Refactor av1_encode_tiles_mt function
Extracted fragments of code in av1_encode_tiles_mt
function into separate functions to facilitate
row based multi-threading in encoder.
Change-Id: I9b705553ec4d79ff779e15f3c1f2f4624e582188
diff --git a/av1/encoder/ethread.c b/av1/encoder/ethread.c
index 1b1d3b6..d1616ac 100644
--- a/av1/encoder/ethread.c
+++ b/av1/encoder/ethread.c
@@ -47,98 +47,130 @@
return 1;
}
-void av1_encode_tiles_mt(AV1_COMP *cpi) {
+static void create_enc_workers(AV1_COMP *cpi, int num_workers) {
AV1_COMMON *const cm = &cpi->common;
- const int tile_cols = cm->tile_cols;
- const int tile_rows = cm->tile_rows;
const AVxWorkerInterface *const winterface = aom_get_worker_interface();
- int num_workers = AOMMIN(cpi->oxcf.max_threads, tile_cols * tile_rows);
- int i;
- av1_init_tile_data(cpi);
+ CHECK_MEM_ERROR(cm, cpi->workers,
+ aom_malloc(num_workers * sizeof(*cpi->workers)));
- // 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)));
- 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->pc_tree = NULL;
- av1_setup_pc_tree(cm, thread_data->td);
-
- CHECK_MEM_ERROR(cm, thread_data->td->above_pred_buf,
- (uint8_t *)aom_memalign(
- 16, MAX_MB_PLANE * MAX_SB_SQUARE *
- sizeof(*thread_data->td->above_pred_buf)));
- CHECK_MEM_ERROR(cm, thread_data->td->left_pred_buf,
- (uint8_t *)aom_memalign(
- 16, MAX_MB_PLANE * MAX_SB_SQUARE *
- sizeof(*thread_data->td->left_pred_buf)));
-
- CHECK_MEM_ERROR(
- cm, thread_data->td->wsrc_buf,
- (int32_t *)aom_memalign(
- 16, MAX_SB_SQUARE * sizeof(*thread_data->td->wsrc_buf)));
-
- for (int x = 0; x < 2; x++)
- for (int y = 0; y < 2; y++)
- CHECK_MEM_ERROR(
- cm, thread_data->td->hash_value_buffer[x][y],
- (uint32_t *)aom_malloc(
- AOM_BUFFER_SIZE_FOR_BLOCK_HASH *
- sizeof(*thread_data->td->hash_value_buffer[0][0])));
-
- CHECK_MEM_ERROR(
- cm, thread_data->td->mask_buf,
- (int32_t *)aom_memalign(
- 16, MAX_SB_SQUARE * sizeof(*thread_data->td->mask_buf)));
- // Allocate frame counters in thread data.
- CHECK_MEM_ERROR(cm, thread_data->td->counts,
- aom_calloc(1, sizeof(*thread_data->td->counts)));
-
- // Allocate buffers used by palette coding mode.
- CHECK_MEM_ERROR(
- cm, thread_data->td->palette_buffer,
- aom_memalign(16, sizeof(*thread_data->td->palette_buffer)));
-
- // 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);
- }
- } else {
- num_workers = AOMMIN(num_workers, cpi->num_workers);
- }
-
- for (i = 0; i < num_workers; i++) {
+ for (int i = 0; i < num_workers; i++) {
AVxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = &cpi->tile_thr_data[i];
- worker->hook = (AVxWorkerHook)enc_worker_hook;
+ ++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->pc_tree = NULL;
+ av1_setup_pc_tree(cm, thread_data->td);
+
+ CHECK_MEM_ERROR(cm, thread_data->td->above_pred_buf,
+ (uint8_t *)aom_memalign(
+ 16, MAX_MB_PLANE * MAX_SB_SQUARE *
+ sizeof(*thread_data->td->above_pred_buf)));
+ CHECK_MEM_ERROR(cm, thread_data->td->left_pred_buf,
+ (uint8_t *)aom_memalign(
+ 16, MAX_MB_PLANE * MAX_SB_SQUARE *
+ sizeof(*thread_data->td->left_pred_buf)));
+
+ CHECK_MEM_ERROR(
+ cm, thread_data->td->wsrc_buf,
+ (int32_t *)aom_memalign(
+ 16, MAX_SB_SQUARE * sizeof(*thread_data->td->wsrc_buf)));
+
+ for (int x = 0; x < 2; x++)
+ for (int y = 0; y < 2; y++)
+ CHECK_MEM_ERROR(
+ cm, thread_data->td->hash_value_buffer[x][y],
+ (uint32_t *)aom_malloc(
+ AOM_BUFFER_SIZE_FOR_BLOCK_HASH *
+ sizeof(*thread_data->td->hash_value_buffer[0][0])));
+
+ CHECK_MEM_ERROR(
+ cm, thread_data->td->mask_buf,
+ (int32_t *)aom_memalign(
+ 16, MAX_SB_SQUARE * sizeof(*thread_data->td->mask_buf)));
+ // Allocate frame counters in thread data.
+ CHECK_MEM_ERROR(cm, thread_data->td->counts,
+ aom_calloc(1, sizeof(*thread_data->td->counts)));
+
+ // Allocate buffers used by palette coding mode.
+ CHECK_MEM_ERROR(
+ cm, thread_data->td->palette_buffer,
+ aom_memalign(16, sizeof(*thread_data->td->palette_buffer)));
+
+ // 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);
+ }
+}
+
+static void launch_enc_workers(AV1_COMP *cpi, int num_workers) {
+ const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+ // Encode a frame
+ for (int 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);
+ }
+}
+
+static void sync_enc_workers(AV1_COMP *cpi, int num_workers) {
+ const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+
+ // Encoding ends.
+ for (int i = 0; i < num_workers; i++) {
+ AVxWorker *const worker = &cpi->workers[i];
+ winterface->sync(worker);
+ }
+}
+
+static void accumulate_counters_enc_workers(AV1_COMP *cpi, int num_workers) {
+ for (int i = 0; i < num_workers; i++) {
+ AVxWorker *const worker = &cpi->workers[i];
+ EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
+ cpi->intrabc_used |= thread_data->td->intrabc_used_this_tile;
+ // Accumulate counters.
+ if (i < cpi->num_workers - 1) {
+ av1_accumulate_frame_counts(&cpi->counts, thread_data->td->counts);
+ accumulate_rd_opt(&cpi->td, thread_data->td);
+ cpi->td.mb.txb_split_count += thread_data->td->mb.txb_split_count;
+ }
+ }
+}
+
+static void prepare_enc_workers(AV1_COMP *cpi, AVxWorkerHook hook,
+ int num_workers) {
+ for (int i = 0; i < num_workers; i++) {
+ AVxWorker *const worker = &cpi->workers[i];
+ EncWorkerData *const thread_data = &cpi->tile_thr_data[i];
+
+ worker->hook = hook;
worker->data1 = thread_data;
worker->data2 = NULL;
@@ -167,38 +199,25 @@
if (i < num_workers - 1)
thread_data->td->mb.palette_buffer = thread_data->td->palette_buffer;
}
+}
- // Encode a frame
- for (i = 0; i < num_workers; i++) {
- AVxWorker *const worker = &cpi->workers[i];
- EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
+void av1_encode_tiles_mt(AV1_COMP *cpi) {
+ AV1_COMMON *const cm = &cpi->common;
+ const int tile_cols = cm->tile_cols;
+ const int tile_rows = cm->tile_rows;
+ int num_workers = AOMMIN(cpi->oxcf.max_threads, tile_cols * tile_rows);
- // Set the starting tile for each thread.
- thread_data->start = i;
-
- if (i == cpi->num_workers - 1)
- winterface->execute(worker);
- else
- winterface->launch(worker);
+ av1_init_tile_data(cpi);
+ // Only run once to create threads and allocate thread data.
+ if (cpi->num_workers == 0) {
+ create_enc_workers(cpi, num_workers);
+ } else {
+ num_workers = AOMMIN(num_workers, cpi->num_workers);
}
-
- // 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;
- cpi->intrabc_used |= thread_data->td->intrabc_used_this_tile;
- // Accumulate counters.
- if (i < cpi->num_workers - 1) {
- av1_accumulate_frame_counts(&cpi->counts, thread_data->td->counts);
- accumulate_rd_opt(&cpi->td, thread_data->td);
- cpi->td.mb.txb_split_count += thread_data->td->mb.txb_split_count;
- }
- }
+ prepare_enc_workers(cpi, (AVxWorkerHook)enc_worker_hook, num_workers);
+ launch_enc_workers(cpi, num_workers);
+ sync_enc_workers(cpi, num_workers);
+ accumulate_counters_enc_workers(cpi, num_workers);
}
// Accumulate frame counts. FRAME_COUNTS consist solely of 'unsigned int'
