dtfft_nvrtc_module

Module for managing nvRTC compiled CUDA kernels Each module has only one templated kernel that can be instantiated with different parameters

Uses

Help

Graph Key

Nodes of different colours represent the following:

Solid arrows point from a submodule to the (sub)module which it is descended from. Dashed arrows point from a module or program unit to modules which it uses.
Where possible, edges connecting nodes are given different colours to make them easier to distinguish in large graphs.

Used by

Help

Graph Key

Nodes of different colours represent the following:

Solid arrows point from a submodule to the (sub)module which it is descended from. Dashed arrows point from a module or program unit to modules which it uses.
Where possible, edges connecting nodes are given different colours to make them easier to distinguish in large graphs.

Variables

Type	Visibility	Attributes		Name		Initial
character(len=*),	private,	parameter	::	DEFAULT_KERNEL_NAME	=	"dtfft_kernel"	Basic kernel name

Derived Types

type, public :: nvrtc_module

Class for managing nvRTC compiled CUDA kernels

Components

Type	Visibility	Attributes		Name		Initial
logical,	private		::	is_created	=	.false.	Is module created
character(len=:),	private,	allocatable	::	basic_name			Basic kernel name
integer(kind=int32),	private		::	ndims			Number of dimensions, used only for forward permutation
type(CUmodule),	private		::	cumod			CUDA module
type(nvrtcProgram),	private		::	prog			nvRTC program
type(kernel_type_t),	private		::	kernel_type			Type of kernel
integer(kind=int64),	private		::	base_storage			Number of bytes needed to store single element
type(kernel_config),	private,	allocatable	::	configs(:)			Kernel configurations that this module was compiled for

Type-Bound Procedures

procedure, public, pass(self) :: create	../../ Creates module with given parameters
procedure, public, pass(self) :: destroy	../../ Destroys module and frees resources
procedure, public, pass(self) :: get	../../ Returns kernel ready to be executed
generic, public :: check => check_instance, check_module	../../ Checks if kernel is with given parameters is available in this module
procedure, private, pass(self) :: check_instance	../../ Checks if kernel with given parameters is available in this module
procedure, private, pass(self) :: check_module	../../ Basic check that this module provides kernels of given type

type, private, extends(string) :: codegen_t

Class for generating CUDA code

Components

Type	Visibility	Attributes		Name		Initial
character(len=:),	public,	allocatable	::	raw			String

Type-Bound Procedures

procedure, public, pass(self) :: destroy => destroy_string
procedure, public, pass(self) :: add => add_line	../../ Adds new line to CUDA code

Functions

private function get(self, ndims, kernel_type, base_storage, tile_size, block_rows) result(fun)

Returns kernel ready to be executed

Arguments

Type	Intent		Name
class(nvrtc_module),	intent(in)	::	self	This module
integer(kind=int32),	intent(in)	::	ndims	Number of dimensions, used only for forward permutation
type(kernel_type_t),	intent(in)	::	kernel_type	Type of kernel to build
integer(kind=int64),	intent(in)	::	base_storage	Number of bytes needed to store single element
integer(kind=int32),	intent(in)	::	tile_size	Size of shared memory tile, template parameter
integer(kind=int32),	intent(in)	::	block_rows	Number of rows processed by single thread, template parameter

Return Value type(CUfunction)

Resulting kernel

private function check_instance(self, ndims, kernel_type, base_storage, tile_size, block_rows)

Checks if kernel with given parameters is available in this module

Arguments

Type	Intent		Name
class(nvrtc_module),	intent(in)	::	self	This module
integer(kind=int32),	intent(in)	::	ndims	Number of dimensions
type(kernel_type_t),	intent(in)	::	kernel_type	Type of kernel to build
integer(kind=int64),	intent(in)	::	base_storage	Number of bytes needed to store single element
integer(kind=int32),	intent(in)	::	tile_size	Size of shared memory tile, template parameter
integer(kind=int32),	intent(in)	::	block_rows	Number of rows processed by single thread, template parameter

Return Value logical

private function check_module(self, ndims, kernel_type, base_storage)

Basic check that this module provides kernels of given type

Arguments

Type	Intent		Name
class(nvrtc_module),	intent(in)	::	self	This module
integer(kind=int32),	intent(in)	::	ndims	Number of dimensions
type(kernel_type_t),	intent(in)	::	kernel_type	Type of kernel to build
integer(kind=int64),	intent(in)	::	base_storage	Number of bytes needed to store single element

Return Value logical

private function compile_program(code, prog_name, configs, props) result(prog)

Compiles nvRTC program with given configurations

Arguments

Type	Intent		Name
type(codegen_t),	intent(in)	::	code	CUDA code to compile
character(len=*),	intent(in)	::	prog_name	Basic kernel name
type(kernel_config),	intent(in)	::	configs(:)	Kernel configurations that this module should be compiled for
type(device_props),	intent(in)	::	props	GPU architecture properties

Return Value type(nvrtcProgram)

Resulting nvRTC program

private function get_name_expression(basic_name, tile_dim, block_rows, padding) result(expression)

Generates name expression for given template parameters

Arguments

Type	Intent		Name
character(len=*),	intent(in)	::	basic_name	Basic kernel name
integer(kind=int32),	intent(in)	::	tile_dim	Size of shared memory tile, template parameter
integer(kind=int32),	intent(in)	::	block_rows	Number of rows processed by single thread, template parameter
integer(kind=int32),	intent(in)	::	padding	Padding to avoid shared memory bank conflicts, template parameter

Return Value character(len=c_char), allocatable, (:)

Resulting name expression

private function get_mangled_name(basic_name, prog, tile_dim, block_rows, padding) result(mangled)

Gets mangled name for given template parameters from nvRTC program

Arguments

Type	Intent		Name
character(len=*),	intent(in)	::	basic_name	Basic kernel name
type(nvrtcProgram),	intent(in)	::	prog	nvRTC program
integer(kind=int32),	intent(in)	::	tile_dim	Size of shared memory tile, template parameter
integer(kind=int32),	intent(in)	::	block_rows	Number of rows processed by single thread, template parameter
integer(kind=int32),	intent(in)	::	padding	Padding to avoid shared memory bank conflicts, template parameter

Return Value type(c_ptr)

Mangled kernel name

private function get_code(kernel_name, ndims, base_storage, kernel_type) result(code)

Generates code that will be used to locally tranpose data and prepares to send it to other processes

Arguments

Type	Intent		Name
character(len=*),	intent(in)	::	kernel_name	Name of CUDA kernel
integer(kind=int32),	intent(in)	::	ndims	Number of dimensions
integer(kind=int64),	intent(in)	::	base_storage	Number of bytes needed to store single element
type(kernel_type_t),	intent(in)	::	kernel_type	Type of kernel to generate code for

Return Value type(codegen_t)

Resulting code

Subroutines

private subroutine add_line(self, line)

Adds new line to CUDA code

Arguments

Type	Intent	Optional	Attributes		Name
class(codegen_t),	intent(inout)			::	self	Kernel code
character(len=*),	intent(in)			::	line	Line to add

private subroutine create(self, ndims, kernel_type, base_storage, configs, props)

Creates module with given parameters, compiles nvRTC program and loads it as CUDA module

Arguments

Type	Intent		Name
class(nvrtc_module),	intent(inout)	::	self	This module
integer(kind=int32),	intent(in)	::	ndims	Number of dimensions, used only for forward permutation
type(kernel_type_t),	intent(in)	::	kernel_type	Type of kernel to build
integer(kind=int64),	intent(in)	::	base_storage	Number of bytes needed to store single element
type(kernel_config),	intent(in)	::	configs(:)	Kernel configurations that this module should be compiled for
type(device_props),	intent(in)	::	props	GPU architecture properties

private subroutine destroy(self)

Destroys module and frees resources

Arguments

Type	Intent	Optional	Attributes		Name
class(nvrtc_module),	intent(inout)			::	self

private subroutine set_name_expression(prog, basic_name, tile_dim, block_rows, padding)

Sets name expression for given template parameters to nvRTC program

Arguments

Type	Intent		Name
type(nvrtcProgram),	intent(in)	::	prog	nvRTC program
character(len=*),	intent(in)	::	basic_name	Basic kernel name
integer(kind=int32),	intent(in)	::	tile_dim	Size of shared memory tile, template parameter
integer(kind=int32),	intent(in)	::	block_rows	Number of rows processed by single thread, template parameter
integer(kind=int32),	intent(in)	::	padding	Padding to avoid shared memory bank conflicts, template parameter

dtfft_nvrtc_module Module

Contents

Variables

Derived Types

Functions

Subroutines

Uses

Used by

Variables

Derived Types

type, public :: nvrtc_module

Components

Type-Bound Procedures

type, private, extends(string) :: codegen_t

Components

Type-Bound Procedures

Functions

private function get(self, ndims, kernel_type, base_storage, tile_size, block_rows) result(fun)

Arguments

Return Value type(CUfunction)

private function check_instance(self, ndims, kernel_type, base_storage, tile_size, block_rows)

Arguments

Return Value logical

private function check_module(self, ndims, kernel_type, base_storage)

Arguments

Return Value logical

private function compile_program(code, prog_name, configs, props) result(prog)

Arguments

Return Value type(nvrtcProgram)

private function get_name_expression(basic_name, tile_dim, block_rows, padding) result(expression)

Arguments

Return Value character(len=c_char), allocatable, (:)

private function get_mangled_name(basic_name, prog, tile_dim, block_rows, padding) result(mangled)

Arguments

Return Value type(c_ptr)

private function get_code(kernel_name, ndims, base_storage, kernel_type) result(code)

Arguments

Return Value type(codegen_t)

Subroutines

private subroutine add_line(self, line)

Arguments

private subroutine create(self, ndims, kernel_type, base_storage, configs, props)

Arguments

private subroutine destroy(self)

Arguments

private subroutine set_name_expression(prog, basic_name, tile_dim, block_rows, padding)

Arguments