How jw1912/bullet works
Main Components
a) Core ML Framework Layer:
• Autograd System (/src/autograd/)
• Tensor Operations (/src/tensor/)
• Backend Operations (/src/backend/)
b) Hardware Acceleration Layer:
• CUDA Backend
• HIP Backend
• BLAS Integration Points
c) NNUE-Specific Components:
• Training System (/src/trainer/)
• Input Feature Handlers
• Output Bucket Management
• Network Architecture
┌─ Application Layer ─────────────────────┐
• Examples & Training Scripts │
• ┬────────────────────┘
• ↓
┌─ Domain Layer ────────────────────────────────────────┐
• ┌─ Training Pipeline ─┐ ┌─ NNUE Specific Tools ─┐ │
• Default Trainer │ │ - Input Features │ │
• Schedule Manager │ │ - Output Buckets │ │
• Data Loader │ │ - Network Architecture │ │
• ┘ └─────────────────────────┘ │
• ┬────────────────────────────────┘
• ↓
┌─ Core ML Layer ──────────────────────────────────┐
• ┌─ Autograd ─┐ ┌─ Tensor Ops ─┐ ┌─ Optimiser ┐│
• Graph │ │ - Dense │ │ - AdamW ││
• Ops │ │ - Sparse │ │ - Utils ││
• ┘ └─────────────┘ └────────────┘ │
• ┬───────────────────────────┘
• ↓
┌─ Hardware Acceleration Layer ─────────────────────┐
• ┌─ CUDA ────┐ ┌─ HIP ─────┐ ┌─ BLAS ────────┐ │
• Kernels │ │ Kernels │ │ Integration │ │
• ┘ └───────────┘ └──────────────┘ │
• ┘
Important Relationships to Highlight
• Data flow from input features through the network
• Interaction between autograd and tensor operations
• Demonstrate how the hardware acceleration layer supports upper layers
• Training pipeline flow
Additional Elements
• Version/compatibility information
• Configuration points
• Extension points for custom implementations