TornadoVM

TornadoVM is a plug-in to OpenJDK and other JDK distributions (e.g., GraalVM, Red Hat Mandrel, Amazon Corretto, Microsoft OpenJDK, Azul Zulu) that allows developers to automatically run Java programs on heterogeneous hardware. TornadoVM targets devices compatible with OpenCL, PTX and Level-Zero, which include multi-core CPUs, dedicated GPUs (Intel, NVIDIA, AMD), integrated GPUs (Intel HD Graphics and ARM Mali), and FPGAs (Intel and Xilinx).

TornadoVM has three backends: OpenCL, NVIDIA CUDA PTX, and SPIR-V. Developers can choose which backends to install and run.

TornadoVM does not replace existing VMs, but it rather complements them with the capability of offloading Java code to OpenCL, PTX and SPIR-V, handling memory management between Java and hardware accelerators, and running/coordinating the compute-kernels.

Contents

• Introduction to TornadoVM
  • Main Features
• Installation & Configuration
  • Pre-requisites
  • Supported Platforms
  • Installation
    • A) Automatic Installation
    • B) Manual Installation
      • Linux
      • Windows
      • ARM Mali GPUs and Linux
      • Compilation with Maven
    • Installation for GraalVM 22.3.1 on Linux and OSx
      • 1. Download GraalVM Community Edition 22.3.1
      • 2. Download TornadoVM
      • 3. Install CMAKE (if cmake < 3.6)
        • For Linux:
        • For OSX:
      • 4. Compile TornadoVM with GraalVM
    • TornadoVM for JDK 11+ on Linux and OSx
      • 1. Download a JDK 11+ distribution
      • 2. Download TornadoVM
      • 3. Install CMAKE (if cmake < 3.6)
        • For Linux:
        • For OSX:
      • 4. Compile TornadoVM for JDK 11+
      • Running with JDK 11+
    • TornadoVM for Windows 10/11 using GraalVM
      • 1. Install prerequisites
      • 2. Download TornadoVM
      • 3. Download Graal 22.3.1
      • 4. Install the NVIDIA drivers and CUDA SDK
      • 5. Configure the build
      • 6. Compile TornadoVM
      • 7. Check the installation
    • TornadoVM on ARM Mali GPUs
    • Known issues on Linux
    • Known issues on Windows
    • IDE Code Formatter
      • Using Eclipse and Netbeans
      • Using IntelliJ
    • TornadoVM Maven Projects
    • Versions available
• Running Examples and Benchmarks
  • Running TornadoVM Programs
  • Benchmarking
    • Running all benchmarks with default values
    • Running a specific benchmark
  • Unittests
• Core Programming
  • 1. Expressing Parallelism within Java Methods
    • Loop Parallel API
    • Kernel API
      • Kernel Context
      • KernelContext Features
    • Example
    • Running Multiple Tasks with the Kernel Context
  • 2. Selecting the methods to be accelerated using a Task-Graph API
    • A. Defining copies from the host (main CPU) to the device (accelerator).
      • B. Code definition
      • C. Copy out from the device (accelerator) to the host (main CPU).
  • 3. Execution Plans
  • 4. Obtain the result and the profiler
  • Parallel Reductions
    • Creating reduction tasks
    • Map/Reduce
    • Reduction with dependencies
  • Dynamic Reconfiguration
  • Batch Computing Processing
    • Current Limitations of Batch Processing
  • Migration to TornadoVM v0.15
    • Step 1. TaskSchedule renamed to TaskGraph
      • A. Defining copies from the host (main CPU) to the device (accelerator).
      • B. Code definition
      • C. Copy out from the device (accelerator) to the host (main CPU).
    • Step 2: Create an Immutable Task Graph
    • Step 3: Build, Optimise and Execute an Execution Plan
      • What can we do with an execution plan?
      • How can we optimize an execution plan?
    • Step 4. Obtain the result and the profiler
    • Further reading and examples
• TornadoVM Profiler
  • 1. Enable the Profiler from the Command Line
  • 1. Enabling/Disabling the Profiler using the TornadoExecutionPlan
  • Explanation of all values
    • Note
      • Print timers at the end of the execution
      • Save profiler into a file
      • Parsing Json files
  • Code feature extraction for the OpenCL/PTX generated code
• Benchmarking TornadoVM
  • Benchmarks
  • How to run
    • Example
  • Using JMH
• FPGA Programming in TornadoVM
  • Pre-requisites
  • Step 1: Update/Create the FPGA’s configuration file
    • Example of configuration file for Intel FPGAs (Emulation mode) with the TornadoVM Docker image:
    • Example of configuration file for Intel Nallatech-A385 FPGA (Intel Arria 10 GT1150):
    • Example of configuration file for Xilinx KCU1500:
    • Example of configuration file for Xilinx Alveo U50:
    • Example of configuration file for AWS xilinx_aws-vu9p-f1-04261818_dynamic_5_0:
  • Step 2: Select one of the three FPGA Execution Modes
    • 1. Full JIT Mode
    • 2. Ahead of Time Execution Mode
    • 3. Emulation Mode
      • A) Emulation of an Intel platform:
      • B) Emulation of a Xilinx platform (using Vitis):
• Docker Containers
  • Nvidia GPUs
    • Prerequisites
    • How to run?
    • Using TornadoVM with GraalVM for NVIDIA GPUs
    • Some options
  • Intel Integrated Graphics
    • Prerequisites
    • How to run?
    • Using TornadoVM with GraalVM for Intel Integrated Graphics
• Cloud Deployments
  • 1. Running on AWS for CPUs and GPUs
  • 2. Running on AWS EC2 F1 Xilinx FPGAs
    • Pre-requisites:
    • 1. Install TornadoVM as a CentOS user. The Xilinx FPGA is not exposed to simple users.
    • 2. Follow these steps to get access to the Xilinx FPGA.
    • 3. Update the the FPGA Conguration file
    • 4. Run a program that offloads a task on the FPGA.
    • 5. You can monitor the status of your Amazon FPGA Image.
    • 6. Now that the AFI is available, you can execute the program and run the OpenCL kernel on the AWS FPGA.
• SPIR-V Devices
  • Install Intel oneAPI Level Zero Compute Runtime
  • Install TornadoVM for SPIR-V
  • Running examples with the SPIR-V backend
    • Running DFT from the unit-test suite
      • Disassemble the SPIR-V binary:
• CUDA Devices
  • Prerequisites
    • Driver Installation
  • TornadoVM Installation
  • Addressing Possible issues
  • Testing the CUDA Backend of TornadoVM
• TornadoVM Flags
  • List of TornadoVM Flags:
    • Optimizations
    • Level Zero
• IDE Integration
  • IntelliJ
    • Required Plugins:
    • Run and Debug TornadoVM with Intellij
    • Build/Run Configuration
    • Debug/Run Configuration
• Frequently Asked Questions
  • 1. What can TornadoVM do?
  • 2. Can I use TornadoVM in a commercial application?
  • 3. How can I use it?
    • List of compatible JDKs
      • Windows
      • ARM Mali GPUs and Linux
    • Usage
  • 1. Which programming languages does TornadoVM support?
  • 5. Is TornadoVM a Domain Specific Language (DSL)?
  • 6. Does it support the whole Java Language?
  • 1. Can TornadoVM degrade the performance of my application?
  • 8. Dynamic Reconfiguration? What is this?
  • 9. Does TornadoVM support only OpenCL devices?
  • 10. Why is it called a VM?
  • 11. How does it interact with OpenJDK?
  • 12. How do I know which parts of my application are suitable for acceleration?
  • 13. How can I contribute to TornadoVM?
  • 14. Does TornadoVM support calls to standard Java libraries?
• Unsupported Java features
  • 1. No Recursion
  • 2. No Object Support (*)
  • 3. No Dynamic Memory Allocation (*)
  • 4. No Support for Traps/Exceptions (*)
  • 6. No Support for static TaskGrapshs and Tasks
    • Note
• Resources
  • Videos
  • Presentations
  • Articles
  • Demos & Artefacts
• Publications
  • Full list of publications
• TornadoVM Changelog
  • TornadoVM 0.15
    • Improvements
    • Bug fixes:
  • TornadoVM 0.14.1
    • Improvements
    • Compatibility/Integration
    • Bug Fixes
  • TornadoVM 0.14
    • New Features
    • Compatibility/Integration
    • Improvements/Bug Fixes
  • TornadoVM 0.13
  • TornadoVM 0.12
  • TornadoVM 0.11
  • TornadoVM 0.10
  • TornadoVM 0.9
  • TornadoVM 0.8
  • TornadoVM 0.7
  • TornadoVM 0.6
  • TornadoVM 0.5
  • TornadoVM 0.4
  • TornadoVM 0.3
  • TornadoVM 0.2
  • Tornado 0.1.0