# Copyright (C) 2018-2021 Intel Corporation # SPDX-License-Identifier: Apache-2.0 import os import sys from datetime import datetime from openvino.tools.benchmark.benchmark import Benchmark from openvino.tools.benchmark.parameters import parse_args from openvino.tools.benchmark.utils.constants import MULTI_DEVICE_NAME, HETERO_DEVICE_NAME, CPU_DEVICE_NAME, \ GPU_DEVICE_NAME, MYRIAD_DEVICE_NAME, GNA_DEVICE_NAME, BLOB_EXTENSION from openvino.tools.benchmark.utils.inputs_filling import set_inputs from openvino.tools.benchmark.utils.logging import logger from openvino.tools.benchmark.utils.progress_bar import ProgressBar from openvino.tools.benchmark.utils.utils import next_step, get_number_iterations, process_precision, \ process_help_inference_string, print_perf_counters, dump_exec_graph, get_duration_in_milliseconds, \ get_command_line_arguments, parse_nstreams_value_per_device, parse_devices, get_inputs_info, \ print_inputs_and_outputs_info, get_batch_size, load_config, dump_config from openvino.tools.benchmark.utils.statistics_report import StatisticsReport, averageCntReport, detailedCntReport def main(): # ------------------------------ 1. Parsing and validating input arguments ------------------------------------- next_step() run(parse_args()) def run(args): statistics = None try: if args.number_streams is None: logger.warning(" -nstreams default value is determined automatically for a device. " "Although the automatic selection usually provides a reasonable performance, " "but it still may be non-optimal for some cases, for more information look at README. ") command_line_arguments = get_command_line_arguments(sys.argv) if args.report_type: statistics = StatisticsReport(StatisticsReport.Config(args.report_type, args.report_folder)) statistics.add_parameters(StatisticsReport.Category.COMMAND_LINE_PARAMETERS, command_line_arguments) def is_flag_set_in_command_line(flag): return any(x.strip('-') == flag for x, y in command_line_arguments) device_name = args.target_device devices = parse_devices(device_name) device_number_streams = parse_nstreams_value_per_device(devices, args.number_streams) config = {} if args.load_config: load_config(args.load_config, config) is_network_compiled = False _, ext = os.path.splitext(args.path_to_model) if ext == BLOB_EXTENSION: is_network_compiled = True print("Network is compiled") # ------------------------------ 2. Loading Inference Engine --------------------------------------------------- next_step(step_id=2) benchmark = Benchmark(args.target_device, args.number_infer_requests, args.number_iterations, args.time, args.api_type) ## CPU (MKLDNN) extensions if CPU_DEVICE_NAME in device_name and args.path_to_extension: benchmark.add_extension(path_to_extension=args.path_to_extension) ## GPU (clDNN) Extensions if GPU_DEVICE_NAME in device_name and args.path_to_cldnn_config: if GPU_DEVICE_NAME not in config.keys(): config[GPU_DEVICE_NAME] = {} config[GPU_DEVICE_NAME]['CONFIG_FILE'] = args.path_to_cldnn_config if GPU_DEVICE_NAME in config.keys() and 'CONFIG_FILE' in config[GPU_DEVICE_NAME].keys(): cldnn_config = config[GPU_DEVICE_NAME]['CONFIG_FILE'] benchmark.add_extension(path_to_cldnn_config=cldnn_config) version = benchmark.get_version_info() logger.info(version) # --------------------- 3. Setting device configuration -------------------------------------------------------- next_step() perf_counts = False for device in devices: if device not in config.keys(): config[device] = {} ## Set performance counter if is_flag_set_in_command_line('pc'): ## set to user defined value config[device]['PERF_COUNT'] = 'YES' if args.perf_counts else 'NO' elif 'PERF_COUNT' in config[device].keys() and config[device]['PERF_COUNT'] == 'YES': logger.warning(f"Performance counters for {device} device is turned on. " + "To print results use -pc option.") elif args.report_type in [ averageCntReport, detailedCntReport ]: logger.warning(f"Turn on performance counters for {device} device " + f"since report type is {args.report_type}.") config[device]['PERF_COUNT'] = 'YES' elif args.exec_graph_path is not None: logger.warning(f"Turn on performance counters for {device} device " + "due to execution graph dumping.") config[device]['PERF_COUNT'] = 'YES' else: ## set to default value config[device]['PERF_COUNT'] = 'YES' if args.perf_counts else 'NO' perf_counts = True if config[device]['PERF_COUNT'] == 'YES' else perf_counts def set_throughput_streams(): key = device + "_THROUGHPUT_STREAMS" if device in device_number_streams.keys(): ## set to user defined value supported_config_keys = benchmark.ie.get_metric(device, 'SUPPORTED_CONFIG_KEYS') if key not in supported_config_keys: raise Exception(f"Device {device} doesn't support config key '{key}'! " + "Please specify -nstreams for correct devices in format :,:") config[device][key] = device_number_streams[device] elif key not in config[device].keys() and args.api_type == "async": logger.warning(f"-nstreams default value is determined automatically for {device} device. " + "Although the automatic selection usually provides a reasonable performance," "but it still may be non-optimal for some cases, for more information look at README.") if device != MYRIAD_DEVICE_NAME: ## MYRIAD sets the default number of streams implicitly config[device][key] = device + "_THROUGHPUT_AUTO" if key in config[device].keys(): device_number_streams[device] = config[device][key] if device == CPU_DEVICE_NAME: # CPU supports few special performance-oriented keys # limit threading for CPU portion of inference if args.number_threads and is_flag_set_in_command_line("nthreads"): config[device]['CPU_THREADS_NUM'] = str(args.number_threads) if is_flag_set_in_command_line("enforcebf16") or is_flag_set_in_command_line("enforce_bfloat16"): config[device]['ENFORCE_BF16'] = 'YES' if args.enforce_bfloat16 else 'NO' if is_flag_set_in_command_line('pin'): ## set to user defined value config[device]['CPU_BIND_THREAD'] = args.infer_threads_pinning elif 'CPU_BIND_THREAD' not in config[device].keys(): if MULTI_DEVICE_NAME in device_name and GPU_DEVICE_NAME in device_name: logger.warning(f"Turn off threads pinning for {device} " + "device since multi-scenario with GPU device is used.") config[device]['CPU_BIND_THREAD'] = 'NO' ## for CPU execution, more throughput-oriented execution via streams set_throughput_streams() elif device == GPU_DEVICE_NAME: ## for GPU execution, more throughput-oriented execution via streams set_throughput_streams() if MULTI_DEVICE_NAME in device_name and CPU_DEVICE_NAME in device_name: logger.warning("Turn on GPU trottling. Multi-device execution with the CPU + GPU performs best with GPU trottling hint, " + "which releases another CPU thread (that is otherwise used by the GPU driver for active polling)") config[device]['CLDNN_PLUGIN_THROTTLE'] = '1' elif device == MYRIAD_DEVICE_NAME: set_throughput_streams() config[device]['LOG_LEVEL'] = 'LOG_INFO' elif device == GNA_DEVICE_NAME: if is_flag_set_in_command_line('qb'): if args.qb == 8: config[device]['GNA_PRECISION'] = 'I8' else: config[device]['GNA_PRECISION'] = 'I16' if args.number_threads and is_flag_set_in_command_line("nthreads"): config[device]['GNA_LIB_N_THREADS'] = str(args.number_threads) else: supported_config_keys = benchmark.ie.get_metric(device, 'SUPPORTED_CONFIG_KEYS') if 'CPU_THREADS_NUM' in supported_config_keys and args.number_threads and is_flag_set_in_command_line("nthreads"): config[device]['CPU_THREADS_NUM'] = str(args.number_threads) if 'CPU_THROUGHPUT_STREAMS' in supported_config_keys and args.number_streams and is_flag_set_in_command_line("streams"): config[device]['CPU_THROUGHPUT_STREAMS'] = args.number_streams if 'CPU_BIND_THREAD' in supported_config_keys and args.infer_threads_pinning and is_flag_set_in_command_line("pin"): config[device]['CPU_BIND_THREAD'] = args.infer_threads_pinning perf_counts = perf_counts benchmark.set_config(config) batch_size = args.batch_size if args.cache_dir: benchmark.set_cache_dir(args.cache_dir) topology_name = "" load_from_file_enabled = is_flag_set_in_command_line('load_from_file') or is_flag_set_in_command_line('lfile') if load_from_file_enabled and not is_network_compiled: next_step() print("Skipping the step for loading network from file") next_step() print("Skipping the step for loading network from file") next_step() print("Skipping the step for loading network from file") # --------------------- 7. Loading the model to the device ------------------------------------------------- next_step() start_time = datetime.utcnow() exe_network = benchmark.load_network(args.path_to_model) duration_ms = f"{(datetime.utcnow() - start_time).total_seconds() * 1000:.2f}" logger.info(f"Load network took {duration_ms} ms") if statistics: statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('load network time (ms)', duration_ms) ]) app_inputs_info, _ = get_inputs_info(args.shape, args.layout, args.batch_size, exe_network.input_info) if batch_size == 0: batch_size = 1 elif not is_network_compiled: # --------------------- 4. Read the Intermediate Representation of the network ----------------------------- next_step() start_time = datetime.utcnow() ie_network = benchmark.read_network(args.path_to_model) topology_name = ie_network.name duration_ms = f"{(datetime.utcnow() - start_time).total_seconds() * 1000:.2f}" logger.info(f"Read network took {duration_ms} ms") if statistics: statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('read network time (ms)', duration_ms) ]) # --------------------- 5. Resizing network to match image sizes and given batch --------------------------- next_step() app_inputs_info, reshape = get_inputs_info(args.shape, args.layout, args.batch_size, ie_network.input_info) if reshape: start_time = datetime.utcnow() shapes = { k : v.shape for k,v in app_inputs_info.items() } logger.info( 'Reshaping network: {}'.format(', '.join("'{}': {}".format(k, v) for k, v in shapes.items()))) ie_network.reshape(shapes) duration_ms = f"{(datetime.utcnow() - start_time).total_seconds() * 1000:.2f}" logger.info(f"Reshape network took {duration_ms} ms") if statistics: statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('reshape network time (ms)', duration_ms) ]) # use batch size according to provided layout and shapes batch_size = get_batch_size(app_inputs_info) if args.layout else ie_network.batch_size logger.info(f'Network batch size: {batch_size}') # --------------------- 6. Configuring inputs and outputs of the model -------------------------------------------------- next_step() process_precision(ie_network, app_inputs_info, args.input_precision, args.output_precision, args.input_output_precision) print_inputs_and_outputs_info(ie_network) # --------------------- 7. Loading the model to the device ------------------------------------------------- next_step() start_time = datetime.utcnow() exe_network = benchmark.load_network(ie_network) duration_ms = f"{(datetime.utcnow() - start_time).total_seconds() * 1000:.2f}" logger.info(f"Load network took {duration_ms} ms") if statistics: statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('load network time (ms)', duration_ms) ]) else: next_step() print("Skipping the step for compiled network") next_step() print("Skipping the step for compiled network") next_step() print("Skipping the step for compiled network") # --------------------- 7. Loading the model to the device ------------------------------------------------- next_step() start_time = datetime.utcnow() exe_network = benchmark.import_network(args.path_to_model) duration_ms = f"{(datetime.utcnow() - start_time).total_seconds() * 1000:.2f}" logger.info(f"Import network took {duration_ms} ms") if statistics: statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('import network time (ms)', duration_ms) ]) app_inputs_info, _ = get_inputs_info(args.shape, args.layout, args.batch_size, exe_network.input_info) if batch_size == 0: batch_size = 1 # --------------------- 8. Setting optimal runtime parameters -------------------------------------------------- next_step() # Update number of streams for device in device_number_streams.keys(): key = device + '_THROUGHPUT_STREAMS' device_number_streams[device] = benchmark.ie.get_config(device, key) # Number of requests infer_requests = exe_network.requests # Iteration limit benchmark.niter = get_number_iterations(benchmark.niter, benchmark.nireq, args.api_type) # ------------------------------------ 9. Creating infer requests and filling input blobs ---------------------- next_step() paths_to_input = list() if args.paths_to_input: for path in args.paths_to_input: paths_to_input.append(os.path.abspath(*path) if args.paths_to_input else None) set_inputs(paths_to_input, batch_size, app_inputs_info, infer_requests) if statistics: statistics.add_parameters(StatisticsReport.Category.RUNTIME_CONFIG, [ ('topology', topology_name), ('target device', device_name), ('API', args.api_type), ('precision', "UNSPECIFIED"), ('batch size', str(batch_size)), ('number of iterations', str(benchmark.niter) if benchmark.niter else "0"), ('number of parallel infer requests', str(benchmark.nireq)), ('duration (ms)', str(get_duration_in_milliseconds(benchmark.duration_seconds))), ]) for nstreams in device_number_streams.items(): statistics.add_parameters(StatisticsReport.Category.RUNTIME_CONFIG, [ (f"number of {nstreams[0]} streams", str(nstreams[1])), ]) # ------------------------------------ 10. Measuring performance ----------------------------------------------- output_string = process_help_inference_string(benchmark) next_step(additional_info=output_string) progress_bar_total_count = 10000 if benchmark.niter and not benchmark.duration_seconds: progress_bar_total_count = benchmark.niter progress_bar = ProgressBar(progress_bar_total_count, args.stream_output, args.progress) if args.progress else None duration_ms = f"{benchmark.first_infer(exe_network):.2f}" logger.info(f"First inference took {duration_ms} ms") if statistics: statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('first inference time (ms)', duration_ms) ]) fps, latency_ms, total_duration_sec, iteration = benchmark.infer(exe_network, batch_size, progress_bar) # ------------------------------------ 11. Dumping statistics report ------------------------------------------- next_step() if args.dump_config: dump_config(args.dump_config, config) logger.info(f"Inference Engine configuration settings were dumped to {args.dump_config}") if args.exec_graph_path: dump_exec_graph(exe_network, args.exec_graph_path) if perf_counts: perfs_count_list = [] for ni in range(int(benchmark.nireq)): perfs_count_list.append(exe_network.requests[ni].get_perf_counts()) if args.perf_counts: print_perf_counters(perfs_count_list) if statistics: statistics.dump_performance_counters(perfs_count_list) if statistics: statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('total execution time (ms)', f'{get_duration_in_milliseconds(total_duration_sec):.2f}'), ('total number of iterations', str(iteration)), ]) if MULTI_DEVICE_NAME not in device_name: statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('latency (ms)', f'{latency_ms:.2f}'), ]) statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('throughput', f'{fps:.2f}'), ]) if statistics: statistics.dump() print(f'Count: {iteration} iterations') print(f'Duration: {get_duration_in_milliseconds(total_duration_sec):.2f} ms') if MULTI_DEVICE_NAME not in device_name: print(f'Latency: {latency_ms:.2f} ms') print(f'Throughput: {fps:.2f} FPS') del exe_network next_step.step_id = 0 except Exception as e: logger.exception(e) if statistics: statistics.add_parameters(StatisticsReport.Category.EXECUTION_RESULTS, [ ('error', str(e)), ]) statistics.dump() sys.exit(1)