🔧 Compression Simulation for CXL Memory Systems

This repository provides a C/Python-based simulation framework to evaluate both lossless (features) and lossy (filters) data compression techniques at the CXL flit granularity (256B), as part of a national project on next-generation memory subsystems for AI accelerators.

---

🔬 Project Context

This work was supported by the Technology Innovation Program
(Industrial Strategic Technology Development Program - Public-Private Joint Investment Advanced Semiconductor Talent Development Project)
Project ID: RS-2023-00237136
Title: Development of CXL/DDR5-based memory subsystem for AI accelerators
Funding: Ministry of Trade, Industry & Energy (MOTIE, Korea) (1415187686)

This repository implements simulation code and evaluation logic for the following key deliverables defined in the national project’s performance goals:

---

📌 1. Lossless Compression Rate of DNN Features

• Goal: Apply lossless compression to 16-bit DNN feature maps in 256B CXL flit blocks, and compute compression ratios.
• Key Metric:

  Compression Ratio = (Original Size) / (Compressed Size)

• Target:
  Unlike stream-based compression (e.g., Zhuo Chen 2018), block-based compression faces limited redundancy, but is crucial for random access. This simulation targets a 2:1 lossless compression ratio on flit-sized blocks.
• Evaluation:
  • Simulated via C/Python code (not FPGA)
  • Based on ResNet-50 feature maps
  • Per-layer results are aggregated and logged

---

📌 2. Lossy Compression Rate of DNN Filters

• Goal: Compress FP32 filters using quantization + lossy compression without reducing Top-1 accuracy by more than 1%.
• Key Metric:

  Compression Ratio = (Original FP32 size) / (Compressed size)

• Target:
  Filters contain fewer zeros, making lossless compression inefficient. This simulation seeks 4:1 lossy compression while preserving accuracy.
• Evaluation:
  • Simulated via C/Python
  • Restored filters are used in ResNet-50 inference
  • Accuracy loss ≤ 1% is enforced

---

📦 Environment & Setup

🔧 Prerequisites

• Docker (≥ 20.10)
• Git (optional)

🚀 Quick Start

1. Clone this repository

git clone https://github.com/scalable-arch/CXL-Compression.git
cd CXL-Compression

2. Build Docker image

docker build -t cxl-comp .

3. Run the container

docker run --rm -it -v $(pwd):/workspace cxl-comp

4. Execute the full pipeline

./run_all.sh

🔧 Optional: Run from a Specific Step

You can start from a specific step by providing a step number as an argument:

./run_all.sh <step_number>

For example:

./run_all.sh 5

This will skip steps 1 to 4 and begin from step 5.

---

🪜 Step-by-Step Breakdown

Step	Title	Description
1	Extract Feature Maps	Extract 16-bit feature maps from sample ImageNet images using ResNet-50.
2	Merge Feature Maps Operation-wise	Group feature maps by operation (e.g., conv1, relu) for flit-level compression.
3	Compile Feature Compression Code	Build the C++ lossless compressor used for 128B feature block compression.
4	Run Feature Compression	Compress feature maps and generate compression ratio logs and plots.
5	Compile Weight Compression Libraries	Build C++ shared libraries for lossy compression & masking compression.
6	Run Weight Compression (Simple Truncation)	Apply compression, fixed masks on all weight in blocks and evaluate accuracy drop.
7	Run Weight Compression (Adaptive Truncation)	Apply compression, select masks adaptively per block and evaluate accuracy drop.
8	Print Total Results	Show compression summary and Top-1 / Top-5 accuracy changes after weight truncation.

---

📂 All result logs and plots are saved under the total-result/ directory.

📁 Directory Structure

.
├── extract-data                    # DNN feature map extraction
├── feature-compression             # Feature lossless compression
├── imagenet-sample-images          # ImageNet sample images
├── weight-sensitivity-analysis     # Weight sensitivity analysis
├── weight-compression              # Weight lossy compression
├── total-result                    # Result reports (created by script)
├── Dockerfile                      # Docker build file
├── run_all.sh                      # Full pipeline script
├── images                          # Graph Result Images
├── README.md
└── LICENSE

---

📊 Example Output

Sample log line:

 Block     Mask    Ratio |     Original     Restored  |                            Orig(bin)                           Rest(bin)
==================================================================================================================================

[Block 1] Ratio: 1.1763, Compressed bits: 1741, Mask usage: {'mask20': 13, 'mask19': 51}
     1   mask20 1.1763354394026422 |  -0.01909715  -0.01757812  |  10111100100111000111000110100011  10111100100100000000000000000000
     1   mask20 1.1763354394026422 |  -0.44219688  -0.43750000  |  10111110111000100110011110100001  10111110111000000000000000000000
     1   mask20 1.1763354394026422 |  -0.01033773  -0.00976562  |  10111100001010010101111110010011  10111100001000000000000000000000
     1   mask20 1.1763354394026422 |  -0.36786637  -0.34375000  |  10111110101111000101100011111011  10111110101100000000000000000000
     1   mask20 1.1763354394026422 |   0.00367908   0.00366211  |  00111011011100010001110010110001  00111011011100000000000000000000
     1   mask20 1.1763354394026422 |  -0.09179689  -0.08593750  |  10111101101111000000000000000010  10111101101100000000000000000000
     ...
     1   mask19 1.1763354394026422 |   0.00530118   0.00512695  |  00111011101011011011010110000011  00111011101010000000000000000000
     1   mask19 1.1763354394026422 |  -0.03543944  -0.03515625  |  10111101000100010010100011110011  10111101000100000000000000000000
     1   mask19 1.1763354394026422 |   0.00795496   0.00781250  |  00111100000000100101010110001000  00111100000000000000000000000000
     1   mask19 1.1763354394026422 |  -0.14247002  -0.14062500  |  10111110000100011110001110101001  10111110000100000000000000000000
     1   mask19 1.1763354394026422 |  -0.01383873  -0.01367188  |  10111100011000101011101111010100  10111100011000000000000000000000
     1   mask19 1.1763354394026422 |  -0.01616473  -0.01562500  |  10111100100001000110101111100100  10111100100000000000000000000000
     1   mask19 1.1763354394026422 |   0.00619890   0.00610352  |  00111011110010110010000000101100  00111011110010000000000000000000

[Block 2] Ratio: 1.1763, Compressed bits: 1741, Mask usage: {'mask20': 13, 'mask19': 51}
     2   mask20 1.1763354394026422 |  -0.01067094  -0.00976562  |  10111100001011101101010100101100  10111100001000000000000000000000
     2   mask20 1.1763354394026422 |   0.10300009   0.10156250  |  00111101110100101111000110110110  00111101110100000000000000000000
     2   mask20 1.1763354394026422 |  -0.00293915  -0.00292969  |  10111011010000001001111011010000  10111011010000000000000000000000
     2   mask20 1.1763354394026422 |   0.28020972   0.25000000  |  00111110100011110111011110100110  00111110100000000000000000000000
     2   mask20 1.1763354394026422 |  -0.08703937  -0.08593750  |  10111101101100100100000110110010  10111101101100000000000000000000
     2   mask20 1.1763354394026422 |  -0.01414643  -0.01367188  |  10111100011001111100011001101111  10111100011000000000000000000000
     ...

Summary report:

═════════════════════════════════════════════════════
🔹 8) Print Total Results
═════════════════════════════════════════════════════

📄 result_bf16.log:
Filename                       Blocks            Average
avgpool_all.bin.log               320             1.2584
bn1_all.bin.log                423360             3.5146
bn2_all.bin.log                215600             2.9779
bn3_all.bin.log                862400             3.0428
conv1_all.bin.log              423360             1.4366
conv2_all.bin.log              215600             1.3945
conv3_all.bin.log              862400             1.4864
downsample_all.bin.log         470400             1.8797
fc_all.bin.log                    156             1.2263
maxpool_all.bin.log             31360             4.5169
relu_all.bin.log               987840             3.2385
TOTAL                         4492796             2.4863

📄 Origin Accuracy
[Original FP32] Top-1 Accuracy: 92.97%, Top-5 Accuracy: 98.83%

📄 Simple Truncation with Exponent-BPC Accuracy
[Restored Adaptive] Top-1 Accuracy: 84.57%, Top-5 Accuracy: 96.48%

📊 Top-1 Drop: 8.40%
📉 Top-5 Drop: 2.34%

📄 Adaptive Truncation with Exponent-BPC Accuracy
[Restored Adaptive] Top-1 Accuracy: 90.23%, Top-5 Accuracy: 98.63%

📊 Top-1 Drop: 2.73%
📉 Top-5 Drop: 0.20%

---

👤 Author

• Name: [Boyeol Choi & Jungrae Kim]
• Affiliation: [Scalable Architecture Lab. / Sungkyunkwan University]
• Contact: [Boyeol Choi(gju06051@gmail.com)]

---

🙌 Acknowledgments

• Research inspiration: Bit-Plane Compression Transforming Data for Better Compression in Many-Core Architectures (ISCA)
• This project is supported by Korea's Ministry of Trade, Industry & Energy (MOTIE)