#include <tensor_impl.h>

Inheritance diagram for Tensor::Impl:

Public Member Functions
virtual void	fillAll (float value)=0
	Fills all elements with `value`.

virtual void	fillRand ()=0
	Fills all elements with random values in [-1, 1].

virtual void	print () const =0
	Prints the entire tensor to stdout.

virtual void	print (const std::vector< size_t > &position) const =0
	Prints the block starting at `position` to stdout.

virtual size_t	getNumElements () const =0
	Returns the total number of elements.

virtual Tensor::Shape	getShape () const =0
	Returns the tensor shape.

virtual std::vector< float >	toVector () const =0
	Copies all elements into a flat vector (row-major order).

virtual std::string	toString () const =0
	Returns a string representation of the data.

virtual std::unique_ptr< Tensor::Impl >	clone () const =0
	Deep copies this implementation.

virtual void	copyFromHost (const float *data, size_t count)=0

virtual void	set (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout)=0
	Copies data from `rhsImpl` with `rhsLayout` into `this` with `lhsLayout`.

virtual bool	compare (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout) const =0
	Returns true if the data with `lhsLayout` matches `rhsImpl` with `rhsLayout`.

virtual std::unique_ptr< Tensor::Impl >	add (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout) const =0
	Elementwise addition. Returns a new Impl with the result with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	sub (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout) const =0
	Elementwise subtraction. Returns a new Impl with the result with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	mul (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout) const =0
	Elementwise multiplication. Returns a new Impl with the result with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	div (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout) const =0
	Elementwise division. Returns a new Impl with the result with `outLayout`.

virtual void	iadd (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout)=0
	In-place elementwise addition. Modifies `lhsLayout` in place.

virtual void	isub (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout)=0
	In-place elementwise subtraction. Modifies `lhsLayout` in place.

virtual void	imul (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout)=0
	In-place elementwise multiplication. Modifies `lhsLayout` in place.

virtual void	idiv (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout)=0
	In-place elementwise division. Modifies `lhsLayout` in place.

virtual std::unique_ptr< Tensor::Impl >	sum (const TensorLayout &layout, const TensorLayout &blockLayout, const TensorLayout &outLayout) const =0
	Reduces dimensions [dim, rank) by summation. Output with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	min (const TensorLayout &layout, const TensorLayout &blockLayout, const TensorLayout &outLayout) const =0
	Reduces dimensions [dim, rank) by taking the minimum. Output with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	max (const TensorLayout &layout, const TensorLayout &blockLayout, const TensorLayout &outLayout) const =0
	Reduces dimensions [dim, rank) by taking the maximum. Output with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	prod (const TensorLayout &layout, const TensorLayout &blockLayout, const TensorLayout &outLayout) const =0
	Reduces dimensions [dim, rank) by taking the product. Output with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	norm (const TensorLayout &layout) const =0
	L2 norm of the tensor described by `layout`.

virtual std::unique_ptr< Tensor::Impl >	matmul (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout, size_t batch, size_t m, size_t k, size_t p) const =0

virtual std::unique_ptr< Tensor::Impl >	less (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout) const =0
	Elementwise less than. Returns a tensor of 0.0 / 1.0 with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	lessEqual (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout) const =0
	Elementwise less or equal. Returns a tensor of 0.0 / 1.0 with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	greater (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout) const =0
	Elementwise greater than. Returns a tensor of 0.0 / 1.0 with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	greaterEqual (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout) const =0
	Elementwise greater or equal. Returns a tensor of 0.0 / 1.0 with `outLayout`.

virtual std::unique_ptr< Tensor::Impl >	isClose (const TensorLayout &lhsLayout, const Tensor::Impl *rhsImpl, const TensorLayout &rhsLayout, const TensorLayout &outLayout, float tolerance) const =0
	Elementwise closeness within `tolerance`. Returns a tensor of 0.0 / 1.0 with `outLayout`.

Detailed Description

Abstract interface for backend specific tensor storage and operations.

All data access goes through TensorLayout descriptors, not raw indices.

The caller must ensure layouts are valid, since the implementation trusts that rhsLayout is correct for rhsImpl and lhsLayout is correct for its own memory.

The caller is responsible for broadcasting.

Member Function Documentation

◆ add()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::add	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Elementwise addition. Returns a new Impl with the result with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ clone()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::clone ( ) const

pure virtual

Deep copies this implementation.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ compare()

virtual bool Tensor::Impl::compare	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout
	)		const

pure virtual

Returns true if the data with lhsLayout matches rhsImpl with rhsLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ copyFromHost()

virtual void Tensor::Impl::copyFromHost	(	const float *	data,
		size_t	count
	)

pure virtual

Copies data from a host float array into this backend's storage.

Parameters

data	Source array (must have at least `count` elements).
count	Number of elements to copy.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ div()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::div	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Elementwise division. Returns a new Impl with the result with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ fillAll()

virtual void Tensor::Impl::fillAll ( float value )

pure virtual

Fills all elements with value.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ fillRand()

virtual void Tensor::Impl::fillRand ( )

pure virtual

Fills all elements with random values in [-1, 1].

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ getNumElements()

virtual size_t Tensor::Impl::getNumElements ( ) const

pure virtual

Returns the total number of elements.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ getShape()

virtual Tensor::Shape Tensor::Impl::getShape ( ) const

pure virtual

Returns the tensor shape.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ greater()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::greater	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Elementwise greater than. Returns a tensor of 0.0 / 1.0 with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ greaterEqual()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::greaterEqual	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Elementwise greater or equal. Returns a tensor of 0.0 / 1.0 with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ iadd()

virtual void Tensor::Impl::iadd	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout
	)

pure virtual

In-place elementwise addition. Modifies lhsLayout in place.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ idiv()

virtual void Tensor::Impl::idiv	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout
	)

pure virtual

In-place elementwise division. Modifies lhsLayout in place.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ imul()

virtual void Tensor::Impl::imul	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout
	)

pure virtual

In-place elementwise multiplication. Modifies lhsLayout in place.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ isClose()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::isClose	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout,
		float	tolerance
	)		const

pure virtual

Elementwise closeness within tolerance. Returns a tensor of 0.0 / 1.0 with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ isub()

virtual void Tensor::Impl::isub	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout
	)

pure virtual

In-place elementwise subtraction. Modifies lhsLayout in place.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ less()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::less	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Elementwise less than. Returns a tensor of 0.0 / 1.0 with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ lessEqual()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::lessEqual	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Elementwise less or equal. Returns a tensor of 0.0 / 1.0 with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ matmul()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::matmul	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout,
		size_t	batch,
		size_t	m,
		size_t	k,
		size_t	p
	)		const

pure virtual

Matrix multiplication. The last two dims of each layout are the matrix dims. batch, m, k, p describe the decomposition of the matmul problem.

2D: (m, k) @ (k, p) => (m, p).

3D: (batch, m, k) @ (batch, k, p) => (batch, m, p).

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ max()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::max	(	const TensorLayout &	layout,
		const TensorLayout &	blockLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Reduces dimensions [dim, rank) by taking the maximum. Output with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ min()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::min	(	const TensorLayout &	layout,
		const TensorLayout &	blockLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Reduces dimensions [dim, rank) by taking the minimum. Output with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ mul()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::mul	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Elementwise multiplication. Returns a new Impl with the result with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ norm()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::norm ( const TensorLayout & layout ) const

pure virtual

L2 norm of the tensor described by layout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ print() [1/2]

virtual void Tensor::Impl::print ( ) const

pure virtual

Prints the entire tensor to stdout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ print() [2/2]

virtual void Tensor::Impl::print ( const std::vector< size_t > & position ) const

pure virtual

Prints the block starting at position to stdout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ prod()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::prod	(	const TensorLayout &	layout,
		const TensorLayout &	blockLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Reduces dimensions [dim, rank) by taking the product. Output with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ set()

virtual void Tensor::Impl::set	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout
	)

pure virtual

Copies data from rhsImpl with rhsLayout into this with lhsLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ sub()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::sub	(	const TensorLayout &	lhsLayout,
		const Tensor::Impl *	rhsImpl,
		const TensorLayout &	rhsLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Elementwise subtraction. Returns a new Impl with the result with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ sum()

virtual std::unique_ptr< Tensor::Impl > Tensor::Impl::sum	(	const TensorLayout &	layout,
		const TensorLayout &	blockLayout,
		const TensorLayout &	outLayout
	)		const

pure virtual

Reduces dimensions [dim, rank) by summation. Output with outLayout.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ toString()

virtual std::string Tensor::Impl::toString ( ) const

pure virtual

Returns a string representation of the data.

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

◆ toVector()

virtual std::vector< float > Tensor::Impl::toVector ( ) const

pure virtual

Copies all elements into a flat vector (row-major order).

Implemented in Tensor::CPUImpl, and Tensor::CUDAImpl.

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

src/backend/tensor_impl.h

Public Member Functions

Detailed Description

Member Function Documentation

◆ add()

◆ clone()

◆ compare()

◆ copyFromHost()

◆ div()

◆ fillAll()

◆ fillRand()

◆ getNumElements()

◆ getShape()

◆ greater()

◆ greaterEqual()

◆ iadd()

◆ idiv()

◆ imul()

◆ isClose()

◆ isub()

◆ less()

◆ lessEqual()

◆ matmul()

◆ max()

◆ min()

◆ mul()

◆ norm()

◆ print() [1/2]

◆ print() [2/2]

◆ prod()

◆ set()

◆ sub()

◆ sum()

◆ toString()

◆ toVector()