Code from A Portable, Fast, DCT-based Compressor for AI Accelerators

Overview

This is the repository for the paper A Portable, Fast, DCT-based Compressor for AI Accelerators. The DCT+Chop compressor has been tested on five systems: NVIDIA A100 GPU, Cerebras CS-2, SambaNova SN30, Groq Groqchip, and Graphcore IPU. While this work originally targets training data, the DCT+Chop compressor can be used for any data on these systems. We use benchmarks from SciML-bench.

Versions Tested

• PyTorch 2.0.1
• Cerebras Release 2.0.1
• SambaFlow 1.17
• GroqFlow 4.2.1
• PopTorch (Graphcore) 3.3.0

Directory Overview

• comp_benchmark: directory containing benchmarking scripts to test scaling input data size for the compressor
• compressor: contains entry points for the compressor
• emdenoise: EMDenoise SciML-Bench benchmark
• opticaldamage: OpticalDamage SciML-Bench benchmark
• resnet34: ResNet34 with CIFAR10 dataset benchmark
• slstrcloud: SLSTRCloud SciML-Bench benchmark
• slstrcloud_highres: SLSTRCloud SciML-Bench benchmark with higher resolution data
• utils: utility functions

File Overview

All scripts to run code need a config.txt file. See emdenoise/config-ch4.txt for an example.

In comp_benchmark, each filename to test compression follows the bench_compress_<platform>.py format, while decompression tests follow the bench_<platform>.py format.

Example (SambaNova):

python bench_samba.py --config_path="./config.txt" --num-iterations=10 --compressor="dct"

Each network (emdenoise, opticaldamage, resnet34, slstrcloud) has several scripts formatted as <network>_<platform>.py. Remember to pass the config file for the compressor.

Example (Graphcore):

cd resnet34
python resnet34_graphcore.py --config_path="./config-ch4.txt" --num-iterations=1 --num-epochs=30

Citation

Please use the citation below if you reference this work:

Milan Shah, Xiaodong Yu, Sheng Di, Michela Becchi, and Franck Cappello. 2024. A Portable, Fast, DCT-based Compressor for AI Accelerators. In Proceedings of the 33rd International Symposium on High-Performance Parallel and Distributed Computing (HPDC '24). Association for Computing Machinery, New York, NY, USA, 109–121. https://doi.org/10.1145/3625549.3658662