Skip to content

gpu.md

Latest commit

 

History

History
38 lines (29 loc) · 1.51 KB

gpu.md

File metadata and controls

38 lines (29 loc) · 1.51 KB

GPU backend profile

MIND exposes a stable Core v1 GPU profile that defines the contract for GPU backends. The open-core crate does not ship a GPU implementation, but the API surface and error model are intended to remain stable so downstream tooling can target them.

Device model

  • DeviceKind distinguishes CPU and GPU devices. The CPU variant is implemented in this crate, and the GPU variant is part of the stable profile for downstream backends.
  • GPU execution is allowed to be non-deterministic at the bit level (e.g., due to floating-point associativity). Backends must still preserve the semantic meaning of IR operations.

Target model

  • BackendTarget::Cpu is the default compilation target and is fully supported.
  • BackendTarget::Gpu is defined by the Core v1 GPU profile. No concrete GPU backend ships with this crate, but downstream runtimes can implement it.
  • Downstream tools can inspect the target to decide whether to dispatch to a custom backend via the GPUBackend trait.

Error model

Requesting the GPU target currently returns a structured CompileError::BackendUnavailable error from the pipeline. The mindc CLI surfaces this as:

error[backend]: no backend available for target gpu

This behavior ensures an explicit failure instead of a panic or implicit fallback when the GPU path is selected. The presence of a distinct backend-selection error for unsupported or unavailable GPU targets is part of the Core v1 GPU profile and must be preserved across compatible releases.