Struct GpuTensor 

pub struct GpuTensor<T: GpuFloat> { /* private fields */ }

A tensor residing on a CUDA GPU.

Wraps a CudaBuffer<T> with shape metadata and a reference to the GpuDevice that owns the memory. Created by tensor_to_gpu or the convenience functions cuda / cuda_default.

Convert back to a CPU Tensor with GpuTensor::cpu or the free function tensor_to_cpu.

Implementations

impl<T: GpuFloat> GpuTensor<T>

pub fn shape(&self) -> &[usize]

The shape of this tensor.

pub fn numel(&self) -> usize

Total number of elements.

pub fn device(&self) -> &GpuDevice

The GPU device that holds this tensor’s data.

pub fn buffer(&self) -> &CudaBuffer<T>

Borrow the underlying CudaBuffer.

pub fn ndim(&self) -> usize

Number of dimensions.

pub fn cpu(&self) -> FerrotorchResult<Tensor<T>>

Copy this tensor back to CPU, returning a Tensor<T>.

This is a convenience wrapper around tensor_to_cpu.

impl<T: GpuFloat> GpuTensor<T>

pub fn add(&self, other: &GpuTensor<T>) -> GpuResult<GpuTensor<T>>

Elementwise addition: out[i] = self[i] + other[i].

Uses a PTX kernel for f32; falls back to CPU round-trip for f64.

Errors

• GpuError::LengthMismatch if shapes differ.
• GpuError::DeviceMismatch if tensors are on different devices.
• GpuError::Driver on CUDA runtime errors.

pub fn sub(&self, other: &GpuTensor<T>) -> GpuResult<GpuTensor<T>>

Elementwise subtraction: out[i] = self[i] - other[i].

Uses a PTX kernel for f32; falls back to CPU round-trip for f64.

pub fn mul(&self, other: &GpuTensor<T>) -> GpuResult<GpuTensor<T>>

Elementwise multiplication: out[i] = self[i] * other[i].

Uses a PTX kernel for f32; falls back to CPU round-trip for f64.

pub fn neg(&self) -> GpuResult<GpuTensor<T>>

Elementwise negation: out[i] = -self[i].

Uses a PTX kernel for f32; falls back to CPU round-trip for f64.

pub fn relu(&self) -> GpuResult<GpuTensor<T>>

Elementwise ReLU: out[i] = max(self[i], 0).

Uses a PTX kernel for f32; falls back to CPU round-trip for f64.

pub fn matmul(&self, other: &GpuTensor<T>) -> GpuResult<GpuTensor<T>>

Matrix multiplication: C = self @ other.

Both tensors must be 2-D. self has shape [m, k] and other has shape [k, n]. The result has shape [m, n].

Uses cuBLAS SGEMM for f32 and DGEMM for f64.

Errors

• GpuError::ShapeMismatch if either tensor is not 2-D or if the inner dimensions do not match (self.shape[1] != other.shape[0]).
• GpuError::DeviceMismatch if tensors are on different devices.
• GpuError::Blas on cuBLAS runtime errors.

pub fn conv2d( &self, weight: &GpuTensor<T>, bias: Option<&GpuTensor<T>>, stride: (usize, usize), padding: (usize, usize), ) -> GpuResult<GpuTensor<T>>

2-D convolution: output = conv2d(self, weight, bias).

Uses im2col (CPU) + cuBLAS GEMM (GPU) — no cuDNN required.

self must have shape [B, C_in, H, W] and weight must have shape [C_out, C_in, kH, kW]. bias, if provided, must have shape [C_out]. The result has shape [B, C_out, H_out, W_out].

Currently only supports f32. For f64 tensors, returns GpuError::ShapeMismatch (f64 conv path not yet implemented).

Errors

• GpuError::ShapeMismatch if tensor dimensions are wrong, channel counts don’t match, or if T is not f32.
• GpuError::DeviceMismatch if tensors are on different devices.
• GpuError::Blas on cuBLAS runtime errors.

Trait Implementations

impl<T: GpuFloat> Debug for GpuTensor<T>

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

Formats the value using the given formatter.