SPaDE: Space Partition Aided Design (Space) Exploration (For VexRiscv)

Setup

For offline DSE: Offline datasets are stored in PresampledDataset. You can check examples/VexRiscv_offline.py for a simple example.

For online DSE, you need to install the required tools for the hardware generation and evaluation, such as Vivado or Yosys.

You will need the VexRiscv generator for online DSE as well:

git submodule init
git submodule update Design/riscv/VexRiscv

VexRiscv requires some tools like Verilator, please check the original repo for detailed guide: https://github.com/SpinalHDL/VexRiscv

Run Examples

Offline DSE:

python examples/VexRiscv_offline.py

Online DSE:

python examples/VexRiscv_online.py

Algorithm Settings

SpadeOptimizer is highly configurable, the default setting of our proposed method is as set in the examples:

method = 'ehvi'
weight_method = 'global-hv'
cluster_method = 'hybrid'

If you want to change them, please check the details in Optimizers/DDSP/mcts/mcts.py.

Cite Us

@inproceedings{space_partition_dse_iccd2024,
  title     = {Efficient Microprocessor Design Space Exploration via Space Partitioning},
  author    = {Zijun Jiang and Yangdi Lyu},
  booktitle = {The 42nd IEEE International Conference on Computer Design (ICCD)},
  year      = {2024}
}

Acknowledgement

C. Bai, Q. Sun, J. Zhai, Y. Ma, B. Yu, and M. D. F. Wong, “BOOM-Explorer: RISC-V BOOM Microarchitecture Design Space Exploration”.

https://github.com/baichen318/boom-explorer-public

L. Wang, R. Fonseca, and Y. Tian, “Learning Search Space Partition for Black-box Optimization using Monte Carlo Tree Search”.

https://github.com/facebookresearch/LaMCTS

Y. Zhao, L. Wang, K. Yang, T. Zhang, T. Guo, and Y. Tian, “Multi-objective Optimization by Learning Space Partitions”.

https://github.com/aoiang/LaMOO