module m_memcheck_build !! The real build of x3d2-memcheck (--build): the real case build, its !! measurement and the measured report. use m_common, only: dp, i8 use m_logger, only: log_info, log_warn use m_mesh, only: mesh_t use m_backend_setup, only: backend_setup_t, poisson_uses_2decomp use m_base_case, only: base_case_t use m_case_factory, only: allocate_flow_case, is_supported_flow_case use m_field, only: flist_t use m_ibm, only: ibm_mask_filename use m_monitoring, only: monitoring_default_enabled use m_scalar_series, only: scalar_series_default_enabled use m_case_abl, only: case_abl_t use m_memcheck_context, only: memcheck_ctx_t use m_memcheck_report, only: n_gpu_list, print_rule, print_table_header, & print_table_row, card_shared_skip use m_memcheck_estimate, only: to_gib, num_gpus_unsupported_reason, & requested_unsupported_reason, & estimate_for_num_gpus, & fields_plus_halo_bytes_n, basis_name use m_memory_estimate, only: peak_fields_lookup, output_field_active implicit none private public :: run_build_mode !> The distance the static estimate may be from the measurement before the !> CHECK line of --build says MISMATCH. real(dp), parameter :: CHECK_TOLERANCE = 0.05_dp contains subroutine validate_build_inputs(ctx, ok) !! Checks the known causes of a failed real build: a missing input file, !! a flow case without a dispatch, a missing ibm mask. ok is false and !! the skip line is printed on a failure. type(memcheck_ctx_t), intent(in) :: ctx logical, intent(out) :: ok character(len=16) :: ibm_file logical :: ibm_file_exists, input_exists ok = .false. inquire (file=trim(ctx%input_path), exist=input_exists) if (.not. input_exists) then call log_info('Real build skipped: input file not found: '// & trim(ctx%input_path)) return end if if (.not. is_supported_flow_case(ctx%domain_cfg%flow_case_name)) then call log_info("Real build skipped: flow case '"// & trim(ctx%domain_cfg%flow_case_name)// & "' has no dispatch in x3d2-memcheck") return end if if (ctx%solver_cfg%ibm_on) then ibm_file = ibm_mask_filename([ctx%periodic_x, ctx%periodic_y, & ctx%periodic_z]) inquire (file=trim(ibm_file), exist=ibm_file_exists) if (.not. ibm_file_exists) then call log_info('Real build skipped: ibm_on=T but the matching & &ibm_.bp mask file was not found in the & &working directory; the estimate above stands.') return end if end if ok = .true. end subroutine validate_build_inputs subroutine check_peak_fields_table(ctx, measured, table_ok) !! Warns when peak_fields_lookup differs from the measured peak_fields. type(memcheck_ctx_t), intent(in) :: ctx integer, intent(in) :: measured logical, intent(out) :: table_ok integer :: predicted character(len=256) :: line_buf predicted = peak_fields_lookup(trim(ctx%domain_cfg%flow_case_name), & ctx%solver_cfg%time_intg, & ctx%solver_cfg%n_species, & trim(ctx%les_cfg%model) /= 'none', & ctx%solver_cfg%ibm_on, & ctx%output_vorticity, & ctx%output_qcriterion, & ctx%solver_cfg%lowmem_transeq) table_ok = predicted == measured if (ctx%irank /= 0) return if (.not. table_ok) then write (line_buf, '(a,i0,a,i0,a)') & 'peak_fields_lookup predicted ', predicted, & ' but the real build measured ', measured, & ' - m_memory_estimate''s table is stale for this config and & &should be recalibrated (see src/memory_estimate.f90).' call log_warn(trim(line_buf)) end if end subroutine check_peak_fields_table subroutine report_measured_table(ctx, measured_peak_fields, used_gib, & workspace_gib_measured, requested_verdict) !! Prints the measured table: the num_gpus=1 measurement scaled to every !! GPU count by estimate_for_num_gpus. A supported requested num_gpus !! outside n_gpu_list gets its own row. requested_verdict is the verdict !! of the requested num_gpus. type(memcheck_ctx_t), intent(inout) :: ctx integer, intent(in) :: measured_peak_fields real(dp), intent(in) :: used_gib, workspace_gib_measured character(len=12), intent(out) :: requested_verdict integer :: k, num_gpus, basis real(dp) :: per_gpu_gib, workspace_gib, overhead_gib logical :: any_extrapolated character(len=12) :: verdict character(len=96) :: reason character(len=256) :: line_buf requested_verdict = 'DOES_NOT_FIT' call print_rule('-') write (line_buf, '(a,i0,a,i0,a,f0.2,a,f0.2,a)') & 'Real build (1 GPU, ', & ctx%measured_n_substeps, ' substep(s)): peak_fields measured ', & measured_peak_fields, ', workspace ', workspace_gib_measured, & ' GiB, used ', used_gib, ' GiB' call log_info(trim(line_buf)) ! One line for scripts that compare --extensive runs. if (ctx%extensive_substeps > 0) then write (line_buf, '(a,i0,a,i0,a,f0.3,a,f0.3)') & 'EXTENSIVE result: substeps=', & ctx%measured_n_substeps, ' peak_fields=', measured_peak_fields, & ' used_gib=', used_gib, ' pct_card=', 100._dp*used_gib/ctx%card_gib call log_info(trim(line_buf)) end if call print_table_header() any_extrapolated = .false. do k = 1, size(n_gpu_list) + 1 num_gpus = ctx%domain_cfg%nproc_dir(3) if (k <= size(n_gpu_list)) num_gpus = n_gpu_list(k) if (k > size(n_gpu_list) .and. & (any(n_gpu_list == num_gpus) .or. & len_trim(requested_unsupported_reason(ctx)) > 0)) exit reason = num_gpus_unsupported_reason(ctx, num_gpus) if (len_trim(reason) > 0) then write (line_buf, '(a,i0,a,a,a)') & ' ', num_gpus, ' (skipped: ', & trim(reason), ')' call log_info(trim(line_buf)) cycle end if call estimate_for_num_gpus(ctx, num_gpus, per_gpu_gib, basis, verdict, & workspace_gib, overhead_gib, & used_gib=used_gib, npeak=measured_peak_fields) call print_table_row(num_gpus, [ctx%gdims(1), ctx%gdims(2), & ctx%gdims(3)/num_gpus], workspace_gib, & overhead_gib, per_gpu_gib, ctx%card_gib, & basis_name(basis), verdict) if (num_gpus == ctx%domain_cfg%nproc_dir(3)) requested_verdict = verdict if (num_gpus > 1) any_extrapolated = .true. end do call print_rule('-') if (any_extrapolated) & call log_info('(num_gpus>1 rows extrapolate the single-GPU measurement & &by 1/num_gpus.)') end subroutine report_measured_table subroutine build_mesh(ctx, dims_in, mesh) !! Mesh of the real build at dims_in. type(memcheck_ctx_t), intent(in) :: ctx integer, intent(in) :: dims_in(3) type(mesh_t), intent(out) :: mesh mesh = mesh_t(dims_in, [1, 1, 1], ctx%domain_cfg%L_global, & ctx%domain_cfg%BC_x, ctx%domain_cfg%BC_y, & ctx%domain_cfg%BC_z, ctx%domain_cfg%stretching, & ctx%domain_cfg%beta, & use_2decomp=poisson_uses_2decomp( & ctx%solver_cfg%poisson_solver_type)) end subroutine build_mesh subroutine drive_case(ctx, flow_case) !! Drive the built flow case the way run() does so the allocator reaches !! its high-water mark; records measured_n_substeps and the output flags. type(memcheck_ctx_t), intent(inout) :: ctx class(base_case_t), intent(inout) :: flow_case type(flist_t), allocatable :: curr(:), deriv(:) integer :: n_substeps, i_substep call flow_case%set_curr_fields(curr, deriv) ! Mirrors run()'s pre-loop call to the case-specific postprocess hook. call flow_case%postprocess(0, 0._dp) ! One sub-stage reaches the allocator high-water mark of a time step. ! --extensive drives n sub-stages, which re-checks the allocator ! only, not the per-iteration IO paths. n_substeps = 1 if (ctx%extensive_substeps > 0) n_substeps = ctx%extensive_substeps do i_substep = 1, n_substeps call flow_case%substep(curr, deriv, i_substep) end do ctx%measured_n_substeps = n_substeps ! The per-iteration postprocessing of run(): keep_pressure and the ! derived fields are allocated here, not in postprocess(). ctx%output_vorticity = output_field_active( & flow_case%io_mgr%snapshot_mgr%config, 'vorticity') ctx%output_qcriterion = output_field_active( & flow_case%io_mgr%snapshot_mgr%config, 'qcriterion') call flow_case%finish_step(n_substeps, ctx%output_vorticity, & ctx%output_qcriterion) end subroutine drive_case subroutine build_and_measure(ctx, dims_in, npeak, free_end_bytes) !! Builds the real flow case at dims_in on this GPU, drives one substep !! (n under --extensive) and returns the allocator mark npeak and the !! free device memory free_end_bytes at that point. The real case !! constructors re-read argument 1 as the input file, and case_init !! runs the restart logic of the input, as in production. type(memcheck_ctx_t), target, intent(inout) :: ctx integer, intent(in) :: dims_in(3) integer, intent(out) :: npeak integer(i8), intent(out) :: free_end_bytes type(mesh_t), target :: mesh type(backend_setup_t), target :: setup class(base_case_t), allocatable :: flow_case integer(i8) :: total_b npeak = 0 free_end_bytes = 0_i8 ctx%output_vorticity = .false. ctx%output_qcriterion = .false. call build_mesh(ctx, dims_in, mesh) call setup%init(mesh) ! The case build writes no monitoring.csv or turbine disc.adm in the ! working directory. monitoring_default_enabled = .false. scalar_series_default_enabled = .false. call allocate_flow_case(ctx%domain_cfg%flow_case_name, setup%backend, & mesh, setup%host_allocator, flow_case) ! The ABL profile file is written by sample() only when the profile ! diagnostics are on: switch them off so postprocess(0) leaves it alone. select type (flow_case) type is (case_abl_t) flow_case%diagnostics%cfg%profile_start_iter = -1 end select ! The solver construction settled the FFT path: record it. call ctx%device%capture_solver_info(flow_case) call drive_case(ctx, flow_case) npeak = setup%allocator%next_id call ctx%device%mem_info(total_b, free_end_bytes) end subroutine build_and_measure subroutine run_real_build(ctx, measured_peak_fields, used_gib) !! Runs the real build in this process, after the FFT probe has destroyed !! its plans. used_gib is the card memory the build holds: the context at !! the start plus the fall of the free memory since then. type(memcheck_ctx_t), target, intent(inout) :: ctx integer, intent(out) :: measured_peak_fields real(dp), intent(out) :: used_gib integer(i8) :: free_end_bytes call build_and_measure(ctx, [ctx%gdims(1), ctx%gdims(2), ctx%gdims(3)], & measured_peak_fields, free_end_bytes) used_gib = to_gib(ctx%device%footprint_bytes(free_end_bytes)) end subroutine run_real_build subroutine run_build_mode(ctx) !! The real build: a case build and one substep on 1 GPU in this process, !! measured with cudaMemGetInfo. The estimate stands when the !! num_gpus=1 row is DOES_NOT_FIT or the inputs are not valid. type(memcheck_ctx_t), intent(inout) :: ctx real(dp) :: used_gib, workspace_gib_measured real(dp) :: pct_error, estimate_ng1_excl_io_gib character(len=8) :: check_word integer :: measured_peak_fields logical :: inputs_ok, table_ok character(len=12) :: measured_verdict character(len=256) :: line_buf if (trim(ctx%ng1_verdict) == 'DOES_NOT_FIT') then call print_rule('-') call log_info('Real build skipped: the num_gpus=1 estimate is & &DOES_NOT_FIT; the estimate above stands.') return end if call validate_build_inputs(ctx, inputs_ok) if (.not. inputs_ok) return call run_real_build(ctx, measured_peak_fields, used_gib) call check_peak_fields_table(ctx, measured_peak_fields, table_ok) ! Another process changes its card memory or NVML cannot identify it. ! Keep the estimate when the measurement is not reliable if (card_shared_skip(emit_line=table_ok)) return workspace_gib_measured = & to_gib(fields_plus_halo_bytes_n(ctx, 1, measured_peak_fields)) call report_measured_table(ctx, measured_peak_fields, used_gib, & workspace_gib_measured, measured_verdict) ! An unsupported nproc_dir stays UNSUPPORTED. if (len_trim(requested_unsupported_reason(ctx)) == 0) & ctx%final_verdict = measured_verdict ! The CHECK line compares the num_gpus=1 estimate without the IO staging ! term, which the real build never measures, with the measured memory. estimate_ng1_excl_io_gib = ctx%estimate_ng1_gib - ctx%io_staging_ng1_gib pct_error = 100._dp*(estimate_ng1_excl_io_gib - used_gib)/used_gib if (abs(estimate_ng1_excl_io_gib - used_gib) <= & CHECK_TOLERANCE*used_gib) then check_word = 'OK' else check_word = 'MISMATCH' end if write (line_buf, '(a,f0.2,a,f0.2,a,sp,f0.1,ss,a,a,a)') & 'CHECK num_gpus=1: estimated ', & estimate_ng1_excl_io_gib, ' GiB, measured ', used_gib, ' GiB (', & pct_error, '%) - ', trim(check_word), ' (tolerance 5%)' call log_info(trim(line_buf)) if (ctx%io_staging_ng1_gib > 0._dp) then write (line_buf, '(a,f0.1,a)') & ' (GPU-aware IO staging ', & ctx%io_staging_ng1_gib*1024._dp, ' MiB excluded from CHECK: the real & &build performs no snapshot/checkpoint write.)' call log_info(trim(line_buf)) end if end subroutine run_build_mode end module m_memcheck_build