Homec4science
Diffusion COSOlver (improved_sugama) Browse

cosolver/basis_transformation/MPI_T4/FM/FMinterval/SampleFMinterval.f95
cd1286e2d410improved_sugama

Last Change
Raw File

SampleFMinterval.f95
View Options

! This is a sample program using the FMZM and FM_INTERVAL_ARITHMETIC modules for doing
! interval arithmetic using the FM_INTERVAL derived type.
! The output is saved in file SampleFMinterval.out. A comparison file, SampleFMinterval.chk,
! is provided showing the expected output from machines using 64-bit double precision and IEEE
! arithmetic. This would give about 16 significant digit accuracy for a stable calculation.
! When run on other computers, all the multiple precision results should be the same, and the
! results from the machine precision (d.p.) calculations will be different.
! The program checks all the results and the last line of the output file should be
! "All results were ok."
! Sample 3 below uses an array-valued function of type FM_INTERVAL.
! The function is defined here in a module with an explicit interface.
MODULE EXP_SUM_MOD
INTERFACE EXP_SUM
MODULE PROCEDURE EXP_SUM3
END INTERFACE
CONTAINS
FUNCTION EXP_SUM3(R_FM)
! Sample function usage for type FM_INTERVAL.
! The test function is exp(x) = 1 + x + x**2/2! + x**3/3! + ...
! summed for the three values of x in array R_FM
! Note that functions returning an FM_INTERVAL variable need a call to FM_ENTER_USER_FUNCTION
! upon entry to the routine and one to FM_EXIT_USER_FUNCTION upon exit, where the argument in
! each case is the function name (EXP_SUM3 here).
! To keep from wasting memory, local variables like S should have the SAVE attribute.
USE FMZM
USE FM_INTERVAL_ARITHMETIC
IMPLICIT NONE
TYPE (FM) :: R_FM(3)
TYPE (FM_INTERVAL) :: EXP_SUM3(3)
TYPE (FM_INTERVAL), SAVE :: S, T, X
INTEGER :: J, K
CALL FM_ENTER_USER_FUNCTION(EXP_SUM3)
DO J = 1, 3
S = 1
T = 1
X = R_FM(J)
DO K = 1, 1000
T = T * X / K
S = S + T
IF (ABS(T) < TO_FM('1.0E-75')) EXIT
ENDDO
EXP_SUM3(J) = S
ENDDO
CALL FM_EXIT_USER_FUNCTION(EXP_SUM3)
END FUNCTION EXP_SUM3
END MODULE EXP_SUM_MOD
PROGRAM SAMPLE_INTERVAL
USE EXP_SUM_MOD
USE FM_INTERVAL_ARITHMETIC
IMPLICIT NONE
! Declare the multiple precision variables.
! (FM) for multiple precision reals
! (FM_INTERVAL) for multiple precision real intervals
TYPE (FM), SAVE :: DIGITS_LOST_FM, ERROR_FM, FACT_FM, R_FM(3), S_FM, S2_FM, T_FM, X_FM, X2_FM
TYPE (FM_INTERVAL), SAVE :: FACT_FM_INTERVAL, S_FM_INTERVAL, T_FM_INTERVAL, &
X_FM_INTERVAL, X2_FM_INTERVAL, EXP_SUM_INTERVAL(3)
! TYPE (FM_INTERVAL), EXTERNAL :: EXP_SUM
! Declare the other variables (not multiple precision).
CHARACTER(80) :: ST1, STF
INTEGER :: E(3), J, K, KOUT, NERROR
DOUBLE PRECISION :: FACT, S, T, X, X2
! Write output to the file SampleFMinterval.out.
KOUT = 18
OPEN (KOUT,FILE='SampleFMinterval.out')
NERROR = 0
! -----------------------------------------------------------------------------------------Sample 1
! One of the common uses for multiple precision and also interval arithmetic is to
! test the accuracy and stability of an algorithm.
! Here is a sum that theoretically converges to the Bessel function
! J(1,x) for x = 35.
! 1. Try it using double precision.
! Printing the partial sums each 5 terms shows that this formula is unstable for
! x = 35, since some of the partial sums are more than 1.0e+14 times larger than
! the final sum. This makes it seem that we have lost at least 14 significant
! digits to cancellation.
! For this example it is fairly clear from the double precision output that the
! final value of S is not accurate, but that might not be easy to see for a more
! complicated calculation.
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Sample 1. Unstable summation.'
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' 1. Do the sum in double precision.'
WRITE (KOUT,*) ' '
S = 0
X = 35.0D0/2
X2 = -(X**2)
FACT = 1
DO K = 0, 70
T = X / ( (K+1) * FACT**2 )
IF ( ABS(T) < EPSILON(S)*ABS(S) ) THEN
WRITE (KOUT,"(A,I3,A,ES25.15)") ' K = ',K,' S = ',S
EXIT
ENDIF
S = S + T
X = X * X2
FACT = FACT * (K+1)
IF (MOD(K,5) == 0) THEN
WRITE (KOUT,"(A,I3,A,ES25.15)") ' K = ',K,' S = ',S
ENDIF
ENDDO
IF (ABS(S-4.399D-2) > 3.0D-3) THEN
NERROR = NERROR + 1
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Error in case 1 (or double precision accuracy is not 53 bits).'
WRITE (KOUT,*) ' '
ENDIF
! 2. Try it using multiple precision, with 20, 30, 40, and 50 significant digits.
! To measure the error each time, compute it first with 100 digits.
! The error is measured in ulps (units in the last place), since the actual
! accuracy is slightly more than the number of digits requested.
! When FM uses the default large base (10^7 is typical for 64-bit double precision)
! this test will show about 18 (base 10) digits lost during the calculation.
! The reason for the "at least" in the descriptions below is that if the base is
! 10^7, then the first word of the multiple precision number can have from 1 to 7
! base 10 digits. So asking for 20 digit precision with CALL FM_SET(20) gives
! 5 digits base 10^7, since we want a few guard digits past 20, and using 4 digits
! base 10^7 would guarantee only 1 + 3*7 = 22 decimal digits. With 5 digits every
! intermediate value in the computation will have from 29 to 35 significant digits.
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' 2. Use FM with increasing precision.'
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Setting precision to J digits via CALL FM_SET(J) will actually set the'
WRITE (KOUT,*) ' equivalent number of decimal significant digits slightly higher than J.'
WRITE (KOUT,*) ' For example, if the base used internally in FM is 10**7, then asking for'
WRITE (KOUT,*) ' 20 digits with CALL FM_SET(20) gives at least 29 significant digits.'
WRITE (KOUT,*) ' CALL FM_SET(30) gives at least 36 significant digits.'
WRITE (KOUT,*) ' CALL FM_SET(40) gives at least 50 significant digits.'
WRITE (KOUT,*) ' CALL FM_SET(50) gives at least 57 significant digits.'
WRITE (KOUT,*) ' '
CALL FM_SET(100)
S2_FM = 0
X_FM = 35.0D0/2
X2_FM = -(X_FM**2)
FACT_FM = 1
DO K = 0, 700
T_FM = X_FM / ( (K+1) * FACT_FM**2 )
IF ( ABS(T_FM) < EPSILON(S2_FM)*ABS(S2_FM) ) EXIT
S2_FM = S2_FM + T_FM
X_FM = X_FM * X2_FM
FACT_FM = FACT_FM * (K+1)
ENDDO
DO J = 20, 50, 10
CALL FM_SET(J)
S_FM = 0
X_FM = 35.0D0/2
X2_FM = -(X_FM**2)
FACT_FM = 1
DO K = 0, 700
T_FM = X_FM / ( (K+1) * FACT_FM**2 )
S_FM = S_FM + T_FM
IF ( ABS(T_FM) < EPSILON(S_FM)*ABS(S_FM) ) THEN
WRITE (STF,*) ' F',J+3,'.',J
CALL FM_FORM(TRIM(STF),S_FM,ST1)
WRITE (KOUT,"(5X,I3,A,I3,A,A)") J,' digits,',K,' terms gave S_FM = ',TRIM(ST1)
CALL FM_ULP(S_FM,T_FM)
! Since S2_FM was computed at a different precision than S_FM, we should
! round it to the current precision. If we knew the two values of the FM
! internal variable NDIG that were used in computing S2_FM and S_FM, the
! standard FM rounding routine FM_EQU could be used. Here we used FM_SET
! to ask for slightly more than 100 and J decimal digits for S2_FM and S_FM,
! so the routine ROUND_FM in this program gets those values of NDIG and does
! the rounding.
CALL ROUND_FM(S2_FM,X2_FM,100,J)
ERROR_FM = ABS( (S_FM-X2_FM)/T_FM )
DIGITS_LOST_FM = NINT(LOG10(ERROR_FM))
WRITE (KOUT,"(A,I3,A)") ' This calculation lost about ', &
TO_INT(DIGITS_LOST_FM),' digits.'
EXIT
ENDIF
X_FM = X_FM * X2_FM
FACT_FM = FACT_FM * (K+1)
ENDDO
ENDDO
IF (ABS(S_FM-TO_FM('.04399094217962563996969897065974247192700503984511')) > 1.0D-35) THEN
NERROR = NERROR + 1
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Error in case 2.'
WRITE (KOUT,*) ' '
ENDIF
! 3. Sometimes we want to measure the errors using base 2 arithmetic in FM,
! to more accurately reflect what is happening in the d.p. calculation.
! Set FM to use base 2, and do the calculation with 53 bits of precision
! (64-bit d.p.), then 73, 93, 113 bits.
! We want exact control over the base and precision, so use FM_SETVAR
! instead of FM_SET.
! This shows a loss of 14 or 15 (base 10) digits when using base 2.
! This is typical of the comparison between using FM with the default large base
! and base 2. Normalization error is larger with a large base, but the 30 s.d.
! calculation in case 2 gets about the same accuracy as the 113-bit calculation
! in case 3, and using base 2 is much slower.
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' 3. Use FM with increasing precision in base 2.'
WRITE (KOUT,*) ' '
CALL FM_SETVAR(" MBASE = 2 ")
CALL FM_SETVAR(" NDIG = 150 ")
S2_FM = 0
X_FM = 35.0D0/2
X2_FM = -(X_FM**2)
FACT_FM = 1
DO K = 0, 700
T_FM = X_FM / ( (K+1) * FACT_FM**2 )
IF ( ABS(T_FM) < EPSILON(S2_FM)*ABS(S2_FM) ) EXIT
S2_FM = S2_FM + T_FM
X_FM = X_FM * X2_FM
FACT_FM = FACT_FM * (K+1)
ENDDO
DO J = 53, 113, 20
WRITE (STF,*) ' NDIG =',J
CALL FM_SETVAR(TRIM(STF))
S_FM = 0
X_FM = 35.0D0/2
X2_FM = -(X_FM**2)
FACT_FM = 1
DO K = 0, 700
T_FM = X_FM / ( (K+1) * FACT_FM**2 )
S_FM = S_FM + T_FM
IF ( ABS(T_FM) < EPSILON(S_FM)*ABS(S_FM) ) THEN
WRITE (STF,*) ' F',NINT(J*0.301)+3,'.',NINT(J*0.301)+1
CALL FM_FORM(TRIM(STF),S_FM,ST1)
WRITE (KOUT,"(A,I3,A,I3,A,A)") ' Using ',J,' bits,',K, &
' terms gave S_FM = ',TRIM(ST1)
CALL FM_ULP(S_FM,T_FM)
! Since S2_FM was computed at a different precision than S_FM, we should
! round it to the current precision. For this case we have explicitly set
! NDIG instead of using FM_SET as in case 2 above, so we use FM_EQU to do
! the rounding.
CALL FM_EQU(S2_FM,X2_FM,150,J)
ERROR_FM = ABS( (S_FM-X2_FM)/T_FM )
DIGITS_LOST_FM = NINT(LOG10(ERROR_FM))
WRITE (KOUT,"(A,I3,A)") ' This calculation lost about ', &
TO_INT(DIGITS_LOST_FM),' decimal digits.'
EXIT
ENDIF
X_FM = X_FM * X2_FM
FACT_FM = FACT_FM * (K+1)
ENDDO
ENDDO
IF (ABS(S_FM-TO_FM('.0439909421796256399686302351876196')) > 1.0D-20) THEN
NERROR = NERROR + 1
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Error in case 3.'
WRITE (KOUT,*) ' '
ENDIF
! 4. A second way to check an algorithm's stability is to re-do the calculation
! at the same precision but with different rounding modes.
! Use base 2 with 53 bits and round down, then round symmetrically, then round up.
! The results show the three values have no digits of agreement, confirming the
! loss of about 15 or 16 s.d. in the ones rounded down and up.
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' 4. Use FM with different rounding modes in base 2.'
WRITE (KOUT,*) ' '
CALL FM_SETVAR(" MBASE = 2 ")
CALL FM_SETVAR(" NDIG = 53 ")
DO J = 1, 3
IF (J == 1) THEN
CALL FM_SETVAR(" KROUND = -1 ")
ELSE IF (J == 2) THEN
CALL FM_SETVAR(" KROUND = 1 ")
ELSE IF (J == 3) THEN
CALL FM_SETVAR(" KROUND = 2 ")
ENDIF
S_FM = 0
X_FM = 35.0D0/2
X2_FM = -(X_FM**2)
FACT_FM = 1
DO K = 0, 700
T_FM = X_FM / ( (K+1) * FACT_FM**2 )
S_FM = S_FM + T_FM
IF ( ABS(T_FM) < EPSILON(S_FM)*ABS(S_FM) ) THEN
IF (J == 1) THEN
STF = 'rounding left ,'
ELSE IF (J == 2) THEN
STF = 'rounding symmetrically,'
ELSE IF (J == 3) THEN
STF = 'rounding right ,'
ENDIF
CALL FM_FORM(' F20.17 ',S_FM,ST1)
WRITE (KOUT,"(A,I4,A,A,I3,A,A)") ' Using',53,' bits, ',TRIM(STF),K, &
' terms gave S_FM = ',TRIM(ST1)
EXIT
ENDIF
X_FM = X_FM * X2_FM
FACT_FM = FACT_FM * (K+1)
ENDDO
R_FM(J) = S_FM
ENDDO
ERROR_FM = MAX( ABS((R_FM(1)-R_FM(2))/R_FM(1)) , ABS((R_FM(1)-R_FM(3))/R_FM(1)) , &
ABS((R_FM(2)-R_FM(3))/R_FM(2)) )
J = -NINT(LOG10(ERROR_FM))
WRITE (KOUT,"(A,I3,A)") ' These agree to about',J,' decimal digits.'
IF (ABS(S_FM-TO_FM('.0519420065663800')) > 1.0D-4) THEN
NERROR = NERROR + 1
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Error in case 4.'
WRITE (KOUT,*) ' '
ENDIF
! Use base 2 with 113 bits and round down, then round symmetrically, then round up.
! Now the three values agree to about 18 decimal digits, which is again consistent
! with a loss of about 15 or 16 s.d. in the ones rounded down and up.
WRITE (KOUT,*) ' '
CALL FM_SETVAR(" MBASE = 2 ")
CALL FM_SETVAR(" NDIG = 113 ")
DO J = 1, 3
IF (J == 1) THEN
CALL FM_SETVAR(" KROUND = -1 ")
ELSE IF (J == 2) THEN
CALL FM_SETVAR(" KROUND = 1 ")
ELSE IF (J == 3) THEN
CALL FM_SETVAR(" KROUND = 2 ")
ENDIF
S_FM = 0
X_FM = 35.0D0/2
X2_FM = -(X_FM**2)
FACT_FM = 1
DO K = 0, 700
T_FM = X_FM / ( (K+1) * FACT_FM**2 )
S_FM = S_FM + T_FM
IF ( ABS(T_FM) < EPSILON(S_FM)*ABS(S_FM) ) THEN
IF (J == 1) THEN
STF = 'rounding left ,'
ELSE IF (J == 2) THEN
STF = 'rounding symmetrically,'
ELSE IF (J == 3) THEN
STF = 'rounding right ,'
ENDIF
CALL FM_FORM(' F20.17 ',S_FM,ST1)
WRITE (KOUT,"(A,I4,A,A,I3,A,A)") ' Using',113,' bits, ',TRIM(STF),K, &
' terms gave S_FM = ',TRIM(ST1)
EXIT
ENDIF
X_FM = X_FM * X2_FM
FACT_FM = FACT_FM * (K+1)
ENDDO
R_FM(J) = S_FM
ENDDO
ERROR_FM = MAX( ABS((R_FM(1)-R_FM(2))/R_FM(1)) , ABS((R_FM(1)-R_FM(3))/R_FM(1)) , &
ABS((R_FM(2)-R_FM(3))/R_FM(2)) )
J = -NINT(LOG10(ERROR_FM))
WRITE (KOUT,"(A,I3,A)") ' These agree to about',J,' decimal digits.'
IF (ABS(S_FM-TO_FM('.0439909421796257')) > 1.0D-4) THEN
NERROR = NERROR + 1
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Error in case 4.'
WRITE (KOUT,*) ' '
ENDIF
! 5. A third way to check an algorithm's stability is to re-do the calculation
! at the same precision but using interval arithmetic.
! Use base 2 with 53 bits.
! The result shows the two endpoints of the interval having opposite signs
! indicating a worst-case loss of all 16 s.d. during the calculation.
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' 5. Use FM with interval arithmetic in base 2.'
WRITE (KOUT,*) ' '
CALL FM_SETVAR(" MBASE = 2 ")
CALL FM_SETVAR(" NDIG = 53 ")
S_FM_INTERVAL = 0
X_FM_INTERVAL = 35.0D0/2
X2_FM_INTERVAL = -(X_FM_INTERVAL**2)
FACT_FM_INTERVAL = 1
DO K = 0, 700
T_FM_INTERVAL = X_FM_INTERVAL / ( (K+1) * FACT_FM_INTERVAL**2 )
S_FM_INTERVAL = S_FM_INTERVAL + T_FM_INTERVAL
IF ( ABS(T_FM_INTERVAL) < EPSILON(S_FM_INTERVAL)*ABS(S_FM_INTERVAL) ) THEN
CALL FM_INTERVAL_FORM(' F20.16 ',S_FM_INTERVAL,ST1)
WRITE (KOUT,"(A,I4,A,I3,A,A)") ' Using',53,' bits, ',K, &
' terms gave S_FM_INTERVAL = ',TRIM(ST1)
EXIT
ENDIF
X_FM_INTERVAL = X_FM_INTERVAL * X2_FM_INTERVAL
FACT_FM_INTERVAL = FACT_FM_INTERVAL * (K+1)
ENDDO
ERROR_FM = ABS((RIGHT_ENDPOINT(S_FM_INTERVAL)-LEFT_ENDPOINT(S_FM_INTERVAL)) / &
RIGHT_ENDPOINT(S_FM_INTERVAL))
J = -NINT(LOG10(ERROR_FM))
WRITE (KOUT,"(A,I3,A)") ' The two endpoints agree to about',J,' decimal digits.'
! Use base 2 with 113 bits.
! The result shows the two endpoints of the interval agree to about 18 s.d.
! indicating a worst-case loss of about 16 s.d. during the calculation.
WRITE (KOUT,*) ' '
CALL FM_SETVAR(" MBASE = 2 ")
CALL FM_SETVAR(" NDIG = 113 ")
S_FM_INTERVAL = 0
X_FM_INTERVAL = 35.0D0/2
X2_FM_INTERVAL = -(X_FM_INTERVAL**2)
FACT_FM_INTERVAL = 1
DO K = 0, 700
T_FM_INTERVAL = X_FM_INTERVAL / ( (K+1) * FACT_FM_INTERVAL**2 )
S_FM_INTERVAL = S_FM_INTERVAL + T_FM_INTERVAL
IF ( ABS(T_FM_INTERVAL) < EPSILON(S_FM_INTERVAL)*ABS(S_FM_INTERVAL) ) THEN
CALL FM_INTERVAL_FORM(' F20.16 ',S_FM_INTERVAL,ST1)
WRITE (KOUT,"(A,I4,A,I3,A,A)") ' Using',113,' bits, ',K, &
' terms gave S_FM_INTERVAL = ',TRIM(ST1)
EXIT
ENDIF
X_FM_INTERVAL = X_FM_INTERVAL * X2_FM_INTERVAL
FACT_FM_INTERVAL = FACT_FM_INTERVAL * (K+1)
ENDDO
ERROR_FM = ABS((RIGHT_ENDPOINT(S_FM_INTERVAL)-LEFT_ENDPOINT(S_FM_INTERVAL)) / &
RIGHT_ENDPOINT(S_FM_INTERVAL))
J = -NINT(LOG10(ERROR_FM))
WRITE (KOUT,"(A,I3,A)") ' The two endpoints agree to about',J,' decimal digits.'
IF (ABS(S_FM_INTERVAL-TO_FM('.0439909421796257')) > 1.0D-4) THEN
NERROR = NERROR + 1
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Error in case 5.'
WRITE (KOUT,*) ' '
ENDIF
IF (NERROR == 0) THEN
WRITE (KOUT,"(//A)") ' Summary:'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' For this calculation all four methods for measuring the degree of'
WRITE (KOUT,"(A)") ' instability worked well. When done with double precision carrying'
WRITE (KOUT,"(A)") ' 16 significant digits and using the default symmetric rounding,'
WRITE (KOUT,"(A)") ' only 1 digit remained correct at the end of the sum.'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' Interval arithmetic is probably the strongest of these checks, and'
WRITE (KOUT,"(A)") ' using FM arithmetic with 30 digits and a large base (method 2) is'
WRITE (KOUT,"(A)") ' the fastest of these methods for getting the sum correct to full'
WRITE (KOUT,"(A)") ' double precision accuracy.'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' Comparing the last value of S in method 1 with the first value of'
WRITE (KOUT,"(A)") ' S_FM in method 3 should show whether this compiler carries extra'
WRITE (KOUT,"(A)") ' digits while evaluating expressions like X / ( (K+1) * FACT**2 ).'
WRITE (KOUT,"(A)") ' If the two values are the same, no extra digits are carried in d.p.'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' '
ENDIF
! -----------------------------------------------------------------------------------------Sample 2
! Here is a recurrence that is seriously unstable.
! It is not a realistic problem that would come up in a practical application, but
! is designed as a counter-example to show that just carrying much higher precision
! cannot be proved to always cure numerical instability.
! Reference: William Kahan -- (2006)
! "How Futile are Mindless Assessments of Roundoff in Floating-Point Computation?"
! http://www.cs.berkeley.edu/~wkahan/Mindless.pdf
! After 60 steps the result without any rounding errors would be very close to 5,
! but rounding errors cause the result to converge to 100 instead.
CALL FM_SET(30)
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Sample 2. Unstable recurrence.'
WRITE (KOUT,*) ' '
WRITE (KOUT,"(A)") ' 1. Use non-interval FM arithmetic with 30 digits.'
WRITE (KOUT,*) ' '
X_FM = 4
X2_FM = TO_FM('4.25')
DO J = 1, 60
T_FM = 108 - ( 815 - 1500/X_FM ) / X2_FM
X_FM = X2_FM
X2_FM = T_FM
IF (MOD(J,5) == 0) THEN
CALL FM_FORM(' F20.14 ',X2_FM,ST1)
WRITE (KOUT,"(A,I2,A,A)") ' After ',J, &
' terms with 30 digit accuracy, the result is',TRIM(ST1)
ENDIF
ENDDO
WRITE (KOUT,*) ' '
WRITE (KOUT,"(A)") ' 2. Use interval arithmetic with 30 digit accuracy.'
WRITE (KOUT,*) ' '
X_FM_INTERVAL = 4
X2_FM_INTERVAL = TO_FM_INTERVAL('4.25')
DO J = 1, 60
T_FM_INTERVAL = 108 - ( 815 - 1500/X_FM_INTERVAL ) / X2_FM_INTERVAL
X_FM_INTERVAL = X2_FM_INTERVAL
X2_FM_INTERVAL = T_FM_INTERVAL
CALL FM_INTERVAL_FORM(' F20.16 ',X2_FM_INTERVAL,ST1)
WRITE (KOUT,"(A,I3,A,A)") ' After',J,' terms the result is',TRIM(ST1)
IF (LEFT_ENDPOINT(X_FM_INTERVAL) <= 0 .AND. RIGHT_ENDPOINT(X_FM_INTERVAL) >= 0) EXIT
IF (J > 10 .AND. J < 25) THEN
IF (ABS(LEFT_ENDPOINT(X2_FM_INTERVAL)-5) > 0.01 .OR. &
ABS(RIGHT_ENDPOINT(X2_FM_INTERVAL)-5) > 0.01) THEN
NERROR = NERROR + 1
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Error in sample 2, case 2.'
WRITE (KOUT,*) ' '
EXIT
ENDIF
ENDIF
ENDDO
IF (NERROR == 0) THEN
WRITE (KOUT,"(//A)") ' Summary:'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' The general solution of this recurrence has a term that causes'
WRITE (KOUT,"(A)") ' the values to converge to 100, but for these initial conditions'
WRITE (KOUT,"(A)") ' 4, 4.25, the specific solution has a coefficient of zero on that'
WRITE (KOUT,"(A)") ' term, so this sequence converges to 5 mathematically.'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' When rounding errors occur in later x-values, they introduce a'
WRITE (KOUT,"(A)") ' very small but nonzero amount of the term that causes convergence'
WRITE (KOUT,"(A)") ' to 100, and that term grows rapidly and soon swamps the rest of'
WRITE (KOUT,"(A)") ' the solution.'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' Interval arithmetic tracks the growing uncertainty in the x-values,'
WRITE (KOUT,"(A)") ' and when an interval gets big enough to include zero, dividing by'
WRITE (KOUT,"(A)") ' that interval is undefined and the result is'
WRITE (KOUT,"(A)") ' [ -overflow , +overflow ].'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' Interval arithmetic works better than a sequence of increasing'
WRITE (KOUT,"(A)") ' precision FM results here, since comparing FM results at 30, 40, 50'
WRITE (KOUT,"(A)") ' digits gives 100 each time.'
WRITE (KOUT,"(A)") ' '
ENDIF
! -----------------------------------------------------------------------------------------Sample 3
! Sum an unstable series.
! exp(x) = 1 + x + x**2/2! + x**3/3! + ...
! converges mathematically for all x, but is unstable for negative x.
S_FM_INTERVAL = 1
T_FM_INTERVAL = 1
R_FM = (/ -25, -30, -35 /)
EXP_SUM_INTERVAL = EXP_SUM3(R_FM)
E = (/ -12, -9, -3 /)
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' '
WRITE (KOUT,*) ' Sample 3. Unstable sum.'
DO J = 1, 3
CALL FM_INTERVAL_FORM(' ES25.16 ',EXP_SUM_INTERVAL(J),ST1)
CALL FM_FORM(' ES25.16 ',EXP(R_FM(J)),STF)
WRITE (KOUT,"(/A,F6.2,A,A)") ' For x = ', TO_DP(R_FM(J)), ' The sum gave ', ST1
WRITE (KOUT,"(20X,A,A)") ' correct = ', STF
! Check the results.
X_FM = LEFT_ENDPOINT(EXP_SUM_INTERVAL(J))
X2_FM = RIGHT_ENDPOINT(EXP_SUM_INTERVAL(J))
S_FM = EXP(R_FM(J))
T_FM = ABS( (X2_FM - X_FM ) / S_FM )
IF (.NOT.(S_FM > X_FM) .OR. .NOT.(S_FM < X2_FM) .OR. .NOT.(T_FM < TO_FM(10)**E(J))) THEN
NERROR = NERROR + 1
EXIT
ENDIF
ENDDO
IF (NERROR == 0) THEN
WRITE (KOUT,"(//A)") ' Summary:'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' As the input x becomes more negative, the instability increases.'
WRITE (KOUT,"(A)") ' This is shown as the left and right endpoints of the interval'
WRITE (KOUT,"(A)") ' result agree to fewer digits.'
WRITE (KOUT,"(A)") ' '
WRITE (KOUT,"(A)") ' The mathematically correct value of the sum lies within the'
WRITE (KOUT,"(A)") ' interval in each case.'
WRITE (KOUT,"(A)") ' '
ENDIF
IF (NERROR == 0) THEN
WRITE (* ,"(//A,A/)") ' All results were ok. (The output is in file ', &
'SampleFMinterval.out)'
WRITE (KOUT,"(//A/)") ' All results were ok.'
ELSE
WRITE (* ,"(//I3,A,A/)") NERROR,' error(s) found. (The output is in file ', &
'SampleFMinterval.out)'
WRITE (KOUT,"(//I3,A/)") NERROR,' error(s) found.'
ENDIF
STOP
END PROGRAM SAMPLE_INTERVAL
SUBROUTINE ROUND_FM(A,B,J1,J2)
USE FMVALS
USE FMZM
IMPLICIT NONE
! A was computed with precision defined by CALL FM_SET(J1).
! B will be returned with the value of A rounded to precision defined by CALL FM_SET(J2).
! Do not use B in the calling program at any precision higher than J2.
TYPE (FM) :: A, B
INTEGER :: J1, J2, NDIG1, NDIG2, NSAVE
NSAVE = NDIG
CALL FM_SET(J1)
NDIG1 = NDIG
CALL FM_SET(J2)
NDIG2 = NDIG
CALL FM_EQU(A,B,NDIG1,NDIG2)
NDIG = NSAVE
END SUBROUTINE ROUND_FM
c4science · Help