OASIS reproducible sum method from P. Worley. More...

Functions/Subroutines
subroutine, public	oasis_reprosum_setopts (repro_sum_use_ddpdd_in, repro_sum_rel_diff_max_in, repro_sum_recompute_in, repro_sum_master, repro_sum_logunit)
	Set runtime options. More...

subroutine, public	oasis_reprosum_calc (arr, arr_gsum, nsummands, dsummands, nflds, ddpdd_sum, arr_gbl_max, arr_gbl_max_out, arr_max_levels, arr_max_levels_out, gbl_max_nsummands, gbl_max_nsummands_out, gbl_count, repro_sum_validate, repro_sum_stats, rel_diff, commid)
	Compute the global sum of each field in "arr" using the indicated communicator with a reproducible yet scalable implementation based on a fixed point algorithm. An alternative is to use an "almost always reproducible" floating point algorithm. More...

subroutine	oasis_reprosum_int (arr, arr_gsum, nsummands, dsummands, nflds, arr_max_shift, arr_gmax_exp, max_levels, max_level, validate, recompute, omp_nthreads, mpi_comm)
	Compute the global sum of each field in "arr" using the indicated communicator with a reproducible yet scalable implementation based on a fixed point algorithm. The accuracy of the fixed point algorithm is controlled by the number of "levels" of integer expansion, the maximum value of which is specified by max_level. More...

logical function, public	oasis_reprosum_tolexceeded (name, nflds, master, logunit, rel_diff)
	Test whether distributed sum exceeds tolerance and print out a warning message. More...

subroutine	oasis_reprosum_ddpdd (arr, arr_gsum, nsummands, dsummands, nflds, mpi_comm)
	Compute the global sum of each field in "arr" using the indicated communicator with a reproducible yet scalable implementation based on He and Ding's implementation of the double-double algorithm. More...

subroutine, private	ddpdd (dda, ddb, len, itype)

subroutine, private	split_indices (total, num_pieces, ibeg, iend)

Variables
logical, public	oasis_reprosum_recompute = .false.

real(r8), public	oasis_reprosum_reldiffmax = -1.0_r8

logical	repro_sum_use_ddpdd = .false.

logical	detailed_timing = .false.

Detailed Description

OASIS reproducible sum method from P. Worley.

Function/Subroutine Documentation

◆ ddpdd()

subroutine, private mod_oasis_reprosum::ddpdd	(	complex(r8), dimension(len), intent(in)	dda,
		complex(r8), dimension(len), intent(inout)	ddb,
		integer, intent(in)	len,
		integer, intent(in)	itype
	)

private

Definition at line 1354 of file mod_oasis_reprosum.F90.

◆ oasis_reprosum_calc()

subroutine, public mod_oasis_reprosum::oasis_reprosum_calc	(	real(r8), dimension(dsummands,nflds), intent(in)	arr,
		real(r8), dimension(nflds), intent(out)	arr_gsum,
		integer, intent(in)	nsummands,
		integer, intent(in)	dsummands,
		integer, intent(in)	nflds,
		logical, intent(in), optional	ddpdd_sum,
		real(r8), dimension(nflds), intent(in), optional	arr_gbl_max,
		real(r8), dimension(nflds), intent(out), optional	arr_gbl_max_out,
		integer, dimension(nflds), intent(in), optional	arr_max_levels,
		integer, dimension(nflds), intent(out), optional	arr_max_levels_out,
		integer, intent(in), optional	gbl_max_nsummands,
		integer, intent(out), optional	gbl_max_nsummands_out,
		integer, intent(in), optional	gbl_count,
		logical, intent(in), optional	repro_sum_validate,
		integer, dimension(5), intent(inout), optional	repro_sum_stats,
		real(r8), dimension(2,nflds), intent(out), optional	rel_diff,
		integer, intent(in), optional	commid
	)

Compute the global sum of each field in "arr" using the indicated communicator with a reproducible yet scalable implementation based on a fixed point algorithm. An alternative is to use an "almost always reproducible" floating point algorithm.

Parameters

[in]	nsummands	number of local summands
[in]	dsummands	declared first dimension
[in]	nflds	number of fields
[in]	arr	input array
[out]	arr_gsum	global means
[in]	ddpdd_sum	use ddpdd algorithm instead of fixed precision algorithm
[in]	arr_gbl_max	upper bound on max(abs(arr))
[out]	arr_gbl_max_out	calculated upper bound on max(abs(arr))
[in]	arr_max_levels	maximum number of levels of integer expansion to use
[out]	arr_max_levels_out	output of number of levels of integer expansion to used
[in]	gbl_max_nsummands	maximum of nsummand over all processes
[out]	gbl_max_nsummands_out	calculated maximum nsummands over all processes
[in]	gbl_count	was total number of summands; now is ignored; use gbl_max_nsummands instead
[in]	repro_sum_validate	flag enabling/disabling testing that gmax and max_levels are accurate/sufficient. Default is enabled.
[in,out]	repro_sum_stats	increment running totals for (1) one-reduction repro_sum (2) two-reduction repro_sum (3) both types in one call (4) nonrepro_sum (5) global max nsummands reduction
[out]	rel_diff	relative and absolute differences between fixed and floating point sums
[in]	commid	MPI communicator

Definition at line 267 of file mod_oasis_reprosum.F90.

◆ oasis_reprosum_ddpdd()

subroutine mod_oasis_reprosum::oasis_reprosum_ddpdd	(	real(r8), dimension(dsummands,nflds), intent(in)	arr,
		real(r8), dimension(nflds), intent(out)	arr_gsum,
		integer, intent(in)	nsummands,
		integer, intent(in)	dsummands,
		integer, intent(in)	nflds,
		integer, intent(in)	mpi_comm
	)

private

Compute the global sum of each field in "arr" using the indicated communicator with a reproducible yet scalable implementation based on He and Ding's implementation of the double-double algorithm.

Parameters

[in]	nsummands	number of local summands
[in]	dsummands	declared first dimension
[in]	nflds	number of fields
[in]	arr	input array
[in]	mpi_comm	MPI subcommunicator
[out]	arr_gsum	global sums

Definition at line 1281 of file mod_oasis_reprosum.F90.

◆ oasis_reprosum_int()

subroutine mod_oasis_reprosum::oasis_reprosum_int	(	real(r8), dimension(dsummands,nflds), intent(in)	arr,
		real(r8), dimension(nflds), intent(out)	arr_gsum,
		integer, intent(in)	nsummands,
		integer, intent(in)	dsummands,
		integer, intent(in)	nflds,
		integer, intent(in)	arr_max_shift,
		integer, dimension(nflds), intent(in)	arr_gmax_exp,
		integer, dimension(nflds), intent(in)	max_levels,
		integer, intent(in)	max_level,
		logical, intent(in)	validate,
		logical, intent(out)	recompute,
		integer, intent(in)	omp_nthreads,
		integer, intent(in)	mpi_comm
	)

private

Compute the global sum of each field in "arr" using the indicated communicator with a reproducible yet scalable implementation based on a fixed point algorithm. The accuracy of the fixed point algorithm is controlled by the number of "levels" of integer expansion, the maximum value of which is specified by max_level.

Parameters

[in]	nsummands	number of local summands
[in]	dsummands	declared first dimension
[in]	nflds	number of fields
[in]	arr_max_shift	maximum safe exponent for value < 1 (so that sum does not overflow)
[in]	arr_gmax_exp	exponents of global maxima
[in]	max_levels	maximum number of levels of integer expansion
[in]	max_level	maximum value in max_levels
[in]	omp_nthreads	number of OpenMP threads
[in]	mpi_comm	MPI subcommunicator
[in]	arr	input array
[in]	validate	flag indicating that accuracy of solution generated from arr_gmax_exp and max_levels should be tested
[out]	recompute	flag indicating that either the upper bounds are inaccurate, or max_levels and arr_gmax_exp do not generate accurate enough sums
[out]	arr_gsum	global means

Definition at line 751 of file mod_oasis_reprosum.F90.

◆ oasis_reprosum_setopts()

subroutine, public mod_oasis_reprosum::oasis_reprosum_setopts	(	logical, intent(in), optional	repro_sum_use_ddpdd_in,
		real(r8), intent(in), optional	repro_sum_rel_diff_max_in,
		logical, intent(in), optional	repro_sum_recompute_in,
		logical, intent(in), optional	repro_sum_master,
		integer, intent(in), optional	repro_sum_logunit
	)

Set runtime options.

Parameters

[in]	repro_sum_use_ddpdd_in	Use DDPDD algorithm instead of fixed precision algorithm
[in]	repro_sum_rel_diff_max_in	maximum permissible difference between reproducible and nonreproducible sums
[in]	repro_sum_recompute_in	recompute using different algorithm when difference between reproducible and nonreproducible sums is too great
[in]	repro_sum_master	flag indicating whether this process should output log messages
[in]	repro_sum_logunit	unit number for log messages

Definition at line 105 of file mod_oasis_reprosum.F90.

◆ oasis_reprosum_tolexceeded()

logical function, public mod_oasis_reprosum::oasis_reprosum_tolexceeded	(	character(len=*), intent(in)	name,
		integer, intent(in)	nflds,
		logical, intent(in)	master,
		integer, intent(in), optional	logunit,
		real(r8), dimension(2,nflds), intent(in)	rel_diff
	)

Test whether distributed sum exceeds tolerance and print out a warning message.

Parameters

[in]	name	distributed sum identifier
[in]	nflds	number of fields
[in]	master	process that will write warning messages?
[in]	logunit	unit warning messages written to
[in]	rel_diff	relative and absolute differences between fixed and floating point sums

Definition at line 1195 of file mod_oasis_reprosum.F90.

◆ split_indices()

subroutine, private mod_oasis_reprosum::split_indices	(	integer, intent(in)	total,
		integer, intent(in)	num_pieces,
		integer, dimension(num_pieces), intent(out)	ibeg,
		integer, dimension(num_pieces), intent(out)	iend
	)

private

Definition at line 1394 of file mod_oasis_reprosum.F90.

Variable Documentation

◆ detailed_timing

logical mod_oasis_reprosum::detailed_timing = .false.

private

Definition at line 89 of file mod_oasis_reprosum.F90.

◆ oasis_reprosum_recompute

logical, public mod_oasis_reprosum::oasis_reprosum_recompute = .false.

Definition at line 73 of file mod_oasis_reprosum.F90.

◆ oasis_reprosum_reldiffmax

real(r8), public mod_oasis_reprosum::oasis_reprosum_reldiffmax = -1.0_r8

Definition at line 75 of file mod_oasis_reprosum.F90.

◆ repro_sum_use_ddpdd

logical mod_oasis_reprosum::repro_sum_use_ddpdd = .false.

private

Definition at line 88 of file mod_oasis_reprosum.F90.

Functions/Subroutines

Variables

Detailed Description

Function/Subroutine Documentation

◆ ddpdd()

◆ oasis_reprosum_calc()

◆ oasis_reprosum_ddpdd()

◆ oasis_reprosum_int()

◆ oasis_reprosum_setopts()

◆ oasis_reprosum_tolexceeded()

◆ split_indices()

Variable Documentation

◆ detailed_timing

◆ oasis_reprosum_recompute

◆ oasis_reprosum_reldiffmax

◆ repro_sum_use_ddpdd