vp9_decoder.c b/vp9/decoder/vp9_decoder.c index 1d254d2db..7bef265b8 100644 --- 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; diff --git a/vp9/decoder/vp9_decoder.h b/vp9/decoder/vp9_decoder.h index 25b7339ed..1415019a1 100644 --- a/vp9/decoder/vp9_decoder.h +++ b/vp9/decoder/vp9_decoder.h @@ -15,12 +15,11 @@ #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_dthread.h" +#include "vp9/decoder/vp9_reader.h" #ifdef __cplusplus extern "C" { @@ -33,6 +32,13 @@ typedef struct TileData { 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); diff --git a/vp9/decoder/vp9_dthread.c b/vp9/decoder/vp9_dthread.c deleted file mode 100644 index 3d2d0dd2e..000000000 --- a/vp9/decoder/vp9_dthread.c +++ /dev/null @@ -1,272 +0,0 @@ -/* - * 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); - } -} diff --git a/vp9/decoder/vp9_dthread.h b/vp9/decoder/vp9_dthread.h deleted file mode 100644 index 664aaa32a..000000000 --- a/vp9/decoder/vp9_dthread.h +++ /dev/null @@ -1,63 +0,0 @@ -/* - * 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_ |