3 ms·
Floats can be subnormal, and different bit patterns can represent the same value. It's not as easy as just comparing bits. Here is some cursed C code to play w
by marginalia_nu 4y ago
Floats can be subnormal, and different bit patterns can represent the same value. It's not as easy as just comparing bits.
Here is some cursed C code to play with:
#include <stdio.h>
#include <float.h>
#include <math.h>
int main() {
int a = 0x1000002;
int b = 0x2000001;
float* fa = (float*)&a;
float* fb = (float*)&b;
printf("value-equals? %d\n", *fa == *fb);
printf("fabs-equals? %d\n", fabs(*fa-*fb)<FLT_EPSILON);
}
May print something like
value-equals? 0
fabs-equals? 1
- jbay808 4y agoI'm not sure how subnormal numbers are relevant. For binary floats, the only duplicate values are (+0, -0) and the NaNs, and in both of those cases it's really up to you to decide if those are really representing the same value or not.
- marginalia_nu 4y agoI'm arguing against the notion that floats are just bit patterns in memory. If you base your operator== on such a notion, you will get very weird behavior.
- eesmith 4y agoCould you explain why that output is cursed? I get the same output using the following (I switched to the portable memcpy solution.) #include <stdio.h> #include <float.h> #include <math.h> int main() { int a = 0x361dc5df; // 2.351e-6f int b = 0x361e3e22; // 2.358e-6f; float fa, fb; memcpy(&fa, &a, 4); memcpy(&fb, &b, 4); printf("value-equals? %d\n", fa == fb); printf("fabs-equals? %d\n", fabs(fa-fb)<FLT_EPSILON); } Since FLT_EPSILON is 1.19209290E-07F , I would expect a lot of small, close numbers to give the same output, even without touching subnormal values.