fpgaHART

Prerequisites

Create a python environment with python version >=3.9 and install the dependencies from requirements.txt file.

pip install -r requirements.txt

Library basic usage (modelling and optimization of CNNs)

python main.py {model_name} {device_name} {optimization-type} {optimization-target}

Mandatory arguments

• model_name: Name of the CNN model to be optimized. The model should be provided in an .onnx file and placed under the folder models/. The value of this parameters should be the name of the .onnx file without the suffix.
• device_name: Name of the FPGA device to be used for the optimization process. The supported devices are zc706, zcu104-106, zcu102, vc707, vc709, vus440.
• optimization-type: This argument accepts only three values network, partition or layer. It dictates the type of optimization to be performed. Whether to optimize the complete model, a single partition or a specific layer of the model.
• optimization-target: This argument accepts only two values throughput or latency. Dictates the target of the optimization process.

Additional arguments

• --singlethreaded: Whether to use a single-threaded solution or not.
• --se_block: Whether to treat squeeze excitation as a block/layer or not.
• --plot_layers: Whether to plot design points per layer or not.
• --gap_approx: Whether to use historical data as an approximation for GlobalAveragePooling (GAP) layers or not during the optimization process.
• --sweep: Whether to run a (wandb) sweep of the design space or not.
• --enable_wandb: Whether to enable wandb support or not.
• --profile: Whether to profile the whole program or not.

Backend code generation

./backend_pipeline.sh (only works for *throughput* optimization-target)

Automates the process of optimization data generation and creation of the backend code files for HLS. At the start of the scripts there are some variables that need to be defined.

• MODEL_NAME: Same as model_name argument
• EXECUTION_TYPE: Same as optimization-type argument
• TARGET: Same as optimization-target argument
• CONFIG_FILE: The output file path of the the optimization process (only for throughput optimization-target)
• PARTITION_NAME: The name of the final folder containing the partitions of the model
• HLS_PARENT_DIR: The path of the HLS folder in which the backend code will be generated

./layer_hls_project.sh (only works for *throughput* optimization-target)

The arguments are the same as in the backend_pipeline.sh case.

Configuration files

The configuration files for specifying the FPGA device characteristics (config_fpga.ini), the wandb and optimizer parameters (config_optimizer.yaml), and the pytorch/onnx supported layers and parameters (config_pytorch.ini) are all under the following path $fpga-hart/fpga_hart/config/.