Merge "Moving loopfilter call to vp9_decode_frame()."

9 files changed, 36 insertions, 30 deletions

diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c
index 3ac5a0577..5b43e233b 100644
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -1224,7 +1224,8 @@ void vp9_loop_filter_rows(const YV12_BUFFER_CONFIG *frame_buffer,
   }
 }
 
-void vp9_loop_filter_frame(VP9_COMMON *cm, MACROBLOCKD *xd,
+void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
+                           VP9_COMMON *cm, MACROBLOCKD *xd,
                            int frame_filter_level,
                            int y_only, int partial_frame) {
   int start_mi_row, end_mi_row, mi_rows_to_filter;
@@ -1238,7 +1239,7 @@ void vp9_loop_filter_frame(VP9_COMMON *cm, MACROBLOCKD *xd,
   }
   end_mi_row = start_mi_row + mi_rows_to_filter;
   vp9_loop_filter_frame_init(cm, frame_filter_level);
-  vp9_loop_filter_rows(cm->frame_to_show, cm, xd,
+  vp9_loop_filter_rows(frame, cm, xd,
                        start_mi_row, end_mi_row,
                        y_only);
 }
diff --git a/vp9/common/vp9_loopfilter.h b/vp9/common/vp9_loopfilter.h
index 97ae9d22d..83463c544 100644
--- a/vp9/common/vp9_loopfilter.h
+++ b/vp9/common/vp9_loopfilter.h
@@ -104,7 +104,8 @@ void vp9_loop_filter_init(struct VP9Common *cm);
 // calls this function directly.
 void vp9_loop_filter_frame_init(struct VP9Common *cm, int default_filt_lvl);
 
-void vp9_loop_filter_frame(struct VP9Common *cm,
+void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
+                           struct VP9Common *cm,
                            struct macroblockd *mbd,
                            int filter_level,
                            int y_only, int partial_frame);
diff --git a/vp9/decoder/vp9_decodeframe.c b/vp9/decoder/vp9_decodeframe.c
index f5aabe0ee..9dc0cf1a3 100644
--- a/vp9/decoder/vp9_decodeframe.c
+++ b/vp9/decoder/vp9_decodeframe.c
@@ -676,13 +676,13 @@ static void setup_frame_size_with_refs(VP9_COMMON *cm,
 }
 
 static void decode_tile(VP9Decoder *pbi, const TileInfo *const tile,
-                        vp9_reader *r) {
+                        int do_loopfilter_inline, vp9_reader *r) {
   const int num_threads = pbi->max_threads;
   VP9_COMMON *const cm = &pbi->common;
   int mi_row, mi_col;
   MACROBLOCKD *xd = &pbi->mb;
 
-  if (pbi->do_loopfilter_inline) {
+  if (do_loopfilter_inline) {
     LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
     lf_data->frame_buffer = get_frame_new_buffer(cm);
     lf_data->cm = cm;
@@ -702,7 +702,7 @@ static void decode_tile(VP9Decoder *pbi, const TileInfo *const tile,
       decode_partition(cm, xd, tile, mi_row, mi_col, r, BLOCK_64X64);
     }
 
-    if (pbi->do_loopfilter_inline) {
+    if (do_loopfilter_inline) {
       const int lf_start = mi_row - MI_BLOCK_SIZE;
       LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
 
@@ -723,7 +723,7 @@ static void decode_tile(VP9Decoder *pbi, const TileInfo *const tile,
     }
   }
 
-  if (pbi->do_loopfilter_inline) {
+  if (do_loopfilter_inline) {
     LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
 
     vp9_worker_sync(&pbi->lf_worker);
@@ -811,7 +811,8 @@ static void get_tile_buffers(VP9Decoder *pbi,
 
 static const uint8_t *decode_tiles(VP9Decoder *pbi,
                                    const uint8_t *data,
-                                   const uint8_t *data_end) {
+                                   const uint8_t *data_end,
+                                   int do_loopfilter_inline) {
   VP9_COMMON *const cm = &pbi->common;
   const int aligned_cols = mi_cols_aligned_to_sb(cm->mi_cols);
   const int tile_cols = 1 << cm->log2_tile_cols;
@@ -846,7 +847,7 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
       vp9_tile_init(&tile, cm, tile_row, col);
       setup_token_decoder(buf->data, data_end, buf->size, &cm->error, &r,
                           pbi->decrypt_cb, pbi->decrypt_state);
-      decode_tile(pbi, &tile, &r);
+      decode_tile(pbi, &tile, do_loopfilter_inline, &r);
 
       if (last_tile)
         end = vp9_reader_find_end(&r);
@@ -1306,6 +1307,8 @@ int vp9_decode_frame(VP9Decoder *pbi,
   const int tile_rows = 1 << cm->log2_tile_rows;
   const int tile_cols = 1 << cm->log2_tile_cols;
   YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
+  const int do_loopfilter_inline = tile_rows == 1 && tile_cols == 1 &&
+                                   cm->lf.filter_level;
   xd->cur_buf = new_fb;
 
   if (!first_partition_size) {
@@ -1322,9 +1325,7 @@ int vp9_decode_frame(VP9Decoder *pbi,
     vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                        "Truncated packet or corrupt header length");
 
-  pbi->do_loopfilter_inline =
-      (cm->log2_tile_rows | cm->log2_tile_cols) == 0 && cm->lf.filter_level;
-  if (pbi->do_loopfilter_inline && pbi->lf_worker.data1 == NULL) {
+  if (do_loopfilter_inline && pbi->lf_worker.data1 == NULL) {
     CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
                     vpx_memalign(32, sizeof(LFWorkerData)));
     pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
@@ -1357,7 +1358,8 @@ int vp9_decode_frame(VP9Decoder *pbi,
       cm->frame_parallel_decoding_mode) {
     *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
   } else {
-    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
+    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end,
+                               do_loopfilter_inline);
   }
 
   new_fb->corrupted |= xd->corrupted;
@@ -1386,5 +1388,16 @@ int vp9_decode_frame(VP9Decoder *pbi,
   if (cm->refresh_frame_context)
     cm->frame_contexts[cm->frame_context_idx] = cm->fc;
 
+  // Loopfilter
+  if (!do_loopfilter_inline) {
+    // If multiple threads are used to decode tiles, then we use those threads
+    // to do parallel loopfiltering.
+    if (pbi->num_tile_workers) {
+      vp9_loop_filter_frame_mt(new_fb, pbi, cm, cm->lf.filter_level, 0, 0);
+    } else {
+      vp9_loop_filter_frame(new_fb, cm, &pbi->mb, cm->lf.filter_level, 0, 0);
+    }
+  }
+
   return 0;
 }
diff --git a/vp9/decoder/vp9_decoder.c b/vp9/decoder/vp9_decoder.c
index 8f72aa48f..6f310c765 100644
--- a/vp9/decoder/vp9_decoder.c
+++ b/vp9/decoder/vp9_decoder.c
@@ -279,16 +279,6 @@ int vp9_receive_compressed_data(VP9Decoder *pbi,
 
   swap_frame_buffers(pbi);
 
-  if (!pbi->do_loopfilter_inline) {
-    // If multiple threads are used to decode tiles, then we use those threads
-    // to do parallel loopfiltering.
-    if (pbi->num_tile_workers) {
-      vp9_loop_filter_frame_mt(pbi, cm, cm->lf.filter_level, 0, 0);
-    } else {
-      vp9_loop_filter_frame(cm, &pbi->mb, cm->lf.filter_level, 0, 0);
-    }
-  }
-
   vp9_clear_system_state();
 
   cm->last_width = cm->width;
diff --git a/vp9/decoder/vp9_decoder.h b/vp9/decoder/vp9_decoder.h
index 66b0f15f9..d6110c47e 100644
--- a/vp9/decoder/vp9_decoder.h
+++ b/vp9/decoder/vp9_decoder.h
@@ -39,7 +39,6 @@ typedef struct VP9Decoder {
 
   int decoded_key_frame;
 
-  int do_loopfilter_inline;  // apply loopfilter to available rows immediately
   VP9Worker lf_worker;
 
   VP9Worker *tile_workers;
diff --git a/vp9/decoder/vp9_dthread.c b/vp9/decoder/vp9_dthread.c
index 019b0ffe8..5fe5ed79a 100644
--- a/vp9/decoder/vp9_dthread.c
+++ b/vp9/decoder/vp9_dthread.c
@@ -132,8 +132,8 @@ static int loop_filter_row_worker(void *arg1, void *arg2) {
 
 // VP9 decoder: Implement multi-threaded loopfilter that uses the tile
 // threads.
-void vp9_loop_filter_frame_mt(VP9Decoder *pbi,
-                              VP9_COMMON *cm,
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+                              VP9Decoder *pbi, VP9_COMMON *cm,
                               int frame_filter_level,
                               int y_only, int partial_frame) {
   VP9LfSync *const lf_sync = &pbi->lf_row_sync;
@@ -184,7 +184,7 @@ void vp9_loop_filter_frame_mt(VP9Decoder *pbi,
     worker->hook = (VP9WorkerHook)loop_filter_row_worker;
 
     // Loopfilter data
-    lf_data->frame_buffer = get_frame_new_buffer(cm);
+    lf_data->frame_buffer = frame;
     lf_data->cm = cm;
     lf_data->xd = pbi->mb;
     lf_data->start = i;
diff --git a/vp9/decoder/vp9_dthread.h b/vp9/decoder/vp9_dthread.h
index 8738ceebd..c3b7a293b 100644
--- a/vp9/decoder/vp9_dthread.h
+++ b/vp9/decoder/vp9_dthread.h
@@ -48,7 +48,8 @@ void vp9_loop_filter_alloc(struct VP9Common *cm, VP9LfSync *lf_sync,
 void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows);
 
 // Multi-threaded loopfilter that uses the tile threads.
-void vp9_loop_filter_frame_mt(struct VP9Decoder *pbi,
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+                              struct VP9Decoder *pbi,
                               struct VP9Common *cm,
                               int frame_filter_level,
                               int y_only, int partial_frame);
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 5d3844758..64747638c 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -1605,7 +1605,7 @@ static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
   }
 
   if (lf->filter_level > 0) {
-    vp9_loop_filter_frame(cm, xd, lf->filter_level, 0, 0);
+    vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
   }
 
   vp9_extend_frame_inner_borders(cm->frame_to_show);
diff --git a/vp9/encoder/vp9_picklpf.c b/vp9/encoder/vp9_picklpf.c
index a4b280c53..53284656e 100644
--- a/vp9/encoder/vp9_picklpf.c
+++ b/vp9/encoder/vp9_picklpf.c
@@ -38,7 +38,8 @@ static int try_filter_frame(const YV12_BUFFER_CONFIG *sd, VP9_COMP *const cpi,
   VP9_COMMON *const cm = &cpi->common;
   int filt_err;
 
-  vp9_loop_filter_frame(cm, &cpi->mb.e_mbd, filt_level, 1, partial_frame);
+  vp9_loop_filter_frame(cm->frame_to_show, cm, &cpi->mb.e_mbd, filt_level, 1,
+                        partial_frame);
   filt_err = vp9_get_y_sse(sd, cm->frame_to_show);
 
   // Re-instate the unfiltered frame