I have some experience with gfortran compiler but no experience whatsoever with ifort, so I should be doing something terribly wrong in what follows. I have recently installed parallel_studio_xe_2016_update1 in my laptop running Debian (Jessie release). I have prepared a very simple program with a very naive approach to computing Pi using Leibniz series. Probably is one of the worst ways of computing Pi, but I just wanted to check vectorization and the stability of the program. The program has a version for single precision and another version for double precision.
The single precision is the following:
PROGRAM VECT_TEST_SINGLE
IMPLICIT NONE
INTEGER, PARAMETER :: SP = KIND(1.0)
INTEGER, PARAMETER :: DP = KIND(1.0D0)
!
REAL(KIND=SP) :: Pi_app
REAL(KIND=DP) :: time_start, time_end, time_tot
INTEGER :: I,J
!
time_tot = 0.0_dp
Pi_app = 0.0_SP
!
DO J = 0, 10
Pi_app = 0.0_SP
!
CALL CPU_TIME(time_start)
!
DO I = 0, (2**J)*100000
Pi_app = Pi_app + (-1.0_SP)**I/REAL(2*I+1,SP)
END DO
!
CALL CPU_TIME(time_end)
!
time_tot = time_tot + time_end-time_start
!
PRINT*, (2**J)*100000, time_tot, 4.0_SP*Pi_app, 4.0_SP*Pi_app-ACOS(-1.0_SP)
!
ENDDO
!
END PROGRAM VECT_TEST_SINGLEWhile for double precision the program has minor variations.
PROGRAM VECT_TEST_DOUBLE
IMPLICIT NONE
INTEGER, PARAMETER :: DP = KIND(1.0D0)
!
REAL(KIND=DP) :: Pi_app
REAL(KIND=DP) :: time_start, time_end, time_tot
INTEGER :: I,J
!
time_tot = 0.0_dp
Pi_app = 0.0_DP
!
DO J = 0, 10
Pi_app = 0.0_DP
CALL CPU_TIME(time_start)
!
DO I = 0, (2**J)*100000
Pi_app = Pi_app + (-1.0_DP)**I/REAL(2*I+1,DP)
END DO
!
CALL CPU_TIME(time_end)
!
time_tot = time_tot + time_end-time_start
!
PRINT*, (2**J)*100000, time_tot, 4.0_DP*Pi_app, 4.0_DP*Pi_app-ACOS(-1.0_DP)
!
ENDDO
!
END PROGRAM VECT_TEST_DOUBLEIn both cases the output lines is the numbers of terms summed of the series, the time taken to do the sum and the approx. value of Pi and the difference with ACOS(-1). I compiled the programs with the following Makefile for gfortran:
# OPT=-O3 FC = gfortran FLAGS = -march=native -fopt-info-vec-missed -fopt-info-vec-optimized all: test_vectorize_Single test_vectorize_Double .FORCE: test_vectorize_Single: test_vectorize_Single.f90 Makefile .FORCE $(FC) $(OPT) $(FLAGS) -o $@_$(FC) $< test_vectorize_Double: test_vectorize_Double.f90 Makefile .FORCE $(FC) $(OPT) $(FLAGS) -o $@_$(FC) $< clean: rm -f *.o *.s *.exe *.lst *.ppm test_vectorize_Single_$(FC) test_vectorize_Double_$(FC)
and a different Makefile for ifort:
#
OPT=-fast
INFO=-qopt-report=3
FLAGS = -L${MKLROOT}/lib/intel64 -lmkl_intel_lp64 -lmkl_core -lmkl_sequential -lpthread -lm
FC = ifort
all: test_vectorize_Single test_vectorize_Double
.FORCE:
test_vectorize_Single: test_vectorize_Single.f90 Makefile_ifort .FORCE
$(FC) $(OPT) $(INFO) -o $@_$(FC) $< $(FLAGS)
test_vectorize_Double: test_vectorize_Double.f90 Makefile_ifort .FORCE
$(FC) $(OPT) $(INFO) -o $@_$(FC) $< $(FLAGS)
clean:
rm -f *.o *.s *.exe *.lst *.ppm test_vectorize_Single_$(FC) test_vectorize_Double_$(FC)
I find the results surprising, specially for single precision:
$ time ./test_vectorize_Single_gfortran
100000 0.0000000000000000 3.14160585 1.31130219E-05
200000 0.0000000000000000 3.14160132 8.58306885E-06
400000 4.0000000000000001E-003 3.14159846 5.72204590E-06
800000 8.0000000000000002E-003 3.14159727 4.52995300E-06
1600000 2.0000000000000000E-002 3.14159703 4.29153442E-06
3200000 3.2000000000000001E-002 3.14159703 4.29153442E-06
6400000 5.1999999999999991E-002 3.14159703 4.29153442E-06
12800000 8.3999999999999991E-002 3.14159703 4.29153442E-06
25600000 0.14799999999999999 3.14159679 4.05311584E-06
51200000 0.26799999999999990 3.14159679 4.05311584E-06
102400000 0.50400000000000000 3.14159679 4.05311584E-06
real 0m0.506s
user 0m0.504s
sys 0m0.000s
$ time ./test_vectorize_Single_ifort
100000 8.000000000000000E-003 3.141608 1.5497208E-05
200000 2.800000000000000E-002 3.141596 3.5762787E-06
400000 5.200000000000000E-002 3.141694 1.0156631E-04
800000 8.799999999999999E-002 3.141549 -4.4107437E-05
1600000 0.160000000000000 3.142432 8.3971024E-04
3200000 0.304000000000000 3.140750 -8.4304810E-04
6400000 0.604000000000000 3.147671 6.0782433E-03
12800000 1.28000000000000 3.121287 -2.0306110E-02
25600000 2.56000000000000 3.064596 -7.6996565E-02
51200000 5.20000000000000 3.031954 -0.1096389
102400000 10.5840000000000 3.031954 -0.1096389
real 0m10.582s
user 0m10.580s
sys 0m0.004s
The ifort program took way much longer to execute and the results were completely off the mark compared to the previous case. In the double precision case results are quite similar though still ifort compiled executable took much longer than the alternative:
$ time ./test_vectorize_Double_gfortran
100000 0.0000000000000000 3.1416026534897203 9.9998999272266076E-006
200000 4.0000000000000001E-003 3.1415976535647618 4.9999749687223982E-006
400000 8.0000000000000002E-003 3.1415951535834941 2.4999937009440032E-006
800000 1.6000000000000000E-002 3.1415939035881548 1.2499983617075827E-006
1600000 2.7999999999999997E-002 3.1415932785894536 6.2499966047013800E-007
3200000 5.1999999999999991E-002 3.1415929660895618 3.1249976872871343E-007
6400000 8.0000000000000016E-002 3.1415928098397976 1.5625000449048798E-007
12800000 0.14000000000000001 3.1415927317148120 7.8125018898589360E-008
25600000 0.25200000000000000 3.1415926926522508 3.9062457712901733E-008
51200000 0.47999999999999998 3.1415926731216950 1.9531901873648394E-008
102400000 0.93599999999999994 3.1415926633549081 9.7651149388866543E-009
real 0m0.938s
user 0m0.936s
sys 0m0.000s
[pts/0][curro.kimoshi: src]$ time ./test_vectorize_Double_ifort
100000 1.200000000000000E-002 3.14160265348968 9.999899887258579E-006
200000 2.800000000000000E-002 3.14159765356467 4.999974872355040E-006
400000 5.599999999999999E-002 3.14159515358340 2.499993601912109E-006
800000 0.108000000000000 3.14159390358802 1.249998222263571E-006
1600000 0.192000000000000 3.14159327858927 6.249994810580972E-007
3200000 0.364000000000000 3.14159296608931 3.124995213710235E-007
6400000 0.720000000000000 3.14159280983955 1.562497522478168E-007
12800000 1.44800000000000 3.14159273171429 7.812449842603542E-008
25600000 2.93600000000000 3.14159269265213 3.906233603245823E-008
51200000 5.98000000000000 3.14159267312262 1.953282380284804E-008
102400000 12.2040000000000 3.14159266335631 9.766513375808472E-009
real 0m12.203s
user 0m12.208s
sys 0m0.000s
I am probably doing something wrong or using wrong options, any explanation of what's going on, suggestion, or remark will be very welcomed.