Struct CudaBackend 

pub struct CudaBackend { /* private fields */ }

CUDA backend (stub) — placeholder when cuda feature is not enabled.

Status: Stub. All methods return errors. Enable the cuda feature for the real implementation backed by cuTENSOR + cudarc.

Examples

// Aspirational API — enable `cuda` feature for real backend.
use tenferro_prims::{CudaBackend, BackendRegistry};

let mut registry = BackendRegistry::new();
registry.load_cutensor("/usr/lib/libcutensor.so").unwrap();

Implementations

impl CudaBackend

pub fn resolve_conj<T: Scalar + Conjugate>( _ctx: &mut CudaContext, src: &Tensor<T>, ) -> Tensor<T>

Materialize a lazily-conjugated tensor on GPU.

Status: Stub fallback. If data is CPU-accessible, this materializes conjugation into a new non-conjugated tensor. Otherwise it returns a clone of src.

Trait Implementations

impl Drop for CudaBackend

Available on non-crate feature cuda only.

fn drop(&mut self)

Executes the destructor for this type.

impl<S: Scalar> TensorAnalyticPrims<Standard<S>> for CudaBackend

Available on non-crate feature cuda only.

type Plan = ()

type Context = CudaContext

fn plan( _ctx: &mut Self::Context, desc: &AnalyticPrimsDescriptor, _shapes: &[&[usize]], ) -> Result<Self::Plan>

Plan an analytic-family operation for the given input/output shapes.

fn execute( _ctx: &mut Self::Context, _plan: &Self::Plan, _alpha: S, _inputs: &[&Tensor<S>], _beta: S, _output: &mut Tensor<S>, ) -> Result<()>

Execute a previously planned analytic-family operation.

fn has_analytic_support(_desc: AnalyticPrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl<Input> TensorComplexRealPrims<Input> for CudaBackend

where Input: ComplexFloat + Scalar, Input::Real: Scalar,

Available on non-crate feature cuda only.

type Real = <Input as ComplexFloat>::Real

The real-valued output scalar type.

type Plan = ()

Backend plan type.

type Context = CudaContext

Backend execution context.

fn plan( _ctx: &mut Self::Context, desc: &ComplexRealPrimsDescriptor, _shapes: &[&[usize]], ) -> Result<Self::Plan>

Plan a complex-to-real unary operation for the given input/output shapes.

fn execute( _ctx: &mut Self::Context, _plan: &Self::Plan, _alpha: Input::Real, _inputs: &[&Tensor<Input>], _beta: Input::Real, _output: &mut Tensor<Self::Real>, ) -> Result<()>

Execute a previously planned complex-to-real unary operation.

fn has_complex_real_support(_desc: ComplexRealPrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl<Input> TensorComplexScalePrims<Input> for CudaBackend

where Input: ComplexFloat + Scalar, Input::Real: Scalar + Send + Sync,

Available on non-crate feature cuda only.

type Plan = ()

Backend plan type.

type Context = CudaContext

Backend execution context.

fn plan( _ctx: &mut Self::Context, desc: &ComplexScalePrimsDescriptor, _shapes: &[&[usize]], ) -> Result<Self::Plan>

Plan a complex-by-real pointwise operation for the given shapes.

fn execute( _ctx: &mut Self::Context, _plan: &Self::Plan, _alpha: Input, _lhs: &Tensor<Input>, _rhs: &Tensor<Input::Real>, _beta: Input, _output: &mut Tensor<Input>, ) -> Result<()>

Execute a previously planned complex-by-real pointwise operation.

fn has_complex_scale_support(_desc: ComplexScalePrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl<S: Scalar> TensorIndexingPrims<Standard<S>> for CudaBackend

Available on non-crate feature cuda only.

type Plan = ()

Backend-specific execution plan.

type Context = CudaContext

Backend execution context.

fn plan( _ctx: &mut Self::Context, desc: &IndexingPrimsDescriptor, _shapes: &[&[usize]], ) -> Result<Self::Plan>

Plan an indexing operation for the given input/index/output shapes.

fn execute( _ctx: &mut Self::Context, _plan: &Self::Plan, _inputs: &[&Tensor<S>], _indices: &Tensor<i64>, _output: &mut Tensor<S>, ) -> Result<()>

Execute a previously planned indexing operation.

fn has_indexing_support(_desc: IndexingPrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl<S: Scalar + NumCast> TensorMetadataCastPrims<S> for CudaBackend

Available on non-crate feature cuda only.

type Plan = MetadataCastPrimsDescriptor

Backend plan type.

type Context = CudaContext

Backend execution context.

fn plan( _ctx: &mut CudaContext, desc: &MetadataCastPrimsDescriptor, _shapes: &[&[usize]], ) -> Result<Self::Plan>

Plan a metadata-to-scalar bridge operation.

fn execute( _ctx: &mut CudaContext, _plan: &Self::Plan, _alpha: S, _inputs: &[MetadataScalarTensorRef<'_, S>], _beta: S, _output: &mut Tensor<S>, ) -> Result<()>

Execute a previously planned metadata-to-scalar bridge operation.

fn has_metadata_cast_support(_desc: MetadataCastPrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl TensorMetadataPrims for CudaBackend

Available on non-crate feature cuda only.

type Plan = MetadataPrimsDescriptor

Backend plan type.

type Context = CudaContext

Backend execution context.

fn plan( _ctx: &mut CudaContext, desc: &MetadataPrimsDescriptor, _inputs: &[MetadataTensorRef<'_>], _output: MetadataTensorMut<'_>, ) -> Result<Self::Plan>

Plan a metadata-family operation for the given input and output tensor handles.

fn execute( _ctx: &mut CudaContext, _plan: &Self::Plan, _inputs: &[MetadataTensorRef<'_>], _output: MetadataTensorMut<'_>, ) -> Result<()>

Execute a previously planned metadata-family operation in overwrite mode.

fn has_metadata_support(_desc: MetadataPrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl TensorRngPrims<Standard<f64>> for CudaBackend

Available on non-crate feature cuda only.

type Plan = (RngPrimsDescriptor, Vec<usize>)

Backend-specific execution plan.

type Context = CudaContext

Backend execution context.

fn plan( _ctx: &mut Self::Context, desc: &RngPrimsDescriptor, _shapes: &[&[usize]], ) -> Result<Self::Plan>

Plan a tensor RNG operation for the given output shape.

fn execute( _ctx: &mut Self::Context, _plan: &Self::Plan, _generator: &mut Generator, _output: &mut Tensor<f64>, ) -> Result<()>

Execute a previously planned RNG operation.

fn has_rng_support(_desc: RngPrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl TensorRngPrims<Standard<i32>> for CudaBackend

Available on non-crate feature cuda only.

type Plan = (RngPrimsDescriptor, Vec<usize>)

Backend-specific execution plan.

type Context = CudaContext

Backend execution context.

fn plan( _ctx: &mut Self::Context, desc: &RngPrimsDescriptor, _shapes: &[&[usize]], ) -> Result<Self::Plan>

Plan a tensor RNG operation for the given output shape.

fn execute( _ctx: &mut Self::Context, _plan: &Self::Plan, _generator: &mut Generator, _output: &mut Tensor<i32>, ) -> Result<()>

Execute a previously planned RNG operation.

fn has_rng_support(_desc: RngPrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl<S: Scalar> TensorScalarPrims<Standard<S>> for CudaBackend

Available on non-crate feature cuda only.

type Plan = ()

type Context = CudaContext

fn plan( _ctx: &mut Self::Context, desc: &ScalarPrimsDescriptor, _shapes: &[&[usize]], ) -> Result<Self::Plan>

Plan a scalar-family operation for the given input/output shapes.

fn execute( _ctx: &mut Self::Context, _plan: &Self::Plan, _alpha: S, _inputs: &[&Tensor<S>], _beta: S, _output: &mut Tensor<S>, ) -> Result<()>

Execute a previously planned scalar-family operation.

fn has_scalar_support(_desc: ScalarPrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl<S: Scalar> TensorSemiringCore<Standard<S>> for CudaBackend

Available on non-crate feature cuda only.

type Plan = CudaPlan<S>

Backend-specific plan type.

type Context = CudaContext

Backend-specific execution context.

fn plan( _ctx: &mut CudaContext, _desc: &SemiringCoreDescriptor, _shapes: &[&[usize]], ) -> Result<CudaPlan<S>>

Plan a semiring-core operation.

fn execute( _ctx: &mut CudaContext, _plan: &CudaPlan<S>, _alpha: S, _inputs: &[&Tensor<S>], _beta: S, _output: &mut Tensor<S>, ) -> Result<()>

Execute a semiring-core operation.

impl<S: Scalar> TensorSemiringFastPath<Standard<S>> for CudaBackend

Available on non-crate feature cuda only.

type Plan = CudaPlan<S>

Backend-specific plan type.

type Context = CudaContext

Backend-specific execution context.

fn plan( _ctx: &mut CudaContext, _desc: &SemiringFastPathDescriptor, _shapes: &[&[usize]], ) -> Result<CudaPlan<S>>

Plan an optional semiring fast path.

fn execute( _ctx: &mut CudaContext, _plan: &CudaPlan<S>, _alpha: S, _inputs: &[&Tensor<S>], _beta: S, _output: &mut Tensor<S>, ) -> Result<()>

Execute an optional semiring fast path.

fn has_fast_path(_desc: SemiringFastPathDescriptor) -> bool

Query whether the optional path is available.

impl<S: Scalar + PartialOrd> TensorSortPrims<Standard<S>> for CudaBackend

Available on non-crate feature cuda only.

type Plan = ()

Backend-specific execution plan.

type Context = CudaContext

Backend execution context.

fn plan( _ctx: &mut Self::Context, desc: &SortPrimsDescriptor, _shapes: &[&[usize]], ) -> Result<Self::Plan>

Plan a sort operation for the given input shape.

fn execute( _ctx: &mut Self::Context, _plan: &Self::Plan, _input: &Tensor<S>, _values_out: &mut Tensor<S>, _indices_out: &mut Tensor<i64>, ) -> Result<()>

Execute a previously planned sort operation.

fn has_sort_support(_desc: &SortPrimsDescriptor) -> bool

Report whether the backend advertises support for the given descriptor.

impl Send for CudaBackend

Available on non-crate feature cuda only.

Safety

CudaBackend can be safely sent across threads because:

• The _handle is an opaque pointer to a cuTENSOR handle
• The _lib (libloading::Library) is thread-safe after loading
• The handle is read-only after construction
• Drop clears the handle before the library is unloaded, preventing use-after-free

impl Sync for CudaBackend

Available on non-crate feature cuda only.

Safety

CudaBackend can be safely shared across threads because:

• The cuTENSOR handle is designed for concurrent use from multiple threads
• The library handle (_lib) is read-only after construction
• Symbol lookup via dlsym is thread-safe on POSIX systems
• Drop uses &mut self, ensuring exclusive access during cleanup