libvpx: Pull from upstream

source/libvpx/vp9/decoder/vp9_dthread.c

 /*
  *  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_onyxd_int.h"
-#include "vpx_mem/vpx_mem.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;
@@ skipping 63 matching lines @@
                                 VP9_COMMON *const cm, MACROBLOCKD *const xd,
                                 int start, int stop, int y_only,
                                 VP9LfSync *const lf_sync, int num_lf_workers) {
   const int num_planes = y_only ? 1 : MAX_MB_PLANE;
   int r, c;  // SB row and col
   LOOP_FILTER_MASK lfm;
   const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
 
   for (r = start; r < stop; r += num_lf_workers) {
     const int mi_row = r << MI_BLOCK_SIZE_LOG2;
-    MODE_INFO **mi_8x8 = cm->mi_grid_visible + mi_row * cm->mode_info_stride;
+    MODE_INFO **mi_8x8 = cm->mi_grid_visible + mi_row * cm->mi_stride;
 
     for (c = 0; c < sb_cols; ++c) {
       const int mi_col = c << MI_BLOCK_SIZE_LOG2;
       int plane;
 
       sync_read(lf_sync, r, c);
 
-      setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
-      vp9_setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mode_info_stride,
-                     &lfm);
+      vp9_setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
+      vp9_setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mi_stride, &lfm);
 
       for (plane = 0; plane < num_planes; ++plane) {
         vp9_filter_block_plane(cm, &xd->plane[plane], mi_row, &lfm);
       }
 
       sync_write(lf_sync, r, c, sb_cols);
     }
   }
 }
 
 // Row-based multi-threaded loopfilter hook
 static int loop_filter_row_worker(void *arg1, void *arg2) {
   TileWorkerData *const tile_data = (TileWorkerData*)arg1;
   LFWorkerData *const lf_data = &tile_data->lfdata;
 
   loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, &lf_data->xd,
                       lf_data->start, lf_data->stop, lf_data->y_only,
                       lf_data->lf_sync, lf_data->num_lf_workers);
   return 1;
 }
 
 // VP9 decoder: Implement multi-threaded loopfilter that uses the tile
 // threads.
-void vp9_loop_filter_frame_mt(VP9D_COMP *pbi,
+void vp9_loop_filter_frame_mt(VP9Decoder *pbi,
                               VP9_COMMON *cm,
                               MACROBLOCKD *xd,
                               int frame_filter_level,
                               int y_only, int partial_frame) {
   // 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(pbi->oxcf.max_threads & ~1, tile_cols);
   int i;
 
   // Allocate memory used in thread synchronization.
   // This always needs to be done even if frame_filter_level is 0.
   if (!cm->current_video_frame || cm->last_height != cm->height) {
     VP9LfSync *const lf_sync = &pbi->lf_row_sync;
 
     if (cm->last_height != cm->height) {
       const int aligned_last_height =
           ALIGN_POWER_OF_TWO(cm->last_height, MI_SIZE_LOG2);
       const int last_sb_rows =
           mi_cols_aligned_to_sb(aligned_last_height >> MI_SIZE_LOG2) >>
           MI_BLOCK_SIZE_LOG2;
 
       vp9_loop_filter_dealloc(lf_sync, last_sb_rows);
     }
 
     vp9_loop_filter_alloc(cm, lf_sync, sb_rows, cm->width);
   }
 
   if (!frame_filter_level) return;
 
   vp9_loop_filter_frame_init(cm, frame_filter_level);
 
   // Initialize cur_sb_col to -1 for all SB rows.
   vpx_memset(pbi->lf_row_sync.cur_sb_col, -1,
              sizeof(*pbi->lf_row_sync.cur_sb_col) * sb_rows);
 
   // Set up loopfilter thread data.
-  for (i = 0; i < pbi->num_tile_workers; ++i) {
+  // The decoder is using num_workers instead of pbi->num_tile_workers
+  // because it has been observed that using more threads on the
+  // loopfilter, than there are tile columns in the frame 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 = &pbi->tile_workers[i];
     TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
     LFWorkerData *const lf_data = &tile_data->lfdata;
 
     worker->hook = (VP9WorkerHook)loop_filter_row_worker;
 
     // Loopfilter data
     lf_data->frame_buffer = get_frame_new_buffer(cm);
     lf_data->cm = cm;
     lf_data->xd = pbi->mb;
     lf_data->start = i;
     lf_data->stop = sb_rows;
     lf_data->y_only = y_only;   // always do all planes in decoder
 
     lf_data->lf_sync = &pbi->lf_row_sync;
-    lf_data->num_lf_workers = pbi->num_tile_workers;
+    lf_data->num_lf_workers = num_workers;
 
     // Start loopfiltering
-    if (i == pbi->num_tile_workers - 1) {
+    if (i == num_workers - 1) {
       vp9_worker_execute(worker);
     } else {
       vp9_worker_launch(worker);
     }
   }
 
   // Wait till all rows are finished
-  for (i = 0; i < pbi->num_tile_workers; ++i) {
+  for (i = 0; i < num_workers; ++i) {
     vp9_worker_sync(&pbi->tile_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;
@@ skipping 56 matching lines @@
     vpx_memset(lf_sync, 0, sizeof(*lf_sync));
   }
 #else
   (void)rows;
   if (lf_sync != NULL) {
     vpx_free(lf_sync->cur_sb_col);
     vpx_memset(lf_sync, 0, sizeof(*lf_sync));
   }
 #endif  // CONFIG_MULTITHREAD
 }