fp-cnn

An end-to-end framework to map a CNN (TensorFlow) onto FPGA (OpenCL).
Link to our repo: https://github.com/yo96/fp-cnn.

How to generate the accelerator

• Specify the network configuration

  The network configuration file is model/<your_model_name>/nn_configs.txt where <your_model_name> is the name of the model. We have provided a reference network configuration for MNIST under model/MNIST.

  nn_configs.txt should be generated by our Tensorflow API wrapper. During our project, we used the nn-quant library from Ritchie to build the model and dump weights, therefore we did not include that part in this repo. We will provide the wrapper after nn-quant, which is currently private, becomes public.

• Provide the weights file

  The network configuration file is model/<your_model_name>/_dump.npy where <your_model_name> is the name of the model. We have provided a reference binary weight file for MNIST under model/MNIST.

  _dump.npy should be generated by --save-activations option of the nn-quant library. It is basically a dictionary of numpy arrays. The key for an array that stores the weights for a specific layer should end with '/w'. For exampe, conv1/w, last/w are valid keys for weights. During our project, we got the permission from Ritchie to use this library but it is private. You can access this library after it becomes public.

  Please note that due to CMSX file size limit, we cannot upload the _dump.npy file to the system. You can download the complete project at https://github.com/yo96/fp-cnn.

• Run the generation framework

  source utils/gen.sh

  The above command will generate a new SDAccel project under src/<you_model_name>/. All files should be ready and you can directly run software/hardware emulation there.

Other projects in this repo

under src directory there are also a number of projects that we created when we are developing this framework:

• conv_example: an example project that contains only functional-level kernels.
• conv: initial project created when developing the compute kernel. It has functional-level load_fmap, load_wts, and a dummy output kernel.
• load_fmap: initial project created when developing the load_fmap kernel. It has a dummy output kernel.
• conv_pool: a project that tests the integration of conv, acc_relu, and pooling kernels.
• conv_pool_test: a project that tests the integration of load_fmap, conv,acc_relu, and pooling kernels.
• verif_conv: a projcet that verifies the device result of a conv layer against the intermediate result directly dumped from tensorflow.
• verif_fc: a projcet that verifies the device result of a fully-connected layer against the intermediate result directly dumped from tensorflow.
• vadd: a hello world SDAccel project.
• measure_mem: a testing project that measures the effective DDR bandwidth.

How to use the makefile inside the SDAccel project directory

• Software emulation

make -f sdaccel.mk run_cpu_em

• Hardware emulation

make -f sdaccel.mk run_hw_emu

• Run on board

make -f sdaccel.mk run_hw

• Clean

make -f sdaccel.mk cleanall