def build_default_flags(self):
flag_dict = {}
flag_dict["verbose"] = self.verbose
# Handle plugins
if self.name in plugin_map:
plugins = plugin_map[self.name]
for plugin in plugins:
self.check_file_exists(plugin)
flag_dict["plugins"] = ",".join(plugins)
# Generate flags for logfile names.
log_dir = self.get_full_log_dir()
if not os.path.exists(log_dir):
os.makedirs(log_dir)
flag_dict["logfile_outdir"] = log_dir
flag_dict["logfile_prefix"] = "mlperf_log_"
# Handle performance sample count
perf_sample_count = dict_get(self.args, "performance_sample_count", None)
if perf_sample_count is not None:
flag_dict["performance_sample_count"] = perf_sample_count
elif benchmark_qsl_size_map[self.name] > 0:
flag_dict["performance_sample_count"] = benchmark_qsl_size_map[self.name]
else:
flag_dict["performance_sample_count"] = self.args["gpu_batch_size"]
# Handle custom arguments
for arg in self.flag_builder_custom_args:
val = dict_get(self.args, arg, None)
if val is not None:
flag_dict[arg] = val
return flag_dict
def build_scenario_specific_flags(self):
flag_dict = {}
prefix = self.qps_prefix
if self.scenario == SCENARIOS.SingleStream:
scenario_keys = common_args.SINGLE_STREAM_PARAMS
elif self.scenario == SCENARIOS.Offline:
scenario_keys = common_args.OFFLINE_PARAMS
elif self.scenario == SCENARIOS.MultiStream:
scenario_keys = common_args.MULTI_STREAM_PARAMS
elif self.scenario == SCENARIOS.Server:
scenario_keys = common_args.SERVER_PARAMS
else:
scenario_keys = []
raise RuntimeError("Unknown Scenario \"{}\"".format(self.scenario))
for arg in scenario_keys:
val = dict_get(self.args, prefix + arg, None)
if val is None:
raise ValueError("Missing required key {:}".format(prefix +
arg))
flag_dict[arg] = val
# Handle RUN_ARGS
for arg in scenario_keys:
val = dict_get(self.args, arg, None)
if val is not None:
flag_dict[arg] = val
return flag_dict
def __init__(self, args):
workspace_size = dict_get(args, "workspace_size", default=(5 << 30))
logging.info("Using workspace size: {:,}".format(workspace_size))
super().__init__(args,
name=BENCHMARKS.BERT,
workspace_size=workspace_size)
self.bert_config_path = "code/bert/tensorrt/bert_config.json"
self.seq_len = 384 # default sequence length
self.batch_size = dict_get(args, "batch_size", default=1)
self.num_profiles = 1
if 'gpu_inference_streams' in args:
# use gpu_inference_streams to determine the number of duplicated profiles
# in the engine when not using lwis mode
self.num_profiles = args['gpu_inference_streams']
self.is_int8 = args['precision'] == 'int8'
if self.is_int8:
self.model_path = dict_get(
args,
"model_path",
default="build/models/bert/bert_large_v1_1_fake_quant.onnx")
else:
self.model_path = dict_get(
args,
"model_path",
default="build/models/bert/bert_large_v1_1.onnx")
self.bert_config = BertConfig(self.bert_config_path)
self.enable_interleaved = False
if self.is_int8 and 'enable_interleaved' in args:
self.enable_interleaved = args['enable_interleaved']
# Small-Tile GEMM Plugin
# Since it doesn't support interleaved format, two options are mutually exclusive
self.use_small_tile_gemm_plugin = self.args.get(
"use_small_tile_gemm_plugin", False)
self.gemm_plugin_fairshare_cache_size = self.args.get(
"gemm_plugin_fairshare_cache_size", -1)
if self.enable_interleaved and self.use_small_tile_gemm_plugin:
assert False, "Small-Tile GEMM Plugin doesn't support interleaved format."
# Query system id for architecture
self.system = get_system()
self.gpu_arch = self.system.arch
if self.batch_size > 512:
# tactics selection is limited at very large batch sizes
self.builder_config.max_workspace_size = 7 << 30
if 'nx' in self.system.gpu.lower():
# use 1GB only for XavierNX
self.builder_config.max_workspace_size = 1 << 30
def _build_custom_flags(self, flag_dict):
# Rename gpu_batch_size to batch_size
batch_size = dict_get(self.args, "gpu_batch_size", default=None)
flag_dict["batch_size"] = batch_size
flag_dict["gpu_batch_size"] = None
# Rename use_graphs to cuda_graph
use_graphs = dict_get(self.args, "use_graphs", default=False)
flag_dict["cuda_graph"] = use_graphs
flag_dict["use_graphs"] = None
# Rename max_seq_length to hp_max_seq_length
max_seq_length = dict_get(self.args, "max_seq_length", default=None)
flag_dict["hp_max_seq_length"] = max_seq_length
flag_dict["max_seq_length"] = None
# Handle more harness_rnnt knobs
no_pipelined = dict_get(self.args, "nopipelined_execution", default=False)
flag_dict["pipelined_execution"] = not no_pipelined
flag_dict["nopipelined_execution"] = None
# Handle more harness_rnnt knobs : disable batch sorting by sequence length
no_sorting = dict_get(self.args, "nobatch_sorting", default=False)
flag_dict["batch_sorting"] = not no_sorting
flag_dict["nobatch_sorting"] = None
# Handle yet another harness_rnnt knob: turning off DALI preprocessing for debug
no_dali = dict_get(self.args, "noenable_audio_processing", default=False)
flag_dict["enable_audio_processing"] = not no_dali
flag_dict["noenable_audio_processing"] = None
# Handle yet another harness_rnnt knob: disable DALI's scatter gather kernel
no_copy_kernel = dict_get(self.args, "nouse_copy_kernel", default=False)
flag_dict["use_copy_kernel"] = not no_copy_kernel
flag_dict["nouse_copy_kernel"] = None
# Rename gpu_inference_streams to streams_per_gpu
num_inference = dict_get(self.args, "gpu_inference_streams", default=None)
flag_dict["streams_per_gpu"] = num_inference
flag_dict["gpu_inference_streams"] = None
audio_fp16_input = dict_get(self.args, "audio_fp16_input", default=True)
flag_dict["audio_fp16_input"] = audio_fp16_input
start_from_device = dict_get(self.args, "start_from_device", default=False)
flag_dict["start_from_device"] = start_from_device
audio_input_suffix = "fp16" if audio_fp16_input else "fp32"
flag_dict["audio_serialized_pipeline_file"] = "build/bin/dali" + "/dali_pipeline_gpu_" + audio_input_suffix + ".pth"
argstr = args_to_string(flag_dict) + " --scenario {:} --model {:}".format(self.scenario, self.name)
# Handle engine dir
argstr += " --engine_dir={:}".format(self.engine_dir)
return argstr
def __init__(self, args, name="", skip_file_checks=False):
self.args = args
self.name = name
self.verbose = dict_get(args, "verbose", default=None)
if self.verbose:
logging.info("===== Harness arguments for {:} =====".format(name))
for key in args:
logging.info("{:}={:}".format(key, args[key]))
self.system_id = args["system_id"]
self.scenario = args["scenario"]
self.config_ver = args["config_ver"]
self.engine_dir = "./build/engines/{:}/{:}/{:}".format(
self.system_id, self.name, self.scenario)
self.precision = args["precision"]
# Detect devices used to set field prefixes
self.has_gpu = dict_get(args, "gpu_batch_size",
default=None) is not None
self.has_dla = dict_get(args, "dla_batch_size",
default=None) is not None
self.qps_prefix = ""
if self.has_gpu and self.has_dla:
self.qps_prefix = "concurrent_"
elif self.has_gpu:
self.qps_prefix = "gpu_"
elif self.has_dla:
self.qps_prefix = "dla_"
# Check if we actually need to execute the harness
self.generate_conf_files_only = False
if dict_get(self.args, "generate_conf_files_only", False):
logging.info("Only generating measurements/ configuration entries")
self.generate_conf_files_only = True
self.args["generate_conf_files_only"] = None
# Enumerate engine files
# Engine not needed if we are only generating measurements/ entries
self.skip_file_checks = skip_file_checks or self.generate_conf_files_only
self.gpu_engine = None
self.dla_engine = None
self.enumerate_engines()
# Enumerate harness executable
self.executable = self._get_harness_executable()
self.check_file_exists(self.executable)
self.use_jemalloc = False
self.env_vars = os.environ.copy()
self.flag_builder_custom_args = []
def _get_submission_benchmark_name(self):
full_benchmark_name = self.name
if dict_get(self.args, "accuracy_level", "99%") == "99.9%":
full_benchmark_name += "-99.9"
elif self.name in BENCHMARKS.HIGH_ACC_ENABLED:
full_benchmark_name += "-99"
return full_benchmark_name
def __init__(self, args):
workspace_size = dict_get(args, "workspace_size", default=(1 << 30))
logging.info("Use workspace_size: {:}".format(workspace_size))
super().__init__(args,
name=BENCHMARKS.ResNet50,
workspace_size=workspace_size)
# Model path
self.model_path = dict_get(
args,
"model_path",
default="code/resnet50/tensorrt/ofa_autosinian_is176.onnx")
logging.info("Using AutoSinian optimized once-for-all network")
self.cache_file = None
self.need_calibration = False
if self.precision == "int8":
# Get calibrator variables
calib_batch_size = dict_get(self.args,
"calib_batch_size",
default=1)
calib_max_batches = dict_get(self.args,
"calib_max_batches",
default=500)
force_calibration = dict_get(self.args,
"force_calibration",
default=False)
cache_file = dict_get(
self.args,
"cache_file",
default="code/resnet50/tensorrt/calibrator.cache")
preprocessed_data_dir = dict_get(self.args,
"preprocessed_data_dir",
default="build/preprocessed_data")
calib_data_map = dict_get(self.args,
"calib_data_map",
default="data_maps/imagenet/cal_map.txt")
calib_image_dir = os.path.join(preprocessed_data_dir,
"imagenet/ResNet50/fp32")
# Set up calibrator
self.calibrator = RN50Calibrator(
calib_batch_size=calib_batch_size,
calib_max_batches=calib_max_batches,
force_calibration=force_calibration,
cache_file=cache_file,
image_dir=calib_image_dir,
calib_data_map=calib_data_map)
self.builder_config.int8_calibrator = self.calibrator
self.cache_file = cache_file
self.need_calibration = force_calibration or not os.path.exists(
cache_file)
def __init__(self, args):
workspace_size = dict_get(args, "workspace_size", default=(5 << 30))
logging.info("Use workspace_size: {:}".format(workspace_size))
super().__init__(args,
name=BENCHMARKS.BERT,
workspace_size=workspace_size)
self.bert_config_path = "code/bert/tensorrt/bert_config.json"
self.seq_len = 384 # default sequence length
assert 'batch_size' in args, 'batch_size is not specified'
self.batch_size = args['batch_size']
self.num_profiles = 1
if 'gpu_inference_streams' in args:
# use gpu_inference_streams to determine the number of duplicated profiles
# in the engine when not using lwis mode
self.num_profiles = args['gpu_inference_streams']
self.is_int8 = args['precision'] == 'int8'
if self.is_int8:
self.model_path = dict_get(
args,
"model_path",
default="build/models/bert/bert_large_v1_1_fake_quant.onnx")
else:
self.model_path = dict_get(
args,
"model_path",
default="build/models/bert/bert_large_v1_1.onnx")
self.bert_config = BertConfig(self.bert_config_path)
self.enable_il = False
if self.is_int8 and 'enable_interleaved' in args:
self.enable_il = args['enable_interleaved']
if self.batch_size > 512:
# tactics selection is limited at very large batch sizes
self.builder_config.max_workspace_size = 7 << 30
if 'nx' in self.system_id.lower():
# use 1GB only for XavierNX
self.builder_config.max_workspace_size = 1 << 30
def get_system_name(self):
override_system_name = dict_get(self.args, "system_name", default=None)
if override_system_name not in {None, ""}:
return override_system_name
system_name = self.system_id
for kw in system_name_map.keys():
if kw in self.system_id:
system_name = "_".join([system_name_map[kw], system_name])
break
return "{:}_TRT{:}".format(system_name, TENSORRT_VERSION)
def get_system_name(self):
override_system_name = dict_get(self.args, "system_name", default=None)
if override_system_name not in {None, ""}:
return override_system_name
system_name = self.system_id
for kw, prepend_name in system_name_map:
if kw in self.system_id:
system_name = "_".join([prepend_name, system_name])
break
full_system_name = "{:}_TRT{:}".format(system_name, TENSORRT_VERSION)
return self._append_config_ver_name(full_system_name)
def build_scenario_specific_flags(self):
"""Return flags specific to current scenario."""
flag_dict = {}
prefix = self.qps_prefix
scenario_keys = common_args.getScenarioMetricArgs(self.scenario)
for arg in scenario_keys:
val = dict_get(self.args, prefix + arg, None)
if val is None:
raise ValueError("Missing required key {:}".format(prefix +
arg))
flag_dict[arg] = val
# Handle RUN_ARGS
for arg in scenario_keys:
val = dict_get(self.args, arg, None)
if val is not None:
flag_dict[arg] = val
return flag_dict
def __init__(self, args):
"""Set up the config and calibrator for DLRM. Does not initialize."""
workspace_size = dict_get(args, "workspace_size", default=(4 << 30))
logging.info("Using workspace size: {:,}".format(workspace_size))
super().__init__(args, name=BENCHMARKS.DLRM, workspace_size=workspace_size)
with open("code/dlrm/tensorrt/mlperf_40m.limit.json") as f:
self.dlrm_config = json.load(f)
logging.info("DLRM config: {:}".format(self.dlrm_config))
self.num_numerical_inputs = self.dlrm_config["num_numerical_features"]
self.num_features = len(self.dlrm_config["categorical_feature_sizes"])
self.num_interactions = (self.num_features + 1) * self.num_features // 2
self.embedding_size = self.dlrm_config["embedding_dim"]
self.embedding_rows = self.dlrm_config["categorical_feature_sizes"]
self.embedding_rows_bound = 40000000
self.embedding_rows = [min(i, self.embedding_rows_bound) for i in self.embedding_rows]
self.embedding_rows_total = np.sum(np.array(self.embedding_rows))
self.bottom_mlp_channels = self.dlrm_config["bottom_mlp_sizes"]
self.bottom_mlp_names = ["bot_l.0", "bot_l.2", "bot_l.4"]
self.output_padding = self.args.get("output_padding_granularity", 32)
self.top_mlp_input_size = (self.num_interactions + self.embedding_size + self.output_padding - 1) // self.output_padding * self.output_padding
self.top_mlp_channels = self.dlrm_config["top_mlp_sizes"]
self.top_mlp_names = ["top_l.0", "top_l.2", "top_l.4", "top_l.6", "top_l.8"]
self.model_filepath = "build/models/dlrm/tb00_40M.pt"
self.embedding_weights_binary_filepath = "build/models/dlrm/40m_limit/dlrm_embedding_weights_int8_v3.bin"
self.model_without_embedding_weights_filepath = "build/models/dlrm/40m_limit/model_test_without_embedding_weights_v3.pt"
self.row_frequencies_binary_filepath = "build/models/dlrm/40m_limit/row_frequencies.bin"
self.row_frequencies_src_dir = "build/models/dlrm/40m_limit/row_freq"
self.embedding_weights_on_gpu_part = self.args.get("embedding_weights_on_gpu_part", 1.0)
self.use_row_frequencies = True if self.embedding_weights_on_gpu_part < 1.0 else False
self.num_profiles = self.args.get("gpu_inference_streams", 1)
self.use_small_tile_gemm_plugin = self.args.get("use_small_tile_gemm_plugin", False)
self.gemm_plugin_fairshare_cache_size = self.args.get("gemm_plugin_fairshare_cache_size", -1)
self.enable_interleaved_top_mlp = self.args.get("enable_interleaved_top_mlp", False)
if self.precision == "fp16":
self.apply_flag(trt.BuilderFlag.FP16)
elif self.precision == "int8":
self.apply_flag(trt.BuilderFlag.INT8)
if self.precision == "int8":
# Get calibrator variables
calib_batch_size = dict_get(self.args, "calib_batch_size", default=512)
calib_max_batches = dict_get(self.args, "calib_max_batches", default=500)
force_calibration = dict_get(self.args, "force_calibration", default=False)
cache_file = dict_get(self.args, "cache_file", default="code/dlrm/tensorrt/calibrator.cache")
preprocessed_data_dir = dict_get(self.args, "preprocessed_data_dir", default="build/preprocessed_data")
calib_data_dir = os.path.join(preprocessed_data_dir, "criteo/full_recalib/val_data_128000")
# Set up calibrator
self.calibrator = DLRMCalibrator(calib_batch_size=calib_batch_size, calib_max_batches=calib_max_batches,
force_calibration=force_calibration, cache_file=cache_file, data_dir=calib_data_dir)
self.builder_config.int8_calibrator = self.calibrator
self.cache_file = cache_file
self.need_calibration = force_calibration or not os.path.exists(cache_file)
else:
self.need_calibration = False
def initialize(self):
"""
Parse the processed model to create the network.
"""
# Create network.
self.network = self.builder.create_network(
1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
channel_idx = 1
# Input shape
input_tensor_dim = [-1] + self.input_volume_dim
input_tensor_dim.insert(channel_idx, self.num_input_channel)
# Parse from onnx file.
parser = trt.OnnxParser(self.network, self.logger)
model = self.preprocess_onnx(onnx.load(self.model_path))
success = parser.parse(onnx._serialize(model))
if not success:
raise RuntimeError(
"3D-Unet onnx model parsing failed! Error: {:}".format(
parser.get_error(0).desc()))
# Set input/output tensor dtype and formats
input_tensor = self.network.get_input(0)
output_tensor = self.network.get_output(0)
input_tensor.shape = input_tensor_dim
if self.input_dtype == "int8":
input_tensor.dtype = trt.int8
elif self.input_dtype == "fp16":
input_tensor.dtype = trt.float16
elif self.input_dtype == "fp32":
input_tensor.dtype = trt.float32
if self.input_format == "linear":
input_tensor.allowed_formats = 1 << int(trt.TensorFormat.LINEAR)
elif self.input_format == "dhwc8":
input_tensor.allowed_formats = 1 << int(trt.TensorFormat.DHWC8)
elif self.input_format == "cdhw32":
input_tensor.allowed_formats = 1 << int(trt.TensorFormat.CDHW32)
# Always use FP16 output
# workaround for calibration not working with the identity layer properly
force_calibration = dict_get(self.args,
"force_calibration",
default=False)
output_tensor.dtype = trt.float16 if force_calibration == False else trt.float32
output_tensor.allowed_formats = 1 << int(trt.TensorFormat.LINEAR)
self.initialized = True
def __init__(self, args):
workspace_size = dict_get(args, "workspace_size", default=(2 << 30))
logging.info("Using workspace size: {:,}".format(workspace_size))
super().__init__(args,
name=BENCHMARKS.SSDResNet34,
workspace_size=workspace_size)
# Model path
self.model_path = dict_get(
args,
"model_path",
default="build/models/SSDResNet34/resnet34-ssd1200.pytorch")
if self.precision == "int8":
force_calibration = dict_get(self.args,
"force_calibration",
default=False)
calib_batch_size = dict_get(self.args,
"calib_batch_size",
default=10)
calib_max_batches = dict_get(self.args,
"calib_max_batches",
default=50)
cache_file = dict_get(
self.args,
"cache_file",
default="code/ssd-resnet34/tensorrt/calibrator.cache")
preprocessed_data_dir = dict_get(self.args,
"preprocessed_data_dir",
default="build/preprocessed_data")
calib_data_map = dict_get(self.args,
"calib_data_map",
default="data_maps/coco/cal_map.txt")
calib_image_dir = os.path.join(preprocessed_data_dir,
"coco/train2017/SSDResNet34/fp32")
self.calibrator = SSDResNet34EntropyCalibrator(
calib_image_dir, cache_file, calib_batch_size,
calib_max_batches, force_calibration, calib_data_map)
self.builder_config.int8_calibrator = self.calibrator
self.cache_file = cache_file
def __init__(self, args):
workspace_size = dict_get(args, "workspace_size", default=(2 << 31))
logging.info("Use workspace_size: {:}".format(workspace_size))
super().__init__(args,
name="ssd-mobilenet",
workspace_size=workspace_size)
# Model path
self.model_path = dict_get(
args,
"model_path",
default="build/models/SSDMobileNet/frozen_inference_graph.pb")
if self.precision == "int8":
calib_batch_size = dict_get(self.args,
"calib_batch_size",
default=1)
calib_max_batches = dict_get(self.args,
"calib_max_batches",
default=500)
force_calibration = dict_get(self.args,
"force_calibration",
default=False)
cache_file = dict_get(
self.args,
"cache_file",
default="code/ssd-mobilenet/tensorrt/calibrator.cache")
preprocessed_data_dir = dict_get(self.args,
"preprocessed_data_dir",
default="build/preprocessed_data")
calib_data_map = dict_get(self.args,
"calib_data_map",
default="data_maps/coco/cal_map.txt")
calib_image_dir = os.path.join(preprocessed_data_dir,
"coco/train2017/SSDMobileNet/fp32")
self.calibrator = SSDMobileNetEntropyCalibrator(
calib_batch_size, calib_max_batches, force_calibration,
cache_file, calib_image_dir, calib_data_map)
self.builder_config.int8_calibrator = self.calibrator
self.cache_file = cache_file
def __init__(self, args):
workspace_size = dict_get(args, "workspace_size", default=(1 << 30))
logging.info("Use workspace_size: {:}".format(workspace_size))
super().__init__(args,
name=BENCHMARKS.ResNet50,
workspace_size=workspace_size)
# Model path
self.model_path = dict_get(
args,
"model_path",
default="code/resnet50/tensorrt/resnet50_inspur_open.onnx")
if self.precision == "int8":
# Get calibrator variables
calib_batch_size = dict_get(self.args,
"calib_batch_size",
default=1)
calib_max_batches = dict_get(self.args,
"calib_max_batches",
default=500)
force_calibration = dict_get(self.args,
"force_calibration",
default=False)
cache_file = dict_get(
self.args,
"cache_file",
default="code/resnet50/tensorrt/calibrator.cache")
calib_data_map = dict_get(self.args,
"calib_data_map",
default="data_maps/imagenet/cal_map.txt")
calib_image_dir = "build/data/imagenet"
# Set up calibrator
self.calibrator = RN50Calibrator(
calib_batch_size=calib_batch_size,
calib_max_batches=calib_max_batches,
force_calibration=force_calibration,
cache_file=cache_file,
image_dir=calib_image_dir,
calib_data_map=calib_data_map)
self.builder_config.int8_calibrator = self.calibrator
self.cache_file = cache_file
self.need_calibration = force_calibration or not os.path.exists(
cache_file)
def _build_custom_flags(self, flag_dict):
# Handle use_jemalloc
self.use_jemalloc = dict_get(flag_dict, "use_jemalloc", False)
flag_dict['use_jemalloc'] = None
argstr = args_to_string(flag_dict) + " --scenario " + self.scenario + " --model " + self.name
return argstr
def __init__(self, args, name="", workspace_size=(1 << 30)):
"""
Constructor
:param args: arguments represented by a dictionary
:param name: name of the benchmark
"""
self.name = name
self.args = args
# Configuration variables
self.verbose = dict_get(args, "verbose", default=False)
if self.verbose:
logging.info("========= BenchmarkBuilder Arguments =========")
for arg in args:
logging.info("{:}={:}".format(arg, args[arg]))
self.system_id = args["system_id"]
self.scenario = args["scenario"]
self.config_ver = args["config_ver"]
self.engine_dir = "./build/engines/{:}/{:}/{:}".format(
self.system_id, self.name, self.scenario)
# Set up logger, builder, and network.
self.logger = trt.Logger(
trt.Logger.VERBOSE if self.verbose else trt.Logger.INFO)
trt.init_libnvinfer_plugins(self.logger, "")
self.builder = trt.Builder(self.logger)
self.builder_config = self.builder.create_builder_config()
self.builder_config.max_workspace_size = workspace_size
if dict_get(args, "verbose_nvtx", default=False):
self.builder_config.profiling_verbosity = trt.ProfilingVerbosity.VERBOSE
# Precision variables
self.input_dtype = dict_get(args, "input_dtype", default="fp32")
self.input_format = dict_get(args, "input_format", default="linear")
self.precision = dict_get(args, "precision", default="int8")
self.clear_flag(trt.BuilderFlag.TF32)
if self.precision == "fp16":
self.apply_flag(trt.BuilderFlag.FP16)
elif self.precision == "int8":
self.apply_flag(trt.BuilderFlag.INT8)
# Device variables
self.device_type = "gpu"
self.dla_core = args.get("dla_core", None)
if self.dla_core is not None:
logging.info("Using DLA: Core {:}".format(self.dla_core))
self.device_type = "dla"
self.apply_flag(trt.BuilderFlag.GPU_FALLBACK)
self.builder_config.default_device_type = trt.DeviceType.DLA
self.builder_config.DLA_core = int(self.dla_core)
if self.scenario == SCENARIOS.SingleStream:
self.batch_size = 1
elif self.scenario in [
SCENARIOS.Server, SCENARIOS.Offline, SCENARIOS.MultiStream
]:
self.batch_size = self.args.get("batch_size", 1)
else:
raise ValueError("Invalid scenario: {:}".format(self.scenario))
# Currently, TRT has limitation that we can only create one execution
# context for each optimization profile. Therefore, create more profiles
# so that LWIS can create multiple contexts.
self.num_profiles = self.args.get("gpu_copy_streams", 4)
self.initialized = False
def __init__(self, args):
workspace_size = dict_get(args, "workspace_size", default=(8 << 30))
super().__init__(args,
name=BENCHMARKS.UNET,
workspace_size=workspace_size)
# input channel
self.num_input_channel = 4
# input volume dimension
self.input_volume_dim = [224, 224, 160]
# use InstNorm3D plugin
self.use_instnorm3d_plugin = True
# use pixelShuffle plugin
self.enable_pixelshuffle3d_plugin = True
self.enable_pixelshuffle3d_plugin_concat_fuse = True
# Deconv->Conv conversion
self.use_conv_for_deconv = True
self.pixel_shuffle_cdwh = True # If false, do dhwc
# use last layer plugin
self.use_conv3d1x1x1k4_plugin = True
# Model is imported from ONNX
self.model_path = dict_get(
args,
"model_path",
default="build/models/3d-unet/3dUNetBraTS.onnx")
force_calibration = dict_get(self.args,
"force_calibration",
default=False)
# Calibrator
if self.precision == "int8" or force_calibration:
self.apply_flag(trt.BuilderFlag.INT8)
preprocessed_data_dir = dict_get(
self.args,
"preprocessed_data_dir",
default="build/preprocessed_data/brats/calibration")
calib_batch_size = dict_get(self.args,
"calib_batch_size",
default=2)
calib_max_batches = dict_get(self.args,
"calib_max_batches",
default=20)
calib_data_map = dict_get(self.args,
"calib_data_map",
default="data_maps/brats/cal_map.txt")
calib_volume_dir = os.path.join(preprocessed_data_dir,
"brats_npy/fp32")
input_shape = [self.num_input_channel] + self.input_volume_dim
cache_file = dict_get(
self.args,
"cache_file",
default="code/3d-unet/tensorrt/calibrator.cache")
self.calibrator = UNet3DLegacyCalibrator(
calib_volume_dir, cache_file, calib_batch_size,
calib_max_batches, force_calibration, calib_data_map,
input_shape)
assert self.calibrator, "Calibrator is not init'ed"
assert self.calibrator.get_algorithm(
) == trt.CalibrationAlgoType.LEGACY_CALIBRATION, "Calibrator type is not Legacy"
self.builder_config.int8_calibrator = self.calibrator
self.cache_file = cache_file
# TRT builder flag
if self.precision == "fp16":
self.apply_flag(trt.BuilderFlag.FP16)
elif self.precision == "int8":
self.apply_flag(trt.BuilderFlag.FP16)
self.apply_flag(trt.BuilderFlag.INT8)
