DeFiNES: A DSE Framework Enabling Fast Exploration of the Depth-first Scheduling Space for DNN Accelerators
Our artifact provides a guide to replicate the primary experiments (case study 1) demonstrated in this paper. Since case study 2 and 3 are equivalently performing case study 1 multiple times (for different workload and HW architecture combinations) and will take a long time to run and generate multi-Gigabytes of data, we here focus on case study 1.
The included materials are the source code of DeFiNES and the scripts to auto-run the experiments, collect data, and make the plots. In the end, we also provide the useful information on experiment customization, i.e. users can use DeFiNES to carry out their own DNN accelerator-schedule design space exploration, considering both layer-by-layer and depth-first scheduling possibilities.
-
Install Anaconda environment
-
Download DeFiNES (
git clonethis repo) -
Use a terminal or an Anaconda Prompt for the following steps:
cdinto the DeFiNES repo- Create the conda environment from the environment.yml file
conda env create -f environment.yml - Activate the new environment:
conda activate DeFiNESenv
After the previous Installation is done, two steps are required to run the case study 1 and reproduce the overall comparison result (Fig. 12) and the detailed analysis results (Fig. 13, Fig. 14, Fig. 15, Fig. 6(left), Fig. 9 and Fig. 10).
Note that the below two steps need to run in sequence in the DeFiNES repo.
Run
python main_artifact.py
It applies 108 depth-first scheduling options (3 modes with 6×6 X-Dim and Y-Dim tile size combinations) for processing FSRCNN on Meta-proto-like DF architecture.
18 hours with the default setting (using 1 CPU thread
and set loma_lpf_limit=8). loma_lpf_limit is
a speed-quality tradeoff tuning knob. User can change
its value in main_artifact.py. The larger it is, the
longer the program runs, and possibly the better the result
found. For all the experiments in the paper, we set it to
8 to guarantee the best results can be found.
For testing purpose, users can set loma_lpf_limit=6, the total runtime will be reduced dramatically from
18 hours to 45 minutes, while some design points’ best
found energy will increase by a few percents. So, the
figures plotted in this case will be slightly different than
the original ones in the paper.
When the program finishes, an overall energy and latency comparison figure will be plotted for these 108 depth-first scheduling options (Fig. 12), and 108 result pickle files (.pkl) will be saved under the result_saving_path defined in the main_artifact.py (by default, it is .\result_pickle_files\).
Run
python plot_artifact.py
It extracts the required information from the generated result pickle files and makes the plots.
1 minute
Multiple detailed analysis figures: Fig. 13, Fig. 14, Fig. 15, Fig. 6(left), Fig. 9 and Fig. 10.
In the end, all the plots will be saved to
.\result_plot\as PDF files.
The goal of this work is to provide an open-source framework for DNN accelerator architecture-schedule optimization, which allows users to plug in their own setting files and perform customized design space exploration experiments. For this, users need to provide DeFiNES with the inputs listed in Fig. 5: a workload, a HW architecture, and some depth-first scheduling parameters (in which the Fuse depth is set automatically by DeFiNES, thus no need to provide).
This work has 5 workloads (inside of each folder) and 10 HW architectures modelled for the case study 2 and 3. Users can perform experiments on them, modify them to try new design options, or create own workloads and/or HW architectures following the same data formats in these example setting files. You can find more details on how to set the input files here.
Some example commands users can run (Use python main.py --help to see what each argument means):
python main.py --accelerator inputs.HW.Edge_TPU_like --workload inputs.WL.Edge_TPU_like.workload_mccnn --dfmode 1 --tilex 16 --tiley 8
python main.py --accelerator inputs.HW.Edge_TPU_like --workload inputs.WL.Edge_TPU_like.workload_mobilenetv1 --dfmode 1 --tilex 7 --tiley 7
python main.py --accelerator inputs.HW.Meta_prototype --workload inputs.WL.Meta_prototype.workload_fsrcnn --dfmode 123 --tilex 96 --tiley 72
The results are saved as pickle files (.pkl) in the pre-defined result_saving_path. User can use/modify the functions provided in plot_artifact.py and
plot_helper_funcs.py to extract various data from the pickle files and visualize the results.
We are continually improving the framework, and welcoming all questions and feedback.
We hope our tool can help other researchers to better explore and understand the vast DNN accelerator architecture-and-scheduling design space and can offer the best design solutions.