magmadnn::MemoryManager< T > Class Template Reference

Public Member Functions
	MemoryManager (unsigned int size, memory_t mem_type, device_t device_id)
	MemoryManager (const MemoryManager &that)
MemoryManager &	operator= (const MemoryManager &that)
	~MemoryManager ()
magmadnn_error_t	copy_from (const MemoryManager< T > &src, unsigned int begin_idx, unsigned int size)
magmadnn_error_t	copy_from (const MemoryManager< T > &src, unsigned int size)
magmadnn_error_t	copy_from (const MemoryManager< T > &src)
magmadnn_error_t	copy_from_host (T *src, unsigned int begin_idx, unsigned int size)
magmadnn_error_t	sync (bool gpu_was_modified=true)
magmadnn_error_t	set_device (device_t device_id)
T	get (unsigned int idx) const
void	set (unsigned int idx, T val)
T *	get_host_ptr ()
T *	get_ptr ()
unsigned int	get_size () const
memory_t	get_memory_type () const

Constructor & Destructor Documentation

MemoryManager() [1/2]

template<typename T >

magmadnn::MemoryManager< T >::MemoryManager	(	unsigned int	size,
		memory_t	mem_type,
		device_t	device_id
	)

MemoryManager class to keep track of a memory address across devices.

Parameters

size	the size of the memory to allocate/manage
mem_type	what memory type will this data belong to
device_id	what device will the data reside on (preferred if mem_type is CUDA_MANAGED)

Here is the call graph for this function:

MemoryManager() [2/2]

template<typename T >

magmadnn::MemoryManager< T >::MemoryManager

(

const MemoryManager< T > &

that

)

Copy Constructor

Parameters

that

Here is the call graph for this function:

~MemoryManager()

template<typename T >

magmadnn::MemoryManager< T >::~MemoryManager< T >

(

)

Destroys the memory manager object and releases all its data.

Member Function Documentation

copy_from() [1/3]

template<typename T >

magmadnn_error_t magmadnn::MemoryManager< T >::copy_from	(	const MemoryManager< T > &	src,
		unsigned int	begin_idx,
		unsigned int	size
	)

Copies the data from src memory manager into the pointer here. Asserts that src and this have the same size.

Parameters

src	the memorymanager to copy data from

Returns

the error code (0 - no error, 1 - src ptr not allocated)

Here is the caller graph for this function:

copy_from() [2/3]

template<typename T >

magmadnn_error_t magmadnn::MemoryManager< T >::copy_from	(	const MemoryManager< T > &	src,
		unsigned int	size
	)

Copies the data from src memory manager into the pointer here. Asserts that src and this have the same size.

Parameters

src	the memorymanager to copy data from

Returns

the error code (0 - no error, 1 - src ptr not allocated)

Here is the call graph for this function:

copy_from() [3/3]

template<typename T >

magmadnn_error_t magmadnn::MemoryManager< T >::copy_from

(

const MemoryManager< T > &

src

)

Copies the data from src memory manager into the pointer here. Asserts that src and this have the same size.

Parameters

src	the memorymanager to copy data from

Returns

the error code (0 - no error, 1 - src ptr not allocated)

Here is the call graph for this function:

copy_from_host()

template<typename T >

magmadnn_error_t magmadnn::MemoryManager< T >::copy_from_host	(	T *	src,
		unsigned int	begin_idx,
		unsigned int	size
	)

copies memory from a host ptr into this memorymanager. will throw an error if it reaches the end of src allocated mem before this is filled.

Parameters

src	the array to copy into this.

Returns

the error code (0 - good, 1 - not enough memory)

get()

template<typename T >

T magmadnn::MemoryManager< T >::get

(

unsigned int

idx

)

const

Returns the value at idx. Error if idx is out of range.

Parameters

idx	index to retrieve

Returns

the value at index idx.

get_host_ptr()

template<typename T >

T * magmadnn::MemoryManager< T >::get_host_ptr

(

)

returns a CPU pointer to the data.

Returns

cpu pointer

get_memory_type()

template<typename T>

memory_t magmadnn::MemoryManager< T >::get_memory_type ( ) const

inline

Returns the memory type of this memory manager.

Returns

memory_t

get_ptr()

template<typename T >

T * magmadnn::MemoryManager< T >::get_ptr

(

)

Returns a pointer to whatever memory type this is using. For MANAGED memory type it returns the device pointer.

Returns

the data ptr

get_size()

template<typename T>

unsigned int magmadnn::MemoryManager< T >::get_size ( ) const

inline

Returns the size of this memorymanager

Returns

unsigned int the size of this memory manager

operator=()

template<typename T >

MemoryManager< T > & magmadnn::MemoryManager< T >::operator=

(

const MemoryManager< T > &

that

)

Copy assignment operator.

Parameters

that

Returns

MemoryManager&

Here is the call graph for this function:

set()

template<typename T >

void magmadnn::MemoryManager< T >::set	(	unsigned int	idx,
		T	val
	)

Sets the value at idx to val. Error if idx is out of range.

Parameters

idx	index to set
val	value to set at idx

set_device()

template<typename T >

magmadnn_error_t magmadnn::MemoryManager< T >::set_device

(

device_t

device_id

)

Changes the device this memory manager points to. Note that the memory type is still the same, but the device_id will be different.

Returns

an error code (0 - ok)

Here is the caller graph for this function:

sync()

template<typename T >

magmadnn_error_t magmadnn::MemoryManager< T >::sync

(

bool

gpu_was_modified = true

)

If MANAGED or CUDA_MANAGED this ensures that data is the same on all devices. It will wait for any gpu kernels to finish before copying data. If HOST or DEVICE memory this does nothing.

Parameters

gpu_was_modified

If true then data will be copied from gpu to cpu, else if false vice-versa. By default true.

Returns

an error code

---

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

• include/memory/memorymanager.h
• src/memory/memorymanager.cpp