Implementation of a PI Kernel for CUDA. More...

#include <cuda/pi_cuda.hpp>

Collaboration diagram for _pi_kernel:

[legend]

Classes
struct	ArgumentInfo

struct	arguments
	Structure that holds the arguments to the kernel. More...

struct	Hash

Public Types
using	native_type = CUfunction

using	native_type = hipFunction_t

Public Member Functions
	_pi_kernel (CUfunction func, CUfunction funcWithOffsetParam, const char *name, pi_program program, pi_context ctxt)

	~_pi_kernel ()

pi_program	get_program () const noexcept

pi_uint32	increment_reference_count () noexcept

pi_uint32	decrement_reference_count () noexcept

pi_uint32	get_reference_count () const noexcept

native_type	get () const noexcept

native_type	get_with_offset_parameter () const noexcept

bool	has_with_offset_parameter () const noexcept

pi_context	get_context () const noexcept

const char *	get_name () const noexcept

pi_uint32	get_num_args () const noexcept
	Returns the number of arguments, excluding the implicit global offset. More...

void	set_kernel_arg (int index, size_t size, const void *arg)

void	set_kernel_local_arg (int index, size_t size)

void	set_implicit_offset_arg (size_t size, std::uint32_t *implicitOffset)

const arguments::args_index_t &	get_arg_indices () const

pi_uint32	get_local_size () const noexcept

void	clear_local_size ()

	_pi_kernel ()

	_pi_kernel (hipFunction_t func, hipFunction_t funcWithOffsetParam, const char *name, pi_program program, pi_context ctxt)

	_pi_kernel (hipFunction_t func, const char *name, pi_program program, pi_context ctxt)

	~_pi_kernel ()

pi_program	get_program () const noexcept

pi_uint32	increment_reference_count () noexcept

pi_uint32	decrement_reference_count () noexcept

pi_uint32	get_reference_count () const noexcept

native_type	get () const noexcept

native_type	get_with_offset_parameter () const noexcept

bool	has_with_offset_parameter () const noexcept

pi_context	get_context () const noexcept

const char *	get_name () const noexcept

pi_uint32	get_num_args () const noexcept
	Returns the number of arguments, excluding the implicit global offset. More...

void	set_kernel_arg (int index, size_t size, const void *arg)

void	set_kernel_local_arg (int index, size_t size)

void	set_implicit_offset_arg (size_t size, std::uint32_t *implicitOffset)

arguments::args_index_t	get_arg_indices () const

pi_uint32	get_local_size () const noexcept

void	clear_local_size ()

	_pi_kernel (ze_kernel_handle_t Kernel, bool OwnZeKernel, pi_program Program)

pi_result	initialize ()

bool	hasIndirectAccess ()

Public Attributes
native_type	function_

native_type	functionWithOffsetParam_

std::string	name_

pi_context	context_

pi_program	program_

std::atomic_uint32_t	refCount_

size_t	reqdThreadsPerBlock_ [REQD_THREADS_PER_BLOCK_DIMENSIONS]

struct _pi_kernel::arguments	args_

ze_kernel_handle_t	ZeKernel

bool	OwnZeKernel

pi_program	Program

std::unordered_set< std::pair< void const, MemAllocRecord > , Hash >	MemAllocs

std::atomic< pi_uint32 >	SubmissionsCount

std::vector< ArgumentInfo >	PendingArguments

ZeCache< ZeStruct< ze_kernel_properties_t > >	ZeKernelProperties

ZeCache< std::string >	ZeKernelName

Static Public Attributes
static constexpr pi_uint32	REQD_THREADS_PER_BLOCK_DIMENSIONS = 3u

Detailed Description

Implementation of a PI Kernel for CUDA.

Implementation of a PI Kernel for HIP.

PI Kernels are used to set kernel arguments, creating a state on the Kernel object for a given invocation. This is not the case of CUFunction objects, which are simply passed together with the arguments on the invocation. The PI Kernel implementation for CUDA stores the list of arguments, argument sizes and offsets to emulate the interface of PI Kernel, saving the arguments for the later dispatch. Note that in PI API, the Local memory is specified as a size per individual argument, but in CUDA only the total usage of shared memory is required since it is not passed as a parameter. A compiler pass converts the PI API local memory model into the CUDA shared model. This object simply calculates the total of shared memory, and the initial offsets of each parameter.

PI Kernels are used to set kernel arguments, creating a state on the Kernel object for a given invocation. This is not the case of HIPFunction objects, which are simply passed together with the arguments on the invocation. The PI Kernel implementation for HIP stores the list of arguments, argument sizes and offsets to emulate the interface of PI Kernel, saving the arguments for the later dispatch. Note that in PI API, the Local memory is specified as a size per individual argument, but in HIP only the total usage of shared memory is required since it is not passed as a parameter. A compiler pass converts the PI API local memory model into the HIP shared model. This object simply calculates the total of shared memory, and the initial offsets of each parameter.

Definition at line 774 of file pi_cuda.hpp.

Member Typedef Documentation

◆ native_type [1/2]

using _pi_kernel::native_type = hipFunction_t

Definition at line 733 of file pi_hip.hpp.

◆ native_type [2/2]

using _pi_kernel::native_type = CUfunction

Definition at line 775 of file pi_cuda.hpp.

Constructor & Destructor Documentation

◆ _pi_kernel() [1/5]

_pi_kernel::_pi_kernel	(	CUfunction	func,
		CUfunction	funcWithOffsetParam,
		const char *	name,
		pi_program	program,
		pi_context	ctxt
	)

inline

Note: this code assumes that there is only one device per context

Definition at line 872 of file pi_cuda.hpp.

◆ ~_pi_kernel() [1/2]

_pi_kernel::~_pi_kernel ( )

inline

Definition at line 886 of file pi_cuda.hpp.

References context_, cuda_piContextRelease(), cuda_piProgramRelease(), and program_.

◆ _pi_kernel() [2/5]

_pi_kernel::_pi_kernel ( )

inline

Definition at line 218 of file pi_esimd_emulator.hpp.

◆ _pi_kernel() [3/5]

_pi_kernel::_pi_kernel	(	hipFunction_t	func,
		hipFunction_t	funcWithOffsetParam,
		const char *	name,
		pi_program	program,
		pi_context	ctxt
	)

inline

Definition at line 827 of file pi_hip.hpp.

◆ _pi_kernel() [4/5]

_pi_kernel::_pi_kernel	(	hipFunction_t	func,
		const char *	name,
		pi_program	program,
		pi_context	ctxt
	)

inline

Definition at line 835 of file pi_hip.hpp.

◆ ~_pi_kernel() [2/2]

_pi_kernel::~_pi_kernel ( )

inline

Definition at line 839 of file pi_hip.hpp.

References context_, hip_piContextRelease(), hip_piProgramRelease(), and program_.

◆ _pi_kernel() [5/5]

_pi_kernel::_pi_kernel	(	ze_kernel_handle_t	Kernel,
		bool	OwnZeKernel,
		pi_program	Program
	)

inline

Definition at line 1526 of file pi_level_zero.hpp.

Member Function Documentation

◆ clear_local_size() [1/2]

void _pi_kernel::clear_local_size ( )

inline

Definition at line 890 of file pi_hip.hpp.

References args_, and _pi_kernel::arguments::clear_local_size().

◆ clear_local_size() [2/2]

void _pi_kernel::clear_local_size ( )

inline

Definition at line 937 of file pi_cuda.hpp.

References args_, and _pi_kernel::arguments::clear_local_size().

◆ decrement_reference_count() [1/2]

pi_uint32 _pi_kernel::decrement_reference_count ( )

inlinenoexcept

Definition at line 848 of file pi_hip.hpp.

References refCount_.

◆ decrement_reference_count() [2/2]

pi_uint32 _pi_kernel::decrement_reference_count ( )

inlinenoexcept

Definition at line 895 of file pi_cuda.hpp.

References refCount_.

◆ get() [1/2]

native_type _pi_kernel::get ( ) const

inlinenoexcept

Definition at line 852 of file pi_hip.hpp.

References function_.

◆ get() [2/2]

native_type _pi_kernel::get ( ) const

inlinenoexcept

Definition at line 899 of file pi_cuda.hpp.

References function_.

◆ get_arg_indices() [1/2]

arguments::args_index_t _pi_kernel::get_arg_indices ( ) const

inline

Definition at line 884 of file pi_hip.hpp.

References args_, and _pi_kernel::arguments::get_indices().

◆ get_arg_indices() [2/2]

const arguments::args_index_t& _pi_kernel::get_arg_indices ( ) const

inline

Definition at line 931 of file pi_cuda.hpp.

References args_, and _pi_kernel::arguments::get_indices().

◆ get_context() [1/2]

pi_context _pi_kernel::get_context ( ) const

inlinenoexcept

Definition at line 862 of file pi_hip.hpp.

References context_.

◆ get_context() [2/2]

pi_context _pi_kernel::get_context ( ) const

inlinenoexcept

Definition at line 909 of file pi_cuda.hpp.

References context_.

◆ get_local_size() [1/2]

pi_uint32 _pi_kernel::get_local_size ( ) const

inlinenoexcept

Definition at line 888 of file pi_hip.hpp.

References args_, and _pi_kernel::arguments::get_local_size().

◆ get_local_size() [2/2]

pi_uint32 _pi_kernel::get_local_size ( ) const

inlinenoexcept

Definition at line 935 of file pi_cuda.hpp.

References args_, and _pi_kernel::arguments::get_local_size().

◆ get_name() [1/2]

const char* _pi_kernel::get_name ( ) const

inlinenoexcept

Definition at line 864 of file pi_hip.hpp.

References name_.

◆ get_name() [2/2]

const char* _pi_kernel::get_name ( ) const

inlinenoexcept

Definition at line 911 of file pi_cuda.hpp.

References name_.

◆ get_num_args() [1/2]

pi_uint32 _pi_kernel::get_num_args ( ) const

inlinenoexcept

Returns the number of arguments, excluding the implicit global offset.

Note this only returns the current known number of arguments, not the real one required by the kernel, since this cannot be queried from the HIP Driver API

Definition at line 870 of file pi_hip.hpp.

References args_, and _pi_kernel::arguments::indices_.

◆ get_num_args() [2/2]

pi_uint32 _pi_kernel::get_num_args ( ) const

inlinenoexcept

Returns the number of arguments, excluding the implicit global offset.

Note this only returns the current known number of arguments, not the real one required by the kernel, since this cannot be queried from the CUDA Driver API

Definition at line 917 of file pi_cuda.hpp.

References args_, and _pi_kernel::arguments::indices_.

◆ get_program() [1/2]

pi_program _pi_kernel::get_program ( ) const

inlinenoexcept

Definition at line 844 of file pi_hip.hpp.

References program_.

◆ get_program() [2/2]

pi_program _pi_kernel::get_program ( ) const

inlinenoexcept

Definition at line 891 of file pi_cuda.hpp.

References program_.

◆ get_reference_count() [1/2]

pi_uint32 _pi_kernel::get_reference_count ( ) const

inlinenoexcept

Definition at line 850 of file pi_hip.hpp.

References refCount_.

◆ get_reference_count() [2/2]

pi_uint32 _pi_kernel::get_reference_count ( ) const

inlinenoexcept

Definition at line 897 of file pi_cuda.hpp.

References refCount_.

◆ get_with_offset_parameter() [1/2]

native_type _pi_kernel::get_with_offset_parameter ( ) const

inlinenoexcept

Definition at line 854 of file pi_hip.hpp.

References functionWithOffsetParam_.

◆ get_with_offset_parameter() [2/2]

native_type _pi_kernel::get_with_offset_parameter ( ) const

inlinenoexcept

Definition at line 901 of file pi_cuda.hpp.

References functionWithOffsetParam_.

◆ has_with_offset_parameter() [1/2]

bool _pi_kernel::has_with_offset_parameter ( ) const

inlinenoexcept

Definition at line 858 of file pi_hip.hpp.

References functionWithOffsetParam_.

◆ has_with_offset_parameter() [2/2]

bool _pi_kernel::has_with_offset_parameter ( ) const

inlinenoexcept

Definition at line 905 of file pi_cuda.hpp.

References functionWithOffsetParam_.

◆ hasIndirectAccess()

bool _pi_kernel::hasIndirectAccess ( )

inline

Definition at line 1534 of file pi_level_zero.hpp.

◆ increment_reference_count() [1/2]

pi_uint32 _pi_kernel::increment_reference_count ( )

inlinenoexcept

Definition at line 846 of file pi_hip.hpp.

References refCount_.

◆ increment_reference_count() [2/2]

pi_uint32 _pi_kernel::increment_reference_count ( )

inlinenoexcept

Definition at line 893 of file pi_cuda.hpp.

References refCount_.

◆ initialize()

pi_result _pi_kernel::initialize ( )

Definition at line 5021 of file pi_level_zero.cpp.

References _pi_program::Context, PI_CALL, piContextRetain(), piProgramRetain(), _pi_queue::Properties, and ZE_CALL_NOCHECK.

◆ set_implicit_offset_arg() [1/2]

void _pi_kernel::set_implicit_offset_arg	(	size_t	size,
		std::uint32_t *	implicitOffset
	)

inline

Definition at line 880 of file pi_hip.hpp.

References args_, and _pi_kernel::arguments::set_implicit_offset().

◆ set_implicit_offset_arg() [2/2]

void _pi_kernel::set_implicit_offset_arg	(	size_t	size,
		std::uint32_t *	implicitOffset
	)

inline

Definition at line 927 of file pi_cuda.hpp.

References args_, and _pi_kernel::arguments::set_implicit_offset().

◆ set_kernel_arg() [1/2]

void _pi_kernel::set_kernel_arg	(	int	index,
		size_t	size,
		const void *	arg
	)

inline

Definition at line 872 of file pi_hip.hpp.

References _pi_kernel::arguments::add_arg(), and args_.

◆ set_kernel_arg() [2/2]

void _pi_kernel::set_kernel_arg	(	int	index,
		size_t	size,
		const void *	arg
	)

inline

Definition at line 919 of file pi_cuda.hpp.

References _pi_kernel::arguments::add_arg(), and args_.

◆ set_kernel_local_arg() [1/2]

void _pi_kernel::set_kernel_local_arg	(	int	index,
		size_t	size
	)

inline

Definition at line 876 of file pi_hip.hpp.

References _pi_kernel::arguments::add_local_arg(), and args_.

◆ set_kernel_local_arg() [2/2]

void _pi_kernel::set_kernel_local_arg	(	int	index,
		size_t	size
	)

inline

Definition at line 923 of file pi_cuda.hpp.

References _pi_kernel::arguments::add_local_arg(), and args_.

Member Data Documentation

◆ args_

struct _pi_kernel::arguments _pi_kernel::args_

Referenced by clear_local_size(), get_arg_indices(), get_local_size(), get_num_args(), set_implicit_offset_arg(), set_kernel_arg(), and set_kernel_local_arg().

◆ context_

pi_context _pi_kernel::context_

Definition at line 780 of file pi_cuda.hpp.

Referenced by get_context(), and ~_pi_kernel().

◆ function_

native_type _pi_kernel::function_

Definition at line 777 of file pi_cuda.hpp.

Referenced by get().

◆ functionWithOffsetParam_

native_type _pi_kernel::functionWithOffsetParam_

Definition at line 778 of file pi_cuda.hpp.

Referenced by get_with_offset_parameter(), and has_with_offset_parameter().

◆ MemAllocs

std::unordered_set<std::pair<void *const, MemAllocRecord> *, Hash> _pi_kernel::MemAllocs

Definition at line 1572 of file pi_level_zero.hpp.

Referenced by piKernelRelease().

◆ name_

std::string _pi_kernel::name_

Definition at line 779 of file pi_cuda.hpp.

Referenced by get_name().

◆ OwnZeKernel

bool _pi_kernel::OwnZeKernel

Definition at line 1545 of file pi_level_zero.hpp.

Referenced by piKernelRelease().

◆ PendingArguments

std::vector<ArgumentInfo> _pi_kernel::PendingArguments

Definition at line 1595 of file pi_level_zero.hpp.

Referenced by piEnqueueKernelLaunch(), and piextKernelSetArgMemObj().

◆ Program

pi_program _pi_kernel::Program

Definition at line 1548 of file pi_level_zero.hpp.

Referenced by piEnqueueKernelLaunch(), piKernelGetInfo(), and piKernelRelease().

◆ program_

pi_program _pi_kernel::program_

Definition at line 781 of file pi_cuda.hpp.

Referenced by get_program(), and ~_pi_kernel().

◆ refCount_

std::atomic_uint32_t _pi_kernel::refCount_

Definition at line 782 of file pi_cuda.hpp.

Referenced by decrement_reference_count(), get_reference_count(), and increment_reference_count().

◆ REQD_THREADS_PER_BLOCK_DIMENSIONS

constexpr pi_uint32 _pi_kernel::REQD_THREADS_PER_BLOCK_DIMENSIONS = 3u

staticconstexpr

Definition at line 784 of file pi_cuda.hpp.

◆ reqdThreadsPerBlock_

size_t _pi_kernel::reqdThreadsPerBlock_[REQD_THREADS_PER_BLOCK_DIMENSIONS]

Definition at line 785 of file pi_cuda.hpp.

◆ SubmissionsCount

std::atomic<pi_uint32> _pi_kernel::SubmissionsCount

Definition at line 1583 of file pi_level_zero.hpp.

Referenced by piKernelRelease().

◆ ZeKernel

ze_kernel_handle_t _pi_kernel::ZeKernel

Definition at line 1541 of file pi_level_zero.hpp.

Referenced by piEnqueueKernelLaunch(), piextKernelGetNativeHandle(), piextKernelSetArgSampler(), piKernelGetGroupInfo(), piKernelGetInfo(), piKernelRelease(), piKernelSetArg(), and piKernelSetExecInfo().

◆ ZeKernelName

ZeCache<std::string> _pi_kernel::ZeKernelName

Definition at line 1599 of file pi_level_zero.hpp.

Referenced by piKernelGetInfo().

◆ ZeKernelProperties

ZeCache<ZeStruct<ze_kernel_properties_t> > _pi_kernel::ZeKernelProperties

Definition at line 1598 of file pi_level_zero.hpp.

Referenced by piKernelGetGroupInfo(), piKernelGetInfo(), and piKernelGetSubGroupInfo().

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

plugins/cuda/pi_cuda.hpp
plugins/hip/pi_hip.hpp
plugins/level_zero/pi_level_zero.hpp
plugins/esimd_emulator/pi_esimd_emulator.hpp
plugins/level_zero/pi_level_zero.cpp

Classes

Public Types

Public Member Functions

Public Attributes

Static Public Attributes

Detailed Description

Member Typedef Documentation

◆ native_type [1/2]

◆ native_type [2/2]

Constructor & Destructor Documentation

◆ _pi_kernel() [1/5]

◆ ~_pi_kernel() [1/2]

◆ _pi_kernel() [2/5]

◆ _pi_kernel() [3/5]

◆ _pi_kernel() [4/5]

◆ ~_pi_kernel() [2/2]

◆ _pi_kernel() [5/5]

Member Function Documentation

◆ clear_local_size() [1/2]

◆ clear_local_size() [2/2]

◆ decrement_reference_count() [1/2]

◆ decrement_reference_count() [2/2]

◆ get() [1/2]

◆ get() [2/2]

◆ get_arg_indices() [1/2]

◆ get_arg_indices() [2/2]

◆ get_context() [1/2]

◆ get_context() [2/2]

◆ get_local_size() [1/2]

◆ get_local_size() [2/2]

◆ get_name() [1/2]

◆ get_name() [2/2]

◆ get_num_args() [1/2]

◆ get_num_args() [2/2]

◆ get_program() [1/2]

◆ get_program() [2/2]

◆ get_reference_count() [1/2]

◆ get_reference_count() [2/2]

◆ get_with_offset_parameter() [1/2]

◆ get_with_offset_parameter() [2/2]

◆ has_with_offset_parameter() [1/2]

◆ has_with_offset_parameter() [2/2]

◆ hasIndirectAccess()

◆ increment_reference_count() [1/2]

◆ increment_reference_count() [2/2]

◆ initialize()

◆ set_implicit_offset_arg() [1/2]

◆ set_implicit_offset_arg() [2/2]

◆ set_kernel_arg() [1/2]

◆ set_kernel_arg() [2/2]

◆ set_kernel_local_arg() [1/2]

◆ set_kernel_local_arg() [2/2]

Member Data Documentation

◆ args_

◆ context_

◆ function_

◆ functionWithOffsetParam_

◆ MemAllocs

◆ name_

◆ OwnZeKernel

◆ PendingArguments

◆ Program

◆ program_

◆ refCount_

◆ REQD_THREADS_PER_BLOCK_DIMENSIONS

◆ reqdThreadsPerBlock_

◆ SubmissionsCount

◆ ZeKernel

◆ ZeKernelName

◆ ZeKernelProperties