问题描述
IEEE 754-2008:
7.5 下溢
当检测到微小的非零结果时,应发出下溢异常信号。对于二进制格式,这应该是:
a) 舍入后 — 当计算出非零结果时,就好像指数范围是无界的一样 将严格介于 ±bemin 或
b) 舍入前 — 当计算出非零结果时,就好像指数范围和 无界精度将严格介于 ±bemin 之间。
实现者应选择检测微小的方式,但应以相同的方式检测基数为 2 的所有操作的微小,包括二进制舍入属性下的转换操作。
但是,C11 和 C17..C2x(工作草案 - 2020 年 2 月 5 日,n2479.pdf
)都没有提及微小:
$ pdfgrep.exe -i 'tininess' ISO-IEC-9899-2011.pdf n2479.pdf --color never
<nothing>
困惑。
问题:为什么没有 FLT_tininesS
宏 (-1 -- indeterminable,0 -- after rounding,1 -- before rounding
)?
更新。问题原因:像往常一样:一些 FP 测试(测试 FP 操作生成的结果的正确性)失败,因为预期引发的异常是 FE_INEXACT
,而实际引发的异常是 FE_INEXACT
和 {{1 }}。然后结果是 HW FE_UNDERFLOW
。因此,逻辑问题出现了:“如何确定是在四舍五入之前还是在四舍五入之后或不可确定的?”。由于在编译时无法确定,所以需要在运行时确定。
解决方法
以下程序可以确定是在四舍五入之前还是之后报告微小。
#include <fenv.h>
#include <float.h>
_Static_assert(FLT_RADIX == 2,"This program expects binary floating-point.");
typedef float Float;
enum { // Change FLT prefix according to type set for Float,above.
Precision = FLT_MANT_DIG,// Number of bits in significand.
MinimumExponent = FLT_MIN_EXP-1,// Minimum normal exponent.
/* The -1 is due to C's definition of floating-point exponents being
for significands in [1/2,1) instead of [1,2).
*/
};
// Use the following if your compiler supports it. Not all do.
//#pragma STDC FENV_ACCESS ON
// Report true iff a*b reports underflow.
static _Bool ProductUnderflows(Float a,Float b)
{
feclearexcept(FE_ALL_EXCEPT);
volatile Float c;
c = a*b;
return fetestexcept(FE_UNDERFLOW);
}
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
int main(void)
{
if (fesetround(FE_TONEAREST) != 0)
{
fprintf(stderr,"Error,cannot set rounding mode to nearest.\n");
exit(EXIT_FAILURE);
}
/* Find the least positive integer that does not divide the number of bits
in a significand (also called p or the precision of the type).
*/
int q = 1;
while (Precision % q == 0)
++q;
// Set a to a string of q bits after the radix point.
Float a = 1 - ldexp(1,-q);
/* Consider 1/a. This necessarily rounds down and sets b to a repeating
pattern of a 1 bit followed by q-1 0 bits.
To see that it rounds down,consider the binary representation of the
mathematical quotient 1/a. It is a repeating pattern of a 1 bit
followed by q-1 0 bits. So the 1 bits land at offsets from the first 1
bit of q,2q,3q,and so on. So they only land at multiples of q. And
we know p is not a multiple of q,so there is no 1 bit at the position
p bits beyond the leading bit. In other words,the first bit that is
does not fit in the p-bit significand is 0. So the residue being
discarded during rounding is less than 1/2 ULP,so round-to-nearest
rounds down.
We set b to 1/a scaled so that a*b is just below the normal range.
Then the mathematical product of a and b has a significand of
ceil(p/q)*q 1 bits,which is greater than p,so the product must be
rounded to fit in a signifcand. In round-to-nearest-ties-to-even mode,it will round upward,so the floating-point product of a and b will be
the smallest normal number. Therefore,there is an underflow if
tininess is detected before rounding but not if it is detected after
rounding.
*/
Float b = ldexp(1/a,MinimumExponent);
printf("a = %a.\n",a);
printf("b = %a.\n",b);
/* Test that we hit the boundary correctly: (a/2)*b underflows but
(2*a)*b does not. Also test that underflow reporting works.
*/
if (!ProductUnderflows(a/2,b))
{
fprintf(stderr,"Internal error,%a * %a -> %a is expected to underflow but did not.\n",a/2,b,(a/2)*b);
exit(EXIT_FAILURE);
}
if (ProductUnderflows(2*a,%a * %a -> %a is expected not to underflow but did.\n",2*a,(2*a)*b);
exit(EXIT_FAILURE);
}
// Test whether tininess is detected before or after rounding.
printf("Tininess is detected %s rounding.\n",ProductUnderflows(a,b) ? "before" : "after");
}