An interface to CUDA linrary to facilitate working with CUDA, such as memory allocation, copy data to and from device, etc. This class contains all public static functions and serves as a namespace. More...

#include <cuda_api.h>

Static Public Member Functions
static ArrayType *	alloc (const size_t array_size)
	Allocates memory on gpu device. This function creates a pointer and returns it.

static void	alloc (ArrayType *&device_array, const size_t array_size)
	Allocates memory on gpu device. This function uses an existing given pointer.

static void	alloc_bytes (void *&device_array, const size_t num_bytes)
	Allocates memory on gpu device. This function uses an existing given pointer.

static void	copy_to_device (const ArrayType host_array, const size_t array_size, ArrayType device_array)
	Copies memory on host to device memory.

static void	del (void *device_array)
	Deletes memory on gpu device if its pointer is not `NULL`, then sets the pointer to `NULL`.

static void	set_device (int device_id)
	Sets the current device in multi-gpu applications.

static int	get_device ()
	Gets the current device in multi-gpu applications.

Detailed Description

template<typename ArrayType>
class CudaAPI< ArrayType >

An interface to CUDA linrary to facilitate working with CUDA, such as memory allocation, copy data to and from device, etc. This class contains all public static functions and serves as a namespace.

Definition at line 35 of file cuda_api.h.

Member Function Documentation

◆ alloc() [1/2]

template<typename ArrayType >

void CudaAPI< ArrayType >::alloc	(	ArrayType *&	device_array,
		const size_t	array_size
	)

static

Allocates memory on gpu device. This function uses an existing given pointer.

Parameters

[in,out]	device_array	A pointer to the device memory to be allocated
[in]	array_size	Size of the array to be allocated.

Definition at line 79 of file cuda_api.cu.

{
    // Check if overflowing might make array_size negative if array_size is
    // a signed type. For unsigned type, we have no clue at this point.
    assert(array_size > 0);
 
    // Check if computing num_bytes will not overflow size_t (unsigned int)
    size_t max_index = std::numeric_limits<size_t>::max();
    if (max_index / sizeof(ArrayType) < array_size)
    {
        std::cerr << "The size of array in bytes exceeds the maximum " \
                  << "integer limit, which is: " << max_index << ". The " \
                  << "array size is: " << array_size << ", and the size of " \
                  << "data type is: " << sizeof(ArrayType) << "-bytes." \
                  << std::endl;
        abort();
    }
 
    size_t num_bytes = array_size * sizeof(ArrayType);
    cudaError_t error = cudaMalloc(&device_array, num_bytes);
    assert(error == cudaSuccess);
}

References cudaMalloc().

Here is the call graph for this function:

◆ alloc() [2/2]

template<typename ArrayType >

ArrayType * CudaAPI< ArrayType >::alloc ( const size_t array_size )

static

Allocates memory on gpu device. This function creates a pointer and returns it.

Parameters

[in] array_size Size of the array to be allocated.

Returns: A pointer to the allocated 1D array on device.

Definition at line 39 of file cuda_api.cu.

{
    // Check if overflowing might make array_size negative if array_size is
    // a signed type. For unsigned type, we have no clue at this point.
    assert(array_size > 0);
 
    // Check if computing num_bytes will not overflow size_t (unsigned int)
    size_t max_index = std::numeric_limits<size_t>::max();
    if (max_index / sizeof(ArrayType) < array_size)
    {
        std::cerr << "The size of array in bytes exceeds the maximum " \
                  << "integer limit, which is: " << max_index << ". The " \
                  << "array size is: " << array_size << ", and the size of " \
                  << "data type is: " << sizeof(ArrayType) << "-bytes." \
                  << std::endl;
        abort();
    }
 
    ArrayType* device_array;
    size_t num_bytes = array_size * sizeof(ArrayType);
    cudaError_t error = cudaMalloc(&device_array, num_bytes);
    assert(error == cudaSuccess);
 
    return device_array;
}

References cudaMalloc().

Referenced by cuCSCMatrix< DataType >::copy_host_to_device(), cuCSRMatrix< DataType >::copy_host_to_device(), cuDenseMatrix< DataType >::copy_host_to_device(), cu_golub_kahn_bidiagonalization(), and cu_lanczos_tridiagonalization().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ alloc_bytes()

template<typename ArrayType >

void CudaAPI< ArrayType >::alloc_bytes	(	void *&	device_array,
		const size_t	num_bytes
	)

static

Allocates memory on gpu device. This function uses an existing given pointer.

Parameters

[in,out]	device_array	A pointer to the device memory to be allocated
[in]	num_bytes	Number of bytes of the array to be allocated.

Definition at line 118 of file cuda_api.cu.

{
    // Check if overflowing might make num_bytes negative if size_t is
    // a signed type. For unsigned type, we have no clue at this point.
    assert(num_bytes > 0);
 
    cudaError_t error = cudaMalloc(&device_array, num_bytes);
    assert(error == cudaSuccess);
}

References cudaMalloc().

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

Here is the call graph for this function:

Here is the caller graph for this function:

◆ copy_to_device()

template<typename ArrayType >

void CudaAPI< ArrayType >::copy_to_device	(	const ArrayType *	host_array,
		const size_t	array_size,
		ArrayType *	device_array
	)

static

Copies memory on host to device memory.

Parameters

[in]	host_array	Pointer of 1D array memory on host
[in]	array_size	The size of array on host.
[out]	device_array	Pointer to the destination memory on device.

Definition at line 145 of file cuda_api.cu.

{
    size_t num_bytes = array_size * sizeof(ArrayType);
    cudaError_t error = cudaMemcpy(device_array, host_array, num_bytes,
                                   cudaMemcpyHostToDevice);
    assert(error == cudaSuccess);
}

References cudaMemcpy().

Referenced by cuCSCMatrix< DataType >::copy_host_to_device(), cuCSRMatrix< DataType >::copy_host_to_device(), cuDenseMatrix< DataType >::copy_host_to_device(), cu_golub_kahn_bidiagonalization(), cu_lanczos_tridiagonalization(), and cuOrthogonalization< DataType >::orthogonalize_vectors().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ del()

template<typename ArrayType >

void CudaAPI< ArrayType >::del ( void * device_array )

static

Deletes memory on gpu device if its pointer is not NULL, then sets the pointer to NULL.

Parameters

[in,out] device_array A pointer to memory on device to be deleted. This pointer will be set to NULL.

Definition at line 169 of file cuda_api.cu.

{
    if (device_array != NULL)
    {
        cudaError_t error = cudaFree(device_array);
        assert(error == cudaSuccess);
        device_array = NULL;
    }
}

References cudaFree().

Referenced by cuCSCMatrix< DataType >::allocate_buffer(), cuCSRMatrix< DataType >::allocate_buffer(), cu_golub_kahn_bidiagonalization(), cu_lanczos_tridiagonalization(), cuCSCMatrix< DataType >::~cuCSCMatrix(), cuCSRMatrix< DataType >::~cuCSRMatrix(), and cuDenseMatrix< DataType >::~cuDenseMatrix().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ get_device()

template<typename ArrayType >

int CudaAPI< ArrayType >::get_device ( )

static

Gets the current device in multi-gpu applications.

Returns: device_id The id of the current device. The id is a number from 0 to num_gpu_devices-1

Definition at line 209 of file cuda_api.cu.

{
    int device_id = -1;
    cudaError_t error = cudaGetDevice(&device_id);
    assert(error == cudaSuccess);
 
    return device_id;
}

References cudaGetDevice().

Referenced by cuAffineMatrixFunction< DataType >::_add_scaled_vector(), cuCSCMatrix< DataType >::dot(), cuCSRMatrix< DataType >::dot(), cuDenseMatrix< DataType >::dot(), cuCSCMatrix< DataType >::dot_plus(), cuCSRMatrix< DataType >::dot_plus(), cuDenseMatrix< DataType >::dot_plus(), cuLinearOperator< DataType >::get_cublas_handle(), cuCSCMatrix< DataType >::transpose_dot(), cuCSRMatrix< DataType >::transpose_dot(), cuDenseMatrix< DataType >::transpose_dot(), cuCSCMatrix< DataType >::transpose_dot_plus(), cuCSRMatrix< DataType >::transpose_dot_plus(), and cuDenseMatrix< DataType >::transpose_dot_plus().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ set_device()

template<typename ArrayType >

void CudaAPI< ArrayType >::set_device ( int device_id )

static

Sets the current device in multi-gpu applications.

Parameters

[in] device_id The id of the device to switch to. The id is a number from 0 to num_gpu_devices-1

Definition at line 191 of file cuda_api.cu.

{
    cudaError_t error = cudaSetDevice(device_id);
    assert(error == cudaSuccess);
}

References cudaSetDevice().

Referenced by cuCSCMatrix< DataType >::copy_host_to_device(), cuCSRMatrix< DataType >::copy_host_to_device(), cuDenseMatrix< DataType >::copy_host_to_device(), cuTraceEstimator< DataType >::cu_trace_estimator(), cuLinearOperator< DataType >::initialize_cublas_handle(), cuLinearOperator< DataType >::initialize_cusparse_handle(), cuCSCMatrix< DataType >::~cuCSCMatrix(), cuCSRMatrix< DataType >::~cuCSRMatrix(), cuDenseMatrix< DataType >::~cuDenseMatrix(), and cuLinearOperator< DataType >::~cuLinearOperator().

Here is the call graph for this function:

Here is the caller graph for this function:

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

/home/runner/work/imate/imate/imate/_cuda_utilities/cuda_api.h
/home/runner/work/imate/imate/imate/_cuda_utilities/cuda_api.cu

Static Public Member Functions

Detailed Description

Member Function Documentation

◆ alloc() [1/2]

◆ alloc() [2/2]

◆ alloc_bytes()

◆ copy_to_device()

◆ del()

◆ get_device()

◆ set_device()