A wrapper around a TensorLike that adds buffer ownership. More...

#include <TensorOwner.hpp>

Public Types
using	ValueType = typename TensorType::ValueType

using	Type = typename TensorType::Type

Public Member Functions
template<typename... Args>
	TensorOwner (sp::UniquePtr< Type > &&inBuffer, Args &&... args)
	Construct from a unique pointer, and take ownership of it. Behaviour is undefined if the buffer is too small. More...

	TensorOwner ()=default
	Creates an uninitialised TensorOwner. More...

	operator TensorType () const
	Implicitly convert to the underlying TensorLike. More...

auto	get () const
	Explicitly convert to the underlying TensorLike. More...

template<typename std::enable_if<!std::is_same_v< TensorType, typename TensorType::ConstTensor >, int >::type = 0>
	operator typename TensorType::ConstTensor () const
	Implicitly convert to the underlying TensorLike. More...

auto	data ()
	Allow access to the buffer (since a common use-case is initialisation). More...

auto	data () const

Static Public Attributes
constexpr static int	Rank = TensorType::Rank

Detailed Description

template<typename TensorType>
class sp::TensorOwner< TensorType >

A wrapper around a TensorLike that adds buffer ownership.

Typically, this is used indirectly via libSCALE's Device or Host objects. Something like:

auto gpu = sp::Device::get(0);
 
// Allocate a tensor on a specific GPU
auto myTensor = gpu.allocateTensorLike<sp::Tensor<__device float, 2>>({1, 2, 3});
 
// Works with any TensorLike...
auto myTriangular = gpu.allocateTensorLike<sp::TriangularMatrix<__device float>>(4);

Objects allocated in this way will automatically be unallocated when they go out of scope, and will implicitly convert to the underlying TensorLike type. Usually, you'll create some TensorOwner<T>s to own the memory, and pass them (with an implicit conversion) to functions/kernels that accept T to do the actual processing.