void av1_decoder_remove(AV1Decoder *pbi) {
int i;
if (!pbi) return;
aom_get_worker_interface()->end(&pbi->lf_worker);
aom_free(pbi->lf_worker.data1);
aom_free(pbi->tile_data);
for (i = 0; i < pbi->num_tile_workers; ++i) {
AVxWorker *const worker = &pbi->tile_workers[i];
aom_get_worker_interface()->end(worker);
}
aom_free(pbi->tile_worker_data);
aom_free(pbi->tile_worker_info);
aom_free(pbi->tile_workers);
if (pbi->num_tile_workers > 0) {
av1_loop_filter_dealloc(&pbi->lf_row_sync);
}
#if CONFIG_ACCOUNTING
aom_accounting_clear(&pbi->accounting);
#endif
aom_free(pbi);
}
void av1_encode_tiles_mt(AV1_COMP *cpi) {
AV1_COMMON *const cm = &cpi->common;
const int tile_cols = cm->tile_cols;
const AVxWorkerInterface *const winterface = aom_get_worker_interface();
const int num_workers = AOMMIN(cpi->oxcf.max_threads, tile_cols);
int i;
av1_init_tile_data(cpi);
// Only run once to create threads and allocate thread data.
if (cpi->num_workers == 0) {
CHECK_MEM_ERROR(cm, cpi->workers,
aom_malloc(num_workers * sizeof(*cpi->workers)));
CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
aom_calloc(num_workers, sizeof(*cpi->tile_thr_data)));
for (i = 0; i < num_workers; i++) {
AVxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = &cpi->tile_thr_data[i];
++cpi->num_workers;
winterface->init(worker);
thread_data->cpi = cpi;
if (i < num_workers - 1) {
// Allocate thread data.
CHECK_MEM_ERROR(cm, thread_data->td,
aom_memalign(32, sizeof(*thread_data->td)));
av1_zero(*thread_data->td);
// Set up pc_tree.
thread_data->td->leaf_tree = NULL;
thread_data->td->pc_tree = NULL;
av1_setup_pc_tree(cm, thread_data->td);
// Set up variance tree if needed.
if (cpi->sf.partition_search_type == VAR_BASED_PARTITION)
av1_setup_var_tree(cm, thread_data->td);
// Allocate frame counters in thread data.
CHECK_MEM_ERROR(cm, thread_data->td->counts,
aom_calloc(1, sizeof(*thread_data->td->counts)));
// Create threads
if (!winterface->reset(worker))
aom_internal_error(&cm->error, AOM_CODEC_ERROR,
"Tile encoder thread creation failed");
} else {
// Main thread acts as a worker and uses the thread data in cpi.
thread_data->td = &cpi->td;
}
winterface->sync(worker);
}
}
for (i = 0; i < num_workers; i++) {
AVxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data;
worker->hook = (AVxWorkerHook)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.
if (thread_data->td != &cpi->td) {
thread_data->td->mb = cpi->td.mb;
thread_data->td->rd_counts = cpi->td.rd_counts;
}
if (thread_data->td->counts != &cpi->common.counts) {
memcpy(thread_data->td->counts, &cpi->common.counts,
sizeof(cpi->common.counts));
}
#if CONFIG_PALETTE
// Allocate buffers used by palette coding mode.
if (cpi->common.allow_screen_content_tools && i < num_workers - 1) {
MACROBLOCK *x = &thread_data->td->mb;
CHECK_MEM_ERROR(cm, x->palette_buffer,
aom_memalign(16, sizeof(*x->palette_buffer)));
}
#endif // CONFIG_PALETTE
}
// Encode a frame
for (i = 0; i < num_workers; i++) {
AVxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
// Set the starting tile for each thread.
thread_data->start = i;
if (i == cpi->num_workers - 1)
winterface->execute(worker);
else
winterface->launch(worker);
}
// Encoding ends.
for (i = 0; i < num_workers; i++) {
AVxWorker *const worker = &cpi->workers[i];
winterface->sync(worker);
}
for (i = 0; i < num_workers; i++) {
AVxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
// Accumulate counters.
if (i < cpi->num_workers - 1) {
av1_accumulate_frame_counts(&cm->counts, thread_data->td->counts);
accumulate_rd_opt(&cpi->td, thread_data->td);
}
}
}
AV1Decoder *av1_decoder_create(BufferPool *const pool) {
AV1Decoder *volatile const pbi = aom_memalign(32, sizeof(*pbi));
AV1_COMMON *volatile const cm = pbi ? &pbi->common : NULL;
if (!cm) return NULL;
av1_zero(*pbi);
if (setjmp(cm->error.jmp)) {
cm->error.setjmp = 0;
av1_decoder_remove(pbi);
return NULL;
}
cm->error.setjmp = 1;
CHECK_MEM_ERROR(cm, cm->fc,
(FRAME_CONTEXT *)aom_memalign(32, sizeof(*cm->fc)));
CHECK_MEM_ERROR(cm, cm->frame_contexts,
(FRAME_CONTEXT *)aom_memalign(
32, FRAME_CONTEXTS * sizeof(*cm->frame_contexts)));
memset(cm->fc, 0, sizeof(*cm->fc));
memset(cm->frame_contexts, 0, FRAME_CONTEXTS * sizeof(*cm->frame_contexts));
pbi->need_resync = 1;
once(initialize_dec);
// Initialize the references to not point to any frame buffers.
memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
memset(&cm->next_ref_frame_map, -1, sizeof(cm->next_ref_frame_map));
cm->current_video_frame = 0;
pbi->ready_for_new_data = 1;
pbi->common.buffer_pool = pool;
cm->bit_depth = AOM_BITS_8;
cm->dequant_bit_depth = AOM_BITS_8;
cm->alloc_mi = av1_dec_alloc_mi;
cm->free_mi = av1_dec_free_mi;
cm->setup_mi = av1_dec_setup_mi;
av1_loop_filter_init(cm);
#if CONFIG_NCOBMC_ADAPT_WEIGHT
get_default_ncobmc_kernels(cm);
#endif // CONFIG_NCOBMC_ADAPT_WEIGHT
#if CONFIG_AOM_QM
aom_qm_init(cm);
#endif
#if CONFIG_LOOP_RESTORATION
av1_loop_restoration_precal();
#endif // CONFIG_LOOP_RESTORATION
#if CONFIG_ACCOUNTING
pbi->acct_enabled = 1;
aom_accounting_init(&pbi->accounting);
#endif
cm->error.setjmp = 0;
aom_get_worker_interface()->init(&pbi->lf_worker);
return pbi;
}
int av1_receive_compressed_data(AV1Decoder *pbi, size_t size,
const uint8_t **psource) {
AV1_COMMON *volatile const cm = &pbi->common;
BufferPool *volatile const pool = cm->buffer_pool;
RefCntBuffer *volatile const frame_bufs = cm->buffer_pool->frame_bufs;
const uint8_t *source = *psource;
int retcode = 0;
cm->error.error_code = AOM_CODEC_OK;
if (size == 0) {
// This is used to signal that we are missing frames.
// We do not know if the missing frame(s) was supposed to update
// any of the reference buffers, but we act conservative and
// mark only the last buffer as corrupted.
//
// TODO(jkoleszar): Error concealment is undefined and non-normative
// at this point, but if it becomes so, [0] may not always be the correct
// thing to do here.
if (cm->frame_refs[0].idx > 0) {
assert(cm->frame_refs[0].buf != NULL);
cm->frame_refs[0].buf->corrupted = 1;
}
}
pbi->ready_for_new_data = 0;
// Find a free buffer for the new frame, releasing the reference previously
// held.
// Check if the previous frame was a frame without any references to it.
// Release frame buffer if not decoding in frame parallel mode.
if (!cm->frame_parallel_decode && cm->new_fb_idx >= 0 &&
frame_bufs[cm->new_fb_idx].ref_count == 0)
pool->release_fb_cb(pool->cb_priv,
&frame_bufs[cm->new_fb_idx].raw_frame_buffer);
// Find a free frame buffer. Return error if can not find any.
cm->new_fb_idx = get_free_fb(cm);
if (cm->new_fb_idx == INVALID_IDX) return AOM_CODEC_MEM_ERROR;
// Assign a MV array to the frame buffer.
cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
pbi->hold_ref_buf = 0;
if (cm->frame_parallel_decode) {
AVxWorker *const worker = pbi->frame_worker_owner;
av1_frameworker_lock_stats(worker);
frame_bufs[cm->new_fb_idx].frame_worker_owner = worker;
// Reset decoding progress.
pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
pbi->cur_buf->row = -1;
pbi->cur_buf->col = -1;
av1_frameworker_unlock_stats(worker);
} else {
pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
}
if (setjmp(cm->error.jmp)) {
const AVxWorkerInterface *const winterface = aom_get_worker_interface();
int i;
cm->error.setjmp = 0;
pbi->ready_for_new_data = 1;
// Synchronize all threads immediately as a subsequent decode call may
// cause a resize invalidating some allocations.
winterface->sync(&pbi->lf_worker);
for (i = 0; i < pbi->num_tile_workers; ++i) {
winterface->sync(&pbi->tile_workers[i]);
}
lock_buffer_pool(pool);
// Release all the reference buffers if worker thread is holding them.
if (pbi->hold_ref_buf == 1) {
int ref_index = 0, mask;
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);
++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);
}
pbi->hold_ref_buf = 0;
}
// Release current frame.
decrease_ref_count(cm->new_fb_idx, frame_bufs, pool);
unlock_buffer_pool(pool);
aom_clear_system_state();
return -1;
}
static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
struct macroblockd_plane *planes, int start,
int stop, int y_only, AVxWorker *workers,
int nworkers, AV1LfSync *lf_sync) {
#if CONFIG_EXT_PARTITION
printf(
"STOPPING: This code has not been modified to work with the "
"extended coding unit size experiment");
exit(EXIT_FAILURE);
#endif // CONFIG_EXT_PARTITION
const AVxWorkerInterface *const winterface = aom_get_worker_interface();
// Number of superblock rows and cols
const int sb_rows = mi_rows_aligned_to_sb(cm) >> cm->mib_size_log2;
// Decoder may allocate more threads than number of tiles based on user's
// input.
const int tile_cols = cm->tile_cols;
const int num_workers = AOMMIN(nworkers, tile_cols);
int i;
if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
num_workers > lf_sync->num_workers) {
av1_loop_filter_dealloc(lf_sync);
av1_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
}
// 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.
#if CONFIG_PARALLEL_DEBLOCKING
// Initialize cur_sb_col to -1 for all SB rows.
memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
// Filter all the vertical edges in the whole frame
for (i = 0; i < num_workers; ++i) {
AVxWorker *const worker = &workers[i];
LFWorkerData *const lf_data = &lf_sync->lfdata[i];
worker->hook = (AVxWorkerHook)loop_filter_ver_row_worker;
worker->data1 = lf_sync;
worker->data2 = lf_data;
// Loopfilter data
av1_loop_filter_data_reset(lf_data, frame, cm, planes);
lf_data->start = start + i * cm->mib_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]);
}
memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
// Filter all the horizontal edges in the whole frame
for (i = 0; i < num_workers; ++i) {
AVxWorker *const worker = &workers[i];
LFWorkerData *const lf_data = &lf_sync->lfdata[i];
worker->hook = (AVxWorkerHook)loop_filter_hor_row_worker;
worker->data1 = lf_sync;
worker->data2 = lf_data;
// Loopfilter data
av1_loop_filter_data_reset(lf_data, frame, cm, planes);
lf_data->start = start + i * cm->mib_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]);
}
#else // CONFIG_PARALLEL_DEBLOCKING
// Initialize cur_sb_col to -1 for all SB rows.
memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
for (i = 0; i < num_workers; ++i) {
AVxWorker *const worker = &workers[i];
LFWorkerData *const lf_data = &lf_sync->lfdata[i];
worker->hook = (AVxWorkerHook)loop_filter_row_worker;
worker->data1 = lf_sync;
worker->data2 = lf_data;
// Loopfilter data
av1_loop_filter_data_reset(lf_data, frame, cm, planes);
lf_data->start = start + i * cm->mib_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]);
}
#endif // CONFIG_PARALLEL_DEBLOCKING
}