Sidenote that isn't exaclty this issue, but related:

There appears to be a glibc bug in strtold on ppc64 which means that it is not possible to read in small float128 values from strings, eg 5e-296 which is not even a subnormal double. This means it is not possible to consistently round-trip float128 values through load/savetxt.

Here is an example running numpy 1.13:

>>> np.float128('5.0e-292')/100   # we can create small values with algebra
5e-294
>>> np.float128('5e-294')  # but can't create them from a string
ValueError: invalid literal for long double: 5e-294

Here's a little C program which does the same:

C program

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>

int main(){
    double d;
    long double x;
    char str[1024], *end;

    d = strtod("5.0e-294", &end); /* doubles are no problem */
    if(errno != 0) {
        printf("error: %s\n", strerror(errno));
    }
    printf("double:      %.40g\n", d);

    x = strtold("5.0e-292", &end);
    if(errno != 0) {
        printf("error: %s\n", strerror(errno));
    }
    printf("long double: %.40Lg\n", x);
    x = x/100;
    printf("/100:        %.40Lg\n", x);

    snprintf(str, 1024, "%.40Lg", x);
    x = 0;  /* reset to 0 to see if it gets overwritten */
    errno = 0;
    x = strtold(str, &end); /* long doubles fail */
    if(errno != 0) {
        printf("error: %s\n", strerror(errno));
    }
    printf("save/load:   %.40Lg\n", x);
}

output on ppc64 (be):

double:      5.0000000000000002566363886578393283723e-294
long double: 4.999999999999999999999999999999967711586e-292
/100:        5.000000000000000000000000000002042787298e-294
error: Numerical result out of range
save/load:   5.000000000000000000000000000002042787298e-294

So it seems to succesfully read in the final value, but mistakenly sets errno.

Uh oh!

Float128 repr is wrong on ppc64 #10429

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions