Struct Session 

pub struct Session<'s> { /* private fields */ }

Public CubeCL extension session.

Not constructible by users; obtained only from CudaExecSession::with_cubecl. The session borrows the exact tenferro CubeCL client for the request scope and is !Send + !Sync by construction, so the execution authority cannot migrate to another thread.

Implementations

impl<'s> Session<'s>

pub fn client(&self) -> &ComputeClient<CubeclCudaRuntime>

Borrow the tenferro CubeCL client.

pub fn tensor_binding<T>( &self, tensor: &TypedTensor<T, impl TensorRank>, op: &'static str, ) -> Result<TensorBinding<CubeclCudaRuntime>>

where T: CubeElement + TensorScalar + Clone,

Build a CubeCL tensor binding for a GPU-backed tensor.

Errors

Returns crate::Error::RuntimeState when the tensor is not resident on this session’s runtime/device, or crate::Error::Validation when its layout cannot be bound.

pub fn array_arg<T>( &self, tensor: &TypedTensor<T, impl TensorRank>, op: &'static str, ) -> Result<ArrayArg<CubeclCudaRuntime>>

where T: CubeElement + TensorScalar + Clone,

Build a CubeCL array argument for a GPU-backed tensor.

Errors

Returns crate::Error::RuntimeState when the tensor is not resident on this session’s runtime/device, or crate::Error::Validation when its layout cannot be bound.

pub fn alloc_output<T>(&self, shape: &[usize]) -> Result<TypedTensor<T>>

where T: CubeElement + TensorScalar + Clone + Send + Sync + 'static,

Allocate a dense GPU tensor on the session’s device.

Errors

Returns crate::Error::Validation when the shape product overflows, or crate::Error::BackendSource when allocation fails.

pub fn alloc_zero_output<T>(&self, shape: &[usize]) -> Result<TypedTensor<T>>

where T: CubeElement + CubePrimitive + TensorScalar + Clone + Send + Sync + 'static,

Allocate a dense GPU tensor zero-filled with the session’s device.

Reuses the backend’s fill-zero structural kernel; it never uploads a host tensor or exposes a device pointer to the caller.

Errors

Returns crate::Error::Validation when the shape product, output byte length, or launch count overflows, crate::Error::RuntimeState when the output is not resident, or crate::Error::BackendSource when allocation or backend resource inspection fails.

Examples

use tenferro_gpu::cuda::cubecl::Session;

fn check(session: &Session<'_>) -> tenferro_tensor::Result<()> {
    let _ = session.alloc_zero_output::<f32>(&[4])?;
    Ok(())
}

pub fn fill_zero_write(&self, output: TensorWrite<'_>) -> Result<()>

Overwrite a caller-owned CUDA destination with an exact +0.0.

This is the beta = 0 reset for re-executing an accumulate-form operation into storage the caller keeps: the previous contents are never read, so a stale NaN or Inf cannot survive and -0.0 is normalized to +0.0, which 0 * y cannot promise. A compact destination (including a nonzero-offset region view) is filled by one stream-ordered memset, so nothing is allocated and no kernel is compiled; a strided region view runs one native fill kernel over its own coordinates and leaves every element outside the region untouched.

Supports F32, F64, I32, I64, C32, C64, and compact Bool destinations.

Errors

Returns crate::Error::RuntimeState when the destination is not resident on this session’s runtime, crate::Error::Validation when the destination layout or its byte span cannot be represented, crate::Error::Unsupported for a strided Bool destination, and crate::Error::BackendSource when the fill cannot be enqueued.

Examples

use tenferro_gpu::cuda::cubecl::Session;
use tenferro_tensor::{Tensor, TensorWrite};

fn reset(session: &Session<'_>) -> tenferro_tensor::Result<()> {
    let destination = session.alloc_output::<f32>(&[4])?;
    let mut destination = Tensor::from_typed::<f32>(destination);
    session.fill_zero_write(TensorWrite::from_tensor(&mut destination))
}

pub fn scale_tensor_write( &self, output: TensorWrite<'_>, factor: f64, ) -> Result<()>

Scale a mutable CUDA tensor in place by a real factor.

Supports F32, F64, C32, and C64 payloads; the factor is interpreted as a real scalar for complex payloads. This is the op-family scale primitive used for normalization after a vendor transform; it never exposes a device pointer to the caller.

Errors

Returns crate::Error::RuntimeState when the tensor is not resident on this session’s runtime, a typed layout error for a non-zero-offset discontinuous view, or the typed unsupported-dtype error for other payloads.

pub fn cube_count_1d(&self, len: usize) -> Result<CubeCount>

Return the cube count for a one-dimensional kernel domain of len.

Errors

Returns crate::Error::Validation carrying ValidationError::InvalidArgument when the one-dimensional launch for len elements would require more than u32::MAX workgroups.

pub fn cube_dim_1d(&self) -> CubeDim

Return the standard one-dimensional CubeCL launch dimension.