| 1 |
/* { dg-do compile } */ |
| 2 |
/* { dg-options "-O2 -mavx512fp16 -mavx512vl" } */ |
| 3 |
|
| 4 |
/* { dg-final { scan-assembler-times "vaddph" 2 { target { ! ia32 } } } } */ |
| 5 |
/* { dg-final { scan-assembler-times "vsubph" 2 { target { ! ia32 } } } } */ |
| 6 |
/* { dg-final { scan-assembler-times "vmulph" 2 { target { ! ia32 } } } } */ |
| 7 |
/* { dg-final { scan-assembler-times "vdivph" 2 { target { ! ia32 } } } } */ |
| 8 |
|
| 9 |
#define DO_PRAGMA(X) _Pragma(#X) |
| 10 |
|
| 11 |
#define VEC_OP_VV(size, op, name) \ |
| 12 |
void \ |
| 13 |
__attribute__ ((noinline, noclone, optimize("tree-slp-vectorize"))) \ |
| 14 |
vecop_v##size##hf##name (_Float16 * restrict dst, \ |
| 15 |
_Float16 * restrict src1, _Float16 * restrict src2) \ |
| 16 |
{ \ |
| 17 |
int i; \ |
| 18 |
DO_PRAGMA (GCC unroll size) \ |
| 19 |
for (i = 0; i < size; i++) \ |
| 20 |
dst[i] = src1[i] op src2[i]; \ |
| 21 |
} |
| 22 |
|
| 23 |
VEC_OP_VV(4, +, add) |
| 24 |
VEC_OP_VV(2, +, add) |
| 25 |
VEC_OP_VV(4, -, sub) |
| 26 |
VEC_OP_VV(2, -, sub) |
| 27 |
VEC_OP_VV(4, *, mul) |
| 28 |
VEC_OP_VV(2, *, mul) |
| 29 |
VEC_OP_VV(4, /, div) |
| 30 |
VEC_OP_VV(2, /, div) |