Automatic Operator Generation for Open AI frameworks: MindSpore/AKG
summary
An automatic kernel generator called AKG compiles MindSpore graphs (accepting MindIR/TVM hybridizer and custom ops) into LLVM IR for CPU, CUDA for GPU and CCO for NPU, applying polyhedral scheduling plus target-specific passes. It fuses operators, finds tile sizes with symbolic tiling, and injects non‑linear constraints to guide scheduling. Measured results show average speedups around 30%, up to 136% on recommendation networks and up to 7× on some operators; remaining challenges are LP scalability, sparse support and target-specific heuristics.
AI-written summary · about
chapters (7)
- 0:00Start
- 3:42Outline
- 5:37Operator Fusion
- 8:21Result Overview
- 11:28Overview of the Architecture
- 12:08Operator Normalization
- 18:19Challenges of Political Scheduling for Machine Learning
related talks
08mMeet MindSpore, the new open source AI framework!
32mPanel Discussion: The AI Kernel...
19mAI Kernel Generation: What's working, what's not, what's next
06mThe Building Blocks of Agentic Al
12mCoding Agents for Compiler Construction: Beyond the...
28mScaling AI Infrastructure
11mKeynote: Open Source Stack Empowers the AI Era
16mProduction MLOps for Spam Detection: Real-Time Multi-Modal AI