Collaboration diagram for dl::TensorImpl:

[legend]

Public Member Functions
TensorPtr	to (Device const &other) const noexcept
	Creates a copy of this tensor on the requested device and returns a pointer to it. If the new device is the same as the device the tensor is already on, the tensor will be copied.

Device const &	device () const noexcept
	Returns the device this tensor is stored on.

void	setRequiresGrad (bool requiresGrad) noexcept
	Set this tensors requirements for a gradient.

bool	requiresGrad () const noexcept
	Returns true iff this tensor requires a gradient, i.e., needs to be updated during backpropagation.

void	backward (bool enableAutodiff=false) noexcept

const TensorPtr &	gradient () const noexcept

void	discardGradient () noexcept

virtual std::ostream &	writeToStream (std::ostream &stream) const noexcept=0

virtual bool	operator== (const TensorPtr &other) const noexcept=0

virtual bool	allclose (const TensorPtr &other, float rtol=1e-5, float atol=1e-8) const noexcept=0

virtual TensorPtr	add (const TensorPtr &other) const noexcept=0

virtual TensorPtr	sub (const TensorPtr &other) const noexcept=0

virtual TensorPtr	mul (const TensorPtr &other) const noexcept=0

virtual TensorPtr	div (const TensorPtr &other) const noexcept=0

virtual TensorPtr	fma (const TensorPtr &factor, const TensorPtr &summand) const noexcept=0
	Performs "fused multiply and add".

virtual TensorPtr	matmul (const TensorPtr &other) const noexcept=0

virtual TensorPtr	transpose (std::vector< size_t > &&permutation) const noexcept=0

virtual TensorPtr	pow (float exponent) const noexcept=0

virtual TensorPtr	exp () const noexcept=0

virtual TensorPtr	log () const noexcept=0

virtual TensorPtr	sqrt () const noexcept=0

virtual TensorPtr	rsqrt () const noexcept=0

virtual TensorPtr	mean () const noexcept=0

virtual TensorPtr	mean (int dim, bool keepdim) const noexcept=0

virtual TensorPtr	sum () const noexcept=0

virtual TensorPtr	sum (int dim, bool keepdim) const noexcept=0

virtual TensorPtr	min () const noexcept=0

virtual TensorPtr	min (int dim, bool keepdim) const noexcept=0

virtual TensorPtr	min (const TensorPtr &other) const noexcept=0
	Computes the element-wise minimum between this tensor and the other.

virtual TensorPtr	max () const noexcept=0

virtual TensorPtr	max (int dim, bool keepdim) const noexcept=0

virtual TensorPtr	max (const TensorPtr &other) const noexcept=0

virtual TensorPtr	var (DOF dof) const noexcept=0

virtual TensorPtr	var (int dim, DOF dof) const noexcept=0

virtual TensorPtr	erf () const noexcept=0

virtual void	mul_inplace (const TensorPtr &other) noexcept=0

virtual void	reshape (SShape shape) noexcept=0
	Reshapes the tensor to fit the specified size.

virtual TensorPtr	clone () const noexcept=0

virtual Shape	shape () const noexcept=0

virtual size_t	shape (int dim) const noexcept=0

size_t	numDim () const noexcept

virtual TensorPtr	flatten () const noexcept=0

virtual size_t	toBytes (char *buffer, size_t buflen) const noexcept=0
	Writes this tensor's data into the byte array.

Public Attributes
GradFn	gradfn = nullptr

TensorPtr	grad = nullptr

Protected Member Functions
	TensorImpl (Device const &device, bool requiresGrad) noexcept

Detailed Description

Definition at line 13 of file tensorimpl.hpp.

Member Function Documentation

◆ device()

Device const & dl::TensorImpl::device ( ) const

noexcept

Returns the device this tensor is stored on.

Returns: The device this tensor is stored on.

See also: TensorImpl::to(const Device& other)

◆ discardGradient()

void dl::TensorImpl::discardGradient ( )

inlinenoexcept

Definition at line 63 of file tensorimpl.hpp.

                                        {
            gradfn = nullptr;
            grad = nullptr;
        }

◆ fma()

virtual TensorPtr dl::TensorImpl::fma	(	const TensorPtr &	factor,
		const TensorPtr &	summand
	)		const

pure virtualnoexcept

Performs "fused multiply and add".

Multiplies this tensor by factor and adds summand to the result. This specialized function exists since some devices (e.g., CUDA and some SIMD instruction sets) provide such a function.

Parameters

factor	the factor to multiply with this tensor.
summand	the summand to add to the product of this with `factor`.

Returns: the result.

◆ gradient()

const TensorPtr & dl::TensorImpl::gradient ( ) const

inlinenoexcept

Definition at line 62 of file tensorimpl.hpp.

62{ return grad; }

◆ min()

virtual TensorPtr dl::TensorImpl::min ( const TensorPtr & other ) const

pure virtualnoexcept

Computes the element-wise minimum between this tensor and the other.

Parameters

other

Returns: TensorPtr

◆ numDim()

size_t dl::TensorImpl::numDim ( ) const

inlinenoexcept

Definition at line 133 of file tensorimpl.hpp.

133{ return shape().size(); }

std::vector::size

T size(T... args)

◆ requiresGrad()

bool dl::TensorImpl::requiresGrad ( ) const

noexcept

Returns true iff this tensor requires a gradient, i.e., needs to be updated during backpropagation.

Returns: True iff this tensor requires a gradient, i.e., needs to be updated during backpropagation.

See also: TensorImpl::setRequiresGrad()

◆ reshape()

virtual void dl::TensorImpl::reshape ( SShape shape )

pure virtualnoexcept

Reshapes the tensor to fit the specified size.

At most one of the entries in the shape may be -1 and will then be inferred by the remaining shape. The element count may not change through reshaping.

Parameters

shape The new shape for the tensor.

◆ setRequiresGrad()

void dl::TensorImpl::setRequiresGrad ( bool requiresGrad )

noexcept

Set this tensors requirements for a gradient.

Parameters

requiresGrad wether the tensor requires a gradient.

See also: TensorImpl::requiresGrad()

◆ to()

TensorPtr dl::TensorImpl::to ( Device const & other ) const

noexcept

Creates a copy of this tensor on the requested device and returns a pointer to it. If the new device is the same as the device the tensor is already on, the tensor will be copied.

Parameters

other The device to copy the tensor onto.

Returns: The newly created tensor on the specified device.

See also: TensorImpl::device()

◆ toBytes()

virtual size_t dl::TensorImpl::toBytes	(	char *	buffer,
		size_t	buflen
	)		const

pure virtualnoexcept

Writes this tensor's data into the byte array.

Parameters

buffer	The byte to write to. If it is a nullpointer, no data will be written but the number of bytes that the tensor uses will still be returned.
buflen	The maximum number of bytes to write.

Returns: The number of bytes that were written. If the buffer was too small, 0 is returned.

Member Data Documentation

◆ grad

TensorPtr dl::TensorImpl::grad = nullptr

Definition at line 22 of file tensorimpl.hpp.

◆ gradfn

GradFn dl::TensorImpl::gradfn = nullptr

Definition at line 21 of file tensorimpl.hpp.

The documentation for this class was generated from the following file:

dl/tensor/tensorimpl.hpp

Public Member Functions

Public Attributes

Protected Member Functions

Detailed Description

Member Function Documentation

◆ device()

◆ discardGradient()

◆ fma()

◆ gradient()

◆ min()

◆ numDim()

◆ requiresGrad()

◆ reshape()

◆ setRequiresGrad()

◆ to()

◆ toBytes()

Member Data Documentation

◆ grad

◆ gradfn