Wrong result from _ZGVsMxvv_powf

[kawashima-fj]

_ZGVsMxvv_powf returns incorrect values depending on the predicate.

Test program:

#include <stdio.h>
#include <mathlib.h>

int main() {
    svint32_t   i = svindex_s32(2, 1);                 // 2,    3,    4,    5
    svfloat32_t x = svcvt_f32_s32_x(svptrue_b32(), i); // 2.0,  3.0,  4.0,  5.0
    svfloat32_t y = svdup_n_f32(2.0f);                 // 2.0,  2.0,  2.0,  2.0

    svbool_t    p = svptrue_b32();                     // T,    T,    T,    T
    svfloat32_t z = _ZGVsMxvv_powf(x, y, p);           // 4.0,  9.0,  16.0,  25.0
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL1), z));
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL2), z));
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL3), z));
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL4), z));
    printf("----\n");

    p = svptrue_pat_b32(SV_VL2);                       // T,    T,    F,    F
    z = _ZGVsMxvv_powf(x, y, p);                       // 4.0,  9.0,  ---,  ---
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL1), z));
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL2), z));
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL3), z));
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL4), z));
    printf("----\n");

    p = svnot_z(svptrue_b32(), p);                     // F,    F,    T,    T
    z = _ZGVsMxvv_powf(x, y, p);                       // ---,  ---,  16.0,  25.0
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL1), z));
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL2), z));
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL3), z));
    printf("%f\n", svlastb(svptrue_pat_b32(SV_VL4), z));
    printf("----\n");

    return 0;
}

Compilation and execution:

$ gcc -march=armv8.2-a+sve -I$AOR/include -L$AOR/lib svepowf.c -lmathlib -lm
$ ./a.out

Expected output:

4.000000
9.000000
16.000000
25.000000
----
4.000000
9.000000
nan
nan
----
nan
nan
16.000000
25.000000
----

Actual output:

4.000000
9.000000
16.000000
25.000000
----
4.000000
1.000000
nan
nan
----
nan
nan
1.000000
25.000000
----

My environment:

• CPU: Grace (Neoverse V2)
• OS: Ubuntu 24.04 (AArch64)
• AOR: commit cbe9f6b (latest master branch)
• Compiler: GCC 13.3.0 and Clang/LLVM latest development branch

It seems libmvec in Glibc (at least version 2.41 in Ubuntu 25.04) also has this bug.

[blapie]

Thanks a lot for taking the time to report the issue. It seems I am able to reproduce using our ulp program, and any even predicate index, for instance p=2

$> ./build/bin/ulp -f -q -e 3.0 -p 2 _ZGVsMxvv_powf 3.1 x 2.1
_ZGVsMxvv_powf(0x1.8cccccp+1, 0x1.0cccccp+1) got 0x1p+0 want 0x1.585b8p+3 +0.331534 ulp err -1.02353e+07

Btw the bug is not limited to input integer values, which make me think of a flaw in our predicate logic. Will try to fix it ASAP and propagate into glibc.

[blapie]

Issue was closed automatically due to me pushing a fix. But let me know if that actually fixes the bug for you.
We will work towards patching glibc too.

[blapie]

Would you mind sharing in what context you use SVE math routines from AOR and/or GNU libc?

[kawashima-fj]

Thanks for the fix! I've confirmed your patch resolves my original problem.

I'm an LLVM developer. I tested the -fveclib=libmvec flag implemented recently in Clang using Fujitsu Compiler Test Suite and found one of the tests fails. I know AOR is the upstream of libmvec/AArch64 so I also tested using the latest AOR.