static void update_mbgraph_mb_stats(VP9_COMP *cpi, MBGRAPH_MB_STATS *stats,
YV12_BUFFER_CONFIG *buf, int mb_y_offset,
YV12_BUFFER_CONFIG *golden_ref,
const MV *prev_golden_ref_mv,
YV12_BUFFER_CONFIG *alt_ref, int mb_row,
int mb_col) {
MACROBLOCK *const x = &cpi->td.mb;
MACROBLOCKD *const xd = &x->e_mbd;
int intra_error;
VP9_COMMON *cm = &cpi->common;
// FIXME in practice we're completely ignoring chroma here
x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
x->plane[0].src.stride = buf->y_stride;
xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
// do intra 16x16 prediction
intra_error = find_best_16x16_intra(cpi, &stats->ref[INTRA_FRAME].m.mode);
if (intra_error <= 0) intra_error = 1;
stats->ref[INTRA_FRAME].err = intra_error;
// Golden frame MV search, if it exists and is different than last frame
if (golden_ref) {
int g_motion_error;
xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
xd->plane[0].pre[0].stride = golden_ref->y_stride;
g_motion_error =
do_16x16_motion_search(cpi, prev_golden_ref_mv,
&stats->ref[GOLDEN_FRAME].m.mv, mb_row, mb_col);
stats->ref[GOLDEN_FRAME].err = g_motion_error;
} else {
stats->ref[GOLDEN_FRAME].err = INT_MAX;
stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
}
// Do an Alt-ref frame MV search, if it exists and is different than
// last/golden frame.
if (alt_ref) {
int a_motion_error;
xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
xd->plane[0].pre[0].stride = alt_ref->y_stride;
a_motion_error =
do_16x16_zerozero_search(cpi, &stats->ref[ALTREF_FRAME].m.mv);
stats->ref[ALTREF_FRAME].err = a_motion_error;
} else {
stats->ref[ALTREF_FRAME].err = INT_MAX;
stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
}
}
/* If any buffer updating is signaled it should be done here. */
static void swap_frame_buffers(VP9Decoder *pbi) {
int ref_index = 0, mask;
VP9_COMMON *const cm = &pbi->common;
BufferPool *const pool = cm->buffer_pool;
RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
const int old_idx = cm->ref_frame_map[ref_index];
// Current thread releases the holding of reference frame.
decrease_ref_count(old_idx, frame_bufs, pool);
// Release the reference frame in reference map.
if (mask & 1) {
decrease_ref_count(old_idx, frame_bufs, pool);
}
cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
++ref_index;
}
// Current thread releases the holding of reference frame.
for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
const int old_idx = cm->ref_frame_map[ref_index];
decrease_ref_count(old_idx, frame_bufs, pool);
cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
}
pbi->hold_ref_buf = 0;
cm->frame_to_show = get_frame_new_buffer(cm);
--frame_bufs[cm->new_fb_idx].ref_count;
// Invalidate these references until the next frame starts.
for (ref_index = 0; ref_index < 3; ref_index++)
cm->frame_refs[ref_index].idx = -1;
}
/* If any buffer updating is signaled it should be done here. */
static void swap_frame_buffers(VP9D_COMP *pbi) {
int ref_index = 0, mask;
VP9_COMMON *const cm = &pbi->common;
for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
if (mask & 1)
ref_cnt_fb(cm->fb_idx_ref_cnt, &cm->ref_frame_map[ref_index],
cm->new_fb_idx);
++ref_index;
}
cm->frame_to_show = get_frame_new_buffer(cm);
cm->fb_idx_ref_cnt[cm->new_fb_idx]--;
// Invalidate these references until the next frame starts.
for (ref_index = 0; ref_index < 3; ref_index++)
cm->active_ref_idx[ref_index] = INT_MAX;
}
/* If any buffer updating is signaled it should be done here. */
static void swap_frame_buffers(AV1Decoder *pbi) {
int ref_index = 0, mask;
AV1_COMMON *const cm = &pbi->common;
BufferPool *const pool = cm->buffer_pool;
RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
lock_buffer_pool(pool);
for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
const int old_idx = cm->ref_frame_map[ref_index];
// Current thread releases the holding of reference frame.
decrease_ref_count(old_idx, frame_bufs, pool);
// Release the reference frame holding in the reference map for the decoding
// of the next frame.
if (mask & 1) decrease_ref_count(old_idx, frame_bufs, pool);
cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
++ref_index;
}
// Current thread releases the holding of reference frame.
for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
const int old_idx = cm->ref_frame_map[ref_index];
decrease_ref_count(old_idx, frame_bufs, pool);
cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
}
unlock_buffer_pool(pool);
pbi->hold_ref_buf = 0;
cm->frame_to_show = get_frame_new_buffer(cm);
// TODO(zoeliu): To fix the ref frame buffer update for the scenario of
// cm->frame_parellel_decode == 1
if (!cm->frame_parallel_decode || !cm->show_frame) {
lock_buffer_pool(pool);
--frame_bufs[cm->new_fb_idx].ref_count;
unlock_buffer_pool(pool);
}
// Invalidate these references until the next frame starts.
for (ref_index = 0; ref_index < INTER_REFS_PER_FRAME; ref_index++) {
cm->frame_refs[ref_index].idx = INVALID_IDX;
cm->frame_refs[ref_index].buf = NULL;
}
}
/* If any buffer updating is signaled it should be done here. */
static void swap_frame_buffers(VP9Decoder *pbi) {
int ref_index = 0, mask;
VP9_COMMON *const cm = &pbi->common;
for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
if (mask & 1) {
const int old_idx = cm->ref_frame_map[ref_index];
ref_cnt_fb(cm->frame_bufs, &cm->ref_frame_map[ref_index],
cm->new_fb_idx);
if (old_idx >= 0 && cm->frame_bufs[old_idx].ref_count == 0)
cm->release_fb_cb(cm->cb_priv,
&cm->frame_bufs[old_idx].raw_frame_buffer);
}
++ref_index;
}
cm->frame_to_show = get_frame_new_buffer(cm);
cm->frame_bufs[cm->new_fb_idx].ref_count--;
// Invalidate these references until the next frame starts.
for (ref_index = 0; ref_index < 3; ref_index++)
cm->frame_refs[ref_index].idx = INT_MAX;
}
// VP9 decoder: Implement multi-threaded loopfilter that uses the tile
// threads.
void vp9_loop_filter_frame_mt(VP9Decoder *pbi,
VP9_COMMON *cm,
int frame_filter_level,
int y_only, int partial_frame) {
VP9LfSync *const lf_sync = &pbi->lf_row_sync;
// 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) {
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(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
// Set up loopfilter thread data.
// 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 = lf_sync;
lf_data->num_lf_workers = num_workers;
// Start loopfiltering
if (i == num_workers - 1) {
vp9_worker_execute(worker);
} else {
vp9_worker_launch(worker);
}
}
// Wait till all rows are finished
for (i = 0; i < num_workers; ++i) {
vp9_worker_sync(&pbi->tile_workers[i]);
}
}
