diff --git a/samples/advanced/pdegen/amg_d_comm_test.F90 b/samples/advanced/pdegen/amg_d_comm_test.F90
index 883457fa..7b091542 100644
--- a/samples/advanced/pdegen/amg_d_comm_test.F90
+++ b/samples/advanced/pdegen/amg_d_comm_test.F90
@@ -26,7 +26,7 @@
! left over from a previous repetition must be freed explicitly.
!
! Run with:
-! mpirun -np ./amg_d_comm_test [idim] [nrep] [nlev] [itmax]
+! mpirun -np
./amg_d_comm_test [idim] [nrep] [nlev] [itmax] [nwarm]
! [--remap] [--csv=] [--mode=all|uniform|sensitivity]
!
program amg_d_comm_test
@@ -52,7 +52,7 @@ program amg_d_comm_test
integer(psb_ipk_) :: info, my_rank, np
integer(psb_ipk_) :: idim, itmax, itrace, istop
- integer(psb_ipk_) :: s_idx, rep, nrep, nlev, lev, nlev_built, csv_unit
+ integer(psb_ipk_) :: s_idx, rep, nrep, nwarm, nlev, lev, nlev_built, csv_unit
integer(psb_ipk_) :: scheme_type(n_schemes)
integer(psb_ipk_) :: scheme_of_level(max_lev)
character(len=25) :: scheme_name(n_schemes)
@@ -69,7 +69,7 @@ program amg_d_comm_test
afmt = 'CSR'
idim = 40
nrep = 5
- nlev = 3
+ nlev = 3
itmax = 1000
itrace = -1
istop = 2
@@ -95,6 +95,15 @@ program amg_d_comm_test
call read_int_arg(2, nrep, 5)
call read_int_arg(3, nlev, 3)
call read_int_arg(4, itmax,1000)
+ ! Warm-up repetitions, discarded from the statistics.
+ !
+ ! The first run of a scheme is not the same measurement as the ones after it.
+ ! Beyond the usual first-touch effects, whichever scheme creates the first RMA
+ ! window of the program pays the one-off initialisation of the one-sided
+ ! transport -- at 448 ranks that was 1600 s aggregated against 62 s for the
+ ! very same code path in the scheme that ran second. Folded into an average it
+ ! is charged entirely to rma_pull, which is simply not where it belongs.
+ call read_int_arg(5, nwarm, 1)
call parse_flags(do_remap, csv_file, run_mode)
want_csv = (len_trim(csv_file) > 0)
@@ -109,7 +118,7 @@ program amg_d_comm_test
write(psb_out_unit,'("Number of processors : ",i0)') np
write(psb_out_unit,'("Preconditioner : ML, V-cycle, JACOBI, max levels ",i0)') nlev
write(psb_out_unit,'("Iterative method : CG, itmax ",i0,", eps ",es9.2)') itmax, eps
- write(psb_out_unit,'("Repetitions : ",i0)') nrep
+ write(psb_out_unit,'("Repetitions : ",i0," (+ ",i0," warm-up, not averaged)")') nrep, nwarm
write(psb_out_unit,'("Remap active : ",l1)') amg_get_do_remap()
write(psb_out_unit,'("Mode : ",a)') trim(run_mode)
if (want_csv) write(psb_out_unit,'("CSV output : ",a)') trim(csv_file)
@@ -155,9 +164,9 @@ program amg_d_comm_test
end if
do s_idx = 1, n_schemes
scheme_of_level(:) = scheme_type(s_idx)
- do rep = 1, nrep
+ do rep = 1-nwarm, nrep
call run_one('uniform', trim(scheme_name(s_idx)), izero, &
- & scheme_of_level, (rep == 1), info)
+ & scheme_of_level, (rep == 1), (rep <= 0), info)
if (info /= psb_success_) goto 9999
end do
end do
@@ -173,7 +182,7 @@ program amg_d_comm_test
if (nlev_built <= 0) then
! Need one build to learn how many levels the hierarchy actually has.
scheme_of_level(:) = scheme_type(1)
- call run_one('probe', trim(scheme_name(1)), izero, scheme_of_level, .false., info)
+ call run_one('probe', trim(scheme_name(1)), izero, scheme_of_level, .false., .false., info)
if (info /= psb_success_) goto 9999
end if
if (my_rank == psb_root_) then
@@ -186,9 +195,9 @@ program amg_d_comm_test
do s_idx = 2, n_schemes
scheme_of_level(:) = scheme_type(1)
scheme_of_level(lev) = scheme_type(s_idx)
- do rep = 1, nrep
+ do rep = 1-nwarm, nrep
call run_one('sensitivity', trim(scheme_name(s_idx)), lev, &
- & scheme_of_level, .false., info)
+ & scheme_of_level, .false., (rep <= 0), info)
if (info /= psb_success_) goto 9999
end do
end do
@@ -226,17 +235,27 @@ contains
! One full init/build/solve with the given per-level scheme assignment.
! Timings are max-reduced across ranks before being recorded.
!
- subroutine run_one(mode, sname, target_lev, sch_of_lev, verbose, info)
+ subroutine run_one(mode, sname, target_lev, sch_of_lev, verbose, warmup, info)
character(len=*), intent(in) :: mode, sname
integer(psb_ipk_), intent(in) :: target_lev
integer(psb_ipk_), intent(in) :: sch_of_lev(:)
logical, intent(in) :: verbose
+ logical, intent(in) :: warmup
integer(psb_ipk_), intent(out) :: info
real(psb_dpk_) :: t_start, t_init, t_bld, t_comm, t_kry, err
integer(psb_ipk_) :: iter, lv
+ character(len=32) :: mode_out
info = psb_success_
+ ! Warm-up rows are kept, not discarded: they carry the setup cost, which is
+ ! a result in its own right. The tag keeps them out of the averages, since
+ ! the analysis only picks up 'uniform' and 'sensitivity'.
+ if (warmup) then
+ mode_out = 'warmup_'//trim(mode)
+ else
+ mode_out = mode
+ end if
! Fine level: must be set before the build, so that desc_a and every vector
! lazily initialised during the solve pick this scheme up.
@@ -317,13 +336,13 @@ contains
end do
end if
if (trim(mode) /= 'probe') then
- write(psb_out_unit,'(a14,1x,a26,1x,"lev ",i2,1x,"it ",i5,1x,"err ",es12.5,&
+ write(psb_out_unit,'(a18,1x,a26,1x,"lev ",i2,1x,"it ",i5,1x,"err ",es12.5,&
&1x,"bld ",es11.4,1x,"solve ",es11.4)') &
- & trim(mode), trim(sname), target_lev, iter, err, t_bld, t_kry
+ & trim(mode_out), trim(sname), target_lev, iter, err, t_bld, t_kry
end if
if (want_csv .and. header_done .and. (trim(mode) /= 'probe')) then
write(csv_unit,'(a,",",a,",",i0,",",i0,",",i0,",",i0,",",i0,5(",",es16.9),",",i0,",",es16.9,",",l1)') &
- & trim(mode), trim(sname), target_lev, np, idim, nlev_built, rep, &
+ & trim(mode_out), trim(sname), target_lev, np, idim, nlev_built, rep, &
& t_init, t_bld, t_comm, t_kry, t_init+t_bld+t_comm+t_kry, &
& iter, err, do_remap
end if