dtfft_nvrtc_kernel

This module describes NVRTC Kernel class nvrtc_kernel It uses caching of compiled kernels to avoid recompilation similar kernels: nvrtc_cache

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
integer(kind=int32),	public,	parameter	::	DEF_TILE_SIZE	=	16	Default tile size
integer(kind=int8),	public,	parameter	::	KERNEL_TRANSPOSE	=	1	Basic transpose kernel type.
integer(kind=int8),	public,	parameter	::	KERNEL_TRANSPOSE_PACKED	=	2	Transposes data and packs it into contiguous buffer. Should be used only in X-Y 3D plans.
integer(kind=int8),	public,	parameter	::	KERNEL_UNPACK	=	3	Unpacks contiguous buffer.
integer(kind=int8),	public,	parameter	::	KERNEL_UNPACK_SIMPLE_COPY	=	4	Doesn’t actually unpacks anything. Performs `cudaMemcpyAsync` call. Should be used only when backend is `DTFFT_GPU_BACKEND_CUFFTMP`.
integer(kind=int8),	public,	parameter	::	KERNEL_UNPACK_PIPELINED	=	5	Unpacks pack of contiguous buffer recieved from rank.
integer(kind=int8),	public,	parameter	::	KERNEL_UNPACK_PARTIAL	=	6	Unpacks contiguous buffer recieved from everyone except myself.
integer(kind=int32),	private,	parameter	::	MIN_TILE_SIZE	=	8	Minimum tile size. Will launch 2 warps
integer(kind=int32),	private,	parameter	::	TARGET_THREADS_PER_BLOCK	=	DEF_TILE_SIZE*DEF_TILE_SIZE	Maximum number of threads to run in a block (256)
character(len=*),	private,	parameter	::	DEFAULT_KERNEL_NAME	=	"dtfft_kernel"	Basic kernel name
integer(kind=int32),	private,	parameter	::	CACHE_PREALLOC_SIZE	=	10	Number of preallocated cache entries
type(nvrtc_cache),	private,	allocatable, save	::	cache(:)			Cache of compiled kernels
integer(kind=int32),	private,	save	::	cache_size	=	0	Number of entries in cache

Derived Types

type, public :: nvrtc_kernel

nvRTC Compiled kernel class

Components

Type	Visibility	Attributes		Name		Initial
logical,	private		::	is_created	=	.false.	Kernel is created flag.
logical,	private		::	is_dummy	=	.false.	If kernel should do anything or not.
type(CUfunction),	private		::	cuda_kernel			Pointer to CUDA kernel.
type(dim3),	private		::	num_blocks			Grid of blocks.
type(dim3),	private		::	block_size			Thread block.
integer(kind=int8),	private		::	kernel_type			Type of kernel to execute.
type(kernelArgs),	private		::	args			Kernel arguments.
integer(kind=int32),	private,	allocatable	::	pointers(:,:)			Optional pointers that hold info about counts and displacements in `KERNEL_UNPACK_PIPELINED` kernel.

Type-Bound Procedures

procedure, public, pass(self) :: create	../../ Creates kernel
procedure, public, pass(self) :: execute	../../ Executes kernel
procedure, public, pass(self) :: destroy	../../ Destroys kernel

type, private :: kernel_code

Class to build CUDA kernel code

Components

Type	Visibility	Attributes		Name		Initial
character(len=:),	private,	allocatable	::	raw			String that holds CUDA code

Type-Bound Procedures

procedure, public, pass(self) :: to_cstr	../../ Converts Fortran CUDA code to C pointer
procedure, public, pass(self) :: add_line	../../ Adds new line to CUDA code
procedure, public, pass(self) :: destroy => destroy_code	../../ Frees all memory

type, private :: nvrtc_cache

Class to cache compiled kernels

Components

Type	Visibility		Name		Initial
integer(kind=int32),	private	::	ref_count	=	0	Number of references to this kernel
type(CUmodule),	private	::	cuda_module	=	CUmodule(c_null_ptr)	Pointer to CUDA Module.
type(CUfunction),	private	::	cuda_kernel	=	CUfunction(c_null_ptr)	Pointer to CUDA kernel.
integer(kind=int8),	private	::	kernel_type			Type of kernel to execute.
type(dtfft_transpose_t),	private	::	transpose_type			Type of transpose
integer(kind=int32),	private	::	tile_size			Tile size of transpose kernel
integer(kind=int64),	private	::	base_storage			Number of bytes needed to store single element
logical,	private	::	has_inner_loop			If kernel has inner loop

Functions

private function get_tile_size(x, y)

Returns tile size to use in a tranpose kernel

Arguments

Type	Intent	Optional	Attributes		Name
integer(kind=int32),	intent(in)			::	x	Number of elements in x direction
integer(kind=int32),	intent(in)			::	y	Number of elements in y direction

Return Value integer(kind=int32)

private function get_cached_kernel(transpose_type, kernel_type, base_storage, tile_size, has_inner_loop) result(kernel)

Returns cached kernel if it exists. If not returns null pointer.

Arguments

Type	Intent		Name
type(dtfft_transpose_t),	intent(in)	::	transpose_type	Type of transposition to perform
integer(kind=int8),	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	Tile size
logical,	intent(in)	::	has_inner_loop	If kernel has inner loop

Return Value type(CUfunction)

Cached kernel

private function get_true_transpose_type(transpose_type) result(transpose_type_)

Returns generic transpose id. Since X-Y and Y-Z transpositions are symmectric, it returns only one of them. X-Z and Z-X are not symmetric

Arguments

Type	Intent	Optional	Attributes		Name
type(dtfft_transpose_t),	intent(in)			::	transpose_type	Type of transposition to perform

Return Value type(dtfft_transpose_t)

Fixed id of transposition

private function compile_and_cache(comm, dims, transpose_type, kernel_type, base_storage, tile_size, has_inner_loop) result(kernel)

Compiles kernel and caches it. Returns compiled kernel.

Arguments

Type	Intent	Attributes		Name
type(MPI_Comm),	intent(in)		::	comm	MPI Communicator
integer(kind=int32),	intent(in),	target	::	dims(:)	Global dimensions to process
type(dtfft_transpose_t),	intent(in)		::	transpose_type	Type of transposition to perform
integer(kind=int8),	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	Tile size
logical,	intent(in)		::	has_inner_loop	If kernel has inner loop

Return Value type(CUfunction)

Compiled kernel to return

private function get_transpose_kernel_code(kernel_name, ndims, base_storage, transpose_type, enable_packing, enable_multiprocess) result(code)

Generates code that will be used to locally tranpose data and prepares to send it to other processes ndims == 2

Arguments

Type	Intent		Name
character(len=*),	intent(in)	::	kernel_name	Name of CUDA kernel
integer(kind=int8),	intent(in)	::	ndims	Number of dimensions
integer(kind=int64),	intent(in)	::	base_storage	Number of bytes needed to store single element
type(dtfft_transpose_t),	intent(in)	::	transpose_type	Transpose id
logical,	intent(in)	::	enable_packing	If data should be manually packed or not
logical,	intent(in)	::	enable_multiprocess	If thread should process more then one element

Return Value type(kernel_code)

Resulting code

private function get_unpack_kernel_code(kernel_name, base_storage, is_partial) result(code)

Generates code that will be used to unpack data when it is recieved

Arguments

Type	Intent		Name
character(len=*),	intent(in)	::	kernel_name	Name of CUDA kernel
integer(kind=int64),	intent(in)	::	base_storage	Number of bytes needed to store single element
logical,	intent(in)	::	is_partial

Return Value type(kernel_code)

Resulting code

private function get_unpack_pipelined_kernel_code(kernel_name, base_storage) result(code)

Generates code that will be used to partially unpack data when it is recieved from other process

Arguments

Type	Intent	Optional	Attributes		Name
character(len=*),	intent(in)			::	kernel_name	Name of CUDA kernel
integer(kind=int64),	intent(in)			::	base_storage	Number of bytes needed to store single element

Return Value type(kernel_code)

Resulting code

Subroutines

public subroutine clean_unused_cache()

Removes unused modules from cuda context

Arguments

None

private subroutine to_cstr(self, c_code)

Converts Fortran CUDA code to C pointer

Arguments

Type	Intent	Optional	Attributes		Name
class(kernel_code),	intent(in)			::	self	Kernel code
character(len=c_char),	intent(out),		allocatable	::	c_code(:)	C pointer to code

private subroutine add_line(self, line)

Adds new line to CUDA code

Arguments

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

private subroutine destroy_code(self)

Frees all memory

Arguments

Type	Intent	Optional	Attributes		Name
class(kernel_code),	intent(inout)			::	self	Kernel code

private subroutine create_device_pointer(ptr, values)

Allocates memory on a device and copies values to it.

Arguments

Type	Intent	Optional	Attributes		Name
type(c_ptr),	intent(inout)			::	ptr	Device pointer
integer(kind=c_int),	intent(in),		target	::	values(:)	Values to copy

private subroutine get_contiguous_execution_blocks(size, num_blocks, block_sizes)

Arguments

Type	Intent		Name
integer(kind=int32),	intent(in)	::	size	Total amount of iterations required
type(dim3),	intent(out)	::	num_blocks	Grid of blocks.
type(dim3),	intent(out)	::	block_sizes	Thread block.

private subroutine create(self, comm, dims, base_storage, transpose_type, kernel_type, pointers)

Creates kernel

Arguments

Type	Intent	Optional	Attributes		Name
class(nvrtc_kernel),	intent(inout)			::	self	nvRTC Compiled kernel class
type(MPI_Comm),	intent(in)			::	comm	MPI Communicator
integer(kind=int32),	intent(in),		target	::	dims(0:)	Global dimensions to process
integer(kind=int64),	intent(in)			::	base_storage	Number of bytes needed to store single element
type(dtfft_transpose_t),	intent(in)			::	transpose_type	Type of transposition to perform
integer(kind=int8),	intent(in)			::	kernel_type	Type of kernel to build
integer(kind=int32),	intent(in),	optional		::	pointers(:,:)	Optional pointers to unpack kernels

private subroutine execute(self, in, out, stream, source)

Executes kernel on stream

Arguments

Type	Intent	Optional	Attributes		Name
class(nvrtc_kernel),	intent(inout)			::	self	nvRTC Compiled kernel class
real(kind=real32),	intent(in),		target	::	in(:)	Source pointer
real(kind=real32),	intent(in),		target	::	out(:)	Target pointer
type(dtfft_stream_t),	intent(in)			::	stream	CUDA Stream
integer(kind=int32),	intent(in),	optional		::	source	Source rank for pipelined unpacking

private subroutine destroy(self)

Destroys kernel

Arguments

Type	Intent	Optional	Attributes		Name
class(nvrtc_kernel),	intent(inout)			::	self	nvRTC Compiled kernel class

private subroutine mark_unused(kernel)

Takes CUDA kernel as an argument and searches for it in cache If kernel is found than reduces ref_count and return null pointer

Arguments

Type	Intent	Optional	Attributes		Name
type(CUfunction),	intent(inout)			::	kernel	CUDA kernel to search for

private subroutine get_neighbor_function_code(code)

Generated device function that is used to determite id of process that to which data is being sent or from which data has been recieved based on local element coordinate

Arguments

Type	Intent	Optional	Attributes		Name
type(kernel_code),	intent(inout)			::	code	Resulting code

private subroutine get_code_init(kernel_name, base_storage, code, buffer_type)

Generates basic code that is used in all other kernels

Arguments

Type	Intent	Optional	Attributes		Name
character(len=*),	intent(in)			::	kernel_name	Name of CUDA kernel
integer(kind=int64),	intent(in)			::	base_storage	Number of bytes needed to store single element
type(kernel_code),	intent(inout)			::	code	Resulting code
character(len=:),	intent(out),	optional,	allocatable	::	buffer_type	Type of buffer that should be used

dtfft_nvrtc_kernel Module

Contents

Variables

Derived Types

Functions

Subroutines

Uses

Used by

Variables

Derived Types

type, public :: nvrtc_kernel

Components

Type-Bound Procedures

type, private :: kernel_code

Components

Type-Bound Procedures

type, private :: nvrtc_cache

Components

Functions

private function get_tile_size(x, y)

Arguments

Return Value integer(kind=int32)

private function get_cached_kernel(transpose_type, kernel_type, base_storage, tile_size, has_inner_loop) result(kernel)

Arguments

Return Value type(CUfunction)

private function get_true_transpose_type(transpose_type) result(transpose_type_)

Arguments

Return Value type(dtfft_transpose_t)

private function compile_and_cache(comm, dims, transpose_type, kernel_type, base_storage, tile_size, has_inner_loop) result(kernel)

Arguments

Return Value type(CUfunction)

private function get_transpose_kernel_code(kernel_name, ndims, base_storage, transpose_type, enable_packing, enable_multiprocess) result(code)

Arguments

Return Value type(kernel_code)

private function get_unpack_kernel_code(kernel_name, base_storage, is_partial) result(code)

Arguments

Return Value type(kernel_code)

private function get_unpack_pipelined_kernel_code(kernel_name, base_storage) result(code)

Arguments

Return Value type(kernel_code)

Subroutines

public subroutine clean_unused_cache()

Arguments

private subroutine to_cstr(self, c_code)

Arguments

private subroutine add_line(self, line)

Arguments

private subroutine destroy_code(self)

Arguments

private subroutine create_device_pointer(ptr, values)

Arguments

private subroutine get_contiguous_execution_blocks(size, num_blocks, block_sizes)

Arguments

private subroutine create(self, comm, dims, base_storage, transpose_type, kernel_type, pointers)

Arguments

private subroutine execute(self, in, out, stream, source)

Arguments

private subroutine destroy(self)

Arguments

private subroutine mark_unused(kernel)

Arguments

private subroutine get_neighbor_function_code(code)

Arguments

private subroutine get_code_init(kernel_name, base_storage, code, buffer_type)

Arguments