https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111335
Bug ID: 111335
Summary: fmaddpch seems not commutative for operands[1] and
operands[2] due to precision loss
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: crazylht at gmail dot com
Target Milestone: ---
fmaddcph is complex _Float16 fma.
cat test.c
#include <immintrin.h>
#include <cstdio>
void func(_Float16 a[], _Float16 b[], _Float16 c[])
{
const __m128h r0 = _mm_loadu_ph(a);
const __m128h r1 = _mm_loadu_ph(b);
const __m128h r2 = _mm_loadu_ph(c);
const __m128h mul = _mm_fmadd_pch(r0, r1, r2);
printf("%f %f\n", (float)mul[0], (float)mul[1]);
}
int main()
{
_Float16 a[8] = {-0.7949218f16, +0.2739257f16};
_Float16 b[8] = {+0.0010070f16, +0.0015659f16};
_Float16 c[8] = {-0.0010366f16, -0.0018014f16};
func(a, b, c);
return 0;
}
g++ -O0 -march=sapphirerapids test.c, we get fmaddpch a, b, c, and the result
is
-0.002266 -0.002769
g++ -O0 -march=sapphirerapids test.c, we get fmaddpch b, a, c, and the result
is
-0.002266 -0.002771
