Bit-Cigma Experiments

Experiment code for the paper "Bit-Sparsity Aware Acceleration With Compact CSD Code on Generic Matrix Multiplication" (IEEE Transactions on Computers, 2025).

Note

We lost the original codes used in the paper, so we publish the repo here, and the simulator backend Cigma-Sim is a new modified version. Although the output may not exactly reproduce the data presented in the paper, you can still observe the trend and comparison.

Requirements

• Python 3.11+
• PyTorch 2.6+
• CUDA 12.6+ (for GPU experiments)
• Cigma-Sim (tag: v2026.02)

Getting Started

Using Container (Recommended)

Build the development container with all dependencies:

make container_build_devel

Local Installation

Install dependencies:

make install_cigmasim

Prepare Datasets

Download Flickr8k and WMT14 datasets:

make prepare_data

Running Experiments

Comparison Experiments

Run all architecture comparison simulations:

make run_all_compare

Or run individual model comparisons:

make run_alexnet_cycle      # AlexNet
make run_mobilenetv3_cycle  # MobileNetV3
make run_resnet152_cycle    # ResNet-152
make run_vgg16_cycle        # VGG-16
make run_t5_cycle           # T5 (cycle simulation)
make run_t5_value           # T5 (value simulation)

Parameter Sweep Experiments

Run Cigma architecture parameter sweeps:

make run_all_sweep

Or run individually:

make run_sweep_encode         # Encoding method sweep
make run_sweep_window         # Window size sweep
make run_sweep_product_width  # Product width sweep
make run_sweep_segment        # Segment count sweep

Other Commands

make help    # Show all available commands
make clean   # Clean cache and temporary files

Publication

@article{zhu2025bitsparsity,
  title={Bit-Sparsity Aware Acceleration With Compact CSD Code on Generic Matrix Multiplication},
  author={Zhu, Zixuan and Zhou, Xiaolong and Wang, Chundong and Tian, Li and Huang, Zunkai and Zhu, Yongxin},
  journal={IEEE Transactions on Computers},
  volume={74},
  number={2},
  pages={414--426},
  year={2025},
  doi={10.1109/TC.2024.3483632}
}

License

MIT