#include "ghost/config.h"
#include "ghost/types.h"
#include "ghost/core.h"
#include "ghost/util.h"
#include "ghost/sparsemat.h"
#include "ghost/locality.h"
#include "ghost/log.h"
#include "ghost/rand.h"
#include <string.h>
#include <stdlib.h>
#include <libgen.h>
#include <unistd.h>

Include dependency graph for cu_util.c:

Macros
#define	ghost_cu_MAX_DEVICE_NAME_LEN 500

Functions
ghost_error	ghost_cu_init (int dev)
	Initalize CUDA on a given device. More...

ghost_error	ghost_cu_malloc_managed (void **mem, size_t bytesize)
	Allocate managed CUDA device memory. More...

ghost_error	ghost_cu_malloc (void **mem, size_t bytesize)
	Allocate CUDA device memory. More...

ghost_error	ghost_cu_memcpy (void dest, void src, size_t bytesize)
	Memcpy GPU memory. More...

ghost_error	ghost_cu_memcpy2d (void dest, size_t dpitch, const void src, size_t spitch, size_t width, size_t height)
	Memcpy strided GPU memory. Copy height rows of width bytes each. More...

ghost_error	ghost_cu_memset (void *s, int c, size_t n)
	Memset GPU memory. More...

ghost_error	ghost_cu_upload (void devmem, void hostmem, size_t bytesize)
	Upload memory from the the host to the GPU. More...

ghost_error	ghost_cu_upload2d (void dest, size_t dpitch, const void src, size_t spitch, size_t width, size_t height)
	Upload strided memory from the host to the GPU. Copy height rows of width bytes each. More...

ghost_error	ghost_cu_download2d (void dest, size_t dpitch, const void src, size_t spitch, size_t width, size_t height)
	Download strided memory from a GPU to the host. Copy height rows of width bytes each. More...

ghost_error	ghost_cu_download (void hostmem, void devmem, size_t bytesize)
	Download memory from a GPU to the host. More...

ghost_error	ghost_cu_free (void *mem)
	Free GPU memory. More...

ghost_error	ghost_cu_barrier ()
	Wait for any outstanding CUDA kernel. More...

static int	stringcmp (const void x, const void y)

ghost_error	ghost_cu_ndevice (int *devcount)
	Get the number of available GPUs. More...

ghost_error	ghost_cu_gpu_info_create (ghost_gpu_info **devInfo)
	Get information about available GPUs. More...

ghost_error	ghost_cu_malloc_mapped (void **mem, const size_t size)
	Allocate mapped host memory. More...

ghost_error	ghost_cu_malloc_pinned (void **mem, const size_t size)
	Allocate pinned host memory. More...

ghost_error	ghost_cu_device (int *device)
	Get the active GPU. More...

ghost_error	ghost_cu_cublas_handle (ghost_cublas_handle_t *handle)
	Get the cuBLAS handle. More...

ghost_error	ghost_cu_cusparse_handle (ghost_cusparse_handle_t *handle)
	Get the CuSparse handle. More...

ghost_error	ghost_cu_version (int *ver)
	Get the CUDA version. More...

ghost_error	ghost_cu_free_host (void *mem)
	Free host memory which has been allocated with ghost_cu_malloc_pinned() or ghost_cu_malloc_mapped(). More...

ghost_error	ghost_cu_deviceprop_get (ghost_cu_deviceprop *prop)
	Get the CUDA device properties. More...

ghost_error	ghost_cu_rand_generator_get (ghost_cu_rand_generator *gen)

ghost_error	ghost_cu_finalize ()

ghost_error	ghost_cu_memtranspose (int m, int n, void to, int ldto, const void from, int ldfrom, ghost_datatype dt)

Macro Definition Documentation

#define ghost_cu_MAX_DEVICE_NAME_LEN 500

Function Documentation

ghost_error ghost_cu_barrier ( )

Wait for any outstanding CUDA kernel.

Returns: GHOST_SUCCESS on success or an error indicator.

Here is the call graph for this function:

ghost_error ghost_cu_cublas_handle ( ghost_cublas_handle_t * handle )

Get the cuBLAS handle.

Parameters

handle Where to store the handle.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_cusparse_handle ( ghost_cusparse_handle_t * handle )

Get the CuSparse handle.

Parameters

handle Where to store the handle.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_device ( int * device )

Get the active GPU.

Parameters

device Where to store the device number.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_deviceprop_get ( ghost_cu_deviceprop * prop )

Get the CUDA device properties.

Parameters

prop	Where to store the properties.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_download	(	void *	hostmem,
		void *	devmem,
		size_t	bytesize
	)

Download memory from a GPU to the host.

Parameters

hostmem	The host side memory location.
devmem	The device side memory location.
bytesize	Number of bytes.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_download2d	(	void *	dest,
		size_t	dpitch,
		const void *	src,
		size_t	spitch,
		size_t	width,
		size_t	height
	)

Download strided memory from a GPU to the host. Copy height rows of width bytes each.

Parameters

dest	The host memory.
dpitch	The pitch in the host memory.
src	The device memory.
spitch	The pitch in the device memory.
width	The number of bytes per row.
height	The number of rows.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_finalize ( )

ghost_error ghost_cu_free ( void * mem )

Free GPU memory.

Parameters

mem	The memory location.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_free_host ( void * mem )

Free host memory which has been allocated with ghost_cu_malloc_pinned() or ghost_cu_malloc_mapped().

Parameters

mem	The memory location.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_gpu_info_create ( ghost_gpu_info ** gpu_info )

Get information about available GPUs.

Parameters

gpu_info Where to store the GPU info.

Returns: GHOST_SUCCESS on success or an error indicator.

Here is the call graph for this function:

ghost_error ghost_cu_init ( int dev )

Initalize CUDA on a given device.

Parameters

dev	The device number.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_malloc	(	void **	mem,
		size_t	bytesize
	)

Allocate CUDA device memory.

Parameters

mem	Where to store the memory.
bytesize	The number of bytes to allocate.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_malloc_managed	(	void **	mem,
		size_t	bytesize
	)

Allocate managed CUDA device memory.

Parameters

mem	Where to store the memory.
bytesize	The number of bytes to allocate.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_malloc_mapped	(	void **	mem,
		const size_t	size
	)

Allocate mapped host memory.

Parameters

mem	Where to store the memory.
size	The number of bytes to allocate.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_malloc_pinned	(	void **	mem,
		const size_t	size
	)

Allocate pinned host memory.

Parameters

mem	Where to store the memory.
size	The number of bytes to allocate.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_memcpy	(	void *	dest,
		void *	src,
		size_t	bytesize
	)

Memcpy GPU memory.

Parameters

dest	The destination memory location.
src	The source memory location.
bytesize	The number of bytes.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_memcpy2d	(	void *	dest,
		size_t	dpitch,
		const void *	src,
		size_t	spitch,
		size_t	width,
		size_t	height
	)

Memcpy strided GPU memory. Copy height rows of width bytes each.

Parameters

dest	The device memory.
dpitch	The pitch in the device memory.
src	The host memory.
spitch	The pitch in the host memory.
width	The number of bytes per row.
height	The number of rows.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_memset	(	void *	s,
		int	c,
		size_t	n
	)

Memset GPU memory.

Parameters

s	The memory location.
c	What to write to the memory.
n	The number of bytes.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_memtranspose	(	int	m,
		int	n,
		void *	to,
		int	ldto,
		const void *	from,
		int	ldfrom,
		ghost_datatype	dt
	)

ghost_error ghost_cu_ndevice ( int * devcount )

Get the number of available GPUs.

Parameters

devcount Where to store the number.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_rand_generator_get ( ghost_cu_rand_generator * gen )

Here is the call graph for this function:

ghost_error ghost_cu_upload	(	void *	devmem,
		void *	hostmem,
		size_t	bytesize
	)

Upload memory from the the host to the GPU.

Parameters

devmem	The device side memory location.
hostmem	The host side memory location.
bytesize	Number of bytes.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_upload2d	(	void *	dest,
		size_t	dpitch,
		const void *	src,
		size_t	spitch,
		size_t	width,
		size_t	height
	)

Upload strided memory from the host to the GPU. Copy height rows of width bytes each.

Parameters

dest	The device memory.
dpitch	The pitch in the device memory.
src	The host memory.
spitch	The pitch in the host memory.
width	The number of bytes per row.
height	The number of rows.

Returns: GHOST_SUCCESS on success or an error indicator.

ghost_error ghost_cu_version ( int * ver )

Get the CUDA version.

Parameters

ver	Where to store the version.

Returns: GHOST_SUCCESS on success or an error indicator.

Here is the call graph for this function:

static int stringcmp	(	const void *	x,
		const void *	y
	)

static

Macros

Functions

Macro Definition Documentation

Function Documentation