Trait FftBackend 

pub trait FftBackend: BackendSession {
    // Required method
    fn execute_fft(
        &mut self,
        input: &Tensor,
        spec: &FftPlanSpec,
        cache: FftExecutionCache<'_>,
    ) -> Result<Tensor>;

    // Provided method
    fn validate_fft_read_input(
        &self,
        _op: &'static str,
        _input: &TensorRead<'_>,
    ) -> Result<()> { ... }
}

Explicit backend capability required by concrete and traced FFT execution.

Implementations must execute on the input’s existing placement. Unsupported dtypes, layouts, placements, or operations return an error; they must never transfer the tensor or select a different backend.

Examples

Backend surface required by the FFT extension runtime and the built-in concrete FFT dispatch.

This is a backend SPI: third-party backend implementers implement this trait on their execution sessions to provide FFT capability. The concrete [TensorFftExt]/[TensorReadFftExt] methods and [FftExecutor] take an erased &mut dyn BackendSession and dispatch internally to the built-in CPU/CUDA/WebGPU execution sessions; a session without this capability returns a typed Error::Unsupported instead of being accepted as an arbitrary third-party backend (issue #1680 Phase 3).

Examples

use tenferro_cpu::CpuExecSession;
use tenferro_fft::FftBackend;

fn accepts_fft_backend<B: FftBackend>() {}

accepts_fft_backend::<CpuExecSession<'static>>();

A generic tensor backend does not expose the FFT capability at the SPI level:

use tenferro_tensor::{Tensor, TensorBackend};
use tenferro_fft::FftBackend;

fn requires_fft_backend<B: TensorBackend>(backend: &mut B) {
    let _: &mut dyn FftBackend = backend;
}

Required Methods

fn execute_fft( &mut self, input: &Tensor, spec: &FftPlanSpec, cache: FftExecutionCache<'_>, ) -> Result<Tensor>

Execute one validated FFT request on input’s existing placement.

Errors

Returns tenferro_tensor::Error::Unsupported for an unsupported operation, dtype, layout, or placement; tenferro_tensor::Error::Validation for inconsistent input/spec metadata or checked shape arithmetic; and a typed backend or runtime source when plan creation, cache access, or execution fails.

Provided Methods

fn validate_fft_read_input( &self, _op: &'static str, _input: &TensorRead<'_>, ) -> Result<()>

Validate a borrowed runtime input before the extension read path materializes it into a compact tensor.

Backends that can materialize their own placement may keep the default no-op. CPU overrides this to report device inputs as unsupported FFT placement errors instead of leaking a generic host-materialization error.

Errors

Returns tenferro_tensor::Error::Unsupported when this backend cannot consume the borrowed input placement without an explicit transfer.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors