Skip to content

Latest commit

 

History

History

GAMMA

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
__init__.py
__init__.py
 
 
gamma.py
gamma.py
 
 
main.py
main.py
 
 
train.py
train.py
 
 

README.md

Basic Usage

GAMMA: Automating the HW Mapping of DNN Models on Accelerators via Genetic Algorithm

In the basic usage, Gamma will create a map space assuming full flexibility in the underlying hardware.

Basic Parameter

  • fitness1: The first fitness objective (latency/ power/ energy)
  • fitness2: The second fitness objective (latency/ power/ energy)
  • model: The model to run (available model in data/model)
  • singlelayer: The layer index of the selected model, if want to optimize only single layer. If want to optimize all layers, skip this specification.
  • num_pe: Number of PEs
  • l1_size: L1 size (Number of elements)
  • l2_size: L2 size (Number of elements)
  • slevel_min: The minimum number of parallelism
  • slevel_max: The maximum number of parallelism. The number of parallelism will be in the range [slevel_min, slevel_max]
  • hwconfig: Read in HW configuration from file. An example of hwconfig can be found here. An example of using it can be found here
  • epochs: Number of generation for the optimization
  • outdir: The output result directory

Advanced Usage (Constrained Map Space Exploration)

A formalisim of Accelerator Flexiblity

We can also specify mapping constraint if the underlying hardware is not fully flexible. Gamma will create a constraint map-space based on the given mapping constraint.

Advanced Parameter

  • mapping_cstr: The constraint on the map-space.
    • Make a customized mapping constraint files and put it in this path data/mapping_cstr.

For example, let's apply an mapping constraint called dla_map (located at data/mapping_cstr/dla_map.py)

python main.ppy --mapping_cstr dla_map

We give detailed explanations and provide examples in data/mapping_cstr.

Advanced Usage (PE-Mapping Co-exploration)

DiGamma: Domain-aware Genetic Algorithm for HW-Mapping Co-optimization for DNN Accelerators

We can also support PE (HW) and mapping Co-exploration. In this mode, we use area (PE area + buffer area) as constraint.

Advanced Parameter

  • num_pe: To execute in this mode, set num_pe to -1.
  • area_budget: The area budget for compute (PE) and memory (L1 buffer + L2 buffer).
  • pe_limit:The upper-bound for available number of PE.