Struct CudaRuntime 

pub struct CudaRuntime { /* private fields */ }

CubeCL CUDA runtime wrapper.

Examples

use tenferro_gpu::cuda::CudaRuntime;

let _ctor: fn(tenferro_gpu::cuda::CudaDeviceId) ->
    Result<CudaRuntime, tenferro_gpu::cuda::CudaDeviceError> = CudaRuntime::new;
let _sync: fn(&CudaRuntime) -> tenferro_tensor::Result<()> =
    CudaRuntime::synchronize;

Implementations

impl CudaRuntime

pub fn new(device_id: CudaDeviceId) -> Result<Self, CudaDeviceError>

Initialize the CubeCL CUDA runtime on the caller-selected device.

Examples

use tenferro_gpu::{cuda::CudaDeviceError, cuda::CudaDeviceId, cuda::CudaRuntime};

let _ctor: fn(CudaDeviceId) -> Result<CudaRuntime, CudaDeviceError> = CudaRuntime::new;

Errors

Returns CudaDeviceError::Discovery when fallback discovery for an invalid selected ordinal fails, CudaDeviceError::Unavailable when that ordinal is not available, or CudaDeviceError::Initialization when CUDA driver, runtime, context, or CubeCL client initialization fails.

pub fn device_id(&self) -> CudaDeviceId

Return the caller-selected CUDA device identity that this runtime targets.

Examples

use tenferro_gpu::{cuda::CudaDeviceId, cuda::CudaRuntime};

let _device_id: fn(&CudaRuntime) -> CudaDeviceId = CudaRuntime::device_id;

pub fn device_info(&self) -> &CudaDeviceInfo

Return immutable metadata for the runtime’s device.

Examples

use tenferro_gpu::cuda::CudaRuntime;

let _info: fn(&CudaRuntime) -> &tenferro_gpu::cuda::CudaDeviceInfo =
    CudaRuntime::device_info;

pub fn allocation_domain(&self) -> AllocationDomainId

Return the allocation ownership domain of this runtime.

Examples

use tenferro_gpu::cuda::CudaRuntime;

let _domain: fn(&CudaRuntime) -> tenferro_tensor::AllocationDomainId =
    CudaRuntime::allocation_domain;

pub fn supports_extension(&self, capability: GpuExtensionCapability) -> bool

Report whether this CUDA session supports a GPU extension capability.

The CUDA provider supports the full extension vocabulary: external CubeCL kernels, native module loading, runtime compilation (NVRTC), raw stream borrowing, and same-device copy. PeerCopy is reported as a directional query; availability is hardware-dependent and is checked per source/destination pair rather than here.

Examples

use tenferro_gpu::cuda::GpuExtensionCapability;
use tenferro_gpu::cuda::CudaRuntime;

let _supports: fn(&CudaRuntime, GpuExtensionCapability) -> bool =
    CudaRuntime::supports_extension;

pub fn with_current_context<R>( &self, op: &'static str, f: impl FnOnce() -> R, ) -> Result<R>

Run f with the tenferro primary context current on this thread.

Saves the calling thread’s current CUDA device/context, activates the tenferro primary context for the duration of f, and attempts to restore the saved state on every exit path (normal return, Err, and unwind). Restoration is best-effort: a failure to restore the caller’s previous device/context is logged to stderr rather than returned. This is the scoped context authority used by vendor-library lifecycle paths (plan creation/retirement) that run outside a raw-session callback.

Errors

Returns crate::Error::BackendSource when the tenferro primary context cannot be activated (device or context driver failure); a partial activation is best-effort rolled back before the error is returned (rollback failures are discarded).

Examples

use tenferro_gpu::cuda::CudaRuntime;

let _check: fn(&CudaRuntime) -> tenferro_tensor::Result<u64> = |rt| {
    rt.with_current_context("test.context", || 7)
};

pub fn runtime_identity(&self) -> CudaRuntimeIdentity

Return the opaque identity of this exact executable runtime instance.

Examples

use tenferro_gpu::cuda::CudaRuntime;

let _identity: fn(&CudaRuntime) -> tenferro_gpu::cuda::CudaRuntimeIdentity =
    CudaRuntime::runtime_identity;

pub fn synchronize(&self) -> Result<()>

Block the current thread until work submitted to the current CUDA stream completes.

Examples

use tenferro_gpu::cuda::CudaRuntime;

let _sync: fn(&CudaRuntime) -> tenferro_tensor::Result<()> =
    CudaRuntime::synchronize;

Errors

Returns crate::Error::RuntimeState when CubeCL cannot expose the current stream, or crate::Error::BackendSource when CUDA context or stream synchronization fails.

Trait Implementations

impl Clone for CudaRuntime

fn clone(&self) -> CudaRuntime

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for CudaRuntime

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.