This repository contains the reference implementation of LEMON mapper and the scripts to run the experiments in the paper Memory-Aware DNN Algorithm-Hardware Mapping via Integer Linear Programming presented at the 20th ACM International Conference on Computing Frontiers.
- Clone the repository
- Acquire a free academic Gurobi Web License
Now you can decide to install LEMON only or setup the environment to replicate the comparisons against the state-on-the-art presented in the paper.
If you want to install LEMON only you can follow one of the following methods:
- Install the Gurobi license in your system
- Install Timeloop Model and Accelergy for mapping evaluation.
- Install the LEMON package
cd lemon
pip install -e .- Put the downloaded
license.licfile inlemon/docker/put-gurobi-license-here - Build and run the docker image
cd lemon/docker
docker-compose -f docker-compose.lemon.yaml run lemon- Put the downloaded
license.licfile inlemon/docker/put-gurobi-license-here - Open the
lemonfolder in VSCode - Install the Dev Containers extension for VSCode
- Open the Command Palette (
F1), select "Reopen Folder in Container" and select lemon-only
More information about devcontainers in VSCode can be found here.
If you want to run the artifact evaluation you can use the provided docker image. One of the other state-of-the-art mapper is implemented using PyTorch, for this reason a CUDA GPU is required. To make the GPU accessable in a docker container, nvidia-docker2 has to be installed in the host system following the instructions here.
Put the downloaded Gurobi license.lic file in the docker/put-gurobi-license-here folder.
Now it is possible to build the docker image and launch the container either using the following commands:
cd docker
docker-compose -f docker-compose.ae.yaml run lemon-aeor, if you are using VSCode, you can use the provided devcontainer as follows:
- Open the
lemonfolder in VSCode - Install the Dev Containers extension for VSCode
- Open the Command Palette (
F1), select "Reopen Folder in Container" and select lemon-artifact-evaluation
More information about devcontainers in VSCode can be found here.
See evaluation.
For information about the lemon command arguments type:
lemon --help
oooo
`888
888 .ooooo. ooo. .oo. .oo. .ooooo. ooo. .oo.
888 d88' `88b `888P"Y88bP"Y88b d88' `88b `888P"Y88b
888 888ooo888 888 888 888 888 888 888 888
888 888 .o 888 888 888 888 888 888 888
o888o `Y8bod8P' o888o o888o o888o `Y8bod8P' o888o o888o
usage: lemon [-h] [-o OUTPUT] [-lu [LAYER_LOOKUP ...]] [-ew ENERGY_WEIGHT] [-lw LATENCY_WEIGHT] [-ot] [-ob] [-fb] [-sp] [-fp] [-tls TIME_LIMIT_SINGLE]
inputs [inputs ...]
positional arguments:
inputs Input YAML files: architecture, mapspace and workloads required
options:
-h, --help show this help message and exit
-o OUTPUT, --output OUTPUT
Output Directory. Default = ./output
-lu [LAYER_LOOKUP ...], --layer_lookup [LAYER_LOOKUP ...]
Workload ID for each layer. Default = 0 1 2 ... W-1
-ew ENERGY_WEIGHT, --energy_weight ENERGY_WEIGHT
Energy weight in the objective. Default = 1
-lw LATENCY_WEIGHT, --latency_weight LATENCY_WEIGHT
Latency weight in the objective. Default = 10
-ot , --objective_type
MIP model objective type: 'blended' for weighted sum of energy and latency as obj (fastest), 'latency-energy' to first minimize
latency and then energy, 'energy-latency' for minimizing energy first and then latency, 'quadratic' to minimize EDP (this
transforms the model in a MIQP, slowest). Default = blended
-ob, --optimize_bypass
Optimize datatype memory bypass not specified in mapspace file (transforms model in MICQP, slower)
-fb, --fix_bypass Make sure to have same bypass directives for all layers
-sp, --static_partitioning
Make sure buffer partitioning is the same for all the layers
-fp, --fix_permutations
Make sure temporal permutations are the same for all the layers
-tls TIME_LIMIT_SINGLE, --time_limit_single TIME_LIMIT_SINGLE
Search time limt (seconds) when mapping a single layer
LEMON can map all the layers of a DNN model simultaneously. In this case a layer-workload-lookup list has to be provided using the option -lu. Each integer in the provided list represent the workload identifier of a layer. The workload files has to be alphabetically ordered in a folder provided as input. See the evaluation/fixing folder for end-to-end mode usage examples.
The bypass optimizaion mode can be enable using the option -ob. This transforms the problem in a MICQP and could require more time to solve. See the evaluation/bypass folder examples.
LEMON supports four optimization objetives that can be configured using the -ot option:
blended: this is the default one and consists in the linear combination of energy and latency. The weights can be adjusted using the-ewand-lwoptions respectively.latency-energy: first minimize latency and then minimize energyenergy-latency: first minimize energy and then minimize latencyquadratic: minimize the EDP, i.e. the product of latency and energy. This transforms the problem in MIQP and could require more time to solve.
If you find this repository useful please cite:
@inproceedings{russo2023lemon,
title={Memory-Aware DNN Algorithm-Hardware Mapping via Integer Linear Programming},
author={Russo, Enrico and Palesi, Maurizio and Ascia, Giuseppe and Patti, Davide and Monteleone, Salvatore and Catania, Vincenzo},
booktitle={Proceedings of the 20th ACM International Conference on Computing Frontiers},
year={2023},
doi={10.1145/3587135.3592206},
}Enrico Russo, Maurizio Palesi, Giuseppe Ascia, Davide Patti, Salvatore Monteleone, and Vincenzo Catania. 2023. Memory-Aware DNN Algorithm-Hardware Mapping via Integer Linear Programming. In 20th ACM International Conference on Computing Frontiers (CF ’23), May 9–11, 2023, Bologna, Italy. ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3587135.3592206