fml Namespace Reference

Core namespace. More...

Namespaces
	copy
	Copiers.
	dimops
	Row/column operations.
	linalg
	Linear algebra functions.
	nvml
	NVIDIA Management Library (NVML) interface. Exceptions Each function can throw a 'runtime_error' exception.
	stats
	Statistics kernels.

Classes
class	card
	GPU data and methods. More...
class	comm
	MPI communicator data and helpers. More...
class	cpumat
	Matrix class for data held on a single CPU. More...
class	cpuvec
	Vector class for data held on a single CPU. More...
class	gpumat
	Matrix class for data held on a single GPU. More...
class	gpuscalar
class	gpuvec
	Vector class for data held on a single GPU. More...
class	grid
	2-dimensional MPI process grid. More...
class	mpimat
	Matrix class for data distributed over MPI in the 2-d block cyclic format. More...
class	parmat
class	parmat_cpu
class	parmat_gpu
class	unimat
	Base matrix class. Not meant for direct use. Instead see cpumat, gpumat, and mpimat. More...
class	univec
	Base vector class. Not meant for direct use. Instead see cpuvec and gpuvec. More...

Typedefs
typedef std::shared_ptr< fml::card >	card_sp_t

Enumerations
enum	gridshape { PROC_GRID_SQUARE, PROC_GRID_WIDE, PROC_GRID_TALL }
	Supported process grid shapes for 2-dimensional BLACS grids. More...
enum	blacsops { BLACS_SUM, BLACS_MAX, BLACS_MIN }
	Supported operations in reduce/allreduce.

Functions
int	get_device_count ()
	Return number of GPU devices.
card_sp_t	new_card (int id=0)
	Initialize a new card. More...
int	get_device_num (const comm &c)
	Returns the device number of the GPU to use with the calling MPI process. More...

Detailed Description

Core namespace.

Function Documentation

get_device_num()

int fml::get_device_num ( const comm &  c )

inline

Returns the device number of the GPU to use with the calling MPI process.

Parameters

[in]

c

A communicator object.

Returns

The ordinal device number for the calling MPI process to use in, e.g. fml::new_card().

Exceptions
If there are more MPI ranks within any node than there are GPUS, the function will throw a 'runtime_error' exception.

new_card()

card_sp_t fml::new_card ( int  id = 0 )

inline

Initialize a new card.

Parameters

[in]

id

GPU id number.

Returns

Shared pointer to initialized card object.

Exceptions
If the GPU can not be initialized, or if the allocation of one of the handles fails, the method will throw a 'runtime_error' exception.