void init_analyzer() { const int aligned_width = ALIGN_POWER_OF_TWO(info->frame_width, MI_SIZE_LOG2); const int aligned_height = ALIGN_POWER_OF_TWO(info->frame_height, MI_SIZE_LOG2); int mi_cols = aligned_width >> MI_SIZE_LOG2; int mi_rows = aligned_height >> MI_SIZE_LOG2; int mi_length = mi_cols * mi_rows; printf("init_analyzer: %d:%d (mi)\n", mi_cols, mi_rows); analyzer_data.mi_grid.buffer = aom_malloc(sizeof(AnalyzerMI) * mi_length); analyzer_data.mi_grid.length = mi_length; size_t size = aligned_width * aligned_height * 2; AnalyzerImagePlane *planes = analyzer_data.predicted_image.planes; planes[AOM_PLANE_Y].size = size; planes[AOM_PLANE_U].size = size; planes[AOM_PLANE_V].size = size; planes[AOM_PLANE_Y].stride = info->frame_width; planes[AOM_PLANE_U].stride = info->frame_width >> 1; planes[AOM_PLANE_V].stride = info->frame_width >> 1; planes[AOM_PLANE_Y].buffer = aom_malloc(size); planes[AOM_PLANE_U].buffer = aom_malloc(size); planes[AOM_PLANE_V].buffer = aom_malloc(size); }
void vp9_set_mb_mi(VP9_COMMON *cm, int width, int height) { const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2); const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2); cm->mi_cols = aligned_width >> MI_SIZE_LOG2; cm->mi_rows = aligned_height >> MI_SIZE_LOG2; cm->mi_stride = calc_mi_size(cm->mi_cols); cm->mb_cols = (cm->mi_cols + 1) >> 1; cm->mb_rows = (cm->mi_rows + 1) >> 1; cm->MBs = cm->mb_rows * cm->mb_cols; }
static int get_max_tile_cols(VP10_COMP *cpi) { const int aligned_width = ALIGN_POWER_OF_TWO(cpi->oxcf.width, MI_SIZE_LOG2); int mi_cols = aligned_width >> MI_SIZE_LOG2; int min_log2_tile_cols, max_log2_tile_cols; int log2_tile_cols; vp10_get_tile_n_bits(mi_cols, &min_log2_tile_cols, &max_log2_tile_cols); log2_tile_cols = clamp(cpi->oxcf.tile_columns, min_log2_tile_cols, max_log2_tile_cols); return (1 << log2_tile_cols); }
static int get_max_tile_cols(VP9_COMP *cpi) { const int aligned_width = ALIGN_POWER_OF_TWO(cpi->oxcf.width, MI_SIZE_LOG2); int mi_cols = aligned_width >> MI_SIZE_LOG2; int min_log2_tile_cols, max_log2_tile_cols; int log2_tile_cols; vp9_get_tile_n_bits(mi_cols, &min_log2_tile_cols, &max_log2_tile_cols); log2_tile_cols = clamp(cpi->oxcf.tile_columns, min_log2_tile_cols, max_log2_tile_cols); if (cpi->oxcf.target_level == LEVEL_AUTO) { const int level_tile_cols = log_tile_cols_from_picsize_level(cpi->common.width, cpi->common.height); if (log2_tile_cols > level_tile_cols) { log2_tile_cols = VPXMAX(level_tile_cols, min_log2_tile_cols); } } return (1 << log2_tile_cols); }
// 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]); } }
int main(int argc, char **argv) { aom_codec_ctx_t codec; AvxVideoReader *reader = NULL; AvxVideoWriter *writer = NULL; const AvxInterface *decoder = NULL; const AvxVideoInfo *info = NULL; int num_references; int i; aom_codec_pts_t pts; const char *tile_list_file = NULL; exec_name = argv[0]; if (argc != 5) die("Invalid number of arguments."); reader = aom_video_reader_open(argv[1]); if (!reader) die("Failed to open %s for reading.", argv[1]); num_references = (int)strtol(argv[3], NULL, 0); info = aom_video_reader_get_info(reader); // The writer to write out ivf file in tile list OBU, which can be decoded by // AV1 decoder. writer = aom_video_writer_open(argv[2], kContainerIVF, info); if (!writer) die("Failed to open %s for writing", argv[2]); tile_list_file = argv[4]; decoder = get_aom_decoder_by_fourcc(info->codec_fourcc); if (!decoder) die("Unknown input codec."); printf("Using %s\n", aom_codec_iface_name(decoder->codec_interface())); if (aom_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0)) die_codec(&codec, "Failed to initialize decoder."); // Decode anchor frames. aom_codec_control_(&codec, AV1_SET_TILE_MODE, 0); printf("Reading %d reference images.\n", num_references); for (i = 0; i < num_references; ++i) { aom_video_reader_read_frame(reader); size_t frame_size = 0; const unsigned char *frame = aom_video_reader_get_frame(reader, &frame_size); pts = (aom_codec_pts_t)aom_video_reader_get_frame_pts(reader); // Copy references bitstream directly. if (!aom_video_writer_write_frame(writer, frame, frame_size, pts)) die_codec(&codec, "Failed to copy compressed anchor frame."); if (aom_codec_decode(&codec, frame, frame_size, NULL)) die_codec(&codec, "Failed to decode frame."); } // Decode camera frames. aom_codec_control_(&codec, AV1_SET_TILE_MODE, 1); aom_codec_control_(&codec, AV1D_EXT_TILE_DEBUG, 1); FILE *infile = aom_video_reader_get_file(reader); // Record the offset of the first camera image. const FileOffset camera_frame_pos = ftello(infile); printf("Loading compressed frames into memory.\n"); // Count the frames in the lightfield. int num_frames = 0; while (aom_video_reader_read_frame(reader)) { ++num_frames; } if (num_frames < 1) die("Input light field has no frames."); // Read all of the lightfield frames into memory. unsigned char **frames = (unsigned char **)malloc(num_frames * sizeof(unsigned char *)); size_t *frame_sizes = (size_t *)malloc(num_frames * sizeof(size_t)); // Seek to the first camera image. fseeko(infile, camera_frame_pos, SEEK_SET); for (int f = 0; f < num_frames; ++f) { aom_video_reader_read_frame(reader); size_t frame_size = 0; const unsigned char *frame = aom_video_reader_get_frame(reader, &frame_size); frames[f] = (unsigned char *)malloc(frame_size * sizeof(unsigned char)); memcpy(frames[f], frame, frame_size); frame_sizes[f] = frame_size; } printf("Read %d frames.\n", num_frames); // Copy first camera frame for getting camera frame header. This is done // only once. { size_t frame_size = frame_sizes[0]; const unsigned char *frame = frames[0]; pts = num_references; aom_tile_data frame_header_info = { 0, NULL, 0 }; // Need to decode frame header to get camera frame header info. So, here // decoding 1 tile is enough. aom_codec_control_(&codec, AV1_SET_DECODE_TILE_ROW, 0); aom_codec_control_(&codec, AV1_SET_DECODE_TILE_COL, 0); aom_codec_err_t aom_status = aom_codec_decode(&codec, frame, frame_size, NULL); if (aom_status) die_codec(&codec, "Failed to decode tile."); aom_codec_control_(&codec, AV1D_GET_FRAME_HEADER_INFO, &frame_header_info); size_t obu_size_offset = (uint8_t *)frame_header_info.coded_tile_data - frame; size_t length_field_size = frame_header_info.coded_tile_data_size; // Remove ext-tile tile info. uint32_t frame_header_size = (uint32_t)frame_header_info.extra_size - 1; size_t bytes_to_copy = obu_size_offset + length_field_size + frame_header_size; unsigned char *frame_hdr_buf = (unsigned char *)malloc(bytes_to_copy); if (frame_hdr_buf == NULL) die_codec(&codec, "Failed to allocate frame header buffer."); memcpy(frame_hdr_buf, frame, bytes_to_copy); // Update frame header OBU size. size_t bytes_written = 0; if (aom_uleb_encode_fixed_size( frame_header_size, length_field_size, length_field_size, frame_hdr_buf + obu_size_offset, &bytes_written)) die_codec(&codec, "Failed to encode the tile list obu size."); // Copy camera frame header bitstream. if (!aom_video_writer_write_frame(writer, frame_hdr_buf, bytes_to_copy, pts)) die_codec(&codec, "Failed to copy compressed camera frame header."); free(frame_hdr_buf); } // Read out the image format. aom_img_fmt_t ref_fmt = 0; if (aom_codec_control(&codec, AV1D_GET_IMG_FORMAT, &ref_fmt)) die_codec(&codec, "Failed to get the image format"); const int bps = get_image_bps(ref_fmt); if (!bps) die_codec(&codec, "Invalid image format."); // read out the tile size. unsigned int tile_size = 0; if (aom_codec_control(&codec, AV1D_GET_TILE_SIZE, &tile_size)) die_codec(&codec, "Failed to get the tile size"); const unsigned int tile_width = tile_size >> 16; const unsigned int tile_height = tile_size & 65535; // Allocate a buffer to store tile list bitstream. const size_t data_sz = MAX_TILES * ALIGN_POWER_OF_TWO(tile_width, 5) * ALIGN_POWER_OF_TWO(tile_height, 5) * bps / 8; unsigned char *tl_buf = (unsigned char *)malloc(data_sz); if (tl_buf == NULL) die_codec(&codec, "Failed to allocate tile list buffer."); aom_codec_pts_t tl_pts = num_references; const uint8_t output_frame_width_in_tiles_minus_1 = output_frame_width / tile_width - 1; const uint8_t output_frame_height_in_tiles_minus_1 = output_frame_height / tile_height - 1; printf("Reading tile list from file.\n"); char line[1024]; FILE *tile_list_fptr = fopen(tile_list_file, "r"); if (!tile_list_fptr) die_codec(&codec, "Failed to open tile list file."); int num_tiles = 0; TILE_LIST_INFO tiles[MAX_TILES]; while ((fgets(line, 1024, tile_list_fptr)) != NULL) { if (line[0] == 'F' || num_tiles >= MAX_TILES) { // Flush existing tile list and start another, either because we hit a // new render frame or because we've hit our max number of tiles per list. if (num_tiles > 0) { process_tile_list(tiles, num_tiles, tl_pts, frames, frame_sizes, &codec, tl_buf, writer, output_frame_width_in_tiles_minus_1, output_frame_height_in_tiles_minus_1); ++tl_pts; } num_tiles = 0; } if (line[0] == 'F') { continue; } if (sscanf(line, "%d %d %d %d", &tiles[num_tiles].image_idx, &tiles[num_tiles].reference_idx, &tiles[num_tiles].tile_col, &tiles[num_tiles].tile_row) == 4) { if (tiles[num_tiles].image_idx >= num_frames) { die("Tile list image_idx out of bounds: %d >= %d.", tiles[num_tiles].image_idx, num_frames); } if (tiles[num_tiles].reference_idx >= num_references) { die("Tile list reference_idx out of bounds: %d >= %d.", tiles[num_tiles].reference_idx, num_references); } ++num_tiles; } } if (num_tiles > 0) { // Flush out the last tile list. process_tile_list(tiles, num_tiles, tl_pts, frames, frame_sizes, &codec, tl_buf, writer, output_frame_width_in_tiles_minus_1, output_frame_height_in_tiles_minus_1); ++tl_pts; } const int num_tile_lists = (int)(tl_pts - pts); printf("Finished processing tile lists. Num tile lists: %d.\n", num_tile_lists); free(tl_buf); for (int f = 0; f < num_frames; ++f) { free(frames[f]); } free(frame_sizes); free(frames); if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec"); aom_video_writer_close(writer); aom_video_reader_close(reader); return EXIT_SUCCESS; }