CudaInterface< ArrayType > Class Template Reference

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_interface.h>

Static Public Member Functions
static ArrayType *	alloc (const LongIndexType array_size)
	Allocates memory on gpu device. This function creates a pointer and returns it. More...
static void	alloc (ArrayType *&device_array, const LongIndexType array_size)
	Allocates memory on gpu device. This function uses an existing given pointer. More...
static void	alloc_bytes (void *&device_array, const size_t num_bytes)
	Allocates memory on gpu device. This function uses an existing given pointer. More...
static void	copy_to_device (const ArrayType *host_array, const LongIndexType array_size, ArrayType *device_array)
	Copies memory on host to device memory. More...
static void	del (void *device_array)
	Deletes memory on gpu device if its pointer is not NULL, then sets the pointer to NULL. More...
static void	set_device (int device_id)
	Sets the current device in multi-gpu applications. More...
static int	get_device ()
	Gets the current device in multi-gpu applications. More...

Detailed Description

template<typename ArrayType>
class CudaInterface< 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 36 of file cuda_interface.h.

Member Function Documentation

alloc() [1/2]

template<typename ArrayType >

void CudaInterface< ArrayType >::alloc ( ArrayType *&  device_array, const LongIndexType  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 76 of file cuda_interface.cu.

{
    // Check if overflowing might make array_size negative if LongIndexType 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 = static_cast<size_t>(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 * CudaInterface< ArrayType >::alloc ( const LongIndexType  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 36 of file cuda_interface.cu.

{
    // Check if overflowing might make array_size negative if LongIndexType 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 = static_cast<size_t>(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 CudaInterface< 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 115 of file cuda_interface.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 CudaInterface< ArrayType >::copy_to_device ( const ArrayType *  host_array, const LongIndexType  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 142 of file cuda_interface.cu.

{
    size_t num_bytes = static_cast<size_t>(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 CudaInterface< 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 166 of file cuda_interface.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 CudaInterface< 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 206 of file cuda_interface.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 CudaInterface< 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 188 of file cuda_interface.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_interface.h
• /home/runner/work/imate/imate/imate/_cuda_utilities/cuda_interface.cu