Dispatcher Integration

PrivateUse1

Backend menggunakan slot backend:

PrivateUse1

yang disediakan oleh PyTorch.

Registrasi Backend

rename_privateuse1_backend("opencl")

Dispatch Flow

flowchart LR
    A[Tensor Operation] --> B[Dispatcher]
    B --> C{Device Type}
    C -->|CPU| D[CPU Kernel]
    C -->|CUDA| E[CUDA Kernel]
    C -->|OpenCL| F[OpenCL Kernel]

Operator Registration

Operator didaftarkan menggunakan:

• TORCH_LIBRARY
• TORCH_LIBRARY_IMPL

Device Guard

Backend juga memerlukan:

• device guard
• stream handling
• synchronization primitive