Skip to content

Latest commit

 

History

History
86 lines (66 loc) · 5 KB

OpenProjects.md

File metadata and controls

86 lines (66 loc) · 5 KB
layout title nav_order
default
Open Projects
2

Open Projects

We are excited to invite passionate individuals to join us. If you are interested in making meaningful open-source contributions and gaining invaluable experience, contact us to become part of our team in this journey.

1 - Buddy Compiler Frontend

Project 1-1: Add Operation Mappings in Graph Representation

  • Description: Expand the Buddy Compiler’s graph representation registry of operation mappings to ensure a more comprehensive encapsulation of the PyTorch and ONNX interfacing. The focus is to improve the compiler's ability to optimize and execute a broader range of models efficiently.

  • Expected Outcomes:

    • Enhanced integration of PyTorch AtenIR operations, boosting Dynamo Compiler's functionality and efficiency.
    • Integration of ONNX operations to support a broader range of model architectures.
    • Comprehensive support for models including but not limited to LLaMA, Bert, CLIP, Whisper, Stable Diffusion, ResNet, and MobileNet.
    • Replacement of static deep learning models in the buddy-benchmark repository with ones utilizing the Buddy Compiler frontend.

  • Skills Required:

    • Solid experience with PyTorch and ONNX frameworks.
    • Proficiency in MLIR and Python bindings.
    • A strong understanding of compiler design principles and IR.

  • Possible Mentors: Linquan Wei, Yuliang Li

  • Expected Size of Project: 175 hour

  • Rate: Easy

2 - Buddy Compiler Midend

Analyze and optimize workloads of various AI models and multimodal processes to improve operation efficiency on multiple backend platforms.

Project 2-1: Operations Optimization for CPU SIMD Platforms

  • Description: This project involves multi-level optimizations with the Vector Dialect, Affine Dialect, Transform Dialect, etc., targeting the platforms that support X86 AVX and Arm Neon instruction sets.

  • Expected Outcomes:

    • Implement optimization passes in the buddy-mlir repository for deep learning workloads and operations targeting SIMD platforms.
    • Conduct performance comparisons at both the operation level and the model level within the buddy-benchmark repository.
    • Achieve comparable performance with other optimization frameworks such as TVM and IREE for targeted platforms.

  • Skills Required:

    • Proficiency in MLIR infrastructure and C++ programming.
    • In-depth understanding of optimizations for CPU SIMD platforms.
    • Familiarity with the design methodologies and optimization strategies of TVM and IREE.

  • Possible Mentors: Xulin Zhou, Liutong Han, Hongbin Zhang

  • Expected Size of Project: 350 hour

  • Rate: Hard

Project 2-2: Operations Optimization for GPU Platforms

  • Description: This project involves multi-level optimizations with the GPU Dialect, Transform Dialect, Vector Dialect, Affine Dialect, etc., targeting the GPU platforms.
  • Expected Outcomes:
    • Implement optimization passes in the buddy-mlir repository for deep learning workloads and operations targeting GPU platforms.
    • Conduct performance comparisons on GPU platforms at both the operation level and the model level within the buddy-benchmark repository.
    • Achieve comparable performance with other optimization frameworks such as Triton, TVM, and IREE for targeted platforms.
  • Skills Required:
    • Proficiency in MLIR infrastructure and C++ programming.
    • In-depth understanding of optimizations for GPU platforms.
    • Familiarity with the design methodologies and optimization strategies of Triton, TVM, and IREE.
  • Possible Mentors: Zikang Liu
  • Expected Size of Project: 350 hour
  • Rate: Hard

3 - Buddy Compiler Backend

Project 3-1: Enhancement and Development of Gemmini LLVM Backend with JIT Execution Engine

  • Description: Gemmini is a systolic array accelerator. This project aims to refine the existing Gemmini LLVM backend infrastructure and construct a robust Just-In-Time (JIT) execution engine in buddy-mlir repository. The engine is designed to effectively interpret and execute LLVM IR or dialects within the Gemmini simulator/evaluation platform, optimizing for high performance and efficiency.
  • Expected Outcomes:
    • Enhance the optimization capabilities of the Gemmini backend and refine the Gemmini compilation pipeline.
    • Develop a JIT execution engine and runtime libraries specifically for Gemmini.
    • Conduct performance comparisons at both the operation and model levels within the buddy-benchmark repository, ensuring competitive performance with Gemmini software stack.
  • Skills Required:
    • Proficiency in MLIR and LLVM infrastructure and C++ programming.
    • In-depth understanding of existing Gemmini LLVM backend in buddy-mlir repository.
    • Familiarity with the Gemmini accelerator.
  • Possible Mentors: Zhongyu Qin, Hongbin Zhang
  • Expected Size of Project: 175 hour
  • Rate: Intermediate