Exclusive scan (prefix sum) primitive. More...

#include <scan.h>

Inheritance diagram for clogs::Scan:

Collaboration diagram for clogs::Scan:

Public Member Functions
	Scan ()
	Default constructor. More...

	Scan (const cl::Context &context, const cl::Device &device, const Type &type)
	Constructor. More...

	Scan (const cl::Context &context, const cl::Device &device, const ScanProblem &problem)
	Constructor. More...

	~Scan ()
	Destructor.

void	enqueue (const cl::CommandQueue &commandQueue, const cl::Buffer &buffer,::size_t elements, const void offset=NULL, const VECTOR_CLASS< cl::Event > events=NULL, cl::Event *event=NULL)
	Enqueue a scan operation on a command queue (in-place). More...

void	enqueue (const cl::CommandQueue &commandQueue, const cl::Buffer &inBuffer, const cl::Buffer &outBuffer,::size_t elements, const void offset=NULL, const VECTOR_CLASS< cl::Event > events=NULL, cl::Event *event=NULL)
	Enqueue a scan operation on a command queue. More...

void	enqueue (cl_command_queue commandQueue, cl_mem inBuffer, cl_mem outBuffer,::size_t elements, const void offset=NULL, cl_uint numEvents=0, const cl_event events=NULL, cl_event *event=NULL)

void	enqueue (const cl::CommandQueue &commandQueue, const cl::Buffer &buffer,::size_t elements, const cl::Buffer &offsetBuffer, cl_uint offsetIndex, const VECTOR_CLASS< cl::Event > events=NULL, cl::Event event=NULL)
	Enqueue a scan operation on a command queue, with an offset in a buffer (in-place). More...

void	enqueue (const cl::CommandQueue &commandQueue, const cl::Buffer &inBuffer, const cl::Buffer &outBuffer,::size_t elements, const cl::Buffer &offsetBuffer, cl_uint offsetIndex, const VECTOR_CLASS< cl::Event > events=NULL, cl::Event event=NULL)
	Enqueue a scan operation on a command queue, with an offset in a buffer. More...

void	enqueue (cl_command_queue commandQueue, cl_mem inBuffer, cl_mem outBuffer,::size_t elements, cl_mem offsetBuffer, cl_uint offsetIndex, cl_uint numEvents=0, const cl_event events=NULL, cl_event event=NULL)

Public Member Functions inherited from clogs::Algorithm
void	setEventCallback (void(callback)(const cl::Event &, void ), void userData, void(free)(void *)=NULL)
	Set a callback function that will receive a list of all underlying events. More...

template<typename T >
void	setEventCallback (const T &callback)
	This is an overloaded member function, provided for convenience. It differs from the above function only in what argument(s) it accepts.The provided function object will be passed a cl::Event. More...

void	setEventCallback (void(callback)(cl_event, void ), void userData, void(free)(void *)=NULL)

Protected Member Functions
void	enqueue (cl_command_queue commandQueue, cl_mem inBuffer, cl_mem outBuffer,::size_t elements, const void offset, cl_uint numEvents, const cl_event events, cl_event event, cl_int &err, const char &errStr)

void	enqueue (cl_command_queue commandQueue, cl_mem inBuffer, cl_mem outBuffer,::size_t elements, cl_mem offsetBuffer, cl_uint offsetIndex, cl_uint numEvents, const cl_event events, cl_event event, cl_int &err, const char *&errStr)

void	moveAssign (Scan &other)

Protected Member Functions inherited from clogs::Algorithm
void	moveConstruct (Algorithm &other)
	Constructs this by stealing the pointer from other.

detail::Algorithm *	moveAssign (Algorithm &other)
	Sets this by stealing the pointer from other, and returning the previous value.

void	swap (Algorithm &other)
	Swaps the pointers between this and other.

detail::Algorithm *	getDetail () const
	Returns the embedded pointer.

detail::Algorithm *	getDetailNonNull () const
	Returns the embedded pointer. More...

void	setDetail (detail::Algorithm *ptr)
	Set the value of the embedded pointer. More...

Friends
void	swap (Scan &, Scan &)

Detailed Description

Exclusive scan (prefix sum) primitive.

One instance of this class can be reused for multiple scans, provided that

calls to enqueue do not overlap; and
their execution does not overlap.

An instance of the class is specialized to a specific context, device, and type of value to scan. Any CL integral scalar or vector type can be used.

The implementation is based on the reduce-then-scan strategy described at https://sites.google.com/site/duanemerrill/ScanTR2.pdf?attredirects=0

Constructor & Destructor Documentation

clogs::Scan::Scan ( )

Default constructor.

The object cannot be used in this state.

clogs::Scan::Scan	(	const cl::Context &	context,
		const cl::Device &	device,
		const Type &	type
	)

inline

Constructor.

Parameters

context	OpenCL context to use
device	OpenCL device to use.
type	Type of the values to scan.

Exceptions

std::invalid_argument	if type is not an integral type supported on the device.
clogs::InternalError	if there was a problem with initialization.

Deprecated:: This interface is deprecated as it does not scale with future feature additions. Use the constructor taking a clogs::ScanProblem instead.

clogs::Scan::Scan	(	const cl::Context &	context,
		const cl::Device &	device,
		const ScanProblem &	problem
	)

inline

Constructor.

Parameters

context	OpenCL context to use
device	OpenCL device to use.
problem	Description of the specific scan problem.

Exceptions

std::invalid_argument	if problem is not supported on the device or is not initialized.
clogs::InternalError	if there was a problem with initialization.

Member Function Documentation

void clogs::Scan::enqueue	(	const cl::CommandQueue &	commandQueue,
		const cl::Buffer &	buffer,
		::size_t	elements,
		const void *	offset = `NULL`,
		const VECTOR_CLASS< cl::Event > *	events = `NULL`,
		cl::Event *	event = `NULL`
	)

inline

Enqueue a scan operation on a command queue (in-place).

This is equivalent to calling enqueue(commandQueue, buffer, buffer, elements, offset, events, event);

void clogs::Scan::enqueue	(	const cl::CommandQueue &	commandQueue,
		const cl::Buffer &	inBuffer,
		const cl::Buffer &	outBuffer,
		::size_t	elements,
		const void *	offset = `NULL`,
		const VECTOR_CLASS< cl::Event > *	events = `NULL`,
		cl::Event *	event = `NULL`
	)

inline

Enqueue a scan operation on a command queue.

An initial offset may optionally be passed in offset, which will be added to all elements of the result. The pointer must point to the type of element specified to the constructor. If no offset is desired, NULL may be passed instead.

The input and output buffers may be the same to do an in-place scan.

Parameters

commandQueue	The command queue to use.
inBuffer	The buffer to scan.
outBuffer	The buffer to fill with output.
elements	The number of elements to scan.
offset	The offset to add to all elements, or `NULL`.
events	Events to wait for before starting.
event	Event that will be signaled on completion.

Exceptions

cl::Error	If inBuffer is not readable on the device.
cl::Error	If outBuffer is not writable on the device.
cl::Error	If the element range overruns the buffer.
cl::Error	If elements is zero.

Precondition

commandQueue was created with the context and device given to the constructor.

Postcondition

After execution, element i will be replaced by the sum of all elements strictly before i, plus the offset (if any).

void clogs::Scan::enqueue	(	cl_command_queue	commandQueue,
		cl_mem	inBuffer,
		cl_mem	outBuffer,
		::size_t	elements,
		const void *	offset = `NULL`,
		cl_uint	numEvents = `0`,
		const cl_event *	events = `NULL`,
		cl_event *	event = `NULL`
	)

inline

This is an overloaded member function, provided for convenience. It differs from the above function only in what argument(s) it accepts.

void clogs::Scan::enqueue	(	const cl::CommandQueue &	commandQueue,
		const cl::Buffer &	buffer,
		::size_t	elements,
		const cl::Buffer &	offsetBuffer,
		cl_uint	offsetIndex,
		const VECTOR_CLASS< cl::Event > *	events = `NULL`,
		cl::Event *	event = `NULL`
	)

inline

Enqueue a scan operation on a command queue, with an offset in a buffer (in-place).

This is equivalent to calling enqueue(commandQueue, buffer, buffer, elements, offsetBuffer, offsetIndex);

void clogs::Scan::enqueue	(	const cl::CommandQueue &	commandQueue,
		const cl::Buffer &	inBuffer,
		const cl::Buffer &	outBuffer,
		::size_t	elements,
		const cl::Buffer &	offsetBuffer,
		cl_uint	offsetIndex,
		const VECTOR_CLASS< cl::Event > *	events = `NULL`,
		cl::Event *	event = `NULL`
	)

inline

Enqueue a scan operation on a command queue, with an offset in a buffer.

The offset is of the same type as the elements to be scanned, and is stored in a buffer. It is added to all elements of the result. It is legal for the offset to be in the same buffer as the values to scan, and it may even be safely overwritten by the scan (it will be read before being overwritten). This makes it possible to use do multi-pass algorithms with variable output. The counting pass fills in the desired allocations, a scan is used with one extra element at the end to hold the grand total, and the subsequent passes use this extra element as the offset.

The input and output buffers may be the same to do an in-place scan.

Parameters

commandQueue	The command queue to use.
inBuffer	The buffer to scan.
outBuffer	The buffer to fill with output.
elements	The number of elements to scan.
offsetBuffer	Buffer containing a value to add to all elements.
offsetIndex	Index (in units of the scan type) into offsetBuffer.
events	Events to wait for before starting.
event	Event that will be signaled on completion.

Exceptions

cl::Error	If inBuffer is not readable on the device.
cl::Error	If outBuffer is not writable on the device.
cl::Error	If the element range overruns the buffer.
cl::Error	If elements is zero.
cl::Error	If offsetBuffer is not readable.
cl::Error	If offsetIndex overruns offsetBuffer.

Precondition

commandQueue was created with the context and device given to the constructor.

Postcondition

After execution, element i will be replaced by the sum of all elements strictly before i, plus the offset.

void clogs::Scan::enqueue	(	cl_command_queue	commandQueue,
		cl_mem	inBuffer,
		cl_mem	outBuffer,
		::size_t	elements,
		cl_mem	offsetBuffer,
		cl_uint	offsetIndex,
		cl_uint	numEvents = `0`,
		const cl_event *	events = `NULL`,
		cl_event *	event = `NULL`
	)

inline

This is an overloaded member function, provided for convenience. It differs from the above function only in what argument(s) it accepts.

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

/home/bruce/devel/clogs/include/clogs/scan.h

Public Member Functions

Protected Member Functions

Friends

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation