graph TD A[Source Code] --> B[Frontend] B --> C[Parser] C --> D[AST Generation] D --> E[Middle-end] E --> F[Optimization Passes] F --> G[Backend] G --> H[NVIDIA CUDA CodeGen] G --> I[Apple Metal CodeGen] G --> J[AMD ROCm CodeGen] H --> K[Runtime System] I --> K J --> K K --> L[GPU Execution] classDef frontend fill:#e1f5fe classDef middleend fill:#f3e5f5 classDef backend fill:#e8f5e9 classDef runtime fill:#fff3e0 class B,C,D frontend class E,F middleend class G,H,I,J backend class K,L runtime

sequenceDiagram participant User participant Compiler participant GPU_Driver participant GPU User->>Compiler: manifoldscript compile file.ms Compiler->>Compiler: Parse & Generate AST Compiler->>Compiler: Optimize Tensor Operations Compiler->>GPU_Driver: Generate GPU-specific code GPU_Driver->>GPU: Load & Execute Kernels GPU->>GPU_Driver: Execution Results GPU_Driver->>Compiler: Return Results Compiler->>User: Output & Performance Metrics

graph TB subgraph "Host Memory" A[CPU Buffer] --> B[Memory Manager] C[Tensor Metadata] --> B end subgraph "GPU Memory" D[GPU Global Memory] --> E[GPU Allocator] F[Shared Memory] --> E G[Constant Memory] --> E end B -->|Transfer| E E -->|Zero-Copy| D E -->|Optimized| F E -->|Read-Only| G H[Multi-GPU Coordinator] --> E H -->|Cross-GPU| I[PCIe/NVLink] classDef host fill:#e3f2fd classDef gpu fill:#ffebee classDef transfer fill:#f3e5f5 class A,B,C host class D,E,F,G gpu class H,I transfer