dbcsr_mm_sched Module

Fourth layer of the dbcsr matrix-matrix multiplication. It hides the differences between performing calculations on the accelerator device or on the CPU. Modification history: - 2010-02-23 Moved from dbcsr_operations - 2011-11 Moved parameter-stack processing routines to dbcsr_mm_methods. - 2013-01 extensive refactoring (Ole Schuett)


Uses


Variables

Type Visibility Attributes Name Initial
character(len=*), private, parameter :: moduleN = 'dbcsr_mm_sched'
type(stats_type), private, DIMENSION(:), ALLOCATABLE, TARGET, SAVE :: stats_per_thread

Counters for each thread to collect statistics

Derived Types

type, public ::  dbcsr_mm_sched_type

Components

Type Visibility Attributes Name Initial
type(dbcsr_work_type), private, POINTER :: product_wm => Null()
type(dbcsr_mm_accdrv_type), private :: accdrv = dbcsr_mm_accdrv_type()
type(dbcsr_mm_hostdrv_type), private :: hostdrv = dbcsr_mm_hostdrv_type()
logical, private :: avoid_accdrv = .FALSE.
logical, private :: product_wm_cleared = .FALSE.
logical, private :: keep_product_data = .TRUE.
integer, private :: product_wm_orig_datasize = -1

type, private ::  stats_type

Components

Type Visibility Attributes Name Initial
integer(kind=int_8), public :: cpu_num_stacks = 0
integer(kind=int_8), public :: smm_num_stacks = 0
integer(kind=int_8), public :: acc_num_stacks = 0
integer(kind=int_8), public :: cpu_flop = 0
integer(kind=int_8), public :: smm_flop = 0
integer(kind=int_8), public :: acc_flop = 0
integer(kind=int_8), public :: max_cpu_flop = 0
integer(kind=int_8), public :: max_smm_flop = 0
integer(kind=int_8), public :: max_acc_flop = 0
integer(kind=int_8), public, DIMENSION(:, :), ALLOCATABLE :: num_mnk_stacks
integer(kind=int_4), public, DIMENSION(64) :: padding = -1_int_4

Subroutines

private subroutine stats_init(stats)

Initialize a stats_type

Arguments

Type IntentOptional Attributes Name
type(stats_type), intent(inout) :: stats

public subroutine dbcsr_mm_sched_lib_init()

Initialize the library

Arguments

None

public subroutine dbcsr_mm_sched_lib_finalize()

Finalize the library and prints DBCSR statistics

Arguments

None

public subroutine dbcsr_mm_sched_print_statistics(group, output_unit)

Prints DBCSR statistics

Arguments

Type IntentOptional Attributes Name
type(mp_comm_type), intent(in) :: group
integer, intent(in) :: output_unit

private subroutine ensure_product_wm_cleared(this)

Makes sure that the product_wm is cleared.

Arguments

Type IntentOptional Attributes Name
type(dbcsr_mm_sched_type), intent(inout) :: this

public subroutine dbcsr_mm_sched_init(this, product_wm, nlayers, keep_product_data)

Initializes a multiplication cycle for new set of C-blocks.

Arguments

Type IntentOptional Attributes Name
type(dbcsr_mm_sched_type), intent(inout) :: this
type(dbcsr_work_type), POINTER :: product_wm
integer, optional :: nlayers
logical, intent(in) :: keep_product_data

public subroutine dbcsr_mm_sched_finalize(this)

Finalizes a multiplication cycle for a set of C-blocks.

Arguments

Type IntentOptional Attributes Name
type(dbcsr_mm_sched_type), intent(inout) :: this

public subroutine dbcsr_mm_sched_dev2host_init(this)

Finalizes a multiplication cycle for a set of C-blocks.

Arguments

Type IntentOptional Attributes Name
type(dbcsr_mm_sched_type), intent(inout) :: this

public subroutine dbcsr_mm_sched_begin_burst(this)

Signal begin of a burst of calls to dbcsr_mm_sched_process.

Arguments

Type IntentOptional Attributes Name
type(dbcsr_mm_sched_type), intent(inout) :: this

public subroutine dbcsr_mm_sched_end_burst()

Signal end of a burst of calls to dbcsr_mm_sched_process.

Arguments

None

public subroutine dbcsr_mm_sched_barrier()

Signal that previous stacks should be processed first

Arguments

None

public subroutine dbcsr_mm_sched_process(this, left, right, stack_data, stack_fillcount, stack_descr)

Processes a given stack. From here on there is no boundary checking due to assumed-SIZE-arguments. This is useful to check stack parameters, BUT it works only for kind=dp

Arguments

Type IntentOptional Attributes Name
type(dbcsr_mm_sched_type), intent(inout) :: this
type(dbcsr_type), intent(in) :: left
type(dbcsr_type), intent(in) :: right
integer, DIMENSION(:, :), POINTER :: stack_data
integer, POINTER :: stack_fillcount
type(stack_descriptor_type), intent(in) :: stack_descr

public subroutine dbcsr_mm_sched_set_orig_datasize(this, newsize)

Change the datasize of the original workspace buffer

Arguments

Type IntentOptional Attributes Name
type(dbcsr_mm_sched_type), intent(inout) :: this
integer, intent(in) :: newsize

private subroutine stats_add(stats, m, n, k, stacksize_cpu, stacksize_smm, stacksize_acc, nstacks_cpu, nstacks_smm, nstacks_acc, generated_acc_untuned)

Helper-routine used by dbcsr_mm_sched_process to supply statistics.

Arguments

Type IntentOptional Attributes Name
type(stats_type), intent(inout) :: stats
integer, intent(in) :: m
integer, intent(in) :: n
integer, intent(in) :: k
integer(kind=int_8), optional :: stacksize_cpu
integer(kind=int_8), optional :: stacksize_smm
integer(kind=int_8), optional :: stacksize_acc
integer(kind=int_8), optional :: nstacks_cpu
integer(kind=int_8), optional :: nstacks_smm
integer(kind=int_8), optional :: nstacks_acc
logical, optional :: generated_acc_untuned

private subroutine stats_collect_from_threads(report)

Collects statistics from all OpenMP-threads into report

Arguments

Type IntentOptional Attributes Name
type(stats_type), intent(inout) :: report

private subroutine stats_collect_from_ranks(report, group)

Collects statistics from all MPI-ranks

Arguments

Type IntentOptional Attributes Name
type(stats_type), intent(inout) :: report
type(mp_comm_type), intent(in) :: group

private subroutine stats_print_report(report, output_unit)

Prints collected statistics

Arguments

Type IntentOptional Attributes Name
type(stats_type), intent(inout) :: report
integer, intent(in) :: output_unit