Eyeriss Chip simulator

This is a Eyeriss chip simulator that does the same thing as Eyeriss. It was inspired by EyerissF.

However, EyerissF is not a full simulator as it does integrate full Eyeriss mapping strategy into consideration. Also, it does not test a full images but one picture. If you try with the complete minist dataset, EyerissF is not working.

This repo reimplemented and reorganized all functions of EyerissF.

How does Eyeriss Work

Eyeriss is a row stationary DNN accelerator.

If this is the first time you heard about Eyeriss, or you are not very familiar with Eyeriss, the readme of EyerissF gives a good explanation. Please refer to that.

How to Use

Run Lenet5_Hive.py, it outputs inference results of LeNet5 on Mnist dataset. The first there convolution layer are compared with pytorch implementation. Please make sure you have pytorch and numpy installed.

Structure

Source

• Configuration
  • conf.py ( default Eyeriss config features)
• Eyeriss Chip
  • PE.py ( row stationary processing element)
  • EyerissF.py (containing 168 PEs; mapping a Pass of [Weight, IfMap] to each PE)
• Hive
  • Conv2d
  • FullyConnect
  • Activiation
  • Pooling
  • Pre/PostProcess ( Compress and Decompress)

Test

• LeNet5 on Hive.py (test LeNet5 on Mnist, output inference result)
• test IO2.py (test compression and decompression)