Printf width specifier to maintain precision of floating-point value
CFloating PointPrintfC99Floating Point-PrecisionC Problem Overview
Is there a printf width specifier which can be applied to a floating point specifier that would automatically format the output to the necessary number of significant digits such that when scanning the string back in, the original floating point value is acquired?
For example, suppose I print a float to a precision of 2 decimal places:
float foobar = 0.9375;
printf("%.2f", foobar); // prints out 0.94
When I scan the output 0.94, I have no standards-compliant guarantee that I'll get the original 0.9375 floating-point value back (in this example, I probably won't).
I would like a way tell printf to automatically print the floating-point value to the necessary number of significant digits to ensure that it can be scanned back to the original value passed to printf.
I could use some of the macros in float.h to derive the maximum width to pass to printf, but is there already a specifier to automatically print to the necessary number of significant digits -- or at least to the maximum width?
C Solutions
Solution 1 - C
I recommend @Jens Gustedt hexadecimal solution: use %a.
OP wants “print with maximum precision (or at least to the most significant decimal)”.
A simple example would be to print one seventh as in:
#include <float.h>
int Digs = DECIMAL_DIG;
double OneSeventh = 1.0/7.0;
printf("%.*e\n", Digs, OneSeventh);
// 1.428571428571428492127e-01
But let's dig deeper ...
Mathematically, the answer is "0.142857 142857 142857 ...", but we are using finite precision floating point numbers.
Let's assume IEEE 754 double-precision binary.
So the OneSeventh = 1.0/7.0 results in the value below. Also shown are the preceding and following representable double floating point numbers.
OneSeventh before = 0.1428571428571428 214571170656199683435261249542236328125
OneSeventh = 0.1428571428571428 49212692681248881854116916656494140625
OneSeventh after = 0.1428571428571428 769682682968777953647077083587646484375
Printing the exact decimal representation of a double has limited uses.
C has 2 families of macros in <float.h> to help us.
The first set is the number of significant digits to print in a string in decimal so when scanning the string back,
we get the original floating point. There are shown with the C spec's minimum value and a sample C11 compiler.
FLT_DECIMAL_DIG 6, 9 (float) (C11)
DBL_DECIMAL_DIG 10, 17 (double) (C11)
LDBL_DECIMAL_DIG 10, 21 (long double) (C11)
DECIMAL_DIG 10, 21 (widest supported floating type) (C99)
The second set is the number of significant digits a string may be scanned into a floating point and then the FP printed, still retaining the same string presentation. There are shown with the C spec's minimum value and a sample C11 compiler. I believe available pre-C99.
FLT_DIG 6, 6 (float)
DBL_DIG 10, 15 (double)
LDBL_DIG 10, 18 (long double)
The first set of macros seems to meet OP's goal of significant digits. But that macro is not always available.
#ifdef DBL_DECIMAL_DIG
#define OP_DBL_Digs (DBL_DECIMAL_DIG)
#else
#ifdef DECIMAL_DIG
#define OP_DBL_Digs (DECIMAL_DIG)
#else
#define OP_DBL_Digs (DBL_DIG + 3)
#endif
#endif
The "+ 3" was the crux of my previous answer. Its centered on if knowing the round-trip conversion string-FP-string (set #2 macros available C89), how would one determine the digits for FP-string-FP (set #1 macros available post C89)? In general, add 3 was the result.
Now how many significant digits to print is known and driven via <float.h>.
To print N significant decimal digits one may use various formats.
With "%e", the precision field is the number of digits after the lead digit and decimal point.
So - 1 is in order. Note: This -1 is not in the initial int Digs = DECIMAL_DIG;
printf("%.*e\n", OP_DBL_Digs - 1, OneSeventh);
// 1.4285714285714285e-01
With "%f", the precision field is the number of digits after the decimal point.
For a number like OneSeventh/1000000.0, one would need OP_DBL_Digs + 6 to see all the significant digits.
printf("%.*f\n", OP_DBL_Digs , OneSeventh);
// 0.14285714285714285
printf("%.*f\n", OP_DBL_Digs + 6, OneSeventh/1000000.0);
// 0.00000014285714285714285
Note: Many are use to "%f". That displays 6 digits after the decimal point; 6 is the display default, not the precision of the number.
Solution 2 - C
The short answer to print floating point numbers losslessly (such that they can be read back in to exactly the same number, except NaN and Infinity):
- If your type is float: use
printf("%.9g", number). - If your type is double: use
printf("%.17g", number).
Do NOT use %f, since that only specifies how many significant digits after the decimal and will truncate small numbers. For reference, the magic numbers 9 and 17 can be found in float.h which defines FLT_DECIMAL_DIG and DBL_DECIMAL_DIG.
Solution 3 - C
If you are only interested in the bit (resp hex pattern) you could use the %a format. This guarantees you:
> The > default precision suffices for an exact representation of the value if an exact representation in base 2 exists and otherwise is sufficiently large to distinguish values of type double.
I'd have to add that this is only available since C99.
Solution 4 - C
No, there is no such printf width specifier to print floating-point with maximum precision. Let me explain why.
The maximum precision of float and double is variable, and dependent on the actual value of the float or double.
Recall float and double are stored in sign.exponent.mantissa format. This means that there are many more bits used for the fractional component for small numbers than for big numbers.
For example, float can easily distinguish between 0.0 and 0.1.
float r = 0;
printf( "%.6f\n", r ) ; // 0.000000
r+=0.1 ;
printf( "%.6f\n", r ) ; // 0.100000
But float has no idea of the difference between 1e27 and 1e27 + 0.1.
r = 1e27;
printf( "%.6f\n", r ) ; // 999999988484154753734934528.000000
r+=0.1 ;
printf( "%.6f\n", r ) ; // still 999999988484154753734934528.000000
This is because all the precision (which is limited by the number of mantissa bits) is used up for the large part of the number, left of the decimal.
The %.f modifier just says how many decimal values you want to print from the float number as far as formatting goes. The fact that the accuracy available depends on the size of the number is up to you as the programmer to handle. printf can't/doesn't handle that for you.
Solution 5 - C
Simply use the macros from <float.h> and the variable-width conversion specifier (".*"):
float f = 3.14159265358979323846;
printf("%.*f\n", FLT_DIG, f);
Solution 6 - C
I run a small experiment to verify that printing with DBL_DECIMAL_DIG does indeed exactly preserve the number's binary representation. It turned out that for the compilers and C libraries I tried, DBL_DECIMAL_DIG is indeed the number of digits required, and printing with even one digit less creates a significant problem.
#include <float.h>
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
union {
short s[4];
double d;
} u;
void
test(int digits)
{
int i, j;
char buff[40];
double d2;
int n, num_equal, bin_equal;
srand(17);
n = num_equal = bin_equal = 0;
for (i = 0; i < 1000000; i++) {
for (j = 0; j < 4; j++)
u.s[j] = (rand() << 8) ^ rand();
if (isnan(u.d))
continue;
n++;
sprintf(buff, "%.*g", digits, u.d);
sscanf(buff, "%lg", &d2);
if (u.d == d2)
num_equal++;
if (memcmp(&u.d, &d2, sizeof(double)) == 0)
bin_equal++;
}
printf("Tested %d values with %d digits: %d found numericaly equal, %d found binary equal\n", n, digits, num_equal, bin_equal);
}
int
main()
{
test(DBL_DECIMAL_DIG);
test(DBL_DECIMAL_DIG - 1);
return 0;
}
I run this with Microsoft's C compiler 19.00.24215.1 and gcc version 7.4.0 20170516 (Debian 6.3.0-18+deb9u1). Using one less decimal digit halves the number of numbers that compare exactly equal. (I also verified that rand() as used indeed produces about one million different numbers.) Here are the detailed results.
Microsoft C
Tested 999507 values with 17 digits: 999507 found numericaly equal, 999507 found binary equal Tested 999507 values with 16 digits: 545389 found numericaly equal, 545389 found binary equal
GCC
Tested 999485 values with 17 digits: 999485 found numericaly equal, 999485 found binary equal Tested 999485 values with 16 digits: 545402 found numericaly equal, 545402 found binary equal
Solution 7 - C
In one of my comments to an answer I lamented that I've long wanted some way to print all the significant digits in a floating point value in decimal form, in much the same way the as the question asks. Well I finally sat down and wrote it. It's not quite perfect, and this is demo code that prints additional information, but it mostly works for my tests. Please let me know if you (i.e. anyone) would like a copy of the whole wrapper program which drives it for testing.
static unsigned int
ilog10(uintmax_t v);
/*
* Note: As presented this demo code prints a whole line including information
* about how the form was arrived with, as well as in certain cases a couple of
* interesting details about the number, such as the number of decimal places,
* and possibley the magnitude of the value and the number of significant
* digits.
*/
void
print_decimal(double d)
{
size_t sigdig;
int dplaces;
double flintmax;
/*
* If we really want to see a plain decimal presentation with all of
* the possible significant digits of precision for a floating point
* number, then we must calculate the correct number of decimal places
* to show with "%.*f" as follows.
*
* This is in lieu of always using either full on scientific notation
* with "%e" (where the presentation is always in decimal format so we
* can directly print the maximum number of significant digits
* supported by the representation, taking into acount the one digit
* represented by by the leading digit)
*
* printf("%1.*e", DBL_DECIMAL_DIG - 1, d)
*
* or using the built-in human-friendly formatting with "%g" (where a
* '*' parameter is used as the number of significant digits to print
* and so we can just print exactly the maximum number supported by the
* representation)
*
* printf("%.*g", DBL_DECIMAL_DIG, d)
*
*
* N.B.: If we want the printed result to again survive a round-trip
* conversion to binary and back, and to be rounded to a human-friendly
* number, then we can only print DBL_DIG significant digits (instead
* of the larger DBL_DECIMAL_DIG digits).
*
* Note: "flintmax" here refers to the largest consecutive integer
* that can be safely stored in a floating point variable without
* losing precision.
*/
#ifdef PRINT_ROUND_TRIP_SAFE
# ifdef DBL_DIG
sigdig = DBL_DIG;
# else
sigdig = ilog10(uipow(FLT_RADIX, DBL_MANT_DIG - 1));
# endif
#else
# ifdef DBL_DECIMAL_DIG
sigdig = DBL_DECIMAL_DIG;
# else
sigdig = (size_t) lrint(ceil(DBL_MANT_DIG * log10((double) FLT_RADIX))) + 1;
# endif
#endif
flintmax = pow((double) FLT_RADIX, (double) DBL_MANT_DIG); /* xxx use uipow() */
if (d == 0.0) {
printf("z = %.*s\n", (int) sigdig + 1, "0.000000000000000000000"); /* 21 */
} else if (fabs(d) >= 0.1 &&
fabs(d) <= flintmax) {
dplaces = (int) (sigdig - (size_t) lrint(ceil(log10(ceil(fabs(d))))));
if (dplaces < 0) {
/* XXX this is likely never less than -1 */
/*
* XXX the last digit is not significant!!! XXX
*
* This should also be printed with sprintf() and edited...
*/
printf("R = %.0f [%d too many significant digits!!!, zero decimal places]\n", d, abs(dplaces));
} else if (dplaces == 0) {
/*
* The decimal fraction here is not significant and
* should always be zero (XXX I've never seen this)
*/
printf("R = %.0f [zero decimal places]\n", d);
} else {
if (fabs(d) == 1.0) {
/*
* This is a special case where the calculation
* is off by one because log10(1.0) is 0, but
* we still have the leading '1' whole digit to
* count as a significant digit.
*/
#if 0
printf("ceil(1.0) = %f, log10(ceil(1.0)) = %f, ceil(log10(ceil(1.0))) = %f\n",
ceil(fabs(d)), log10(ceil(fabs(d))), ceil(log10(ceil(fabs(d)))));
#endif
dplaces--;
}
/* this is really the "useful" range of %f */
printf("r = %.*f [%d decimal places]\n", dplaces, d, dplaces);
}
} else {
if (fabs(d) < 1.0) {
int lz;
lz = abs((int) lrint(floor(log10(fabs(d)))));
/* i.e. add # of leading zeros to the precision */
dplaces = (int) sigdig - 1 + lz;
printf("f = %.*f [%d decimal places]\n", dplaces, d, dplaces);
} else { /* d > flintmax */
size_t n;
size_t i;
char *df;
/*
* hmmmm... the easy way to suppress the "invalid",
* i.e. non-significant digits is to do a string
* replacement of all dgits after the first
* DBL_DECIMAL_DIG to convert them to zeros, and to
* round the least significant digit.
*/
df = malloc((size_t) 1);
n = (size_t) snprintf(df, (size_t) 1, "%.1f", d);
n++; /* for the NUL */
df = realloc(df, n);
(void) snprintf(df, n, "%.1f", d);
if ((n - 2) > sigdig) {
/*
* XXX rounding the integer part here is "hard"
* -- we would have to convert the digits up to
* this point back into a binary format and
* round that value appropriately in order to
* do it correctly.
*/
if (df[sigdig] >= '5' && df[sigdig] <= '9') {
if (df[sigdig - 1] == '9') {
/*
* xxx fixing this is left as
* an exercise to the reader!
*/
printf("F = *** failed to round integer part at the least significant digit!!! ***\n");
free(df);
return;
} else {
df[sigdig - 1]++;
}
}
for (i = sigdig; df[i] != '.'; i++) {
df[i] = '0';
}
} else {
i = n - 1; /* less the NUL */
if (isnan(d) || isinf(d)) {
sigdig = 0; /* "nan" or "inf" */
}
}
printf("F = %.*s. [0 decimal places, %lu digits, %lu digits significant]\n",
(int) i, df, (unsigned long int) i, (unsigned long int) sigdig);
free(df);
}
}
return;
}
static unsigned int
msb(uintmax_t v)
{
unsigned int mb = 0;
while (v >>= 1) { /* unroll for more speed... (see ilog2()) */
mb++;
}
return mb;
}
static unsigned int
ilog10(uintmax_t v)
{
unsigned int r;
static unsigned long long int const PowersOf10[] =
{ 1LLU, 10LLU, 100LLU, 1000LLU, 10000LLU, 100000LLU, 1000000LLU,
10000000LLU, 100000000LLU, 1000000000LLU, 10000000000LLU,
100000000000LLU, 1000000000000LLU, 10000000000000LLU,
100000000000000LLU, 1000000000000000LLU, 10000000000000000LLU,
100000000000000000LLU, 1000000000000000000LLU,
10000000000000000000LLU };
if (!v) {
return ~0U;
}
/*
* By the relationship "log10(v) = log2(v) / log2(10)", we need to
* multiply "log2(v)" by "1 / log2(10)", which is approximately
* 1233/4096, or (1233, followed by a right shift of 12).
*
* Finally, since the result is only an approximation that may be off
* by one, the exact value is found by subtracting "v < PowersOf10[r]"
* from the result.
*/
r = ((msb(v) * 1233) >> 12) + 1;
return r - (v < PowersOf10[r]);
}
Solution 8 - C
To my knowledge, there is a well diffused algorithm allowing to output to the necessary number of significant digits such that when scanning the string back in, the original floating point value is acquired in dtoa.c written by David Gay, which is available here on Netlib (see also the associated paper). This code is used e.g. in Python, MySQL, Scilab, and many others.