Overview

图像缩放
AMD Vitis Accelerated Libraries Diagram

AMD Vitis™ Accelerated Libraries

The AMD Vitis™ unified software platform includes an extensive set of open-source, performance-optimized libraries that offer out-of-the-box acceleration with minimal to zero code changes to your existing applications.

  • Common Vitis accelerated-libraries, including Solver, Basic Linear Algebra Subroutines (BLAS), Sparse, DSP, and Utilities, offer a set of core functionality for a wide range of diverse applications.
  • Domain-specific Vitis accelerated libraries offer out-of-the-box acceleration for workloads like Vision and Image Codec Processing, Quantitative Finance, HPC, Graph, Database, Data Analytics, Data Compression, and more. 
  • Leverage the rich growing ecosystem of partner-accelerated libraries, framework plug-ins, and accelerated applications to hit the ground running and accelerate your path to production.

Vitis libraries now contain DSP, matrixes, and other functions that are optimized for implementation in the AI Engine portion of Versal™ devices. 

Use in Familiar Programming Languages

Use Vitis accelerated-libraries in commonly used programming languages that you know, like C/C++. Some of these libraries also include Python functions on Level 3, such as the Vitis BLAS library and Vitis Quantitative Finance library. By leveraging AMD platforms as an enabler in your applications, you can work at an application level and focus your core competencies on solving challenging problems in your domain—accelerating your time to insight and enabling faster innovation.

Whether you want to accelerate portions of your existing x86 host application code or want to develop accelerators for deployment on AMD embedded platforms, calling a Vitis accelerated-library API or kernel in your code offers the same level of abstraction as any other software library.

Familiar Programming Languages Diagram
图像缩放
AMD Vitis Accelerated-Libraries Diagram

Scalable and Flexible

Vitis accelerated-libraries are accessible to all developers through GitHub and scalable across all AMD platforms. Develop your applications using these optimized libraries and seamlessly deploy across our platforms at the edge, on premise, or in the cloud without having to reimplement your accelerated application.

With the rapid prototyping and quick evaluation that AMD can bring to your applications, you can use these libraries as plug-and-play accelerators, called directly as APIs in the user application for workloads like Vision and Image Codec Processing, Quantitative Finance, HPC, Graph, Database, and Data Analytics, among others.

When designing custom accelerators for your application, use Vitis library functions as optimized algorithmic building blocks, modify them to suit your specific needs, or use them as references to completely design your own. Choose the flexibility you need!

Combine domain-specific Vitis libraries with pre-optimized deep learning models from the Vitis AI library or the Vitis AI development kit to accelerate your whole application and meet overall system-level functionality and performance goals.

图像缩放
AMD Vitis Library Functions Diagram
图像缩放
AMD Vitis Library Functions Diagram

Vitis Library Functions Optimized for Versal AI Engine 

The AI Engines found in the Versal devices provide very high compute density for vector-based algorithms.

The following libraries have AI Engine additions:

  • Vitis DSP Library
  • Vitis Vision Library
  • Vitis Solver Library

AI Engine code can be found under the "AIE" directories under L1 for AIE-only functions and under L2 for functions that are comprised of both AIE and PL code.

Note: For more details, please refer to the pages of each library.

Library File Organization

图像缩放
AMD Vitis Library File Organization Diagram

Typically, a Vitis library includes three levels (L1/L2/L3) of functions:

L1 Primitives

  • Basic algorithmic functions (HLS functions) for designing kernels
  • Customize or combine with other primitives and kernels
  • Requires build and compile with Vitis tools

L2 Kernels

  • Performance-optimized kernels with required interfaces and compiler directives
  • Customize, if needed, and build accelerator image with Vitis tools
  • Initialization and data transfers managed via integration with Vitis Runtime Library

L3 Software APIs

  • High-level software APIs directly callable in host application
  • APIs call pre-built FPGA images (available for specific platforms)
  • Users can customize and rebuild

Libraries

Vitis Blockchain Solution

First Vitis blockchain mining acceleration solution to power-efficient FPGA-based mining. Outperform the most efficient mining card with 2x mining performance/watt over GPUs.

Vitis DSP Library

Accelerating DSP functions on Versal™ AI Engines such as filters, FFt/iFFT, matrix multiply,  widget API cast, widget real to complex, and DDS/Mixer.

Vitis Vision Library

Develop and deploy accelerated computer vision and image processing algorithms for ADAS, video surveillance, data center, etc. on AMD platforms.