#include <ComputeInterface.h>

Public Member Functions
virtual	~ComputeInterface ()=default

virtual void	RunKernel (const KernelDefinitionId id)=0

virtual void	RunKernel (const KernelDefinitionId id, const DimensionVector &globalSize, const DimensionVector &localSize)=0

virtual ComputeActionId	RunKernelAsync (const KernelDefinitionId id, const QueueId queue)=0

virtual ComputeActionId	RunKernelAsync (const KernelDefinitionId id, const QueueId queue, const DimensionVector &globalSize, const DimensionVector &localSize)=0

virtual void	WaitForComputeAction (const ComputeActionId id)=0

virtual void	RunKernelWithProfiling (const KernelDefinitionId id)=0

virtual void	RunKernelWithProfiling (const KernelDefinitionId id, const DimensionVector &globalSize, const DimensionVector &localSize)=0

virtual uint64_t	GetRemainingProfilingRuns (const KernelDefinitionId id) const =0

virtual uint64_t	GetRemainingProfilingRuns () const =0

virtual QueueId	GetDefaultQueue () const =0

virtual std::vector< QueueId >	GetAllQueues () const =0

virtual void	SynchronizeQueue (const QueueId queue)=0

virtual void	SynchronizeQueues ()=0

virtual void	SynchronizeDevice ()=0

virtual const DimensionVector &	GetCurrentGlobalSize (const KernelDefinitionId id) const =0

virtual const DimensionVector &	GetCurrentLocalSize (const KernelDefinitionId id) const =0

virtual const KernelConfiguration &	GetCurrentConfiguration () const =0

virtual KernelRunMode	GetRunMode () const =0

virtual void	ChangeArguments (const KernelDefinitionId id, const std::vector< ArgumentId > &arguments)=0

virtual void	SwapArguments (const KernelDefinitionId id, const ArgumentId &first, const ArgumentId &second)=0

virtual void	UpdateScalarArgument (const ArgumentId &id, const void *data)=0

virtual void	UpdateLocalArgument (const ArgumentId &id, const size_t dataSize)=0

virtual void	UploadBuffer (const ArgumentId &id)=0

virtual TransferActionId	UploadBufferAsync (const ArgumentId &id, const QueueId queue)=0

virtual void	DownloadBuffer (const ArgumentId &id, void *destination, const size_t dataSize=0)=0

virtual TransferActionId	DownloadBufferAsync (const ArgumentId &id, const QueueId queue, void *destination, const size_t dataSize=0)=0

virtual void	UpdateBuffer (const ArgumentId &id, const void *data, const size_t dataSize=0)=0

virtual TransferActionId	UpdateBufferAsync (const ArgumentId &id, const QueueId queue, const void *data, const size_t dataSize=0)=0

virtual void	CopyBuffer (const ArgumentId &destination, const ArgumentId &source, const size_t dataSize=0)=0

virtual TransferActionId	CopyBufferAsync (const ArgumentId &destination, const ArgumentId &source, const QueueId queue, const size_t dataSize=0)=0

virtual void	WaitForTransferAction (const TransferActionId id)=0

virtual void	ResizeBuffer (const ArgumentId &id, const size_t newDataSize, const bool preserveData)=0

virtual void	ClearBuffer (const ArgumentId &id)=0

virtual bool	HasBuffer (const ArgumentId &id)=0

virtual void	GetUnifiedMemoryBufferHandle (const ArgumentId &id, UnifiedBufferMemory &memoryHandle)=0

virtual bool	GetProfiling (const KernelId id)=0

Detailed Description

Interface for customizing kernel runs in order to run some part of computation on CPU, utilize iterative kernel launches, composite kernels and more. In order to use this functionality, custom kernel launcher function must be defined for the corresponding kernel.

Constructor & Destructor Documentation

◆ ~ComputeInterface()

ktt::ComputeInterface::~ComputeInterface ( )

virtualdefault

Compute interface destructor.

Member Function Documentation

◆ ChangeArguments()

void ktt::ComputeInterface::ChangeArguments	(	const KernelDefinitionId	id,
		const std::vector< ArgumentId > &	arguments
	)

pure virtual

Changes kernel arguments for the specified kernel definitions under currently launched kernel.

Parameters

id	Id of kernel definition for which the arguments will be changed. The specified definition must be included in the currently launched kernel.
arguments	Ids of arguments to be used by the specified kernel definition. The order of ids must match the order of arguments inside kernel function. The provided ids must be unique.

◆ ClearBuffer()

void ktt::ComputeInterface::ClearBuffer ( const ArgumentId & id )

pure virtual

Removes compute buffer for the specified vector argument. This method should be used mainly with arguments with ArgumentManagementType set to User.

Parameters

id	Id of vector argument whose buffer will be removed.

◆ CopyBuffer()

void ktt::ComputeInterface::CopyBuffer	(	const ArgumentId &	destination,
		const ArgumentId &	source,
		const size_t	dataSize = `0`
	)

pure virtual

Copies part of the compute buffer of source vector argument to compute buffer of destination vector argument.

Parameters

destination	Id of destination vector argument.
source	Id of source vector argument.
dataSize	Size in bytes of buffer portion which will be copied to destination buffer, starting with the first byte. If zero, the entire buffer will be copied.

◆ CopyBufferAsync()

TransferActionId ktt::ComputeInterface::CopyBufferAsync	(	const ArgumentId &	destination,
		const ArgumentId &	source,
		const QueueId	queue,
		const size_t	dataSize = `0`
	)

pure virtual

Copies part of the compute buffer of source vector argument to compute buffer of destination vector argument. The data will be transferred asynchronously in the specified queue.

Parameters

destination	Id of destination vector argument.
source	Id of source vector argument.
queue	Id of queue in which the command to copy argument will be submitted.
dataSize	Size in bytes of buffer portion which will be copied to destination buffer, starting with the first byte. If zero, the entire buffer will be copied.

Returns: Id of asynchronous action corresponding to the issued data transfer command. The action must be waited for with e.g., WaitForTransferAction(), SynchronizeQueue() methods. Otherwise, problems such as incorrectly recorded transfer durations may occur.

◆ DownloadBuffer()

void ktt::ComputeInterface::DownloadBuffer	(	const ArgumentId &	id,
		void *	destination,
		const size_t	dataSize = `0`
	)

pure virtual

Downloads the specified vector argument from compute buffer.

Parameters

id	Id of vector argument which will be downloaded.
destination	Buffer where the argument data will be downloaded. Its size must be equal or greater than the specified data size.
dataSize	Size in bytes of buffer portion which will be downloaded to specified destination, starting with the first byte. If zero, the entire buffer will be downloaded.

◆ DownloadBufferAsync()

TransferActionId ktt::ComputeInterface::DownloadBufferAsync	(	const ArgumentId &	id,
		const QueueId	queue,
		void *	destination,
		const size_t	dataSize = `0`
	)

pure virtual

Downloads the specified vector argument from compute buffer. The data will be transferred asynchronously in the specified queue.

Parameters

id	Id of vector argument which will be downloaded.
queue	Id of queue in which the command to download argument will be submitted.
destination	Buffer where the argument data will be downloaded. Its size must be equal or greater than the specified data size.
dataSize	Size in bytes of buffer portion which will be downloaded to specified destination, starting with the first byte. If zero, the entire buffer will be downloaded.

Returns: Id of asynchronous action corresponding to the issued data transfer command. The action must be waited for with e.g., WaitForTransferAction(), SynchronizeQueue() methods. Otherwise, problems such as incorrectly recorded transfer durations may occur.

◆ GetAllQueues()

std::vector< QueueId > ktt::ComputeInterface::GetAllQueues ( ) const

pure virtual

Retrieves ids of all available device queues. Number of queues can be specified during tuner creation.

Returns: Ids of all available device queues.

◆ GetCurrentConfiguration()

const KernelConfiguration & ktt::ComputeInterface::GetCurrentConfiguration ( ) const

pure virtual

Returns configuration of the currently launched kernel.

Returns: Configuration of the currently launched kernel. See KernelConfiguration for more information.

◆ GetCurrentGlobalSize()

const DimensionVector & ktt::ComputeInterface::GetCurrentGlobalSize ( const KernelDefinitionId id ) const

pure virtual

Returns global thread size for the specified kernel definition based on the current configuration.

Parameters

id	Id of kernel definition for which the global size will be retrieved. The specified definition must be included in the currently launched kernel.

Returns: Global thread size of the specified kernel definition.

◆ GetCurrentLocalSize()

const DimensionVector & ktt::ComputeInterface::GetCurrentLocalSize ( const KernelDefinitionId id ) const

pure virtual

Returns local thread size for the specified kernel definition based on the current configuration.

Parameters

id	Id of kernel definition for which the local size will be retrieved. The specified definition must be included in the currently launched kernel.

Returns: Local thread size of the specified kernel definition.

◆ GetDefaultQueue()

QueueId ktt::ComputeInterface::GetDefaultQueue ( ) const

pure virtual

Retrieves id of default device queue. All synchronous commands are submitted to this queue.

Returns: Id of default device queue.

◆ GetProfiling()

void ktt::ComputeInterface::GetProfiling ( const KernelId id )

pure virtual

Signals whether profiling is activated in the kernel

Parameters

id	Id of queried kernel.

◆ GetRemainingProfilingRuns() [1/2]

uint64_t ktt::ComputeInterface::GetRemainingProfilingRuns ( ) const

pure virtual

Retrieves number of remaining profiling runs that are needed to collect all the profiling counters for the currently launched kernel. The number is derived from included kernel definitions which have kernel profiling functionality enabled.

Returns: Number of remaining profiling runs. Note that if no profiling runs were run so far, zero is returned.

◆ GetRemainingProfilingRuns() [2/2]

uint64_t ktt::ComputeInterface::GetRemainingProfilingRuns ( const KernelDefinitionId id ) const

pure virtual

Retrieves number of remaining profiling runs that are needed to collect all the profiling counters for the specified kernel definition.

Parameters

id	Id of kernel definition for which the number of remaining profiling runs will be retrieved. The specified definition must be included in the currently launched kernel.

Returns: Number of remaining profiling runs. Note that if no profiling runs were run so far for the specified definition, zero is returned.

◆ GetRunMode()

KernelRunMode ktt::ComputeInterface::GetRunMode ( ) const

pure virtual

Returns run mode of the currently launched kernel.

Returns: Run mode of the currently launched kernel. See KernelRunMode for more information.

◆ GetUnifiedMemoryBufferHandle()

void ktt::ComputeInterface::GetUnifiedMemoryBufferHandle	(	const ArgumentId &	id,
		UnifiedBufferMemory &	memoryHandle
	)

pure virtual

Retrieves memory handle for the specified unified memory argument. The handle can be used to manipulate argument memory on host side. Example usage: ktt::UnifiedBufferMemory memory; GetUnifiedMemoryBufferHandle(..., memory); float* floatArray = static_cast<float*>(memory);

Parameters

id	Id of vector argument whose memory handle will be retrieved.
memoryHandle	Location where the memory handle will be stored.

◆ HasBuffer()

bool ktt::ComputeInterface::HasBuffer ( const ArgumentId & id )

pure virtual

Checks whether compute buffer for the specified vector argument exists.

Parameters

id	Id of vector argument to check.

Returns: True if the buffer for argument exists. False otherwise.

◆ ResizeBuffer()

void ktt::ComputeInterface::ResizeBuffer	(	const ArgumentId &	id,
		const size_t	newDataSize,
		const bool	preserveData
	)

pure virtual

Resizes compute buffer for the specified vector argument.

Parameters

id	Id of vector argument whose buffer will be resized.
newDataSize	Size in bytes for the resized buffer.
preserveData	If true, data from the old buffer will be copied into resized buffer. If false, the old data will be discarded.

◆ RunKernel() [1/2]

void ktt::ComputeInterface::RunKernel ( const KernelDefinitionId id )

pure virtual

Runs the specified kernel definition using thread sizes based on the current configuration.

Parameters

id	Id of kernel definition which will be run. The specified definition must be included in the currently launched kernel.

◆ RunKernel() [2/2]

void ktt::ComputeInterface::RunKernel	(	const KernelDefinitionId	id,
		const DimensionVector &	globalSize,
		const DimensionVector &	localSize
	)

pure virtual

Runs the specified kernel definition using provided thread sizes.

Parameters

id	Id of kernel definition which will be run. The specified definition must be included in the currently launched kernel.
globalSize	Dimensions for global size with which the kernel will be run.
localSize	Dimensions for local size with which the kernel will be run.

◆ RunKernelAsync() [1/2]

ComputeActionId ktt::ComputeInterface::RunKernelAsync	(	const KernelDefinitionId	id,
		const QueueId	queue
	)

pure virtual

Runs the specified kernel definition using thread sizes based on the current configuration. The kernel will be launched asynchronously in the specified queue.

Parameters

id	Id of kernel definition which will be run. The specified definition must be included in the currently launched kernel.
queue	Id of queue in which the command to run kernel will be submitted.

Returns: Id of asynchronous action corresponding to the issued kernel run command. The action must be waited for with e.g., WaitForComputeAction(), SynchronizeQueue() methods. Otherwise, problems such as incorrectly recorded kernel durations may occur.

◆ RunKernelAsync() [2/2]

ComputeActionId ktt::ComputeInterface::RunKernelAsync	(	const KernelDefinitionId	id,
		const QueueId	queue,
		const DimensionVector &	globalSize,
		const DimensionVector &	localSize
	)

pure virtual

Runs the specified kernel definition using provided thread sizes. The kernel will be launched asynchronously in the specified queue.

Parameters

id	Id of kernel definition which will be run. The specified definition must be included in the currently launched kernel.
queue	Id of queue in which the command to run kernel will be submitted.
globalSize	Dimensions for global size with which the kernel will be run.
localSize	Dimensions for local size with which the kernel will be run.

Returns: Id of asynchronous action corresponding to the issued kernel run command. The action must be waited for with e.g., WaitForComputeAction(), SynchronizeQueue() methods. Otherwise, problems such as incorrectly recorded kernel durations may occur.

◆ RunKernelWithProfiling() [1/2]

void ktt::ComputeInterface::RunKernelWithProfiling ( const KernelDefinitionId id )

pure virtual

Runs the specified kernel definition using thread sizes based on the current configuration. Collection of kernel profiling counters will be enabled for this run which means that performance will be decreased. Running kernels with profiling will always cause implicit device synchronization before and after the kernel run is finished.

Parameters

id	Id of kernel definition which will be run. The specified definition must be included in the currently launched kernel.

◆ RunKernelWithProfiling() [2/2]

void ktt::ComputeInterface::RunKernelWithProfiling	(	const KernelDefinitionId	id,
		const DimensionVector &	globalSize,
		const DimensionVector &	localSize
	)

pure virtual

Runs the specified kernel definition using provided thread sizes. Collection of kernel profiling counters will be enabled for this run which means that performance will be decreased. Running kernels with profiling will always cause implicit device synchronization before and after the kernel run is finished.

Parameters

id	Id of kernel definition which will be run. The specified definition must be included in the currently launched kernel.
globalSize	Dimensions for global size with which the kernel will be run.
localSize	Dimensions for local size with which the kernel will be run.

◆ SwapArguments()

void ktt::ComputeInterface::SwapArguments	(	const KernelDefinitionId	id,
		const ArgumentId &	first,
		const ArgumentId &	second
	)

pure virtual

Swaps positions of kernel arguments for the specified kernel definition under currently launched kernel.

Parameters

id	Id of kernel definition for which the arguments will be swapped. The specified definition must be included in the currently launched kernel.
first	Id of the first argument which will be swapped.
second	Id of the second argument which will be swapped.

◆ SynchronizeDevice()

void ktt::ComputeInterface::SynchronizeDevice ( )

pure virtual

Blocks until all commands submitted to device are completed.

◆ SynchronizeQueue()

void ktt::ComputeInterface::SynchronizeQueue ( const QueueId queue )

pure virtual

Blocks until all commands submitted to the specified device queue are completed.

Parameters

queue Id of queue which will be synchronized.

◆ SynchronizeQueues()

void ktt::ComputeInterface::SynchronizeQueues ( )

pure virtual

Blocks until all commands submitted to all device queues are completed.

◆ UpdateBuffer()

void ktt::ComputeInterface::UpdateBuffer	(	const ArgumentId &	id,
		const void *	data,
		const size_t	dataSize = `0`
	)

pure virtual

Updates data in compute buffer of the specified vector argument.

Parameters

id	Id of vector argument which will be updated.
data	Pointer to new data for vector argument. Its size must be equal or greater than the specified data size. The data must have matching kernel argument data type.
dataSize	Size in bytes of buffer portion which will be updated, starting with the first byte. If zero, the entire buffer will be updated.

◆ UpdateBufferAsync()

TransferActionId ktt::ComputeInterface::UpdateBufferAsync	(	const ArgumentId &	id,
		const QueueId	queue,
		const void *	data,
		const size_t	dataSize = `0`
	)

pure virtual

Updates data in compute buffer of the specified vector argument. The data will be transferred asynchronously in the specified queue.

Parameters

id	Id of vector argument which will be updated.
queue	Id of queue in which the command to update argument will be submitted.
data	Pointer to new data for vector argument. Its size must be equal or greater than the specified data size. The data must have matching kernel argument data type.
dataSize	Size in bytes of buffer portion which will be updated, starting with the first byte. If zero, the entire buffer will be updated.

Returns: Id of asynchronous action corresponding to the issued data transfer command. The action must be waited for with e.g., WaitForTransferAction(), SynchronizeQueue() methods. Otherwise, problems such as incorrectly recorded transfer durations may occur.

◆ UpdateLocalArgument()

void ktt::ComputeInterface::UpdateLocalArgument	(	const ArgumentId &	id,
		const size_t	dataSize
	)

pure virtual

Updates the specified local memory argument under currently launched kernel.

Parameters

id	Id of local memory argument which will be updated.
dataSize	New size in bytes for the argument.

◆ UpdateScalarArgument()

void ktt::ComputeInterface::UpdateScalarArgument	(	const ArgumentId &	id,
		const void *	data
	)

pure virtual

Updates the specified scalar argument under currently launched kernel.

Parameters

id	Id of scalar argument which will be updated.
data	Pointer to new data for scalar argument. The data must have matching kernel argument data type.

◆ UploadBuffer()

void ktt::ComputeInterface::UploadBuffer ( const ArgumentId & id )

pure virtual

Uploads the specified vector argument into compute buffer. This method should be used mainly with arguments with ArgumentManagementType set to User.

Parameters

id	Id of vector argument which will be uploaded.

◆ UploadBufferAsync()

TransferActionId ktt::ComputeInterface::UploadBufferAsync	(	const ArgumentId &	id,
		const QueueId	queue
	)

pure virtual

Uploads the specified vector argument into compute buffer. The data will be transferred asynchronously in the specified queue.

Parameters

id	Id of vector argument which will be uploaded.
queue	Id of queue in which the command to upload argument will be submitted.

Returns: Id of asynchronous action corresponding to the issued data transfer command. The action must be waited for with e.g., WaitForTransferAction(), SynchronizeQueue() methods. Otherwise, problems such as incorrectly recorded transfer durations may occur.

◆ WaitForComputeAction()

void ktt::ComputeInterface::WaitForComputeAction ( const ComputeActionId id )

pure virtual

Blocks until the specified compute action is finished.

Parameters

id	Id of compute action to wait for.

◆ WaitForTransferAction()

void ktt::ComputeInterface::WaitForTransferAction ( const TransferActionId id )

pure virtual

Blocks until the specified buffer transfer action is finished.

Parameters

id	Id of transfer action to wait for.

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

Source/Api/ComputeInterface.h

Public Member Functions

Detailed Description

Constructor & Destructor Documentation

◆ ~ComputeInterface()

Member Function Documentation

◆ ChangeArguments()

◆ ClearBuffer()

◆ CopyBuffer()

◆ CopyBufferAsync()

◆ DownloadBuffer()

◆ DownloadBufferAsync()

◆ GetAllQueues()

◆ GetCurrentConfiguration()

◆ GetCurrentGlobalSize()

◆ GetCurrentLocalSize()

◆ GetDefaultQueue()

◆ GetProfiling()

◆ GetRemainingProfilingRuns() [1/2]

◆ GetRemainingProfilingRuns() [2/2]

◆ GetRunMode()

◆ GetUnifiedMemoryBufferHandle()

◆ HasBuffer()

◆ ResizeBuffer()

◆ RunKernel() [1/2]

◆ RunKernel() [2/2]

◆ RunKernelAsync() [1/2]

◆ RunKernelAsync() [2/2]

◆ RunKernelWithProfiling() [1/2]

◆ RunKernelWithProfiling() [2/2]

◆ SwapArguments()

◆ SynchronizeDevice()

◆ SynchronizeQueue()

◆ SynchronizeQueues()

◆ UpdateBuffer()

◆ UpdateBufferAsync()

◆ UpdateLocalArgument()

◆ UpdateScalarArgument()

◆ UploadBuffer()

◆ UploadBufferAsync()

◆ WaitForComputeAction()

◆ WaitForTransferAction()