dtfft_abstract_kernel Module

This module defines abstract_kernel type and its type bound procedures.

The abstract kernel is used in transpose_handle_generic type and is resposible for packing/unpacking/permute operations. The actual implementation of the kernel is deferred to the create_private, execute_private and destroy_private procedures.


Uses

  • module~~dtfft_abstract_kernel~~UsesGraph module~dtfft_abstract_kernel dtfft_abstract_kernel iso_fortran_env iso_fortran_env module~dtfft_abstract_kernel->iso_fortran_env module~dtfft_interface_nvtx dtfft_interface_nvtx module~dtfft_abstract_kernel->module~dtfft_interface_nvtx module~dtfft_parameters dtfft_parameters module~dtfft_abstract_kernel->module~dtfft_parameters module~dtfft_utils dtfft_utils module~dtfft_abstract_kernel->module~dtfft_utils mpi_f08 mpi_f08 module~dtfft_abstract_kernel->mpi_f08 module~dtfft_interface_nvtx->module~dtfft_utils iso_c_binding iso_c_binding module~dtfft_interface_nvtx->iso_c_binding module~dtfft_parameters->iso_fortran_env module~dtfft_parameters->mpi_f08 module~dtfft_parameters->iso_c_binding module~dtfft_utils->iso_fortran_env module~dtfft_utils->module~dtfft_parameters module~dtfft_utils->mpi_f08 module~dtfft_utils->iso_c_binding module~dtfft_errors dtfft_errors module~dtfft_utils->module~dtfft_errors module~dtfft_errors->iso_fortran_env

Used by

  • module~~dtfft_abstract_kernel~~UsedByGraph module~dtfft_abstract_kernel dtfft_abstract_kernel module~dtfft_abstract_backend dtfft_abstract_backend module~dtfft_abstract_backend->module~dtfft_abstract_kernel module~dtfft_kernel_device dtfft_kernel_device module~dtfft_kernel_device->module~dtfft_abstract_kernel module~dtfft_nvrtc_block_optimizer dtfft_nvrtc_block_optimizer module~dtfft_kernel_device->module~dtfft_nvrtc_block_optimizer module~dtfft_nvrtc_module_cache dtfft_nvrtc_module_cache module~dtfft_kernel_device->module~dtfft_nvrtc_module_cache module~dtfft_kernel_host dtfft_kernel_host module~dtfft_kernel_host->module~dtfft_abstract_kernel module~dtfft_nvrtc_block_optimizer->module~dtfft_abstract_kernel module~dtfft_nvrtc_module dtfft_nvrtc_module module~dtfft_nvrtc_module->module~dtfft_abstract_kernel module~dtfft_nvrtc_module->module~dtfft_nvrtc_block_optimizer module~dtfft_nvrtc_module_cache->module~dtfft_abstract_kernel module~dtfft_nvrtc_module_cache->module~dtfft_nvrtc_block_optimizer module~dtfft_nvrtc_module_cache->module~dtfft_nvrtc_module module~dtfft_transpose_handle_generic dtfft_transpose_handle_generic module~dtfft_transpose_handle_generic->module~dtfft_abstract_kernel module~dtfft_transpose_handle_generic->module~dtfft_abstract_backend module~dtfft_transpose_handle_generic->module~dtfft_kernel_device module~dtfft_transpose_handle_generic->module~dtfft_kernel_host module~dtfft_abstract_transpose_handle dtfft_abstract_transpose_handle module~dtfft_transpose_handle_generic->module~dtfft_abstract_transpose_handle module~dtfft_backend_cufftmp_m dtfft_backend_cufftmp_m module~dtfft_transpose_handle_generic->module~dtfft_backend_cufftmp_m module~dtfft_backend_mpi dtfft_backend_mpi module~dtfft_transpose_handle_generic->module~dtfft_backend_mpi module~dtfft_backend_nccl_m dtfft_backend_nccl_m module~dtfft_transpose_handle_generic->module~dtfft_backend_nccl_m module~dtfft_abstract_transpose_handle->module~dtfft_abstract_backend module~dtfft_backend_cufftmp_m->module~dtfft_abstract_backend module~dtfft_backend_mpi->module~dtfft_abstract_backend module~dtfft_backend_nccl_m->module~dtfft_abstract_backend module~dtfft_transpose_plan dtfft_transpose_plan module~dtfft_transpose_plan->module~dtfft_abstract_backend module~dtfft_transpose_plan->module~dtfft_kernel_device module~dtfft_transpose_plan->module~dtfft_transpose_handle_generic module~dtfft_transpose_plan->module~dtfft_abstract_transpose_handle module~dtfft_transpose_handle_datatype dtfft_transpose_handle_datatype module~dtfft_transpose_plan->module~dtfft_transpose_handle_datatype program~test_host_kernels test_host_kernels program~test_host_kernels->module~dtfft_kernel_host module~dtfft_plan dtfft_plan module~dtfft_plan->module~dtfft_transpose_plan module~dtfft_transpose_handle_datatype->module~dtfft_abstract_transpose_handle module~dtfft dtfft module~dtfft->module~dtfft_plan module~dtfft_api dtfft_api module~dtfft_api->module~dtfft_plan

Variables

Type Visibility Attributes Name Initial
type(kernel_type_t), public, parameter :: KERNEL_DUMMY = kernel_type_t(-1)

Dummy kernel, does nothing
type(kernel_type_t), public, parameter :: KERNEL_UNPACK = kernel_type_t(3)

Unpacks contiguous buffer.
type(kernel_type_t), public, parameter :: KERNEL_UNPACK_SIMPLE_COPY = kernel_type_t(4)

Doesn’t actually unpacks anything. Performs cudaMemcpyAsync call. Should be used only when backend is DTFFT_BACKEND_CUFFTMP.
type(kernel_type_t), public, parameter :: KERNEL_UNPACK_PIPELINED = kernel_type_t(5)

Unpacks pack of contiguous buffer recieved from rank.

Unpacks contiguous buffer recieved from everyone except myself. Should be used only when backend is DTFFT_BACKEND_NCCL_PIPELINED.
type(kernel_type_t), public, parameter :: KERNEL_PERMUTE_FORWARD = kernel_type_t(7)
type(kernel_type_t), public, parameter :: KERNEL_PERMUTE_BACKWARD = kernel_type_t(8)
type(kernel_type_t), public, parameter :: KERNEL_PERMUTE_BACKWARD_START = kernel_type_t(9)
type(kernel_type_t), public, parameter :: KERNEL_PERMUTE_BACKWARD_END = kernel_type_t(10)
type(kernel_type_t), public, parameter :: KERNEL_PERMUTE_BACKWARD_END_PIPELINED = kernel_type_t(11)

Unpacks contiguous buffer recieved from everyone except myself. Should be used only when backend is DTFFT_BACKEND_NCCL_PIPELINED.
type(kernel_type_t), private, parameter :: TRANSPOSE_KERNELS(*) = [KERNEL_PERMUTE_FORWARD, KERNEL_PERMUTE_BACKWARD, KERNEL_PERMUTE_BACKWARD_START]

List of all transpose kernel types
type(kernel_type_t), private, parameter :: UNPACK_KERNELS(*) = [KERNEL_PERMUTE_BACKWARD_END, KERNEL_PERMUTE_BACKWARD_END_PIPELINED, KERNEL_UNPACK, KERNEL_UNPACK_PIPELINED]

List of all unpack kernel types

Interfaces

public interface operator(==)

  • private pure elemental function kernel_type_eq(left, right) result(res)

    Arguments

    Type IntentOptional Attributes Name
    type(kernel_type_t), intent(in) :: left
    type(kernel_type_t), intent(in) :: right

    Return Value logical

public interface operator(/=)

  • private pure elemental function kernel_type_ne(left, right) result(res)

    Arguments

    Type IntentOptional Attributes Name
    type(kernel_type_t), intent(in) :: left
    type(kernel_type_t), intent(in) :: right

    Return Value logical


Abstract Interfaces

abstract interface

  • private subroutine create_interface(self, effort, base_storage, force_effort)

    Creates underlying kernel

    Arguments

    Type IntentOptional Attributes Name
    class(abstract_kernel), intent(inout) :: self

    Abstract kernel
    type(dtfft_effort_t), intent(in) :: effort

    Effort level for generating transpose kernels
    integer(kind=int64), intent(in) :: base_storage

    Number of bytes needed to store single element
    logical, intent(in), optional :: force_effort

    Should effort be forced or not

abstract interface

  • private subroutine execute_interface(self, in, out, stream, neighbor)

    Executes underlying kernel

    Arguments

    Type IntentOptional Attributes Name
    class(abstract_kernel), intent(inout) :: self

    Abstract kernel
    real(kind=real32), intent(in), target :: in(:)

    Source buffer, can be device or host pointer
    real(kind=real32), intent(inout), target :: out(:)

    Target buffer, can be device or host pointer
    type(dtfft_stream_t), intent(in) :: stream

    Stream to execute on, used only for device pointers
    integer(kind=int32), intent(in), optional :: neighbor

    Source rank for pipelined unpacking

abstract interface

  • private subroutine destroy_interface(self)

    Destroys underlying kernel

    Arguments

    Type IntentOptional Attributes Name
    class(abstract_kernel), intent(inout) :: self

    Abstract kernel

Derived Types

type, public ::  kernel_type_t

nvRTC Kernel type

Components

Type Visibility Attributes Name Initial
integer(kind=int32), public :: val

type, public, abstract ::  abstract_kernel

Abstract kernel type

Read more…

Components

Type Visibility Attributes Name Initial
logical, public :: is_created = .false.

Kernel is created flag.
logical, public :: is_dummy = .false.

If kernel should do anything or not.
type(kernel_type_t), public :: kernel_type

Type of the kernel
character(len=:), public, allocatable :: kernel_string
integer(kind=int32), public, allocatable :: neighbor_data(:,:)

Neighbor data for pipelined unpacking
integer(kind=int32), public, allocatable :: dims(:)

Local dimensions to process

Type-Bound Procedures

procedure, public, pass(self) :: create ../../

Creates kernel
procedure, public, pass(self) :: execute ../../

Executes kernel
procedure, public, pass(self) :: destroy ../../

Destroys kernel
procedure(create_interface), public, deferred :: create_private ../../

Creates underlying kernel
procedure(execute_interface), public, deferred :: execute_private ../../

Executes underlying kernel
procedure(destroy_interface), public, deferred :: destroy_private ../../

Destroys underlying kernel

Functions

public function get_kernel_string(kernel) result(string)

Gets the string description of a kernel

Arguments

Type IntentOptional Attributes Name
type(kernel_type_t), intent(in) :: kernel

kernel type

Return Value character(len=:), allocatable

kernel string

public pure elemental function is_transpose_kernel(param) result(res)

Arguments

Type IntentOptional Attributes Name
type(kernel_type_t), intent(in) :: param

Return Value logical

public pure elemental function is_unpack_kernel(param) result(res)

Arguments

Type IntentOptional Attributes Name
type(kernel_type_t), intent(in) :: param

Return Value logical

private pure elemental function kernel_type_eq(left, right) result(res)

Arguments

Type IntentOptional Attributes Name
type(kernel_type_t), intent(in) :: left
type(kernel_type_t), intent(in) :: right

Return Value logical

private pure elemental function kernel_type_ne(left, right) result(res)

Arguments

Type IntentOptional Attributes Name
type(kernel_type_t), intent(in) :: left
type(kernel_type_t), intent(in) :: right

Return Value logical


Subroutines

private subroutine create(self, dims, effort, base_storage, kernel_type, neighbor_data, force_effort)

Creates kernel

Arguments

Type IntentOptional Attributes Name
class(abstract_kernel), intent(inout) :: self

Abstract kernel
integer(kind=int32), intent(in) :: dims(:)

Local dimensions to process
type(dtfft_effort_t), intent(in) :: effort

Effort level for generating transpose kernels
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 build
integer(kind=int32), intent(in), optional :: neighbor_data(:,:)

Optional pointers for unpack kernels
logical, intent(in), optional :: force_effort

Should effort be forced or not

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

Executes kernel

Arguments

Type IntentOptional Attributes Name
class(abstract_kernel), intent(inout) :: self

Abstract kernel
real(kind=real32), intent(in) :: in(:)

Source buffer, can be device or host pointer
real(kind=real32), intent(inout) :: out(:)

Target buffer, can be device or host pointer
type(dtfft_stream_t), intent(in) :: stream

Stream to execute on, used only for device pointers
integer(kind=int32), intent(in), optional :: neighbor

Source rank for pipelined unpacking

private subroutine destroy(self)

Destroys kernel

Arguments

Type IntentOptional Attributes Name
class(abstract_kernel), intent(inout) :: self

Abstract kernel