Struct CudaBackendImpl 

pub struct CudaBackendImpl { /* private fields */ }

CUDA implementation of the GpuBackend trait.

Holds one or more GpuDevice handles (currently device 0 only) and delegates every trait method to the corresponding function in crate::kernels, crate::blas, or crate::transfer.

Implementations

impl CudaBackendImpl

pub fn new() -> FerrotorchResult<Self>

Create a new CUDA backend, initializing device 0.

Errors

Returns FerrotorchError::InvalidArgument if CUDA initialization fails (e.g. no GPU available, driver not loaded).

pub fn default_device(&self) -> FerrotorchResult<&Arc<GpuDevice>>

Get the device for ordinal 0 (the default device).

Trait Implementations

impl GpuBackend for CudaBackendImpl

fn as_any(&self) -> &dyn Any

Downcast to &dyn Any for backend-specific access (e.g., getting the underlying GpuDevice for CUDA graph capture).

fn cpu_to_gpu( &self, data: &[u8], elem_size: usize, device: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn cpu_to_gpu_pinned( &self, data: &[u8], elem_size: usize, device: usize, ) -> FerrotorchResult<GpuBufferHandle>

Copy CPU data to GPU via pinned (page-locked) host memory.

fn gpu_to_cpu(&self, handle: &GpuBufferHandle) -> FerrotorchResult<Vec<u8>>

fn clone_buffer( &self, handle: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn alloc_zeros( &self, len: usize, elem_size: usize, device: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn add_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn sub_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn mul_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn neg_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn relu_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn div_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn exp_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn log_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn sqrt_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn pow_f32( &self, a: &GpuBufferHandle, exponent: f32, ) -> FerrotorchResult<GpuBufferHandle>

fn abs_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn sigmoid_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn tanh_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn fused_adam_f32( &self, param: &mut GpuBufferHandle, grad: &GpuBufferHandle, exp_avg: &mut GpuBufferHandle, exp_avg_sq: &mut GpuBufferHandle, beta1: f32, beta2: f32, lr: f32, eps: f32, bc1: f32, bc2: f32, weight_decay: f32, ) -> FerrotorchResult<()>

Fused Adam optimizer step: updates param, exp_avg, and exp_avg_sq in a single kernel launch.

fn maxpool2d_f32( &self, input: &GpuBufferHandle, batch: usize, channels: usize, h_in: usize, w_in: usize, kh: usize, kw: usize, sh: usize, sw: usize, ph: usize, pw: usize, ) -> FerrotorchResult<(GpuBufferHandle, [usize; 4])>

GPU MaxPool2d forward.

fn avgpool2d_f32( &self, input: &GpuBufferHandle, batch: usize, channels: usize, h_in: usize, w_in: usize, kh: usize, kw: usize, sh: usize, sw: usize, ph: usize, pw: usize, ) -> FerrotorchResult<(GpuBufferHandle, [usize; 4])>

GPU AvgPool2d forward.

fn conv2d_f32( &self, input: &GpuBufferHandle, weight: &GpuBufferHandle, bias: Option<&GpuBufferHandle>, input_shape: [usize; 4], weight_shape: [usize; 4], stride: (usize, usize), padding: (usize, usize), ) -> FerrotorchResult<(GpuBufferHandle, [usize; 4])>

GPU Conv2d forward: im2col + GEMM + bias add, entirely on-device.

fn fused_gru_cell_f32( &self, input_gates: &GpuBufferHandle, hidden_gates: &GpuBufferHandle, bias_ih: &GpuBufferHandle, bias_hh: &GpuBufferHandle, hx: &GpuBufferHandle, hidden_size: usize, ) -> FerrotorchResult<(GpuBufferHandle, GpuBufferHandle)>

Fused GRU cell forward: pointwise gate computation on pre-computed gate matrices. Returns (hy_handle, workspace_handle).

fn synchronize(&self, device: usize) -> FerrotorchResult<()>

Synchronize the current stream on the given device, blocking until all enqueued operations have completed.

fn stream_count(&self, device: usize) -> usize

Return the number of streams in the pool for the given device.

fn matmul_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, m: usize, k: usize, n: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn sum_f32( &self, a: &GpuBufferHandle, _len: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn matmul_f64( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, m: usize, k: usize, n: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn broadcast_add_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, a_shape: &[usize], b_shape: &[usize], out_shape: &[usize], ) -> FerrotorchResult<GpuBufferHandle>

fn broadcast_sub_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, a_shape: &[usize], b_shape: &[usize], out_shape: &[usize], ) -> FerrotorchResult<GpuBufferHandle>

fn broadcast_mul_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, a_shape: &[usize], b_shape: &[usize], out_shape: &[usize], ) -> FerrotorchResult<GpuBufferHandle>

fn broadcast_div_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, a_shape: &[usize], b_shape: &[usize], out_shape: &[usize], ) -> FerrotorchResult<GpuBufferHandle>

fn softmax_f32( &self, a: &GpuBufferHandle, rows: usize, cols: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn dropout_f32( &self, a: &GpuBufferHandle, threshold: u32, scale: f32, seed: u32, ) -> FerrotorchResult<GpuBufferHandle>

fn dropout_philox_f32( &self, a: &GpuBufferHandle, threshold: u32, scale: f32, ) -> FerrotorchResult<(GpuBufferHandle, GpuRngState)>

Dropout using the Philox CBRNG for deterministic, reproducible mask generation.

fn transpose_2d_f32( &self, a: &GpuBufferHandle, m: usize, n: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn permute_0213_f32( &self, a: &GpuBufferHandle, d0: usize, d1: usize, d2: usize, d3: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn bmm_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, batch: usize, m: usize, k: usize, n: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn bmm_f16_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, batch: usize, m: usize, k: usize, n: usize, ) -> FerrotorchResult<GpuBufferHandle>

Batched matmul with f16 Tensor Core acceleration. Takes f32 handles, converts to f16 internally, accumulates in f32.

fn gelu_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn gelu_tanh_f32( &self, a: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn gelu_erf_f32(&self, a: &GpuBufferHandle) -> FerrotorchResult<GpuBufferHandle>

fn layernorm_f32( &self, input: &GpuBufferHandle, weight: &GpuBufferHandle, bias: &GpuBufferHandle, rows: usize, cols: usize, eps: f32, ) -> FerrotorchResult<GpuBufferHandle>

fn slice_write_f32( &self, src: &GpuBufferHandle, dst: &mut GpuBufferHandle, n_batch: usize, d: usize, max_len: usize, pos: usize, ) -> FerrotorchResult<()>

fn slice_read_f32( &self, src: &GpuBufferHandle, n_batch: usize, d: usize, len: usize, max_len: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn embed_lookup_f32( &self, idx: &GpuBufferHandle, weight: &GpuBufferHandle, d: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn embed_lookup_batch_f32( &self, indices: &GpuBufferHandle, weight: &GpuBufferHandle, n: usize, d: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn scatter_add_rows_f32( &self, grad_output: &GpuBufferHandle, indices: &GpuBufferHandle, num_embeddings: usize, d: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn scale_f32( &self, a: &GpuBufferHandle, scalar: f32, ) -> FerrotorchResult<GpuBufferHandle>

fn relu_backward_f32( &self, grad: &GpuBufferHandle, input: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn gelu_backward_f32( &self, grad: &GpuBufferHandle, input: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn gelu_backward_tanh_f32( &self, grad: &GpuBufferHandle, input: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn gelu_backward_erf_f32( &self, grad: &GpuBufferHandle, input: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn index_select_1d_f32( &self, input: &GpuBufferHandle, indices: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn scatter_add_1d_f32( &self, grad_output: &GpuBufferHandle, indices: &GpuBufferHandle, input_len: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn masked_fill_f32( &self, input: &GpuBufferHandle, mask: &GpuBufferHandle, value: f32, ) -> FerrotorchResult<GpuBufferHandle>

fn masked_zero_f32( &self, grad: &GpuBufferHandle, mask: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn sigmoid_backward_f32( &self, grad: &GpuBufferHandle, output: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn tanh_backward_f32( &self, grad: &GpuBufferHandle, output: &GpuBufferHandle, ) -> FerrotorchResult<GpuBufferHandle>

fn softmax_backward_f32( &self, grad: &GpuBufferHandle, output: &GpuBufferHandle, cols: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn layernorm_backward_f32( &self, input: &GpuBufferHandle, grad_output: &GpuBufferHandle, weight: &GpuBufferHandle, rows: usize, cols: usize, eps: f32, ) -> FerrotorchResult<(GpuBufferHandle, GpuBufferHandle, GpuBufferHandle)>

fn sum_axis_f32( &self, a: &GpuBufferHandle, shape: &[usize], axis: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn matmul_f16_f32( &self, a: &GpuBufferHandle, b: &GpuBufferHandle, m: usize, k: usize, n: usize, ) -> FerrotorchResult<GpuBufferHandle>

Mixed-precision matmul: cast f32 inputs to f16, multiply, accumulate back to f32. Used by autocast when the category is ReducedPrecision.

fn save_rng_state(&self, device: usize) -> FerrotorchResult<GpuRngState>

Save the current GPU RNG state for a device. Used by checkpoint to ensure dropout masks are identical on recomputation.

fn restore_rng_state(&self, state: GpuRngState) -> FerrotorchResult<()>

Restore a previously saved GPU RNG state for a device.

fn strided_split_f32( &self, input: &GpuBufferHandle, total_along_axis: usize, split_offset: usize, split_size: usize, inner_size: usize, n: usize, ) -> FerrotorchResult<GpuBufferHandle>

fn strided_cat_f32( &self, input: &GpuBufferHandle, output: &mut GpuBufferHandle, total_along_axis: usize, cat_offset: usize, part_size: usize, inner_size: usize, n: usize, ) -> FerrotorchResult<()>

fn add_f64( &self, _a: &GpuBufferHandle, _b: &GpuBufferHandle, ) -> Result<GpuBufferHandle, FerrotorchError>

fn sub_f64( &self, _a: &GpuBufferHandle, _b: &GpuBufferHandle, ) -> Result<GpuBufferHandle, FerrotorchError>

fn mul_f64( &self, _a: &GpuBufferHandle, _b: &GpuBufferHandle, ) -> Result<GpuBufferHandle, FerrotorchError>

fn neg_f64( &self, _a: &GpuBufferHandle, ) -> Result<GpuBufferHandle, FerrotorchError>

fn relu_f64( &self, _a: &GpuBufferHandle, ) -> Result<GpuBufferHandle, FerrotorchError>

fn sum_f64( &self, _a: &GpuBufferHandle, _numel: usize, ) -> Result<GpuBufferHandle, FerrotorchError>

fn has_inf_nan_f32(&self, a: &GpuBufferHandle) -> Result<bool, FerrotorchError>

Check if a GPU buffer contains any inf or NaN values.

fn svd_f32( &self, _a: &GpuBufferHandle, _m: usize, _n: usize, ) -> Result<(GpuBufferHandle, GpuBufferHandle, GpuBufferHandle), FerrotorchError>

fn cholesky_f32( &self, _a: &GpuBufferHandle, _n: usize, ) -> Result<GpuBufferHandle, FerrotorchError>

fn solve_f32( &self, _a: &GpuBufferHandle, _b: &GpuBufferHandle, _n: usize, _nrhs: usize, ) -> Result<GpuBufferHandle, FerrotorchError>

fn qr_f32( &self, _a: &GpuBufferHandle, _m: usize, _n: usize, ) -> Result<(GpuBufferHandle, GpuBufferHandle), FerrotorchError>