| 1 |
typedef __bf16 __m128bf16 __attribute__((__vector_size__(16), __aligned__(16))); |
| 2 |
typedef __bf16 __m256bf16 __attribute__((__vector_size__(32), __aligned__(32))); |
| 3 |
typedef __bf16 __m512bf16 __attribute__((__vector_size__(64), __aligned__(64))); |
| 4 |
|
| 5 |
typedef union |
| 6 |
{ |
| 7 |
float f; |
| 8 |
unsigned int u; |
| 9 |
__bf16 b[2]; |
| 10 |
} unionf_b; |
| 11 |
|
| 12 |
static __bf16 make_f32_bf16 (float f) |
| 13 |
{ |
| 14 |
unionf_b tmp; |
| 15 |
tmp.f = f; |
| 16 |
return tmp.b[1]; |
| 17 |
} |
| 18 |
|
| 19 |
static float make_bf16_f32 (__bf16 bf) |
| 20 |
{ |
| 21 |
unionf_b tmp; |
| 22 |
tmp.u = 0; |
| 23 |
tmp.b[1] = bf; |
| 24 |
return tmp.f; |
| 25 |
} |
| 26 |
|
| 27 |
static int check_bf16 (__bf16 bf1, __bf16 bf2) |
| 28 |
{ |
| 29 |
unionf_b tmp1, tmp2; |
| 30 |
tmp1.u = 0; |
| 31 |
tmp2.u = 0; |
| 32 |
tmp1.b[1] = bf1; |
| 33 |
tmp2.b[1] = bf2; |
| 34 |
return (tmp1.u == tmp2.u); |
| 35 |
} |
| 36 |
|
| 37 |
static int check_bf16_float (__bf16 bf, float f) |
| 38 |
{ |
| 39 |
unionf_b tmp1, tmp2; |
| 40 |
tmp1.u = 0; |
| 41 |
tmp1.b[0] = bf; |
| 42 |
tmp2.f = f; |
| 43 |
tmp2.u >>= 16; |
| 44 |
return (tmp1.u == tmp2.u); |
| 45 |
} |