![[BACK]](/cvsweb/icons/back.gif){kind=icon}
Return to to_double CVS log ![[TXT]](/cvsweb/icons/text.gif){kind=icon}
![[DIR]](/cvsweb/icons/dir.gif){kind=icon}
Up to [CSRG BSD Unix] / 43BSDTahoe / new / help / src / f77|
Return to to_double CVS log | Up to [CSRG BSD Unix] / 43BSDTahoe / new / help / src / f77 |
1.1 root 1: .TI F77/TO_DOUBLE "Sep. 4, 1985"
2: Converting Programs to Double Precision
3:
4: On VAXs, single precision provides about 6-7 digits of accuracy while
5: double precision provides about 15-16 digits of accuracy.
6: Many mathematical subroutine libraries assume
7: double precision arguments.
8: If you have a program written in single precision which needs to be
9: run in double precision, you can either use f77's -r8 flag
10: to compile it as if it were written in double precision,
11: or follow the following steps to convert the source.
12: If you use the -r8 flag, you still must check equivalences and
13: common blocks as described in paragraph 9 below.
14:
15: To convert a program from single to double precision:
16:
17: .IP 1.
18: Add 'implicit double precision (a-h,o-z)'
19: as the first line of the main program (second line if you have
20: a 'program' statement) and as the second line of all other
21: subprograms (subroutines, functions, and block data subprograms).
22:
23: .IP 2.
24: Change all 'real' or 'real*4' declarations to 'double precision'.
25:
26: .IP 3.
27: Change single precision floating point constants to double precision,
28: either by adding 'd0' or changing an 'e' exponent term to 'd', e.g.:
29: .nf
30: 1.2 becomes 1.2d0
31: 3.2e-5 becomes 3.2d-5
32: .fi
33:
34: .IP 4.
35: Change 'complex' or 'complex*8' declarations to 'double complex'.
36:
37: .IP 5.
38: Formats do not need to be changed.
39:
40: .IP 6.
41: Change single precision intrinsic functions to double. Because of the
42: generic function names in f77, the only ones you need to actually change are:
43:
44: .nf
45: single version: double or generic version:
46: ifix() int()
47: sngl() dble()
48: real() dble()
49: float() dble()
50: amax1() max()
51: amin1() min()
52: amod() mod()
53: alog() log()
54: alog10() log10()
55: .fi
56:
57: .IP 7.
58: Change complex intrinsic functions to double complex.
59:
60: .nf
61: complex version: double complex or generic version:
62: cmplx() dcmplx()
63: aimag() imag()
64: csqrt() sqrt()
65: cexp() exp()
66: clog() log()
67: ccos() cos()
68: csin() sin()
69: cabs() abs()
70: .fi
71:
72: .IP 8.
73: If you are passing functions as arguments to other subroutines,
74: you need to change all function names to specific functions.
75: For example:
76:
77: .nf
78: x1 = sqrt(x2)
79: c1 = abs(c2)
80: .fi
81:
82: works fine when x1 and x2 are double and c1 and c2 are double complex.
83: However, if the computation is being done in a subroutine,
84:
85: .nf
86: external sqrt, cabs
87: call subr( x1, x2, sqrt, c1, c2, cabs )
88: .fi
89:
90: it must be changed to:
91:
92: .nf
93: external dsqrt, cdabs
94: call subr( x1, x2, dsqrt, c1, c2, cdabs )
95: .fi
96:
97: .IP 9.
98: Check equivalences common blocks, and dynamic memory allocation.
99: Often these implicitly assume that floating point and integer values
100: use the same amount of storage.
101: For example, if one subroutine declares:
102:
103: common /abc/ key1,vec(5),ival
104:
105: and another declares:
106:
107: common /abc/ dummy(6), ival
108:
109: then the two instances of 'ival' match. If you change 'dummy'
110: and 'vec' from single precision to double precision (either
111: explicitly or via the -r8 flag), the two
112: instances of 'ival' will no longer correspond.
This archive runs on limited infrastructure. Preserving old code on modern bandwidth. Automated agents are requested to crawl responsibly.