From b82192d01982d4213c38e9c398536f300d02b92c Mon Sep 17 00:00:00 2001
From: Dorian Stoll <dorian.stoll@tmsp.io>
Date: Thu, 13 Jun 2024 15:45:16 +0200
Subject: [PATCH] nas-ft: fortran: Remove built-in timekeeping
We are only interested in the runtime of the entire program and that
can be measured more easily.
---
src/benchmarks/nas-ft/fortran/ft.f90 | 148 ++++-----------
src/benchmarks/nas-ft/fortran/ft_data.f90 | 36 ++--
src/benchmarks/nas-ft/fortran/meson.build | 1 -
.../nas-ft/fortran/print_results.f90 | 22 +--
src/benchmarks/nas-ft/fortran/randi8.f90 | 2 +-
src/benchmarks/nas-ft/fortran/timers.f90 | 171 ------------------
6 files changed, 53 insertions(+), 327 deletions(-)
delete mode 100644 src/benchmarks/nas-ft/fortran/timers.f90
diff --git a/src/benchmarks/nas-ft/fortran/ft.f90 b/src/benchmarks/nas-ft/fortran/ft.f90
index cc8ba00c..86452da4 100644
--- a/src/benchmarks/nas-ft/fortran/ft.f90
+++ b/src/benchmarks/nas-ft/fortran/ft.f90
@@ -66,23 +66,16 @@
implicit none
- integer i
-
integer iter
- double precision total_time, mflops
logical verified
character class
!---------------------------------------------------------------------
-! Run the entire problem once to make sure all data is touched.
-! This reduces variable startup costs, which is important for such a
-! short benchmark. The other NPB 2 implementations are similar.
+! Run the entire problem once to make sure all data is touched.
+! This reduces variable startup costs, which is important for such a
+! short benchmark. The other NPB 2 implementations are similar.
!---------------------------------------------------------------------
- do i = 1, t_max
- call timer_clear(i)
- end do
-
call alloc_space
call setup()
@@ -94,57 +87,29 @@
!---------------------------------------------------------------------
! Start over from the beginning. Note that all operations must
-! be timed, in contrast to other benchmarks.
+! be timed, in contrast to other benchmarks.
!---------------------------------------------------------------------
- do i = 1, t_max
- call timer_clear(i)
- end do
-
- call timer_start(T_total)
- if (timers_enabled) call timer_start(T_setup)
-
call compute_indexmap(twiddle, dims(1), dims(2), dims(3))
call compute_initial_conditions(u1, dims(1), dims(2), dims(3))
call fft_init (dims(1))
- if (timers_enabled) call timer_stop(T_setup)
- if (timers_enabled) call timer_start(T_fft)
call fft(1, u1, u0)
- if (timers_enabled) call timer_stop(T_fft)
do iter = 1, niter
- if (timers_enabled) call timer_start(T_evolve)
call evolve(u0, u1, twiddle, dims(1), dims(2), dims(3))
- if (timers_enabled) call timer_stop(T_evolve)
- if (timers_enabled) call timer_start(T_fft)
! call fft(-1, u1, u2)
call fft(-1, u1, u1)
- if (timers_enabled) call timer_stop(T_fft)
- if (timers_enabled) call timer_start(T_checksum)
! call checksum(iter, u2, dims(1), dims(2), dims(3))
call checksum(iter, u1, dims(1), dims(2), dims(3))
- if (timers_enabled) call timer_stop(T_checksum)
end do
call verify(nx, ny, nz, niter, verified, class)
- call timer_stop(t_total)
- total_time = timer_read(t_total)
-
- if( total_time .ne. 0. ) then
- mflops = 1.0d-6*ntotal_f * &
- & (14.8157+7.19641*log(ntotal_f) &
- & + (5.23518+7.21113*log(ntotal_f))*niter) &
- & /total_time
- else
- mflops = 0.0
- endif
call print_results('FT', class, nx, ny, nz, niter, &
- & total_time, mflops, ' floating point', verified, &
+ & ' floating point', verified, &
& npbversion, compiletime, cs1, cs2, cs3, cs4, cs5, cs6, cs7)
- if (timers_enabled) call print_timers()
end
@@ -225,8 +190,8 @@
!---------------------------------------------------------------------
!---------------------------------------------------------------------
-! Fill in array u0 with initial conditions from
-! random number generator
+! Fill in array u0 with initial conditions from
+! random number generator
!---------------------------------------------------------------------
use ft_data
@@ -236,7 +201,7 @@
double complex u0(d1+1, d2, d3)
integer k, j
double precision x0, start, an, dummy, starts(nz)
-
+
start = seed
!---------------------------------------------------------------------
@@ -251,14 +216,14 @@
dummy = randlc(start, an)
starts(k) = start
end do
-
+
!---------------------------------------------------------------------
! Go through by z planes filling in one square at a time.
!---------------------------------------------------------------------
!$omp parallel do default(shared) private(k,j,x0)
- do k = 1, dims(3)
+ do k = 1, dims(3)
x0 = starts(k)
- do j = 1, dims(2)
+ do j = 1, dims(2)
call vranlc(2*nx, x0, a, u0(1, j, k))
end do
end do
@@ -299,7 +264,7 @@
do while (n .gt. 1)
n2 = n/2
if (n2 * 2 .eq. n) then
- dummy = randlc(q, q)
+ dummy = randlc(q, q)
n = n2
else
dummy = randlc(r, q)
@@ -327,8 +292,6 @@
!$ external omp_get_max_threads
debug = .FALSE.
- call check_timer_flag( timers_enabled )
-
write(*, 1000)
niter = niter_default
@@ -354,15 +317,15 @@
! Set up info for blocking of ffts and transposes. This improves
! performance on cache-based systems. Blocking involves
! working on a chunk of the problem at a time, taking chunks
-! along the first, second, or third dimension.
+! along the first, second, or third dimension.
!
! - In cffts1 blocking is on 2nd dimension (with fft on 1st dim)
! - In cffts2/3 blocking is on 1st dimension (with fft on 2nd and 3rd dims)
-! Since 1st dim is always in processor, we'll assume it's long enough
+! Since 1st dim is always in processor, we'll assume it's long enough
! (default blocking factor is 16 so min size for 1st dim is 16)
-! The only case we have to worry about is cffts1 in a 2d decomposition.
-! so the blocking factor should not be larger than the 2nd dimension.
+! The only case we have to worry about is cffts1 in a 2d decomposition.
+! so the blocking factor should not be larger than the 2nd dimension.
!---------------------------------------------------------------------
fftblock = fftblock_default
@@ -373,7 +336,7 @@
return
end
-
+
!---------------------------------------------------------------------
!---------------------------------------------------------------------
@@ -383,8 +346,8 @@
!---------------------------------------------------------------------
!---------------------------------------------------------------------
-! compute function from local (i,j,k) to ibar^2+jbar^2+kbar^2
-! for time evolution exponent.
+! compute function from local (i,j,k) to ibar^2+jbar^2+kbar^2
+! for time evolution exponent.
!---------------------------------------------------------------------
use ft_data
@@ -396,9 +359,9 @@
double precision ap
!---------------------------------------------------------------------
-! basically we want to convert the fortran indices
-! 1 2 3 4 5 6 7 8
-! to
+! basically we want to convert the fortran indices
+! 1 2 3 4 5 6 7 8
+! to
! 0 1 2 3 -4 -3 -2 -1
! The following magic formula does the trick:
! mod(i-1+n/2, n) - n/2
@@ -426,41 +389,6 @@
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- subroutine print_timers()
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- use ft_data
- implicit none
-
- integer i
- double precision t, t_m
- character*25 tstrings(T_max)
- data tstrings / ' total ', &
- & ' setup ', &
- & ' fft ', &
- & ' evolve ', &
- & ' checksum ', &
- & ' fftx ', &
- & ' ffty ', &
- & ' fftz ' /
-
- t_m = timer_read(T_total)
- if (t_m .le. 0.0d0) t_m = 1.0d0
- do i = 1, t_max
- t = timer_read(i)
- write(*, 100) i, tstrings(i), t, t*100.0/t_m
- end do
- 100 format(' timer ', i2, '(', A16, ') :', F9.4, ' (',F6.2,'%)')
- return
- end
-
-
-
!---------------------------------------------------------------------
!---------------------------------------------------------------------
@@ -518,7 +446,6 @@
logd1 = ilog2(d1)
- if (timers_enabled) call timer_start(T_fftx)
!$omp parallel do default(shared) private(i,j,k,jj,y1,y2,jn) &
!$omp& shared(is,logd1,d1) collapse(2)
do k = 1, d3
@@ -530,7 +457,7 @@
y1(j,i) = x(i,j+jj,k)
enddo
enddo
-
+
call cfftz (is, logd1, d1, y1, y2)
@@ -541,7 +468,6 @@
enddo
enddo
enddo
- if (timers_enabled) call timer_stop(T_fftx)
return
end
@@ -566,7 +492,6 @@
logd2 = ilog2(d2)
- if (timers_enabled) call timer_start(T_ffty)
!$omp parallel do default(shared) private(i,j,k,ii,y1,y2,in) &
!$omp& shared(is,logd2,d2) collapse(2)
do k = 1, d3
@@ -580,7 +505,7 @@
enddo
call cfftz (is, logd2, d2, y1, y2)
-
+
do j = 1, d2
do i = 1, fftblock
xout(i+ii,j,k) = y1(i,j)
@@ -588,7 +513,6 @@
enddo
enddo
enddo
- if (timers_enabled) call timer_stop(T_ffty)
return
end
@@ -613,7 +537,6 @@
logd3 = ilog2(d3)
- if (timers_enabled) call timer_start(T_fftz)
!$omp parallel do default(shared) private(i,j,k,ii,y1,y2,in) &
!$omp& shared(is) collapse(2)
do j = 1, d2
@@ -635,7 +558,6 @@
enddo
enddo
enddo
- if (timers_enabled) call timer_stop(T_fftz)
return
end
@@ -650,7 +572,7 @@
!---------------------------------------------------------------------
!---------------------------------------------------------------------
-! compute the roots-of-unity array that will be used for subsequent FFTs.
+! compute the roots-of-unity array that will be used for subsequent FFTs.
!---------------------------------------------------------------------
use ft_data
@@ -672,16 +594,16 @@
do j = 1, m
t = pi / ln
-
+
do i = 0, ln - 1
ti = i * t
u(i+ku) = dcmplx (cos (ti), sin(ti))
enddo
-
+
ku = ku + ln
ln = 2 * ln
enddo
-
+
return
end
@@ -695,10 +617,10 @@
!---------------------------------------------------------------------
! Computes NY N-point complex-to-complex FFTs of X using an algorithm due
-! to Swarztrauber. X is both the input and the output array, while Y is a
-! scratch array. It is assumed that N = 2^M. Before calling CFFTZ to
-! perform FFTs, the array U must be initialized by calling CFFTZ with IS
-! set to 0 and M set to MX, where MX is the maximum value of M for any
+! to Swarztrauber. X is both the input and the output array, while Y is a
+! scratch array. It is assumed that N = 2^M. Before calling CFFTZ to
+! perform FFTs, the array U must be initialized by calling CFFTZ with IS
+! set to 0 and M set to MX, where MX is the maximum value of M for any
! subsequent call.
!---------------------------------------------------------------------
@@ -857,7 +779,7 @@
end do
chk = chk/ntotal_f
-
+
write (*, 30) i, chk
30 format (' T =',I5,5X,'Checksum =',1P2D22.12)
sums(i) = chk
@@ -940,7 +862,7 @@
csum_ref(4) = dcmplx(5.077892868474D+02, 5.101336130759D+02)
csum_ref(5) = dcmplx(5.085233095391D+02, 5.104914655194D+02)
csum_ref(6) = dcmplx(5.091487099959D+02, 5.107917842803D+02)
-
+
else if (d1 .eq. 512 .and. &
& d2 .eq. 256 .and. &
& d3 .eq. 256 .and. &
@@ -1115,7 +1037,7 @@
endif
-
+
if (class .ne. 'U') then
if (verified) then
write(*,2000)
diff --git a/src/benchmarks/nas-ft/fortran/ft_data.f90 b/src/benchmarks/nas-ft/fortran/ft_data.f90
index 9f17d49d..a94b1e34 100644
--- a/src/benchmarks/nas-ft/fortran/ft_data.f90
+++ b/src/benchmarks/nas-ft/fortran/ft_data.f90
@@ -22,9 +22,9 @@
! Cache blocking params. These values are good for most
-! RISC processors.
+! RISC processors.
! FFT parameters:
-! fftblock controls how many ffts are done at a time.
+! fftblock controls how many ffts are done at a time.
! The default is appropriate for most cache-based machines
! On vector machines, the FFT can be vectorized with vector
! length equal to the block size, so the block size should
@@ -35,11 +35,11 @@
include 'blk_par.h'
! integer fftblock_default, fftblockpad_default
! parameter (fftblock_default=32, fftblockpad_default=34)
-
+
integer fftblock, fftblockpad
! we need a bunch of logic to keep track of how
-! arrays are laid out.
+! arrays are laid out.
! Note: this serial version is the derived from the parallel 0D case
@@ -61,29 +61,15 @@
! compute residual(1)
! for the 0D, 1D, 2D strategies, the layouts look like xxx
-!
+!
! 0D 1D 2D
! 1: xyz xyz xyz
! the array dimensions are stored in dims(coord, phase)
integer dims(3)
- integer T_total, T_setup, T_fft, T_evolve, T_checksum, &
- & T_fftx, T_ffty, &
- & T_fftz, T_max
- parameter (T_total = 1, T_setup = 2, T_fft = 3, &
- & T_evolve = 4, T_checksum = 5, &
- & T_fftx = 6, &
- & T_ffty = 7, &
- & T_fftz = 8, T_max = 8)
-
-
-
- logical timers_enabled
- external timer_read
- double precision timer_read
external ilog2
integer ilog2
@@ -125,14 +111,14 @@
module ft_fields
!---------------------------------------------------------------------
-! u0, u1, u2 are the main arrays in the problem.
-! Depending on the decomposition, these arrays will have different
-! dimensions. To accomodate all possibilities, we allocate them as
-! one-dimensional arrays and pass them to subroutines for different
+! u0, u1, u2 are the main arrays in the problem.
+! Depending on the decomposition, these arrays will have different
+! dimensions. To accomodate all possibilities, we allocate them as
+! one-dimensional arrays and pass them to subroutines for different
! views
! - u0 contains the initial (transformed) initial condition
! - u1 and u2 are working arrays
-! - twiddle contains exponents for the time evolution operator.
+! - twiddle contains exponents for the time evolution operator.
!---------------------------------------------------------------------
double complex, allocatable :: &
@@ -143,7 +129,7 @@
!---------------------------------------------------------------------
! Large arrays are in module so that they are allocated on the
! heap rather than the stack. This module is not
-! referenced directly anywhere else. Padding is to avoid accidental
+! referenced directly anywhere else. Padding is to avoid accidental
! cache problems, since all array sizes are powers of two.
!---------------------------------------------------------------------
diff --git a/src/benchmarks/nas-ft/fortran/meson.build b/src/benchmarks/nas-ft/fortran/meson.build
index e3e613c4..add5abc1 100644
--- a/src/benchmarks/nas-ft/fortran/meson.build
+++ b/src/benchmarks/nas-ft/fortran/meson.build
@@ -2,7 +2,6 @@ sources = [
'ft.f90',
'ft_data.f90',
'print_results.f90',
- 'timers.f90',
'randi8.f90',
]
diff --git a/src/benchmarks/nas-ft/fortran/print_results.f90 b/src/benchmarks/nas-ft/fortran/print_results.f90
index f6be545c..dc168bcd 100644
--- a/src/benchmarks/nas-ft/fortran/print_results.f90
+++ b/src/benchmarks/nas-ft/fortran/print_results.f90
@@ -1,13 +1,12 @@
subroutine print_results(name, class, n1, n2, n3, niter, &
- & t, mops, optype, verified, npbversion, &
+ & optype, verified, npbversion, &
& compiletime, cs1, cs2, cs3, cs4, cs5, cs6, cs7)
-
+
implicit none
character(len=*) name
character class
integer n1, n2, n3, niter, j
- double precision t, mops
character optype*24, size*15
logical verified
character(len=*) npbversion, compiletime, &
@@ -58,29 +57,20 @@
write (*, 5) niter
5 format(' Iterations = ', 12x, i12)
-
- write (*, 6) t
- 6 format(' Time in seconds = ',12x, f12.2)
if (num_threads .gt. 0) write (*,7) num_threads
7 format(' Total threads = ', 12x, i12)
-
+
if (max_threads .gt. 0) write (*,8) max_threads
8 format(' Avail threads = ', 12x, i12)
- if (num_threads .ne. max_threads) write (*,88)
+ if (num_threads .ne. max_threads) write (*,88)
88 format(' Warning: Threads used differ from threads available')
- write (*,9) mops
- 9 format(' Mop/s total = ',12x, f12.2)
-
- if (num_threads .gt. 0) write (*,10) mops/float( num_threads )
- 10 format(' Mop/s/thread = ', 12x, f12.2)
-
write(*, 11) optype
11 format(' Operation type = ', a24)
- if (verified) then
+ if (verified) then
write(*,12) ' SUCCESSFUL'
else
write(*,12) 'UNSUCCESSFUL'
@@ -115,7 +105,7 @@
write(*, 127) cs7
127 format(' RAND = ', A)
-
+
write (*,130)
130 format(//' Please send all errors/feedbacks to:'// &
& ' NPB Development Team'/ &
diff --git a/src/benchmarks/nas-ft/fortran/randi8.f90 b/src/benchmarks/nas-ft/fortran/randi8.f90
index f8932eda..102bcaa3 100644
--- a/src/benchmarks/nas-ft/fortran/randi8.f90
+++ b/src/benchmarks/nas-ft/fortran/randi8.f90
@@ -47,7 +47,7 @@
! This doesn't work, because the compiler does the calculation in 32
! bits and overflows. No standard way (without f90 stuff) to specify
-! that the rhs should be done in 64 bit arithmetic.
+! that the rhs should be done in 64 bit arithmetic.
! parameter(i246m1=2**46-1)
parameter(d2m46=0.5d0**46)
diff --git a/src/benchmarks/nas-ft/fortran/timers.f90 b/src/benchmarks/nas-ft/fortran/timers.f90
deleted file mode 100644
index 3a50de94..00000000
--- a/src/benchmarks/nas-ft/fortran/timers.f90
+++ /dev/null
@@ -1,171 +0,0 @@
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- module timers
-
- double precision start(64), elapsed(64)
-!$omp threadprivate(start, elapsed)
-
- double precision, external :: elapsed_time
-
- end module timers
-
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- subroutine timer_clear(n)
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- use timers
- implicit none
-
- integer n
-
- elapsed(n) = 0.0
- return
- end
-
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- subroutine timer_start(n)
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- use timers
- implicit none
-
- integer n
-
- start(n) = elapsed_time()
-
- return
- end
-
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- subroutine timer_stop(n)
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- use timers
- implicit none
-
- integer n
-
- double precision t, now
-
- now = elapsed_time()
- t = now - start(n)
- elapsed(n) = elapsed(n) + t
-
- return
- end
-
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- double precision function timer_read(n)
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- use timers
- implicit none
-
- integer n
-
- timer_read = elapsed(n)
-
- return
- end
-
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- double precision function elapsed_time()
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- implicit none
-!$ external omp_get_wtime
-!$ double precision omp_get_wtime
-
- double precision t
- logical mp
-
-! ... Use the OpenMP timer if we can (via C$ conditional compilation)
- mp = .false.
-!$ mp = .true.
-!$ t = omp_get_wtime()
-
- if (.not.mp) then
-! This function must measure wall clock time, not CPU time.
-! Since there is no portable timer in Fortran (77)
-! we call a routine compiled in C (though the C source may have
-! to be tweaked).
- call wtime(t)
-! The following is not ok for "official" results because it reports
-! CPU time not wall clock time. It may be useful for developing/testing
-! on timeshared Crays, though.
-! call second(t)
- endif
-
- elapsed_time = t
-
- return
- end
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- subroutine check_timer_flag( timeron )
-
-!---------------------------------------------------------------------
-!---------------------------------------------------------------------
-
- implicit none
- logical timeron
-
- integer nc, ios
- character(len=20) val
-
- timeron = .false.
-
-! ... Check environment variable "NPB_TIMER_FLAG"
- call get_environment_variable('NPB_TIMER_FLAG', val, nc, ios)
- if (ios .eq. 0) then
- if (nc .le. 0) then
- timeron = .true.
- else if (val(1:1) .ge. '1' .and. val(1:1) .le. '9') then
- timeron = .true.
- else if (val .eq. 'on' .or. val .eq. 'ON' .or. &
- & val .eq. 'yes' .or. val .eq. 'YES' .or. &
- & val .eq. 'true' .or. val .eq. 'TRUE') then
- timeron = .true.
- endif
-
- else
-
-! ... Check if the "timer.flag" file exists
- open (unit=2, file='timer.flag', status='old', iostat=ios)
- if (ios .eq. 0) then
- close(2)
- timeron = .true.
- endif
-
- endif
-
- return
- end
--
GitLab