2020.1 Vitis™ Application Acceleration Development Flow Tutorials

See 2019.2 Vitis Application Acceleration Development Flow Tutorials

Optimizing Accelerated FPGA Applications: Convolution Example

Introduction

The methodology for developing optimized accelerated applications is comprised of two major phases: architecting the application and developing the accelerator to meet your desired performance goals.

  • In the first phase, you make key decisions about the application architecture to determine which software functions to accelerate using FPGA kernels and identify how much parallelism can be achieved and delivered in the code.

  • In the second phase, you implement kernels by structuring the source code, and applying the necessary compiler options and pragmas to create the kernel architecture needed to achieve the performance target.

You begin this tutorial with a baseline application, and profile it to examine the potential for hardware acceleration. The application uses a 2D convolution filter to process multi-channel RGB video stream, a popular multimedia framework that can play, transcode, mux, demux, and filter many audio/video formats. Then, you perform various optimizations on both the host program and kernel side. In this tutorial, you work with the following optimization techniques:

  • Memory transfer optimizations

  • Fixed point data type adoption

  • Dataflow and streams

  • Loop optimization

This tutorial follows the Methodology for Accelerating Applications with the Vitis Unified Software Platform in the Application Acceleration Development flow of the Vitis Unified Software Platform Documentation (UG1416) about how to migrate a CPU-based application to an optimized FPGA-accelerated design. For a deeper understanding, you should review that material as you are working through this tutorial.

Before You Begin

This tutorial requires that the FFmpeg framework is installed on the machine where the steps will be executed. Download and run the following commands:

  • For CentOS:

    sudo yum localinstall --nogpgcheck https://download1.rpmfusion.org/free/el/rpmfusion-free-release-7.noarch.rpm
sudo yum install ffmpeg

  • For Ubuntu:

    sudo apt update
sudo apt install ffmpeg
ffmpeg -version

The labs in this tutorial use:

  • BASH Linux shell commands.

  • 2020.1 Vitis core development kit release and the xilinx_u200_xdma_201830_2 platform. If necessary, it can be easily ported to other versions and platforms.

  • A Makefile that is detailed and contains many steps and variables. For a discussion of the Makefile structure and contents, refer to Understanding the Makefile.

IMPORTANT:

  • Before running any of the examples, make sure you have installed the Vitis core development kit as described in Installation in the Application Acceleration Development flow of the Vitis Unified Software Platform Documentation (UG1416).

  • If you run applications on Xilinx® Alveo™ Data Center accelerator cards, ensure the card and software drivers have been correctly installed by following the instructions on the Alveo Portfolio page.

Accessing the Tutorial Reference Files

  1. To access the reference files, enter the following in a terminal: git clone  http://github.com/Xilinx/Vitis-Tutorials.

  2. Navigate to the convolution-tutorial directory, and then access the design directory.

Next Steps

The following labs walk through the best practices for taking an existing application and optimizing it as an FPGA-accelerated application. This tutorial is divided into several different labs that show the methodology. It is recommended to complete each lab before proceeding to the next.


Return to Main Page — Return to Getting Started Pathway


Copyright© 2020 Xilinx