Opm::cuistl::GPUAwareMPISender< field_type, block_size, OwnerOverlapCopyCommunicationType > Class Template Reference

Derived class of GPUSender that handles MPI made with CUDA aware MPI The copOwnerToAll function uses MPI calls refering to data that resides on the GPU in order to send it directly to other GPUs, skipping the staging step on the CPU. More...

#include <CuOwnerOverlapCopy.hpp>

Inheritance diagram for Opm::cuistl::GPUAwareMPISender< field_type, block_size, OwnerOverlapCopyCommunicationType >:

Public Types
using	X = CuVector< field_type >

Public Member Functions
	GPUAwareMPISender (const OwnerOverlapCopyCommunicationType &cpuOwnerOverlapCopy)
void	copyOwnerToAll (const X &source, X &dest) const override
	copyOwnerToAll will copy source to the CPU, then call OwnerOverlapCopyCommunicationType::copyOwnerToAll on the copied data, and copy the result back to the GPU More...
void	project (X &x) const
	project will project x to the owned subspace More...
void	dot (const X &x, const X &y, field_type &output) const
	dot will carry out the dot product between x and y on the owned indices, then sum up the result across MPI processes. More...
field_type	norm (const X &x) const
	norm computes the l^2-norm of x across processes. More...

Protected Attributes
std::once_flag	m_initializedIndices
std::unique_ptr< CuVector< int > >	m_indicesOwner
std::unique_ptr< CuVector< int > >	m_indicesCopy
const OwnerOverlapCopyCommunicationType &	m_cpuOwnerOverlapCopy

Detailed Description

template<class field_type, int block_size, class OwnerOverlapCopyCommunicationType>
class Opm::cuistl::GPUAwareMPISender< field_type, block_size, OwnerOverlapCopyCommunicationType >

Derived class of GPUSender that handles MPI made with CUDA aware MPI The copOwnerToAll function uses MPI calls refering to data that resides on the GPU in order to send it directly to other GPUs, skipping the staging step on the CPU.

Template Parameters

field_type	is float or double
block_size	is the blocksize of the blockelements in the matrix
OwnerOverlapCopyCommunicationType	is typically a Dune::LinearOperator::communication_type

Member Typedef Documentation

X

template<class field_type , int block_size, class OwnerOverlapCopyCommunicationType >

using Opm::cuistl::GPUAwareMPISender< field_type, block_size, OwnerOverlapCopyCommunicationType >::X = CuVector<field_type>

Constructor & Destructor Documentation

GPUAwareMPISender()

template<class field_type , int block_size, class OwnerOverlapCopyCommunicationType >

Opm::cuistl::GPUAwareMPISender< field_type, block_size, OwnerOverlapCopyCommunicationType >::GPUAwareMPISender ( const OwnerOverlapCopyCommunicationType &  cpuOwnerOverlapCopy )

inline

Member Function Documentation

copyOwnerToAll()

template<class field_type , int block_size, class OwnerOverlapCopyCommunicationType >

void Opm::cuistl::GPUAwareMPISender< field_type, block_size, OwnerOverlapCopyCommunicationType >::copyOwnerToAll ( const X &  source, X &  dest  ) const

inlineoverridevirtual

copyOwnerToAll will copy source to the CPU, then call OwnerOverlapCopyCommunicationType::copyOwnerToAll on the copied data, and copy the result back to the GPU

Parameters

[in]	source
[out]	dest

Implements Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >.

References Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_cpuOwnerOverlapCopy, Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_initializedIndices, and Opm::cuistl::detail::to_int().

dot()

template<class field_type , class OwnerOverlapCopyCommunicationType >

void Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::dot ( const X &  x, const X &  y, field_type &  output  ) const

inlineinherited

dot will carry out the dot product between x and y on the owned indices, then sum up the result across MPI processes.

Parameters

[out]

output

result will be stored here

Note

This uses the same interface as its DUNE equivalent.

References Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::initIndexSet(), Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_cpuOwnerOverlapCopy, Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_indicesOwner, and Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_initializedIndices.

Referenced by Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::norm().

norm()

template<class field_type , class OwnerOverlapCopyCommunicationType >

field_type Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::norm ( const X &  x ) const

inlineinherited

norm computes the l^2-norm of x across processes.

This will compute the dot product of x with itself on owned indices, then sum the result across process and return the square root of the sum.

References Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::dot().

project()

template<class field_type , class OwnerOverlapCopyCommunicationType >

void Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::project ( X &  x ) const

inlineinherited

project will project x to the owned subspace

For each component i which is not owned, x_i will be set to 0

Parameters

[in,out]

x

the vector to project

References Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::initIndexSet(), Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_indicesCopy, and Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_initializedIndices.

Member Data Documentation

m_cpuOwnerOverlapCopy

template<class field_type , class OwnerOverlapCopyCommunicationType >

const OwnerOverlapCopyCommunicationType& Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_cpuOwnerOverlapCopy

protectedinherited

Referenced by Opm::cuistl::GPUObliviousMPISender< field_type, block_size, OwnerOverlapCopyCommunicationType >::copyOwnerToAll(), Opm::cuistl::GPUAwareMPISender< field_type, block_size, OwnerOverlapCopyCommunicationType >::copyOwnerToAll(), and Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::dot().

m_indicesCopy

template<class field_type , class OwnerOverlapCopyCommunicationType >

std::unique_ptr<CuVector<int> > Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_indicesCopy

mutableprotectedinherited

Referenced by Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::project().

m_indicesOwner

template<class field_type , class OwnerOverlapCopyCommunicationType >

std::unique_ptr<CuVector<int> > Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_indicesOwner

mutableprotectedinherited

Referenced by Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::dot().

m_initializedIndices

template<class field_type , class OwnerOverlapCopyCommunicationType >

std::once_flag Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::m_initializedIndices

mutableprotectedinherited

Referenced by Opm::cuistl::GPUAwareMPISender< field_type, block_size, OwnerOverlapCopyCommunicationType >::copyOwnerToAll(), Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::dot(), and Opm::cuistl::GPUSender< field_type, OwnerOverlapCopyCommunicationType >::project().

---

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

• CuOwnerOverlapCopy.hpp