#include <tensor_impl_CUDA.h>

Inheritance diagram for Tensor::CUDAImpl:

Public Member Functions
	CUDAImpl (const Tensor::Shape &shape)

void	fillAll (float value) override
	Fills all elements with `value`.

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

void	print () const override
	Prints the entire tensor to stdout.

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

size_t	getNumElements () const override
	Returns the total number of elements.

Tensor::Shape	getShape () const override
	Returns the tensor shape.

float *	dataPtr () const
	Returns a raw pointer to the device data buffer.

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

std::string	toString () const override
	Returns a string representation of the data.

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

void	copyFromHost (const float *data, size_t count) override

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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 override

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

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

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

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

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

Detailed Description

CUDA implementation of Tensor::Impl, backed by device memory.

All operations launch CUDA kernels using TensorLayout descriptors. The caller is responsible for layout validity, see Tensor::Impl.

Overridden methods follow the same semantics documented in Tensor::Impl.

Member Function Documentation

◆ add()

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

overridevirtual

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

Implements Tensor::Impl.

◆ clone()

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

overridevirtual

Deep copies this implementation.

Implements Tensor::Impl.

◆ compare()

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

overridevirtual

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

Implements Tensor::Impl.

◆ copyFromHost()

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

overridevirtual

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.

Implements Tensor::Impl.

◆ div()

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

overridevirtual

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

Implements Tensor::Impl.

◆ fillAll()

void Tensor::CUDAImpl::fillAll ( float value )

overridevirtual

Fills all elements with value.

Implements Tensor::Impl.

◆ fillRand()

void Tensor::CUDAImpl::fillRand ( )

overridevirtual

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

Implements Tensor::Impl.

◆ getNumElements()

size_t Tensor::CUDAImpl::getNumElements ( ) const

overridevirtual

Returns the total number of elements.

Implements Tensor::Impl.

◆ getShape()

Tensor::Shape Tensor::CUDAImpl::getShape ( ) const

overridevirtual

Returns the tensor shape.

Implements Tensor::Impl.

◆ greater()

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

overridevirtual

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

Implements Tensor::Impl.

◆ greaterEqual()

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

overridevirtual

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

Implements Tensor::Impl.

◆ iadd()

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

overridevirtual

In-place elementwise addition. Modifies lhsLayout in place.

Implements Tensor::Impl.

◆ idiv()

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

overridevirtual

In-place elementwise division. Modifies lhsLayout in place.

Implements Tensor::Impl.

◆ imul()

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

overridevirtual

In-place elementwise multiplication. Modifies lhsLayout in place.

Implements Tensor::Impl.

◆ isClose()

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

overridevirtual

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

Implements Tensor::Impl.

◆ isub()

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

overridevirtual

In-place elementwise subtraction. Modifies lhsLayout in place.

Implements Tensor::Impl.

◆ less()

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

overridevirtual

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

Implements Tensor::Impl.

◆ lessEqual()

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

overridevirtual

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

Implements Tensor::Impl.

◆ matmul()

std::unique_ptr< Tensor::Impl > Tensor::CUDAImpl::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

overridevirtual

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).

Implements Tensor::Impl.

◆ max()

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

overridevirtual

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

Implements Tensor::Impl.

◆ min()

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

overridevirtual

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

Implements Tensor::Impl.

◆ mul()

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

overridevirtual

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

Implements Tensor::Impl.

◆ norm()

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

overridevirtual

L2 norm of the tensor described by layout.

Implements Tensor::Impl.

◆ print() [1/2]

void Tensor::CUDAImpl::print ( ) const

overridevirtual

Prints the entire tensor to stdout.

Implements Tensor::Impl.

◆ print() [2/2]

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

overridevirtual

Prints the block starting at position to stdout.

Implements Tensor::Impl.

◆ prod()

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

overridevirtual

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

Implements Tensor::Impl.

◆ set()

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

overridevirtual

Copies data from rhsImpl with rhsLayout into this with lhsLayout.

Implements Tensor::Impl.

◆ sub()

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

overridevirtual

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

Implements Tensor::Impl.

◆ sum()

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

overridevirtual

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

Implements Tensor::Impl.

◆ toString()

std::string Tensor::CUDAImpl::toString ( ) const

overridevirtual

Returns a string representation of the data.

Implements Tensor::Impl.

◆ toVector()

std::vector< float > Tensor::CUDAImpl::toVector ( ) const

overridevirtual

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

Implements Tensor::Impl.

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

src/backend/cuda/tensor_impl_CUDA.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()