SybilD/TetrisG-SDK

Project Title

An example for TetrisG-SDK mapping algorithm, with NeuroSim.

Description

Title

TetrisG-SDK: Efficient Convolution Layer Mapping with Adaptive Windows for Fast In Memory Computing

Demonstration of TetrisG-SDK

demoG.mp4

Abstract

Shifted-and-Duplicated-Kernel (SDK) mapping is an emerging technique for accelerating convolutional layers in Compute-In-Memory (CIM) architectures. While state-of-the-art SDK variants have achieved decent mapping efficiency, optimizations are still desired to enhance CIM utilization and minimize computing cycles. We propose TetrisG-SDK, a novel framework that exploits adaptive windows to boost mapping performance. These windows can accommodate more input channels, increase array utilization at marginal space, and adjust window shapes for latency. The methodology also leverages grouped convolutions to further decrease computing cycles, and ensure near-lossless network accuracy. Moreover, TetrisG-SDK is the first framework to integrate a validated, system-level CIM hardware simulator for accurate performance and energy estimation from an application perspective. Our experiments show that Tetris-SDK can remarkably accelerate convolution layer mapping. Compared to VW-SDK, it achieves a speed-up of 3.1× for ResNet50 on ImageNet classification and an improvement of overall latency and energy efficiency by 8.3× and 3.1×, respectively. This shows that TetrisG-SDK is a promising solution to map and accelerate Convolutional Neural Networks in CIM hardware.

Flowchart

Getting Started with TetrisG

Dependencies

• Python 3.x

Installing

• For TetrisG-SDK: Any IDE that can run Jupyter Notebook

Executing TetrisG program

• Download the file and open it in IDE
• Run cell by cell or call the main function

python3 TetrisG_SDK.py

Getting Started with NeuroSim

Dependencies

• C++
• Python 3.9

Installing

• For NeuroSim: Ubuntu 20.04 gcc: v9.4.0 glibc: v2.31 NVIDIA Driver Version: 525.60.13 CUDA Version: 12.0 (Refer to DNN+NeuroSim V1.4 https://github.com/neurosim/DNN_NeuroSim_V1.4/tree/main for more support)

Executing NeuroSim Simulator

• Compile the NeuroSim C++ code

cd Inference_pytorch/NeuroSIM
make

• Run the Simulator

cd ..
python inference.py --dataset cifar10 --model DenseNet40 --mode WAGE --inference 1 --cellBit 1 --ADCprecision 6 --subArray 512 --parallelRead 512 | tee benchmark_tetris_group.log

• Refer to Layer 3 for the modified output, for other layers, modify the the Chip.cpp according to the instructions in the comments.

Version History

• 2
  • Added NeuroSim
• 1
  • Added Tetris_with_grouped_convolutions
• 0.3
  • Added flowchart
• 0.2
  • Various bug fixes and optimizations
• 0.1
  • Initial Release