This is the repo for our DAC24 accepted paper: Explainable Fuzzy Neural Network with Multi-Fidelity Reinforcement Learning for Micro-Architecture Design Space Exploration

Pre-request:

• Chipyard 1.10 (https://chipyard.readthedocs.io/en/1.10.0/Chipyard-Basics/Initial-Repo-Setup.html)
• Python 3.10
• Please refer to requirements.txt for python packages
• VCS is needed during the DSE, please make sure vcs is callable in your commmand line.
• McPAT is needed during the DSE, please download McPAT (https://github.com/HewlettPackard/mcpat). Our script will access mcpat from the following path:

RISCV_Model_and_DSE/FNN_DSE/analytic/mcpat/mcpat-1.3.0/mcpat

Usage:

First, set the envrionment variable:

export chipyard_home="Absolute path to RISCV_Model_and_DSE/chipyard"

Then, activate the conda envrionment:

source $chipyard_home/env.sh

Now we can start the dse using FNN with MFRL:

python dse.py

Also, we can run the baselines for comparison.

To use BoomExplorer, run this scrpt:

python dse-dkg.py

To use Scalable Constraint BO, run this scrpt:

python dse-sc.py

To use BagGBRT, run this scrpt:

python dse-gbrt.py

To use ActBoost, run this scrpt:

python dse-ada.py

To use Random Forest, run this scrpt:

python dse-rf.py

Note

We directly modify the sklearn and botorch files to implement the above algorithms. The official sklearn and botorch package can't support these algorithms. Please refer to util/files_to_replace and make the necessary modifications.

A dataset is automatically built under dataset/vcs. The DSE results are stored in the log directory, containing:

• The trained FNN named fnn.pt and its state-dict named fnn-state.pt.
• details.txt shows the detaied DSE process. For example, the following log shows the starting point of this RL episode, then it shows how the design parameters and metrics change in each step of the markov process.

--- episode: 1 ---
Starting point:params: [16, 2, 128, 2, 2, 1, 32, 1, 1, 1, 2]
cpi: 4.738, miss1: 0.18, miss2: 0.45, area: 4.324
--iter:1
update: [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
params: [16, 4, 128, 2, 2, 1, 32, 1, 1, 1, 2]
cpi: 4.654, miss1: 0.15, miss2: 1.78, area: 4.468
--iter:2
update: [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0]
params: [16, 4, 128, 2, 4, 1, 32, 1, 1, 1, 2]
cpi: 4.508, miss1: 0.15, miss2: 1.78, area: 4.572
--iter:3
update: [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
params: [16, 8, 128, 2, 2, 1, 32, 1, 1, 1, 2]
cpi: 4.576, miss1: 0.08, miss2: 2.50, area: 5.36

• final_cpi.txt records the low-fidelity cpi of the design obtained in each episode.
• High-Fidelity.txt records the high-fidelity cpi of the designs obtained in the high-fidelity phase.
• make_config.log is the generation log of chipyard.

To obtain the rules, check the jupyter notebook get_rules.ipynb.

The benchmarks cover the commonly used applications, which are

• dijkstra, fft, stringsearch from mibench (https://github.com/embecosm/mibench)* and
• mm, qsort, vvadd from riscv-test (https://github.com/riscv-software-src/riscv-tests).

Customization

To use a custom benchmark,

• First obtain the static profiling results using ibm-pisa (https://github.com/exabounds/ibm-pisa).

• Then, put the profilings under anlytic/benchmarks, with a folder named after the benchmark.

• Next, use the analytic/fitting.ipynb tool to fit a linear functon for the cache reuse distance, and add it into analytic/Analytic_model_torch.py.

• Finally, include the profiling in dse.py.