transpose_handle_cuda Derived Type

type, public :: transpose_handle_cuda

CUDA Transpose Handle


Inherits

type~~transpose_handle_cuda~~InheritsGraph type~transpose_handle_cuda transpose_handle_cuda type~abstract_backend abstract_backend type~transpose_handle_cuda->type~abstract_backend comm_handle type~dtfft_transpose_t dtfft_transpose_t type~transpose_handle_cuda->type~dtfft_transpose_t transpose_type type~nvrtc_kernel nvrtc_kernel type~transpose_handle_cuda->type~nvrtc_kernel transpose_kernel, unpack_kernel, unpack_kernel2 type~abstract_backend->type~nvrtc_kernel unpack_kernel, unpack_kernel2 MPI_Comm MPI_Comm type~abstract_backend->MPI_Comm comm type~cudaevent cudaEvent type~abstract_backend->type~cudaevent execution_event, copy_event type~dtfft_backend_t dtfft_backend_t type~abstract_backend->type~dtfft_backend_t backend type~dtfft_stream_t dtfft_stream_t type~abstract_backend->type~dtfft_stream_t copy_stream c_ptr c_ptr type~nvrtc_kernel->c_ptr device_pointers type~cufunction CUfunction type~nvrtc_kernel->type~cufunction cuda_kernel type~dim3 dim3 type~nvrtc_kernel->type~dim3 blocks, threads type~kernel_type_t kernel_type_t type~nvrtc_kernel->type~kernel_type_t kernel_type type~kernelargs kernelArgs type~nvrtc_kernel->type~kernelargs kernelParams type~cudaevent->c_ptr event type~cufunction->c_ptr ptr type~dtfft_stream_t->c_ptr stream type~kernelargs->c_ptr ptrs

Inherited by

type~~transpose_handle_cuda~~InheritedByGraph type~transpose_handle_cuda transpose_handle_cuda type~transpose_plan_cuda transpose_plan_cuda type~transpose_plan_cuda->type~transpose_handle_cuda fplans, bplans

Components

Type Visibility Attributes Name Initial
type(dtfft_transpose_t), private :: transpose_type

Type of transposition to perform
logical, private :: has_exchange = .false.

If current handle has exchanges between GPUs
logical, private :: is_pipelined = .false.

If underlying exchanges are pipelined
type(nvrtc_kernel), private :: transpose_kernel

Kernel for data transposition
type(nvrtc_kernel), private :: unpack_kernel

Kernel for unpacking data
type(nvrtc_kernel), private :: unpack_kernel2

Kernel for unpacking data required by NCCL_PIPELINED algorithm
class(abstract_backend), private, allocatable :: comm_handle

Communication handle

Type-Bound Procedures

procedure, public, pass(self) :: create

Creates CUDA Transpose Handle

  • private subroutine create(self, helper, send, recv, effort, base_storage, backend, force_effort)

    Creates CUDA Transpose Handle

    Arguments

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

    CUDA Transpose Handle
    type(backend_helper), intent(in) :: helper

    Backend helper
    type(pencil), intent(in) :: send

    Send pencil
    type(pencil), intent(in) :: recv

    Recv pencil
    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(dtfft_backend_t), intent(in) :: backend

    Backend type
    logical, intent(in), optional :: force_effort

    Should effort be forced or not

procedure, public, pass(self) :: execute

Executes transpose - exchange - unpack

  • private subroutine execute(self, in, out, stream, aux)

    Executes transpose - exchange - unpack

    Arguments

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

    CUDA Transpose Handle
    real(kind=real32), intent(inout) :: in(:)

    Send pointer
    real(kind=real32), intent(inout) :: out(:)

    Recv pointer
    type(dtfft_stream_t), intent(in) :: stream

    Main execution CUDA stream
    real(kind=real32), intent(inout) :: aux(:)

    Aux pointer

procedure, public, pass(self) :: destroy

Destroys CUDA Transpose Handle

  • private subroutine destroy(self)

    Destroys CUDA Transpose Handle

    Arguments

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

    CUDA Transpose Handle

procedure, public, pass(self) :: get_aux_size

Returns number of bytes required by aux buffer

  • private function get_aux_size(self)

    Returns number of bytes required by aux buffer

    Arguments

    Type IntentOptional Attributes Name
    class(transpose_handle_cuda), intent(in) :: self

    CUDA Transpose Handle

    Return Value integer(kind=int64)