Base class for linear operators. This class serves as interface for all derived classes. More...

#include <cu_linear_operator.h>

Inheritance diagram for cuLinearOperator< DataType >:

[legend]

Collaboration diagram for cuLinearOperator< DataType >:

Public Member Functions
	cuLinearOperator ()
	Default constructor.

	cuLinearOperator (const int num_gpu_devices_)
	Constructor with setting `num_rows` and `num_columns`.

virtual	~cuLinearOperator ()
	Destructor.

cublasHandle_t	get_cublas_handle () const
	This function returns a reference to the `cublasHandle_t` object. The object will be created, if it is not created already.

void	set_parameters (DataType *parameters_)
	Sets the scalar parameter `this->parameters`. Parameter is initialized to `NULL`. However, before calling `dot` or `transpose_dot` functions, the parameters must be set.

virtual DataType	get_eigenvalue (const DataType known_parameters, const DataType known_eigenvalue, const DataType inquiry_parameters) const =0

virtual void	dot (const DataType vector, DataType product)=0

virtual void	transpose_dot (const DataType vector, DataType product)=0

Public Member Functions inherited from cLinearOperatorBase
	cLinearOperatorBase ()
	Default constructor.

	cLinearOperatorBase (const LongIndexType num_rows_, const LongIndexType num_columns_)
	Constructor with setting `num_rows` and `num_columns`.

virtual	~cLinearOperatorBase ()
	Destructor.

LongIndexType	get_num_rows () const
	Returns the number of rows of the matrix.

LongIndexType	get_num_columns () const
	Returns the number of columns of the matrix.

IndexType	get_num_parameters () const
	Returns the number of parameters of the linear operator.

FlagType	is_eigenvalue_relation_known () const
	Returns a flag that determines whether a relation between the parameters of the operator and its eigenvalue(s) is known.

virtual void	set_symmetry (FlagType symmetric)=0

Protected Member Functions
int	query_gpu_devices () const
	Before any numerical computation, this method chechs if any gpu device is available on the machine, or notifies the user if nothing was found.

void	initialize_cublas_handle ()
	Creates a `cublasHandle_t` object, if not created already.

void	initialize_cusparse_handle ()
	Creates a `cusparseHandle_t` object, if not created already.

Protected Attributes
int	num_gpu_devices

bool	copied_host_to_device

cublasHandle_t *	cublas_handle

cusparseHandle_t *	cusparse_handle

DataType *	parameters

Protected Attributes inherited from cLinearOperatorBase
const LongIndexType	num_rows

const LongIndexType	num_columns

FlagType	eigenvalue_relation_known

IndexType	num_parameters

Detailed Description

template<typename DataType>
class cuLinearOperator< DataType >

Base class for linear operators. This class serves as interface for all derived classes.

The prefix c in this class's name (and its derivatves), stands for the cpp code, intrast to the cu prefix, which stands for the cuda code. Most derived classes have a cuda counterpart.

See also: cuMatrix, cuAffineMatrixFunction, cLinearOperator cLinearOperatorBase

Definition at line 63 of file cu_linear_operator.h.

Constructor & Destructor Documentation

◆ cuLinearOperator() [1/2]

template<typename DataType >

cuLinearOperator< DataType >::cuLinearOperator ( )

Default constructor.

Definition at line 40 of file cu_linear_operator.cu.

                                            :
 
    // Initializer list
    num_gpu_devices(0),
    copied_host_to_device(false),
    cublas_handle(NULL),
    cusparse_handle(NULL)
{
    // Check any gpu device exists
    this->num_gpu_devices = this->query_gpu_devices();
 
    // Regardless of using dense (cublas) or sparse (cusparse) matrices, the
    // cublas handle should be initialized, since it is needed for the methods
    // in cuVectorOperations
    this->initialize_cublas_handle();
}

References cuLinearOperator< DataType >::initialize_cublas_handle(), cuLinearOperator< DataType >::num_gpu_devices, and cuLinearOperator< DataType >::query_gpu_devices().

Here is the call graph for this function:

◆ cuLinearOperator() [2/2]

template<typename DataType >

cuLinearOperator< DataType >::cuLinearOperator ( const int num_gpu_devices_ )

explicit

Constructor with setting num_rows and num_columns.

Note: For the classed that are virtually derived (virtual inheritance) from this class, this constructor will never be called. Rather, the default constructor is called by the most derived class. Thus, set the member data directly instead of below.

Parameters

[in] num_gpu_devices_ Number of GPU devices to used for parallelization.

Definition at line 73 of file cu_linear_operator.cu.

                                   :
 
    // Initializer list
    num_gpu_devices(0),
    copied_host_to_device(false),
    cublas_handle(NULL),
    cusparse_handle(NULL)
{
    // Check any gpu device exists
    int device_count = this->query_gpu_devices();
 
    // Set number of gpu devices
    if (num_gpu_devices_ == 0)
    {
        this->num_gpu_devices = device_count;
    }
    else if (num_gpu_devices_ > device_count)
    {
        std::cerr << "ERROR: Number of requested gpu devices exceeds the " \
                  << "number of available gpu devices. Nummber of detected " \
                  << "devices are " << device_count << " while the " \
                  << "requested number of devices are " << num_gpu_devices_ \
                  << "." << std::endl;
        abort();
    }
    else
    {
        this->num_gpu_devices = num_gpu_devices_;
    }
 
    // Regardless of using dense (cublas) or sparse (cusparse) matrices, the
    // cublas handle should be initialized, since it is needed for the methods
    // in cuVectorOperations
    this->initialize_cublas_handle();
}

References cuLinearOperator< DataType >::initialize_cublas_handle(), cuLinearOperator< DataType >::num_gpu_devices, and cuLinearOperator< DataType >::query_gpu_devices().

Here is the call graph for this function:

◆ ~cuLinearOperator()

template<typename DataType >

cuLinearOperator< DataType >::~cuLinearOperator ( )

virtual

Destructor.

This function offloads data from GPUs.

Definition at line 120 of file cu_linear_operator.cu.

{
    // cublas handle
    if (this->cublas_handle != NULL)
    {
        // Set the number of threads
        #if defined(USE_OPENMP) && (USE_OPENMP == 1)
            omp_set_num_threads(this->num_gpu_devices);
        #endif
 
        #if defined(USE_OPENMP) && (USE_OPENMP == 1)
        #pragma omp parallel
        #endif
        {
            // Switch to a device with the same device id as the cpu thread id
            unsigned int thread_id;
            #if defined(USE_OPENMP) && (USE_OPENMP == 1)
                thread_id = omp_get_thread_num();
            #else
                thread_id = 0;
            #endif
 
            CudaAPI<DataType>::set_device(thread_id);
 
            cublasStatus_t status = cublasDestroy(
                    this->cublas_handle[thread_id]);
            assert(status == CUBLAS_STATUS_SUCCESS);
        }
 
        // Deallocate arrays of pointers on cpu
        delete[] this->cublas_handle;
        this->cublas_handle = NULL;
    }
 
    // cusparse handle
    if (this->cusparse_handle != NULL)
    {
        // Set the number of threads
        #if defined(USE_OPENMP) && (USE_OPENMP == 1)
            omp_set_num_threads(this->num_gpu_devices);
        #endif
 
        #if defined(USE_OPENMP) && (USE_OPENMP == 1)
        #pragma omp parallel
        #endif
        {
            // Switch to a device with the same device id as the cpu thread id
            unsigned int thread_id;
            #if defined(USE_OPENMP) && (USE_OPENMP == 1)
                thread_id = omp_get_thread_num();
            #else
                thread_id = 0;
            #endif
 
            CudaAPI<DataType>::set_device(thread_id);
 
            cusparseStatus_t status = cusparseDestroy(
                    this->cusparse_handle[thread_id]);
            assert(status == CUSPARSE_STATUS_SUCCESS);
        }
 
        // Deallocate arrays of pointers on cpu
        delete[] this->cusparse_handle;
        this->cusparse_handle = NULL;
    }
}

References cusparseDestroy(), omp_get_thread_num(), omp_set_num_threads(), and CudaAPI< ArrayType >::set_device().

Here is the call graph for this function:

Member Function Documentation

◆ dot()

template<typename DataType >

virtual void cuLinearOperator< DataType >::dot	(	const DataType *	vector,
		DataType *	product
	)

pure virtual

Implemented in cuCSCMatrix< DataType >, cuCSRMatrix< DataType >, cuDenseMatrix< DataType >, cuCSCAffineMatrixFunction< DataType >, cuCSRAffineMatrixFunction< DataType >, and cuDenseAffineMatrixFunction< DataType >.

Referenced by cu_golub_kahn_bidiagonalization(), and cu_lanczos_tridiagonalization().

Here is the caller graph for this function:

◆ get_cublas_handle()

template<typename DataType >

cublasHandle_t cuLinearOperator< DataType >::get_cublas_handle ( ) const

This function returns a reference to the cublasHandle_t object. The object will be created, if it is not created already.

The cublasHandle is needed for the client code (slq method) for vector operations on GPU. However, in this class, the cublasHandle_t might not be needed by it self if the derived class is a sparse matrix, becase the sparse matrix needs only cusparseHandle_t. In case if the cublasHandle_t is not created, it will be created for the purpose of the client codes.

Returns: A void pointer to the cublasHandle_t instance.

Definition at line 205 of file cu_linear_operator.cu.

{
    // Get device id
    int device_id = CudaAPI<DataType>::get_device();
 
    return this->cublas_handle[device_id];
}

References CudaAPI< ArrayType >::get_device().

Referenced by cu_golub_kahn_bidiagonalization(), and cu_lanczos_tridiagonalization().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ get_eigenvalue()

template<typename DataType >

virtual DataType cuLinearOperator< DataType >::get_eigenvalue	(	const DataType *	known_parameters,
		const DataType	known_eigenvalue,
		const DataType *	inquiry_parameters
	)		const

pure virtual

Implemented in cuAffineMatrixFunction< DataType >, and cuMatrix< DataType >.

Referenced by cuTraceEstimator< DataType >::_cu_stochastic_lanczos_quadrature().

Here is the caller graph for this function:

◆ initialize_cublas_handle()

template<typename DataType >

void cuLinearOperator< DataType >::initialize_cublas_handle ( )

protected

Creates a cublasHandle_t object, if not created already.

Definition at line 222 of file cu_linear_operator.cu.

{
    if (this->cublas_handle == NULL)
    {
        // Allocate pointers for each gpu device
        this->cublas_handle = new cublasHandle_t[this->num_gpu_devices];
 
        // Set the number of threads
        #if defined(USE_OPENMP) && (USE_OPENMP == 1)
            omp_set_num_threads(this->num_gpu_devices);
        #endif
 
        #if defined(USE_OPENMP) && (USE_OPENMP == 1)
        #pragma omp parallel
        #endif
        {
            // Switch to a device with the same device id as the cpu thread id
            unsigned int thread_id;
            #if defined(USE_OPENMP) && (USE_OPENMP == 1)
                thread_id = omp_get_thread_num();
            #else
                thread_id = 0;
            #endif
 
            CudaAPI<DataType>::set_device(thread_id);
 
            cublasStatus_t status_create = cublasCreate(
                    &this->cublas_handle[thread_id]);
            assert(status_create == CUBLAS_STATUS_SUCCESS);
 
            // Set tensor core whenever possible (usually for cublasXgemm)
            cublasStatus_t status_set = cublasSetMathMode(
                    this->cublas_handle[thread_id], CUBLAS_TENSOR_OP_MATH);
            assert(status_set == CUBLAS_STATUS_SUCCESS);
        }
    }
}

References cublasSetMathMode(), omp_get_thread_num(), omp_set_num_threads(), and CudaAPI< ArrayType >::set_device().

Referenced by cuDenseAffineMatrixFunction< DataType >::cuDenseAffineMatrixFunction(), cuDenseAffineMatrixFunction< DataType >::cuDenseAffineMatrixFunction(), cuDenseMatrix< DataType >::cuDenseMatrix(), cuLinearOperator< DataType >::cuLinearOperator(), and cuLinearOperator< DataType >::cuLinearOperator().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ initialize_cusparse_handle()

template<typename DataType >

void cuLinearOperator< DataType >::initialize_cusparse_handle ( )

protected

Creates a cusparseHandle_t object, if not created already.

Definition at line 269 of file cu_linear_operator.cu.

{
    if (this->cusparse_handle == NULL)
    {
        // Allocate pointers for each gpu device
        this->cusparse_handle = new cusparseHandle_t[this->num_gpu_devices];
 
        // Set the number of threads
        #if defined(USE_OPENMP) && (USE_OPENMP == 1)
            omp_set_num_threads(this->num_gpu_devices);
        #endif
 
        #if defined(USE_OPENMP) && (USE_OPENMP == 1)
        #pragma omp parallel
        #endif
        {
            // Switch to a device with the same device id as the cpu thread id
            unsigned int thread_id;
            #if defined(USE_OPENMP) && (USE_OPENMP == 1)
                thread_id = omp_get_thread_num();
            #else
                thread_id = 0;
            #endif
 
            CudaAPI<DataType>::set_device(thread_id);
 
            cusparseStatus_t status = cusparseCreate(
                    &this->cusparse_handle[thread_id]);
            assert(status == CUSPARSE_STATUS_SUCCESS);
        }
    }
}

References cusparseCreate(), omp_get_thread_num(), omp_set_num_threads(), and CudaAPI< ArrayType >::set_device().

Referenced by cuCSCAffineMatrixFunction< DataType >::cuCSCAffineMatrixFunction(), cuCSCAffineMatrixFunction< DataType >::cuCSCAffineMatrixFunction(), cuCSCMatrix< DataType >::cuCSCMatrix(), cuCSRAffineMatrixFunction< DataType >::cuCSRAffineMatrixFunction(), cuCSRAffineMatrixFunction< DataType >::cuCSRAffineMatrixFunction(), and cuCSRMatrix< DataType >::cuCSRMatrix().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ query_gpu_devices()

template<typename DataType >

int cuLinearOperator< DataType >::query_gpu_devices ( ) const

protected

Before any numerical computation, this method chechs if any gpu device is available on the machine, or notifies the user if nothing was found.

Returns: Number of gpu available devices.

Definition at line 314 of file cu_linear_operator.cu.

{
    int device_count = 0;
    cudaError_t error = cudaGetDeviceCount(&device_count);
 
    // Error code 38 means no cuda-capable device was detected.
    if ((error != cudaSuccess) || (device_count < 1))
    {
        std::cerr << "ERROR: No cuda-capable GPU device was detected on " \
                  << "this machine. If a cuda-capable GPU device exists, " \
                  << "install its cuda driver. Alternatively, set " \
                  << "'gpu=False' to use cpu instead." \
                  << std::endl;
 
        abort();
    }
 
    return device_count;
}

References cudaGetDeviceCount().

Referenced by cuLinearOperator< DataType >::cuLinearOperator(), and cuLinearOperator< DataType >::cuLinearOperator().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ set_parameters()

template<typename DataType >

void cuLinearOperator< DataType >::set_parameters ( DataType * parameters_ )

Sets the scalar parameter this->parameters. Parameter is initialized to NULL. However, before calling dot or transpose_dot functions, the parameters must be set.

Parameters

[in] parameters_ A pointer to the scalar or array of parameters.

Definition at line 347 of file cu_linear_operator.cu.

{
    this->parameters = parameters_;
}

Referenced by cuTraceEstimator< DataType >::_cu_stochastic_lanczos_quadrature().

Here is the caller graph for this function:

◆ transpose_dot()

template<typename DataType >

virtual void cuLinearOperator< DataType >::transpose_dot	(	const DataType *	vector,
		DataType *	product
	)

pure virtual

Implemented in cuCSCMatrix< DataType >, cuCSRMatrix< DataType >, cuDenseMatrix< DataType >, cuCSCAffineMatrixFunction< DataType >, cuCSRAffineMatrixFunction< DataType >, and cuDenseAffineMatrixFunction< DataType >.

Referenced by cu_golub_kahn_bidiagonalization().

Here is the caller graph for this function:

Member Data Documentation

◆ copied_host_to_device

template<typename DataType >

bool cuLinearOperator< DataType >::copied_host_to_device

protected

Definition at line 100 of file cu_linear_operator.h.

◆ cublas_handle

template<typename DataType >

cublasHandle_t* cuLinearOperator< DataType >::cublas_handle

protected

Definition at line 101 of file cu_linear_operator.h.

◆ cusparse_handle

template<typename DataType >

cusparseHandle_t* cuLinearOperator< DataType >::cusparse_handle

protected

Definition at line 102 of file cu_linear_operator.h.

◆ num_gpu_devices

template<typename DataType >

int cuLinearOperator< DataType >::num_gpu_devices

protected

Definition at line 99 of file cu_linear_operator.h.

Referenced by cuCSCMatrix< DataType >::cuCSCMatrix(), cuCSRMatrix< DataType >::cuCSRMatrix(), cuLinearOperator< DataType >::cuLinearOperator(), and cuLinearOperator< DataType >::cuLinearOperator().

◆ parameters

template<typename DataType >

DataType* cuLinearOperator< DataType >::parameters

protected

Definition at line 103 of file cu_linear_operator.h.

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

/home/runner/work/imate/imate/imate/_cu_linear_operator/cu_linear_operator.h
/home/runner/work/imate/imate/imate/_cu_linear_operator/cu_linear_operator.cu

Public Member Functions

Protected Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ cuLinearOperator() [1/2]

◆ cuLinearOperator() [2/2]

◆ ~cuLinearOperator()

Member Function Documentation

◆ dot()

◆ get_cublas_handle()

◆ get_eigenvalue()

◆ initialize_cublas_handle()

◆ initialize_cusparse_handle()

◆ query_gpu_devices()

◆ set_parameters()

◆ transpose_dot()

Member Data Documentation

◆ copied_host_to_device

◆ cublas_handle

◆ cusparse_handle

◆ num_gpu_devices

◆ parameters