Struct BoolTensor 

pub struct BoolTensor { /* private fields */ }

Contiguous tensor of booleans, device-aware.

Data lives in a TensorStorage<bool> (CPU Vec<bool> or a CUDA GpuBufferHandle); the storage carries its own Device. CPU-resident behaviour is byte-identical to the pre-Phase-3a Arc<Vec<bool>> design.

Implementations

impl BoolTensor

pub fn from_vec( data: Vec<bool>, shape: Vec<usize>, ) -> Result<BoolTensor, FerrotorchError>

Build from a Vec + shape. Errors on numel mismatch.

pub fn from_slice( data: &[bool], shape: &[usize], ) -> Result<BoolTensor, FerrotorchError>

Build from a slice + shape (clones into a fresh Vec).

pub fn zeros(shape: &[usize]) -> BoolTensor

All-false tensor of the given shape.

pub fn ones(shape: &[usize]) -> BoolTensor

All-true tensor of the given shape.

pub fn from_predicate<T>( t: &Tensor<T>, pred: impl Fn(T) -> bool, ) -> Result<BoolTensor, FerrotorchError>

where T: Float,

Build a mask by applying pred to every element of t. Useful for Tensor < 0, Tensor.is_finite(), etc.

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

Logical shape.

pub fn numel(&self) -> usize

Total number of elements.

pub fn ndim(&self) -> usize

Number of dimensions.

pub fn device(&self) -> Device

The device this tensor’s storage resides on.

pub fn is_cuda(&self) -> bool

Returns true if this tensor is on a CUDA GPU.

pub fn data(&self) -> Result<&[bool], FerrotorchError>

Borrow the contiguous buffer as a host slice.

Errors

Returns FerrotorchError::GpuTensorNotAccessible if the tensor is on a GPU (CUDA / XPU) device, or is a meta tensor. This mirrors IntTensor::data and prevents callers from silently reading device memory as a host slice — call Self::to(Device::Cpu) first to transfer.

pub fn gpu_handle(&self) -> Result<&GpuBufferHandle, FerrotorchError>

Get the CUDA buffer handle. Returns Err for CPU (and other non-CUDA) tensors, mirroring Tensor::gpu_handle.

Errors

Returns FerrotorchError::InvalidArgument when the tensor is not CUDA-resident.

pub fn from_gpu_handle(handle: GpuBufferHandle, shape: Vec<usize>) -> BoolTensor

Construct a GPU-resident BoolTensor from a CUDA buffer handle + shape.

The handle must carry the DType::Bool tag — this is what every Phase-3b GPU op returns. Mirrors IntTensor::from_gpu_handle.

pub fn to(&self, device: Device) -> Result<BoolTensor, FerrotorchError>

Move this tensor to device, returning a new tensor.

Reuses the same DType-tagged raw-byte transport as Tensor::to / IntTensor::to: CPU→CUDA uploads via GpuBackend::cpu_to_gpu (handle tagged DType::Bool), CUDA→CPU reads back via gpu_to_cpu. Both are bit-exact. On-device→same-device is a cheap storage clone.

This is the explicit transfer entry point (PyTorch parity — like .cuda() / .cpu()). The host-readback inside the CUDA→CPU arm is the user-requested D2H copy, not a silent fallback.

Errors

• FerrotorchError::DeviceUnavailable if no GPU backend is registered.
• FerrotorchError::InvalidArgument for unsupported device pairs (XPU / MPS are not wired for BoolTensor).

pub fn not(&self) -> BoolTensor

Pointwise NOT. GPU path (resident) when CUDA-resident, else CPU.

pub fn and(&self, other: &BoolTensor) -> Result<BoolTensor, FerrotorchError>

Pointwise AND. Errors on shape/device mismatch.

pub fn or(&self, other: &BoolTensor) -> Result<BoolTensor, FerrotorchError>

Pointwise OR.

pub fn xor(&self, other: &BoolTensor) -> Result<BoolTensor, FerrotorchError>

Pointwise XOR.

pub fn reshape(&self, shape: &[usize]) -> Result<BoolTensor, FerrotorchError>

Reshape (must preserve numel). Metadata-only; the storage is cloned (cheap for CPU; a clone_buffer for GPU — no host readback).

pub fn count_true(&self) -> Result<usize, FerrotorchError>

Number of true elements (CPU only).

Errors

Returns FerrotorchError::GpuTensorNotAccessible for a CUDA-resident tensor: a host count would require a full-buffer D2H copy, which the device-error policy forbids. Use Self::any / Self::all (which run the reduction on-device) or .to(Device::Cpu) explicitly first.

pub fn any(&self) -> Result<bool, FerrotorchError>

True if any element is true.

On CUDA the OR-reduction runs on the GPU (real PTX kernel); only the single reduced byte crosses to the host — that byte IS the scalar result (PyTorch parity: Tensor.any() -> bool), NOT a full-buffer round trip.

pub fn all(&self) -> Result<bool, FerrotorchError>

True if all elements are true.

On CUDA the AND-reduction runs on the GPU; only the single reduced byte crosses to the host (the scalar result, not a buffer round trip).

pub fn gt<T>( a: &Tensor<T>, b: &Tensor<T>, ) -> Result<BoolTensor, FerrotorchError>

where T: Float,

Pointwise > comparing two float tensors of the same shape; produces a BoolTensor of matching shape. Mirrors torch.gt(a, b) returning a bool tensor. (#615)

pub fn lt<T>( a: &Tensor<T>, b: &Tensor<T>, ) -> Result<BoolTensor, FerrotorchError>

where T: Float,

Pointwise <. (#615)

pub fn ge<T>( a: &Tensor<T>, b: &Tensor<T>, ) -> Result<BoolTensor, FerrotorchError>

where T: Float,

Pointwise >=. (#615)

pub fn le<T>( a: &Tensor<T>, b: &Tensor<T>, ) -> Result<BoolTensor, FerrotorchError>

where T: Float,

Pointwise <=. (#615)

pub fn eq_t<T>( a: &Tensor<T>, b: &Tensor<T>, ) -> Result<BoolTensor, FerrotorchError>

where T: Float,

Pointwise ==. (#615)

pub fn ne<T>( a: &Tensor<T>, b: &Tensor<T>, ) -> Result<BoolTensor, FerrotorchError>

where T: Float,

Pointwise !=. (#615)

pub fn gt_int<I>( a: &IntTensor<I>, b: &IntTensor<I>, ) -> Result<BoolTensor, FerrotorchError>

where I: IntElement,

Pointwise a > b over two integer tensors → BoolTensor. (#1185)

pub fn lt_int<I>( a: &IntTensor<I>, b: &IntTensor<I>, ) -> Result<BoolTensor, FerrotorchError>

where I: IntElement,

Pointwise a < b over two integer tensors. (#1185)

pub fn ge_int<I>( a: &IntTensor<I>, b: &IntTensor<I>, ) -> Result<BoolTensor, FerrotorchError>

where I: IntElement,

Pointwise a >= b over two integer tensors. (#1185)

pub fn le_int<I>( a: &IntTensor<I>, b: &IntTensor<I>, ) -> Result<BoolTensor, FerrotorchError>

where I: IntElement,

Pointwise a <= b over two integer tensors. (#1185)

pub fn eq_int<I>( a: &IntTensor<I>, b: &IntTensor<I>, ) -> Result<BoolTensor, FerrotorchError>

where I: IntElement,

Pointwise a == b over two integer tensors. (#1185)

pub fn ne_int<I>( a: &IntTensor<I>, b: &IntTensor<I>, ) -> Result<BoolTensor, FerrotorchError>

where I: IntElement,

Pointwise a != b over two integer tensors. (#1185)

pub fn to_float<T>(&self) -> Result<Tensor<T>, FerrotorchError>

where T: Float,

Convert to a float tensor: true → 1.0, false → 0.0.

On CUDA the cast runs on the GPU and the result stays resident (no host round-trip); on CPU it maps the host slice.

Trait Implementations

impl Clone for BoolTensor

fn clone(&self) -> BoolTensor

Clone the tensor. For CPU storage this clones the Vec; for GPU storage it allocates a new device buffer with the same contents (via the backend’s clone_buffer). Delegates to TensorStorage::clone.

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

Performs copy-assignment from source.

impl Debug for BoolTensor

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

Formats the value using the given formatter.

impl Display for BoolTensor

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

Formats the value using the given formatter.