module m_memcheck_context !! Run state of x3d2-memcheck, passed explicitly to every routine. use m_common, only: dp use m_config, only: domain_config_t, solver_config_t, les_config_t, & checkpoint_config_t use m_logger, only: abort_run use m_mesh, only: classify_boundaries, global_cell_dims, FFT_BC_000, & FFT_BC_010, FFT_BC_100, FFT_BC_110 use m_cuda_memcheck_device, only: cuda_memcheck_device_t implicit none private public :: memcheck_ctx_t, parse_args, read_config, classify_bc, & resolve_gpu_io_mode type :: memcheck_ctx_t type(cuda_memcheck_device_t) :: device character(len=256) :: input_path !> STATIC (--static), DEFAULT (no flag) or BUILD (--build). character(len=7) :: run_mode !> The substep count of --extensive, 0 when it is not given. integer :: extensive_substeps = 0 type(domain_config_t) :: domain_cfg type(solver_config_t) :: solver_cfg type(les_config_t) :: les_cfg type(checkpoint_config_t) :: checkpoint_cfg integer :: gdims(3) integer :: cdims(3) logical :: periodic_x, periodic_y, periodic_z integer :: bc_code logical :: bc_is_000, bc_is_010, bc_is_100, bc_is_110 logical :: multi_gpu_supported integer :: peak_fields real(dp) :: card_gib !> Free memory on the card at the start. real(dp) :: card_free_gib = 0._dp !> Whether snapshot and checkpoint writes stage through a device buffer, !> the write mode name, and the reason when there is no staging. logical :: gpu_io_device_write = .false. character(len=16) :: gpu_io_mode_name = 'auto' character(len=64) :: gpu_io_reason = '' !> The num_gpus=1 row of the estimate: the CHECK line of --build !> compares it with the measurement. The verdict is blank for an !> unsupported row. real(dp) :: estimate_ng1_gib character(len=12) :: ng1_verdict = '' !> The FFT probe plan: 'full' when the estimate uses the full-grid !> probe's result, else 'small' (also when the full-grid probe failed). !> probe_reason says why the full-grid probe has no result, and !> probe_note_shown is set once estimate_for_num_gpus has printed it. character(len=5) :: probe_plan = 'small' character(len=96) :: probe_reason = '' logical :: probe_note_shown = .false. !> The verdict of the requested nproc_dir, which sets the exit code. character(len=12) :: final_verdict !> The IO staging term (GiB) at num_gpus=1. real(dp) :: io_staging_ng1_gib = 0._dp !> The output fields and the substep count the real build measured. logical :: output_vorticity = .false., output_qcriterion = .false. integer :: measured_n_substeps integer :: irank end type memcheck_ctx_t contains subroutine parse_args(ctx) type(memcheck_ctx_t), intent(inout) :: ctx integer :: i, nargs, iostat_n character(len=256) :: arg logical :: static_flag, build_flag nargs = command_argument_count() if (nargs < 1) call abort_run('usage: x3d2-memcheck & &[--static | --build [--extensive ]]', collective=.true.) ! The input file stays argument 1: the real case constructors re-read it. call get_command_argument(1, ctx%input_path) static_flag = .false. build_flag = .false. i = 2 do while (i <= nargs) call get_command_argument(i, arg) select case (trim(arg)) case ('--static') static_flag = .true. case ('--build') build_flag = .true. case ('--extensive') if (i == nargs) call abort_run('x3d2-memcheck: --extensive needs a & &positive integer substep count, e.g. --extensive 1000', & collective=.true.) i = i + 1 call get_command_argument(i, arg) read (arg, *, iostat=iostat_n) ctx%extensive_substeps if (iostat_n /= 0 .or. ctx%extensive_substeps < 1) & call abort_run('x3d2-memcheck: --extensive needs a positive & &integer substep count', collective=.true.) build_flag = .true. case default call abort_run('x3d2-memcheck: unknown flag '//trim(arg)// & ' (expected --static, --build, or --extensive )', & collective=.true.) end select i = i + 1 end do if (static_flag .and. build_flag) & call abort_run('x3d2-memcheck: --static and --build/--extensive are & &mutually exclusive (--static stops at the formula estimate).', & collective=.true.) if (static_flag) then ctx%run_mode = 'STATIC' else if (build_flag) then ctx%run_mode = 'BUILD' else ctx%run_mode = 'DEFAULT' end if end subroutine parse_args subroutine read_config(ctx) type(memcheck_ctx_t), intent(inout) :: ctx call ctx%domain_cfg%read(nml_file=trim(ctx%input_path)) call ctx%solver_cfg%read(nml_file=trim(ctx%input_path)) call ctx%les_cfg%read(nml_file=trim(ctx%input_path)) call ctx%checkpoint_cfg%read(nml_file=trim(ctx%input_path)) ctx%gdims = ctx%domain_cfg%dims_global end subroutine read_config subroutine classify_bc(ctx) !! The BC parse and four-way split of classify_boundaries in m_mesh. type(memcheck_ctx_t), intent(inout) :: ctx logical :: periodic(3) call classify_boundaries(ctx%domain_cfg%BC_x, ctx%domain_cfg%BC_y, & ctx%domain_cfg%BC_z, periodic, ctx%bc_code) ctx%periodic_x = periodic(1) ctx%periodic_y = periodic(2) ctx%periodic_z = periodic(3) ctx%bc_is_000 = ctx%bc_code == FFT_BC_000 ctx%bc_is_010 = ctx%bc_code == FFT_BC_010 ctx%bc_is_100 = ctx%bc_code == FFT_BC_100 ctx%bc_is_110 = ctx%bc_code == FFT_BC_110 ctx%cdims = global_cell_dims(ctx%gdims, periodic) end subroutine classify_bc subroutine resolve_gpu_io_mode(ctx) #ifdef X3D2_ADIOS2_CUDA use m_io_backend, only: gpu_write_mode_from_env, gpu_write_mode_names, & gpu_write_mode_force_host #endif !! Resolves whether snapshot and checkpoint writes stage through a !! device buffer, with the environment reader the solver uses for the !! write mode. type(memcheck_ctx_t), intent(inout) :: ctx #ifdef X3D2_ADIOS2_CUDA integer :: mode mode = gpu_write_mode_from_env() ctx%gpu_io_mode_name = trim(gpu_write_mode_names(mode)) if (mode == gpu_write_mode_force_host) then ctx%gpu_io_device_write = .false. ctx%gpu_io_reason = 'write mode host' else ctx%gpu_io_device_write = .true. end if #else ctx%gpu_io_device_write = .false. ctx%gpu_io_mode_name = 'host' ctx%gpu_io_reason = 'build lacks X3D2_ADIOS2_CUDA' #endif end subroutine resolve_gpu_io_mode end module m_memcheck_context