The SyncedMemory manages the low-level connection between the GPU and host memory. More...

Inheritance diagram for MyCaffe.common.SyncedMemory< T >:

Public Member Functions
	SyncedMemory (CudaDnn< T > cuda, Log log, long lCapacity=0, object tag=null, bool bUseHalfSize=false)
	The SyncedMemory constructor. More...

void	Dispose ()
	Releases all GPU and host resources used. More...

void	Allocate (long lCount, bool bUseHalfSize=false)
	Allocate a number of items in GPU memory and save the handle. More...

void	Allocate (T[] rg, bool bUseHalfSize=false)
	Allocate a number of items and copy the given array into the memory on the GPU. More...

void	Zero ()
	Set all items in the GPU memory up to the Count, to zero. More...

void	ZeroAll ()
	Set all items in the GPU memory up to the Capacity, to zero. More...

void	Set (double dfVal)
	Set all items up to Count to a given value. More...

void	SetAt (int nIdx, T fVal)
	Set a specific item at a given index to a value. More...

T	GetAt (int nIdx)
	Return a value at a given index. More...

long	Copy (SyncedMemory< T > src, long hDstHostBuffer=0)
	Copy another SyncedMemory into this one. More...

SyncedMemory< T >	Clone ()
	Copy this SyncedMemory. More...

void	set_gpu_data (long hData, long lCount, long lOffset)
	Copies a new Memory Pointer within the low-level CudaDnnDLL where a Memory Pointer uses another already allocated block of GPU memory and just indexes into it. More...

void	SetData (T[] rgData, int nCount, bool bSetCount=true)
	Sets the array of host data on the GPU and re-allocates the GPU memory if needed. More...

T[]	update_cpu_data (long lCount=-1)
	Updates the host data by copying the GPU data to the host data. More...

void	set_cpu_data_locally (T[] rg)
	This does not place the data on the GPU - call async_gpu_push() to move it to the GPU. More...

void	async_gpu_push (long hStream, T[] rg)
	Pushes the host data, previously set with set_cpu_data_locally(), to the GPU. More...

Properties
bool	HalfSize `[get]`
	Returns whether or not the sync memory is half-sized memory. More...

object	Tag `[getset]`
	Get/set data associated with the synced memory. More...

int	DeviceID `[get]`
	Returns the Device ID on which the GPU memory of this SyncedMemory was allocated. More...

long	Capacity `[get]`
	Returns the total amount of GPU memory held by this SyncedMemory. More...

long	Count `[getset]`
	Returns the current count of items in this SyncedMemory. Note, the Count may be less than the Capacity. More...

long	gpu_data `[get]`
	Returns the handle to the GPU memory. More...

long	mutable_gpu_data `[get]`
	Returns the mutable handle to GPU data. More...

T[]	cpu_data `[get]`
	Returns the data on the CPU that has already been transferred from GPU to CPU. More...

T[]	mutable_cpu_data `[getset]`
	Get/set the mutable host data. More...

Detailed Description

The SyncedMemory manages the low-level connection between the GPU and host memory.

The GPU memory is represented by a handle into the memory look-up table managed by the low-level CudaDnn DLL. The host memory is copied and stored in a local array of type 'T'.

Template Parameters

T	Specifies the base type float or double. Using float is recommended to conserve GPU memory.

Definition at line 17 of file SyncedMemory.cs.

Constructor & Destructor Documentation

◆ SyncedMemory()

MyCaffe.common.SyncedMemory< T >.SyncedMemory	(	CudaDnn< T >	cuda,
		Log	log,
		long	lCapacity = `0`,
		object	tag = `null`,
		bool	bUseHalfSize = `false`
	)

The SyncedMemory constructor.

Parameters

cuda	Specifies the CudaDnn connection to Cuda.
log	Specifies the Log for output.
lCapacity	Optionally, specifies the capacity of the SyncedMemory (in items).
tag	Optionally, specifies a tag used for debugging (the default = null).
bUseHalfSize	Optionally, specifies to use half size (FP16) for both data and diff. This option is only available when using the float base type 'T'.

Definition at line 38 of file SyncedMemory.cs.

Member Function Documentation

◆ Allocate() [1/2]

void MyCaffe.common.SyncedMemory< T >.Allocate	(	long	lCount,
		bool	bUseHalfSize = `false`
	)

Allocate a number of items in GPU memory and save the handle.

Parameters

lCount	Specifies the number of items.
bUseHalfSize	Optionally, specifis to use half sized memory (default = false). This only applies to the 'float' base type.

Definition at line 92 of file SyncedMemory.cs.

◆ Allocate() [2/2]

void MyCaffe.common.SyncedMemory< T >.Allocate	(	T[]	rg,
		bool	bUseHalfSize = `false`
	)

Allocate a number of items and copy the given array into the memory on the GPU.

Parameters

rg	Specifies the array of items to copy.
bUseHalfSize	Optionally, specifis to use half sized memory (default = false). This only applies to the 'float' base type.

Definition at line 117 of file SyncedMemory.cs.

◆ async_gpu_push()

void MyCaffe.common.SyncedMemory< T >.async_gpu_push	(	long	hStream,
		T[]	rg
	)

Pushes the host data, previously set with set_cpu_data_locally(), to the GPU.

Note, if necessary, this function re-allocates the GPU memory.

Parameters

hStream
rg

Definition at line 446 of file SyncedMemory.cs.

◆ Clone()

SyncedMemory< T > MyCaffe.common.SyncedMemory< T >.Clone ( )

Copy this SyncedMemory.

Returns: A new SynedMemory that is a copy of this one, is returned.

Definition at line 252 of file SyncedMemory.cs.

◆ Copy()

long MyCaffe.common.SyncedMemory< T >.Copy	(	SyncedMemory< T >	src,
		long	hDstHostBuffer = `0`
	)

Copy another SyncedMemory into this one.

Parameters

src	Specifies the SyncedMemory to copy.
hDstHostBuffer	Optionally, specifies a host buffer used to copy between kernels (default = 0, not used).

Returns: When used the dst host buffer handle is returned.

Definition at line 205 of file SyncedMemory.cs.

◆ Dispose()

void MyCaffe.common.SyncedMemory< T >.Dispose ( )

Releases all GPU and host resources used.

Definition at line 77 of file SyncedMemory.cs.

◆ GetAt()

T MyCaffe.common.SyncedMemory< T >.GetAt ( int nIdx )

Return a value at a given index.

Parameters

nIdx	Specifies the index.

Returns: The value at the index is returned.

Definition at line 192 of file SyncedMemory.cs.

◆ Set()

void MyCaffe.common.SyncedMemory< T >.Set ( double dfVal )

Set all items up to Count to a given value.

Parameters

dfVal Specifies the value.

Definition at line 167 of file SyncedMemory.cs.

◆ set_cpu_data_locally()

void MyCaffe.common.SyncedMemory< T >.set_cpu_data_locally ( T[] rg )

This does not place the data on the GPU - call async_gpu_push() to move it to the GPU.

Parameters

rg	Specifies an array of host data.

Definition at line 433 of file SyncedMemory.cs.

◆ set_gpu_data()

void MyCaffe.common.SyncedMemory< T >.set_gpu_data	(	long	hData,
		long	lCount,
		long	lOffset
	)

Copies a new Memory Pointer within the low-level CudaDnnDLL where a Memory Pointer uses another already allocated block of GPU memory and just indexes into it.

Parameters

hData	Specifies a handle to the already allocated GPU memory that the new Memory Pointer will index into.
lCount	Specifies the number of items in this 'virtual memory'
lOffset	Specifies the offset into the GPU data where the Memory Pointer should start.

Definition at line 319 of file SyncedMemory.cs.

◆ SetAt()

void MyCaffe.common.SyncedMemory< T >.SetAt	(	int	nIdx,
		T	fVal
	)

Set a specific item at a given index to a value.

Parameters

nIdx	Specifies the index.
fVal	Specifies the value.

Definition at line 181 of file SyncedMemory.cs.

◆ SetData()

void MyCaffe.common.SyncedMemory< T >.SetData	(	T[]	rgData,
		int	nCount,
		bool	bSetCount = `true`
	)

Sets the array of host data on the GPU and re-allocates the GPU memory if needed.

Parameters

rgData	Specifies the host data to set.
nCount	Specifies the number of items in the host data to set, which may be less than the host data array length.
bSetCount	Optionally, specifies whether or not to set the count. The count is always set when re-allocating the buffer.

Definition at line 359 of file SyncedMemory.cs.

◆ update_cpu_data()

T[] MyCaffe.common.SyncedMemory< T >.update_cpu_data ( long lCount = -1 )

Updates the host data by copying the GPU data to the host data.

Parameters

lCount Optionally, specifies a count (less than Count) to transfer.

Returns: An array of the host data is returned.

Definition at line 409 of file SyncedMemory.cs.

◆ Zero()

void MyCaffe.common.SyncedMemory< T >.Zero ( )

Set all items in the GPU memory up to the Count, to zero.

Definition at line 142 of file SyncedMemory.cs.

◆ ZeroAll()

void MyCaffe.common.SyncedMemory< T >.ZeroAll ( )

Set all items in the GPU memory up to the Capacity, to zero.

Definition at line 154 of file SyncedMemory.cs.

Property Documentation

◆ Capacity

long MyCaffe.common.SyncedMemory< T >.Capacity

get

Returns the total amount of GPU memory held by this SyncedMemory.

Definition at line 290 of file SyncedMemory.cs.

◆ Count

long MyCaffe.common.SyncedMemory< T >.Count

getset

Returns the current count of items in this SyncedMemory. Note, the Count may be less than the Capacity.

Definition at line 298 of file SyncedMemory.cs.

◆ cpu_data

T [] MyCaffe.common.SyncedMemory< T >.cpu_data

get

Returns the data on the CPU that has already been transferred from GPU to CPU.

Definition at line 348 of file SyncedMemory.cs.

◆ DeviceID

int MyCaffe.common.SyncedMemory< T >.DeviceID

get

Returns the Device ID on which the GPU memory of this SyncedMemory was allocated.

Definition at line 282 of file SyncedMemory.cs.

◆ gpu_data

long MyCaffe.common.SyncedMemory< T >.gpu_data

get

Returns the handle to the GPU memory.

Definition at line 307 of file SyncedMemory.cs.

◆ HalfSize

bool MyCaffe.common.SyncedMemory< T >.HalfSize

get

Returns whether or not the sync memory is half-sized memory.

Definition at line 265 of file SyncedMemory.cs.

◆ mutable_cpu_data

T [] MyCaffe.common.SyncedMemory< T >.mutable_cpu_data

getset

Get/set the mutable host data.

When setting the mutable host data, the data is copied to the GPU. When get'ing the host data, the data is transferred from the GPU first and then returned.

Definition at line 385 of file SyncedMemory.cs.

◆ mutable_gpu_data

long MyCaffe.common.SyncedMemory< T >.mutable_gpu_data

get

Returns the mutable handle to GPU data.

Note: This is the same as gpu_data, but is provided for compatibility and readability with the original C++ Caffe code.

Definition at line 335 of file SyncedMemory.cs.

◆ Tag

object MyCaffe.common.SyncedMemory< T >.Tag

getset

Get/set data associated with the synced memory.

Definition at line 273 of file SyncedMemory.cs.

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

C:/Data/Data/SS_Projects/Intelligence/GitHub/MyCaffe/MyCaffe/common/SyncedMemory.cs

Public Member Functions

Properties

Detailed Description

Constructor & Destructor Documentation

◆ SyncedMemory()

Member Function Documentation

◆ Allocate() [1/2]

◆ Allocate() [2/2]

◆ async_gpu_push()

◆ Clone()

◆ Copy()

◆ Dispose()

◆ GetAt()

◆ Set()

◆ set_cpu_data_locally()

◆ set_gpu_data()

◆ SetAt()

◆ SetData()

◆ update_cpu_data()

◆ Zero()

◆ ZeroAll()

Property Documentation

◆ Capacity

◆ Count

◆ cpu_data

◆ DeviceID

◆ gpu_data

◆ HalfSize

◆ mutable_cpu_data

◆ mutable_gpu_data

◆ Tag