Skip to main content

Module backends

candela::docs

Module backends 

Source
Expand description

§Backends

For most of Candela’s life there was no “backend” - there was one implementation: Intel MKL, linked unconditionally and specialized for f64. That scaled badly once new dtypes such as f32 arrived and needed different kernel implementations. Inspired by Burn, the backend became a trait, to ease porting across hardware and kernel implementations.

§The ComputeFor<B> layer

Each backend may specialize a dtype differently - some dtypes are vectorizable, others are not, and that changes from backend to backend. ComputeFor expresses this (backend, dtype) space in a somewhat scalable way, one implementation per pair. It also doubles as a construction gate: only a dtype that implements ComputeFor for a given backend can be constructed against it.

§The backends that ship

Both are CPU, and both are zero-sized policy types - all the state lives in the TensorData buffers that flow through compute.

  • CpuPure - the default. Pure Rust: matrixmultiply for MatMul, straightforward loops over the layout for everything else. No system dependencies, builds on any target Rust supports.
  • CpuMkl - opt-in, behind --features mkl. Routes element-wise math through MKL/VML and matmul through CBLAS, faster on Intel hardware. Requires the MKL libraries on the system.

DefaultBackend is the type alias the constructors infer when none is named, selected by the mkl feature flag: CpuMkl when enabled, CpuPure otherwise.