Modular CNN

This work presents a C++-basedsoftware library of reusable modules to build arbitrary CNNs that sup-port High-Level-Synthesis to be implemented as FPGA hardware acceler-ators for the inference process. Our work demonstrates how parametriza-tion and modularization of basic building blocks of a CNN enable aneasier customization of the hardware to match the software model. Thisproject also works with low-precision parameters throughout the CNN,in order to provide a more resource-efficient implementation

Instructions

As of now, there are only the HLS soruces for each layer

1. Create a Viviado HLS project, add core.cpp and core.h as sources. (check the main function name of core.cpp in each layer to know which name to put in as 'top layer')
2. add test_core.cpp, testUtils.cpp, and testUtils.h as testbench.
3. Modify the parameters in core.h according to each layer needs (i.e. number of kernels, size of kernels, etc.)
4. Compile it, test it if needed with a testbench.
5. Sintetize it and export it as IP
6. Add in Vivado as IP block.
7. The python script in the data_proc folder can be used to convert a hdf5 file (a save file from tensorflow for example) into c header files to use directly with the IP blocks.

Published paper

16th Arc 2020, link coming soon