Program mod1b ! ---------------------------------------------------------------------- ! ********************************************************************** ! *** This program is part of the EuroBen Benchmark *** ! *** Copyright: EuroBen Group p/o *** ! *** Utrecht University, Computational Physics Dept. *** ! *** P.O. Box 80.000 *** ! *** 3508 TA Utrecht *** ! *** The Netherlands *** ! *** *** ! *** Authors of this program: Peter P.M. de Rijk and *** ! *** Aad J. van der Steen *** ! *** Contributed: Autumn 1990 *** ! *** Last change: Autumn 2003 *** ! ********************************************************************** ! Version 4.2 ! ! --- Purpose of this module ! -------------------------- ! This program times some kernels for some vector lengths and ! calculates their MFlop rate. ! ! In addition, n(1/2) and r(infi) are estimated for the same kernels. ! ! Both parameters are obtained by a linear least squares fit on the ! function f(n) = alfa + beta*n ! ! The following holds: r(infi) = 1/beta (as long as in primary cache) ! n(1/2) = alfa / beta ! !----------------------------------------------------------------------- Use numerics Implicit None ! --- Universal Constants Real(l_), Parameter :: zero = 0.0_l_, half = 0.5_l_, & one = 1.0_l_, two = 2.0_l_, & oneneg =-1.0_l_ ! --- Variables used for measurements Real(l_), Allocatable :: x1(:), x2(:), x3(:), x4(:), y(:,:) ! --- Parameters used for no. of measuring methods and lower bound of ! time resolution accepted. Real(l_), Parameter :: timelb=1.0e-12_l_ ! --- Parameters used in the Repetition Factor for the timings Integer :: nfixed, mrep != 10 * nfixed ! --- Variables used to prevent overoptimisation in timings Integer, Allocatable :: indj(:) Integer :: ndummy ! --- Number of kernels to be examined Integer :: nkern ! --- Character variables Character :: symbol*60 ! --- Stuff to be calculated Integer :: nflop, ntrans Real(l_) :: time, rmflop, rtrans, cycle, fpc ! --- Counters and other stuff Integer :: i, k, kn, n, nrep Integer :: omp_get_num_threads, nprocs Real(l_) :: tbegin, tend ! --- Introduce Common block to frustrate overoptimisation. Real(l_) :: s(2), ssw Common /cfake/ s, ssw ! --- External functions Real(l_) :: wclock, dran1 ! --- Initialization ! ---------------------------------------------------------------------- ! --- Call module identification routine Call state ( 'mod1b ' ) ! ---------------------------------------------------------------------- tbegin = wclock() !$omp parallel shared(nprocs) nprocs = omp_get_num_threads() !$omp end parallel Print 9010, nprocs Open( 1, File = 'mod1b.in' ) ! Contains vector lengths & rep. fac. Open( 2, File = 'mod1b.krn' )! Contains flops executed in kernel. Open( 3, File = 'mod1b.cyc' )! Contains clock frequency of proc. ! (in Hz.) ! --- Get clock frequency of processor. Read( 3, * ) cycle ! --- Get number of kernels. nkern = 0 10 Read( 2, *, End = 20 ) symbol, nflop nkern = nkern + 1 Go To 10 20 Rewind ( 2 ) Do k = 1, nkern Read ( 2, * ) symbol, nflop, ntrans, kn Print 9030, kn, symbol, nflop ! --- Get problem size and repetition factor. 30 Read( 1, *, End = 50 ) n, nfixed mrep = 10*nfixed ! --- Initialize index vector to be used against overoptimization. ndummy = -1999 Allocate ( indj(mrep) ) Do i = 1, mrep 40 Continue indj(i) = Int( dran1( ndummy ) * mrep ) + 1 ! --- Avoid problematic cases (just to be sure ...) If( indj(i) mrep ) Go To 40 End Do ! --- Get vector length, allocate, and initialise. nrep = Max( 10, Min( Int(nfixed/n) * 100, mrep ) ) Allocate( x1(4*n), x2(4*n), x3(4*n), x4(4*n), y(4*n,2) ) x1 = half x2 = oneneg x3 = two x4 = half y = one ! --- Time vector operation Call timer( kn, n, x1, x2, x3, x4, y, 4*n, indj, mrep/2, nrep, & time ) ! --- Average MFlop/s If( nflop == 0 ) nflop = 1 rmflop = Real(nflop, l_)*1.0e-6_l_*Real(n, l_) / & Max( time, timelb ) rtrans = ntrans*1.0e-9_l_*Real(n, l_)/Max( time, timelb ) fpc = rmflop*1.0e5_l_/cycle Print 9040, n, time, rmflop, rtrans, fpc, nrep Deallocate( x1, x2, x3, x4, y, indj ) Go To 30 50 Print 9020 Rewind( 1 ) End Do tend = wclock() Print 9050, tend - tbegin ! ---------------------------------------------------------------------- 9010 Format( &' ------------------------------------------------------------'/ &' | Performance measurement of some basic kernels |'/ &' | ======================================================== |'/ &' | A measuring method is used that evades overoptimisation |'/ &' | of the repetition loop that is used for better timing |'/ &' | accuracy by calling the function JSWITCH. The call |'/ &' | of this funtion depends on the iteration counter of |'/ &' | the repetition loop. |'/ &' ------------------------------------------------------------'// &' No. of proc.s = ', i4 ) 9020 Format( 1x, 72('-') ) 9030 Format( //,1x, 72('=')/1x, 'Kernel', i3, ': ', a/ &12x,'No. of flops per Iteration =', i3/ &1x, 72('-')/ &1x, ' Loop ','| ', 'CPU Time', 2x,'|', 2x, 'Average', 4x, '|', $' Average', 4x, '|', ' Flop per ', '|', 2x, 'Repeat |'/ &1x, ' length ','| ', ' sec ', 2x,'|', 2x, 'Mflop/s', 4x, '|', $' GB/s ', 4x, '|', ' cycle ', '|',2x, ' count |'/ &1x, 72('-') ) 9040 Format( 1x, i8, '|', 1pg13.5, '|', g13.5, '|', g13.5, '| ', & f8.4, ' |', i8, ' |' ) 9050 Format( //1x, 'Total execution time: ',g12.5, ' sec.' ) ! ---------------------------------------------------------------------- End Program mod1b
