MojoBLAS

Description

MojoBLAS is a GPU-accelerated implementation of a basic linear algebra subprogram (BLAS) library in the programming language Mojo. This library is an efficient, portable, and easy-to-use BLAS library built directly on Mojo’s native GPU abstractions. It ensures the technical correctness of Level 1, 2, and 3 BLAS routines based on the original FORTRAN 77 specification, support for many floating-point precision types in GPU kernels, multi-platform portability across NVIDIA and AMD hardware, and matches performance in current C++ BLAS libraries.

Installation Instructions

Machines with NVIDIA / AMD GPUs

1. Set up your Pixi coding environment.

• If you don't have pixi already, you can install it with:

  curl -fsSL https://pixi.sh/install.sh | sh
  

2. Restart the terminal

• Close and reopen the terminal.

• If pixi has been successfully installed, you will see the pixi version return with this test:

  pixi --version
  

3. That's it! To run the program, see the 'Usage' section.

Macs

NOTE: Mojo currently has limitied support for Mac GPUs. Thus, we do not expect all tests to succeed on Mac machines. However, to provide full accessibility to our project, installation instructions have been included below. If you want full functionality of MojoBLAS, please consider using NVIDIA or AMD GPUs, which are fully supported by Mojo.

1. Set up your Pixi coding environment.

• If you don't have pixi already, you can install it with:

  curl -fsSL https://pixi.sh/install.sh | sh
  

2. Ensure pixi paths are visible.

• Manually add pixi to PATH.
  • For zsh shell:

    echo 'export PATH="$HOME/.pixi/bin:$PATH"' >> ~/.zshrc
    source ~/.zshrc
    

  • For bash shell:

    echo 'export PATH="$HOME/.pixi/bin:$PATH"' >> ~/.bashrc
    source ~/.bashrc
    

    • Sometimes this command fails to add the pixi path to the login shell, which means this path won't be remembered by future shells. If you are running into an issue where new bash shells can't find the pixi executable, try adding it to your bash profile with:

      echo 'if [ -f ~/.bashrc ]; then . ~/.bashrc; fi' >> ~/.bash_profile
      source ~/.bash_profile
      

3. Check pixi for successful installation:

• If pixi has been successfully installed, you will see the pixi version return with this test:

  pixi --version
  

4. Ensure XCode is downloaded.

• Check your application folder with:

  ls /Applications/Xcode.app
  

• If you see a list of folder contents, you're good to go!
• If not, download XCode from the App Store, and follow these extra instructions:
  • Change the command line tools to point to XCode's toolchain with:

    sudo xcode-select --switch /Applications/Xcode.app/Contents/Developer
    

  • Accept XCode's license agreement:

    sudo xcodebuild -license accept
    

  • Download MetalToolchain:

    sudo xcodebuild -downloadComponent MetalToolchain
    

  • Check it was successfully downloaded with:

    xcrun --find metallib
    

Usage

Machines with NVIDIA / AMD GPUs

1. Run the test with:

   pixi run mojo test-level1.mojo
   

Macs

1. Add arm64 support.

• First, add a workspace platfrom for osx-arm64 systesm by running:

  pixi workspace platform add osx-arm64
  

  • This will add a pixi.lock file for supported osx-arm64 machines.

2. Follow the instructions for NVIDIA / AMD GPUs to run the tests.

Contact Information

For questions about this library, please contact the project leader Tatiana Melnichenko or Alexa Andershock.

Acknowledgements

This project could not have been completed without contributions from Alexa Andershock, Gian Fernandez-Aleman, Tatiana Melnichenko, Jackson Mowry, Holden Roaten, and Jonah Weston.