AIE-kmeans

Towards an HW/SW Co-Design Approach for Versal-Based Kmeans Acceleration

The repo will be populated soon :) Keep in Touch!

Info & Description

This repository contains the source code of the Project "A Hardware/Software Co-Design Approach for Versal-Based Kmeans Acceleration".

	Information
Student	Eleonora Cabai ([email protected])
Supervisor name	Davide Conficconi ([email protected])
Supervisor name	Giuseppe Sorrentino ([email protected])
Board used	AMD-Xilinx Versal VCK5000
Software version	2024.1

Main Structure

The archive contains two different implementations of the K-means Clustering algorithm: MacQueen's and Lloyd's. Each version might have multiple implementations, described in detail in the README file inside each folder. Each implementation is structured as follows:

├─ aie/                # AIE kernel for transfomation
|  └─ src/             # Contains the code
├─ data_movers/        # PL kernels for setting up the AIE and for fetching
├─ mutual_info/        # PL Kernel for computing mutual information
├─ common/             # Common codes and variables
├─ hw/                 # System integration
├─ sw/                 # Host application and dataset
└─ Makefile/           # Top-level Makefile to build and run  

Instructions to build and test project

Step 1 - Clone the repository

Open a terminal, then clone the repository by running the following command

git clone https://github.com/necst/AIE-kmeans.git

Then, move into the repository with

cd AIE-KMEANS

and move in the folder of the selected version with

cd implementation/version

Step 2 - Setup the environment

Before building and/or running the framework, run the following script. Notice that this will source Vitis 2022.1, xrt and devtoolset-7

source ./setup_all.sh

Main Commands

aie

make aie_compile_x86 : compile your code for x86 architecture.
make aie_simulate_x86 : simulate your x86 architecture.
make aie_compile : compile your code for VLIW architecture, as your final hardware for HW ad HW_EMU.
make aie_simulate : simulate your code for VLIW architecture, as your final hardware.
make clean : removes all the output file created by the commands listed above.

data_movers

make compile TARGET=HW/HW_EMU: compiles all your kernel, skipping the ones already compiled.
make run_testbench_setup_aie : compiles and execute the testbench for the kernel setup_aie.
make run_testbench_sink_from_aie : compiles and execute the testbench for the kernel setup_aie.

HW

make all TARGET=HW : builds the hardware linking your components make all TARGET=HW_EMU : builds the hardware emu linking your components make clean: removes all files.

SW

make build_sw : compiles the sw ./setup_emu.sh -s on : enables the hardware emulation ./host_overlay.exe : runs the emulation

General useful commands:

If you need to move your bitstream and executable on the target machine, you may want it prepared in a single folder that contains all the required stuff to be moved. In this case, you can use the

make build_and_pack TARGET=hw/hw_emu : it allows you to pack our build in a single folder. Notice that the hw_emu does not have to be moved on the device, it must be executed on the development machine.