LomaEngine Class Reference

Class that handles optimization of temporal mapping given a: More...

Public Member Functions
def	__init__ (self, *Accelerator accelerator, LayerNode layer, SpatialMappingInternal spatial_mapping, TemporalMappingType mapping_type, int|None loma_lpf_limit=None, **Any kwargs)
	The memory hierarchy from the correct core is extracted from the accelerator. More...
None	set_constraints (self, list[PermutationConstraint] constraints)
Generator[TemporalMapping, None, None]	run (self)
	Runs the LomaEngine. More...
def	get_temporal_loops (self)
	Get all loops that have to be temporally scheduled given layer and spatial mapping. More...
def	update_min_lpf_factor (self, dict[LayerDim, UnrollFactor] loop_sizes)
None	get_prime_factors (self)
	Get the prime factors for all temporal loops. More...
def	compute_nb_permutations (self)
	Compute the number of permutations that will have to be considered given the LPF distribution. More...
def	reduce_static_fps (self)
tuple[int, int]	find_smallest_non_static_pf (self, LayerDim layer_dim)
None	limit_lpfs (self)
	Function to limit the total number of loop prime factors present in this instance. More...
Generator[list[tuple[LayerDim, int]], None, None]	ordering_generator (self)
	Generator that yields all orderings of the temporal loops. More...

Public Attributes
	lpf_limit
	accelerator
	layer
	spatial_mapping
	constraints
	has_constraints
	mapping_type
	memory_hierarchy
	show_progress_bar
	temporal_loop_dim_size
	temporal_loop_pf_counts
	temporal_loop_pf_count_sums
	lpfs
	nb_permutations

Detailed Description

Class that handles optimization of temporal mapping given a:

• layer
• spatial mapping
• a memory hierarchy

This optimization is carried out through loop order based memory allocation. For each ordering of the temporal loops, they are allocated bottom-up to the levels in the memory hierarchy.

See https://ieeexplore.ieee.org/document/9458493 for more details.

Constructor & Destructor Documentation

__init__()

def __init__	(		self,
		*Accelerator	accelerator,
		LayerNode	layer,
		SpatialMappingInternal	spatial_mapping,
		TemporalMappingType	mapping_type,
		int | None	loma_lpf_limit = None,
		**Any	kwargs
	)

The memory hierarchy from the correct core is extracted from the accelerator.

Parameters

accelerator	accelerator to use the memory hierarchy of
layer	layer to generate temporal mappings for
spatial_mapping	SpatialMapping to use
loma_lpf_limit
kwargs	further unused, for ease of calling only

Member Function Documentation

compute_nb_permutations()

def compute_nb_permutations

(

self

)

Compute the number of permutations that will have to be considered given the LPF distribution.

find_smallest_non_static_pf()

tuple[int, int] find_smallest_non_static_pf	(		self,
		LayerDim	layer_dim
	)

Here is the caller graph for this function:

get_prime_factors()

None get_prime_factors

(

self

)

Get the prime factors for all temporal loops.

This is saved in three separate class attributes (temporal_loop_pfs, temporal_loop_pf_counts, temporal_loop_pf_count_sums)

TODO clean up (functions should not change class state variables...)

get_temporal_loops()

def get_temporal_loops

(

self

)

Get all loops that have to be temporally scheduled given layer and spatial mapping.

TODO clean up (and make use of LayerDimSizes methods)

limit_lpfs()

None limit_lpfs

(

self

)

Function to limit the total number of loop prime factors present in this instance.

This function scans the lpfs and while the number of lpfs is greater than self.lpf_limit it:

• picks the loop dimension that has the most lpfs
• merges the smallest two lpfs of that loop dimension (multiplying their values)

Here is the call graph for this function:

ordering_generator()

Generator[list[tuple[LayerDim, int]], None, None] ordering_generator

(

self

)

Generator that yields all orderings of the temporal loops.

Here is the call graph for this function:

reduce_static_fps()

def reduce_static_fps

(

self

)

Here is the caller graph for this function:

run()

Generator[TemporalMapping, None, None] run

(

self

)

Runs the LomaEngine.

Returns

Generator that yields all temporal mappings

Here is the caller graph for this function:

set_constraints()

None set_constraints	(		self,
		list[PermutationConstraint]	constraints
	)

update_min_lpf_factor()

def update_min_lpf_factor	(		self,
		dict[LayerDim, UnrollFactor]	loop_sizes
	)

Member Data Documentation

accelerator

accelerator

constraints

constraints

has_constraints

has_constraints

layer

layer

lpf_limit

lpf_limit

lpfs

lpfs

mapping_type

mapping_type

memory_hierarchy

memory_hierarchy

nb_permutations

nb_permutations

show_progress_bar

show_progress_bar

spatial_mapping

spatial_mapping

temporal_loop_dim_size

temporal_loop_dim_size

temporal_loop_pf_count_sums

temporal_loop_pf_count_sums

temporal_loop_pf_counts

temporal_loop_pf_counts

---

The documentation for this class was generated from the following file:

• /home/runner/work/zigzag/zigzag/zigzag/opt/loma/engine.py