Namespace containing OpenCL backend specific implementation details. Should not directly be used by users. More...

Classes
class	command_queue
	RAII wrapper class around a cl_command_queue. More...

class	context
	RAII wrapper class around a cl_context. More...

class	device_ptr
	Small wrapper class around an OpenCL device pointer together with commonly used device functions. More...

class	error_code
	Class wrapping an OpenCL error code. More...

class	kernel
	RAII wrapper class around a cl_kernel. More...

Enumerations
enum class	compute_kernel_name { q_kernel , svm_kernel , w_kernel , predict_kernel }
	Enum class for all different OpenCL compute kernels. More...

Functions
std::ostream &	operator<< (std::ostream &out, error_code ec)
	Output the error code encapsulated by `ev` to the given output-stream `out`. More...

bool	operator== (error_code lhs, error_code rhs) noexcept
	Compares two error codes for equality. More...

bool	operator!= (error_code lhs, error_code rhs) noexcept
	Compares two error codes for inequality. More...

void	device_assert (error_code code, std::string_view msg="")
	Check the OpenCL error `code`. If `code` signals an error, throw a plssvm::opencl::backend_exception. More...

std::pair< std::vector< context >, target_platform >	get_contexts (target_platform target)
	Returns the context listing all devices matching the target platform `target` and the actually used target platform (only interesting if the provided `target` was automatic). More...

void	device_synchronize (const command_queue &queue)
	Wait for the compute device associated with `queue` to finish. More...

std::string	get_device_name (const command_queue &queue)
	Get the name of the device associated with the OpenCL command queue `queue`. More...

std::vector< std::pair< compute_kernel_name, std::string > >	kernel_type_to_function_names (kernel_function_type kernel)
	Convert the kernel type `kernel` to the device function names and return the plssvm::opencl::detail::compute_kernel_name identifier. More...

std::vector< command_queue >	create_command_queues (const std::vector< context > &contexts, target_platform target, const std::vector< std::pair< compute_kernel_name, std::string >> &kernel_names)
	Create command queues for all devices in the OpenCL `contexts` with respect to `target` given and the associated compute kernels with respect to the given source files and kernel function names. Always builds kernels for `float` and `double` precision floating point types. More...

template<typename... Args>
void	set_kernel_args (cl_kernel kernel, Args... args)
	Set all arguments in the parameter pack `args` for the kernel `kernel`. More...

template<typename... Args>
void	run_kernel (const command_queue &queue, cl_kernel kernel, const std::vector< std::size_t > &grid_size, const std::vector< std::size_t > &block_size, Args &&...args)
	Run the 1D `kernel` on the `queue` with the additional parameters `args`. More...

template<typename... Args>
void	run_kernel (const command_queue &queue, cl_kernel kernel, std::size_t grid_size, std::size_t block_size, Args &&...args)
	Run the 1D `kernel` on the `queue` with the additional parameters `args`. More...

Detailed Description

Namespace containing OpenCL backend specific implementation details. Should not directly be used by users.

Enumeration Type Documentation

◆ compute_kernel_name

enum plssvm::opencl::detail::compute_kernel_name

strong

Enum class for all different OpenCL compute kernels.

Used to distinguish kernels in the plssvm::opencl::detail::command_queue class.

Enumerator
q_kernel	The kernels to generate the `q` vector.
svm_kernel	The main C-SVM kernel.
w_kernel	The predict kernel for the linear kernel function.
predict_kernel	The predict kernels for the polynomial and rbf kernel functions.

Function Documentation

◆ operator<<()

std::ostream& plssvm::opencl::detail::operator<<	(	std::ostream &	out,
		error_code	ec
	)

Output the error code encapsulated by ev to the given output-stream out.

Example output of an error code:

"-1: CL_DEVICE_NOT_FOUND"

Parameters

[in,out]	out	the output-stream to write the error code to
[in]	ec	the error code

Returns: the output-stream

◆ operator==()

bool plssvm::opencl::detail::operator==	(	error_code	lhs,
		error_code	rhs
	)

noexcept

Compares two error codes for equality.

Parameters

[in]	lhs	the first error code
[in]	rhs	the second error code

Returns: true if both error codes are equal, false otherwise ([[nodiscard]])

◆ operator!=()

bool plssvm::opencl::detail::operator!=	(	error_code	lhs,
		error_code	rhs
	)

noexcept

Compares two error codes for inequality.

Parameters

[in]	lhs	the first error code
[in]	rhs	the second error code

Returns: true if both error codes are unequal, false otherwise ([[nodiscard]])

◆ device_assert()

void plssvm::opencl::detail::device_assert	(	error_code	code,
		std::string_view	msg = `""`
	)

Check the OpenCL error code. If code signals an error, throw a plssvm::opencl::backend_exception.

The exception contains the following message: "OpenCL assert 'OPENCL_ERROR_NAME' (OPENCL_ERROR_CODE): OPTIONAL_OPENCL_ERROR_STRING".

Parameters

[in]	code	the OpenCL error code to check
[in]	msg	optional message printed if the error code check failed

Exceptions

plssvm::opencl::backend_exception if the error code signals a failure

◆ get_contexts()

std::pair<std::vector<context>, target_platform> plssvm::opencl::detail::get_contexts ( target_platform target )

Returns the context listing all devices matching the target platform target and the actually used target platform (only interesting if the provided target was automatic).

If the selected target platform is plssvm::target_platform::automatic the selector tries to find devices according to plssvm::determine_default_target_platform.

Parameters

[in] target the target platform for which the devices must match

Returns: the command queues and used target platform ([[nodiscard]])

◆ device_synchronize()

void plssvm::opencl::detail::device_synchronize ( const command_queue & queue )

Wait for the compute device associated with queue to finish.

Parameters

[in] queue the command queue to synchronize

◆ get_device_name()

std::string plssvm::opencl::detail::get_device_name ( const command_queue & queue )

Get the name of the device associated with the OpenCL command queue queue.

Parameters

[in] queue the OpenCL command queue

Returns: the device name ([[nodiscard]])

◆ kernel_type_to_function_names()

std::vector<std::pair<compute_kernel_name, std::string> > plssvm::opencl::detail::kernel_type_to_function_names ( kernel_function_type kernel )

Convert the kernel type kernel to the device function names and return the plssvm::opencl::detail::compute_kernel_name identifier.

Parameters

[in] kernel the kernel type

Returns: the kernel function names with the respective plssvm::opencl::detail::compute_kernel_name identifier ([[nodiscard]])

◆ create_command_queues()

std::vector<command_queue> plssvm::opencl::detail::create_command_queues	(	const std::vector< context > &	contexts,
		target_platform	target,
		const std::vector< std::pair< compute_kernel_name, std::string >> &	kernel_names
	)

Create command queues for all devices in the OpenCL contexts with respect to target given and the associated compute kernels with respect to the given source files and kernel function names. Always builds kernels for float and double precision floating point types.

Manually caches the OpenCL JIT compiled code in the current $TEMP directory (no special means to prevent race conditions are implemented). The cached binaries are reused if:

the cached files already exist
the number of cached files match the number of needed files
the kernel source checksum matches (no changes in the source files since the last caching)

A custom SHA256 implementation is used to detect changes in the OpenCL kernel source files. Additionally, adds the path to the currently used OpenCL library as a comment to the kernel source string (before the checksum calculation) to detect changes in the used OpenCL implementation and trigger a kernel rebuild.

Parameters

[in]	contexts	the used OpenCL contexts
[in]	target	the target platform
[in]	kernel_names	all kernel name for which an OpenCL cl_kernel should be build

Exceptions

plssvm::invalid_file_format_exception if the file couldn't be read using std::ifstream::read

Returns: the command queues with all necessary kernels ([[nodiscard]])

◆ set_kernel_args()

template<typename... Args>

void plssvm::opencl::detail::set_kernel_args	(	cl_kernel	kernel,
		Args...	args
	)

inline

Set all arguments in the parameter pack args for the kernel kernel.

Template Parameters

Args	the types of the arguments

Parameters

[in]	kernel	the OpenCL kernel to set the arguments
[in]	args	the arguments to set

◆ run_kernel() [1/2]

template<typename... Args>

void plssvm::opencl::detail::run_kernel	(	const command_queue &	queue,
		cl_kernel	kernel,
		const std::vector< std::size_t > &	grid_size,
		const std::vector< std::size_t > &	block_size,
		Args &&...	args
	)

inline

Run the 1D kernel on the queue with the additional parameters args.

Template Parameters

Args	the types of the arguments

Parameters

[in]	queue	the command queue on which the kernel should be executed
[in]	kernel	the kernel to run
[in]	grid_size	the number of global work-items (possibly multi-dimensional)
[in]	block_size	the number of work-items that make up a work-group (possibly multi-dimensional)
[in]	args	the arguments to set

◆ run_kernel() [2/2]

template<typename... Args>

void plssvm::opencl::detail::run_kernel	(	const command_queue &	queue,
		cl_kernel	kernel,
		std::size_t	grid_size,
		std::size_t	block_size,
		Args &&...	args
	)

inline

Run the 1D kernel on the queue with the additional parameters args.

Template Parameters

Args	the types of the arguments

Parameters

[in]	queue	the command queue on which the kernel should be executed
[in]	kernel	the kernel to run
[in]	grid_size	the number of global work-items
[in]	block_size	the number of work-items that make up a work-group
[in]	args	the arguments to set

Classes

Enumerations

Functions

Detailed Description

Enumeration Type Documentation

◆ compute_kernel_name

Function Documentation

◆ operator<<()

◆ operator==()

◆ operator!=()

◆ device_assert()

◆ get_contexts()

◆ device_synchronize()

◆ get_device_name()

◆ kernel_type_to_function_names()

◆ create_command_queues()

◆ set_kernel_args()

◆ run_kernel() [1/2]

◆ run_kernel() [2/2]