Floating point subnormals are not handled correctly #1582

New issue

Open

opened 2023-12-17 17:44:55 -06:00 by dpgeorge · 3 comments

dpgeorge commented

2023-12-17 17:44:55 -06:00

(Migrated from github.com)

It seems that subnormal floats are not handled by the ROM float routines, and the pico-sdk wrappers don't account for this.

Code to reproduce on pico-sdk 1.5.1 (EDIT: fixed typo where it used to say one = 0):

#include <math.h>

static void pr(float f_in) {
    volatile float f = f_in;
    const uint8_t *b = (void *)&f;
    printf("%02x:%02x:%02x:%02x\n", b[0], b[1], b[2], b[3]);
}

int main(int argc, char **argv) {
    stdio_init_all();

    union {
        volatile float f;
        volatile uint32_t i;
    } x;
    volatile float zero = 0;
    volatile float one = 1;

    x.i = 29972;
    pr(x.f);
    pr(x.f + zero);
    pr(x.f * one);

    x.i = 2997220;
    pr(x.f);
    pr(x.f + zero);
    pr(x.f * one);

    return 0;
}

Output is:

14:75:00:00
00:00:00:00
00:00:00:00
e4:bb:2d:00
00:00:00:00
00:00:00:00

It should be:

14:75:00:00
14:75:00:00
14:75:00:00
e4:bb:2d:00
e4:bb:2d:00
e4:bb:2d:00

Above are just some simple examples, but most other operations on subnormals also return incorrect results (according to IEEE).

It seems that subnormal floats are not handled by the ROM float routines, and the pico-sdk wrappers don't account for this. Code to reproduce on pico-sdk 1.5.1 (EDIT: fixed typo where it used to say `one = 0`): ```c #include <math.h> static void pr(float f_in) { volatile float f = f_in; const uint8_t *b = (void *)&f; printf("%02x:%02x:%02x:%02x\n", b[0], b[1], b[2], b[3]); } int main(int argc, char **argv) { stdio_init_all(); union { volatile float f; volatile uint32_t i; } x; volatile float zero = 0; volatile float one = 1; x.i = 29972; pr(x.f); pr(x.f + zero); pr(x.f * one); x.i = 2997220; pr(x.f); pr(x.f + zero); pr(x.f * one); return 0; } ``` Output is: ``` 14:75:00:00 00:00:00:00 00:00:00:00 e4:bb:2d:00 00:00:00:00 00:00:00:00 ``` It should be: ``` 14:75:00:00 14:75:00:00 14:75:00:00 e4:bb:2d:00 e4:bb:2d:00 e4:bb:2d:00 ``` Above are just some simple examples, but most other operations on subnormals also return incorrect results (according to IEEE).

lurch commented

2023-12-17 17:58:30 -06:00

(Migrated from github.com)

volatile float one = 0;
Is that a bug, or just a typo? 😉

`volatile float one = 0;` Is that a bug, or just a typo? :wink:

dpgeorge commented

2023-12-17 18:02:05 -06:00

(Migrated from github.com)

Is that a bug, or just a typo?

Ooops! That's a typo. I just tested it again with the fixed typo and the output is as in the original post (ie the subnormal bug is there).

> Is that a bug, or just a typo? Ooops! That's a typo. I just tested it again with the fixed typo and the output is as in the original post (ie the subnormal bug is there).

dpgeorge commented

2023-12-17 18:04:49 -06:00

(Migrated from github.com)