Functions
( tuple[float, float, list[tuple[CostModelEvaluationABC, Any]]]\|tuple[float, float, float, float, list[tuple[CostModelEvaluationABC, Any]]])	get_hardware_performance_zigzag (str\|list[dict[str, Any]]\|ModelProto workload, str accelerator, str mapping, *Literal["loma"]\|Literal["salsa"] temporal_mapping_search_engine="loma", Literal["uneven"]\|Literal["even"] temporal_mapping_type="uneven", str opt="latency", str dump_folder=f"outputs/{datetime.now()}", str\|None pickle_filename=None, int lpf_limit=6, int nb_spatial_mappings_generated=3, bool in_memory_compute=False, bool exploit_data_locality=False, bool enable_mix_spatial_mapping=False, bool loma_show_progress_bar=True)
	ZigZag API: estimates the cost of running the given workload on the given hardware architecture. More...

tuple[float, float, float, float, list[tuple[CostModelEvaluationABC, Any]]]	get_hardware_performance_zigzag_imc (*Any args)
	Overload with type hint. More...

Function Documentation

◆ get_hardware_performance_zigzag()

( tuple[float, float, list[tuple[CostModelEvaluationABC, Any]]] \| tuple[float, float, float, float, list[tuple[CostModelEvaluationABC, Any]]] ) zigzag.api.get_hardware_performance_zigzag	(	str \| list[dict[str, Any]] \| ModelProto	workload,
		str	accelerator,
		str	mapping,
		*Literal["loma"] \| Literal["salsa"]	temporal_mapping_search_engine = `"loma"`,
		Literal["uneven"] \| Literal["even"]	temporal_mapping_type = `"uneven"`,
		str	opt = `"latency"`,
		str	dump_folder = `f"outputs/{datetime.now()}"`,
		str \| None	pickle_filename = `None`,
		int	lpf_limit = `6`,
		int	nb_spatial_mappings_generated = `3`,
		bool	in_memory_compute = `False`,
		bool	exploit_data_locality = `False`,
		bool	enable_mix_spatial_mapping = `False`,
		bool	loma_show_progress_bar = `True`
	)

ZigZag API: estimates the cost of running the given workload on the given hardware architecture.

Parameters

workload	Either a filepath to the workload ONNX or yaml file, an ONNX model.
accelerator	Filepath to accelerator yaml file.
mapping	Filepath to mapping yaml file.
opt	Optimization criterion: either `energy`, `latency` or `EDP`.
dump_folder	Folder where outputs will be saved.
pickle_filename	Filename of pickle dump.
lpf_limit	Determines the number of temporal unrollings that are evaluated.
nb_spatial_mappings_generated	Max nb of spatial mappings automatically generated (if not provided in mapping).
in_memory_compute	Optimizes the run for IMC architectures.
exploit_data_locality	Iff true, an attempt will be made to keep data in lower-level memory in between layers
enable_mix_spatial_mapping	Wether `mixed` spatial mappings will be generated, i.e. unrolling multiple Layer Dimensions in a single Operational Array Dimension.

Here is the caller graph for this function:

tuple[float, float, float, float, list[tuple[CostModelEvaluationABC, Any]]] zigzag.api.get_hardware_performance_zigzag_imc ( *Any args )

Overload with type hint.

Here is the call graph for this function: