FPGA_Arch_Exploration

FPGA_Arch_Exploration is used for explore the routing architecture for FPGAs.

Introduction

This routing architecture is based on VIB architecture. To explore the magnificant design space for VIB architecture, Bayesian Optimization is used for efficient exploration. Bayesian Optimization is a probabilistic strategy for optimizing complex black-box functions across various parameters. At its core, BO constructs a surrogate model to approximate the objective function and utilizes an acquisition function to sample high-quality candidates, balancing exploration and exploitation.

Hyperopt is a hyperparameter optimization framework using the Tree-structured Parzen Estimator (TPE), a prominent Bayesian Optimization variant. It supports exploration in parallel.

Requirements

Hyperopt Installation

Hyperopt is needed for Bayesian Optimization and parallel exploration.

Install Hyperopt using pip:

pip install hyperopt

After installation, verify that Hyperopt is installed correctly:

python -c "import hyperopt; print(hyperopt.__version__)"

Install MongoDB Dependency：

pip install pymongo

VTR Installation

VTR supported VIB architecture can be downloaded from VTR4VIB. Install it in the vtr/ directory of this project.

Install VTR：

cd vtr
make vpr

File Description

blif_files/ consists of benchmarks in .blif format for VPR.

blif_list lists the circuits for architecture exploration evaluation. The circuits in blif_list should also contained in blif_files/.

baseline_result.csv cantains the baseline data such as routing area and critical path delay for each circuit.

alkaidT_vib.xml is the baseline architecture. The exploration is based on it.

Seeker_bayes_seg.py is the main document of the exploration. Various wire types and VIB parameters combinations are generated via Bayesian Optimization and then converted to new architecture description files (.xml). VPR evaluates the routing area and critical path delay of the architectures for exploration iteratively. The optimization goal is the area delay product (ADP) here. get_info.sh is embedded to get the value of area and delay from vpr log file.

run_hyperopt_seg.sh is the interface of the project. The information of the exploration is saved in log.txt and logfile.txt (rough and detailed).

How to Use

First, the user need to change the vpr_dir in run_hyperopt_seg.sh to the specific path of VPR. The value of parallel can be change to the number of parallel of the user.

Then run exploration by:

./run_hyperopt_seg.sh

A folder named workdir is newly created and cantains new-generated architectures and their results.

If the user wants to get the top 10 architectures, just run:

python3 get_result.py > best_archs.txt

The detailed results for each circuit implemented on each FPGA architecture will be listed in best_archs.txt.

Results

results folder contains the top 10 best FPGA architectures and their area, delay and ADP results for "alkaidT" arch exploration. The area, delay and ADP improvement compared to the baseline architecture can reach 20.06%, 11.58% and 29.32% respectively. The detailed data for the 10 archs can be found in result.csv.