| 1 |
/* { dg-do run { target avx2 } } */ |
| 2 |
/* { dg-options "-O1 -mavx512bf16 -fdump-tree-vect-details -fdump-tree-optimized" } */ |
| 3 |
|
| 4 |
extern void abort (void); |
| 5 |
static void do_test (void); |
| 6 |
|
| 7 |
#define DO_TEST do_test |
| 8 |
#define AVX512BF16 |
| 9 |
#include "avx512-check.h" |
| 10 |
|
| 11 |
__bf16 b_128[8], r_abs_128[8], r_neg_128[8]; |
| 12 |
__bf16 b_256[16], r_abs_256[16], r_neg_256[16]; |
| 13 |
|
| 14 |
void |
| 15 |
__attribute__((optimize ("O2"), noinline, noipa, noclone, no_icf)) |
| 16 |
abs_128 (void) |
| 17 |
{ |
| 18 |
for (int i = 0; i < 8; i++) |
| 19 |
r_abs_128[i] = __builtin_fabsf16(b_128[i]); |
| 20 |
} |
| 21 |
|
| 22 |
void |
| 23 |
__attribute__((optimize ("O2"), noinline, noipa, noclone, no_icf)) |
| 24 |
neg_128 (void) |
| 25 |
{ |
| 26 |
for (int i = 0; i < 8; i++) |
| 27 |
r_neg_128[i] = -b_128[i]; |
| 28 |
} |
| 29 |
|
| 30 |
void |
| 31 |
__attribute__((optimize ("O2"), noinline, noipa, noclone, no_icf)) |
| 32 |
abs_256 (void) |
| 33 |
{ |
| 34 |
for (int i = 0; i < 16; i++) |
| 35 |
r_abs_256[i] = __builtin_fabsf16(b_256[i]); |
| 36 |
} |
| 37 |
|
| 38 |
void |
| 39 |
__attribute__((optimize ("O2"), noinline, noipa, noclone, no_icf)) |
| 40 |
neg_256 (void) |
| 41 |
{ |
| 42 |
for (int i = 0; i < 16; i++) |
| 43 |
r_neg_256[i] = -b_256[i]; |
| 44 |
} |
| 45 |
|
| 46 |
void |
| 47 |
check_absneg_results (__bf16 *b, __bf16 *r_abs, __bf16 *r_neg, int len) |
| 48 |
{ |
| 49 |
for (int i = 0; i < len; i++) |
| 50 |
{ |
| 51 |
__bf16 expected_abs = __builtin_fabsf16(b[i]); |
| 52 |
__bf16 expected_neg = -b[i]; |
| 53 |
if (r_abs[i] != expected_abs || r_neg[i] != expected_neg) |
| 54 |
abort (); |
| 55 |
} |
| 56 |
} |
| 57 |
|
| 58 |
static void |
| 59 |
__attribute__ ((noinline, noclone)) |
| 60 |
do_test (void) |
| 61 |
{ |
| 62 |
/* Initialize test values */ |
| 63 |
float float_b[16] = {-1.2f, 3.4f, -5.6f, 7.8f, |
| 64 |
-9.0f, 1.0f, -2.0f, 3.0f, |
| 65 |
-4.0f, -5.0f, 6.0f, 7.0f, |
| 66 |
-8.0f, -9.0f, 10.0f, 11.0f}; |
| 67 |
|
| 68 |
for (int i = 0; i < 8; i++) |
| 69 |
b_128[i] = (__bf16)float_b[i]; |
| 70 |
|
| 71 |
for (int i = 0; i < 16; i++) |
| 72 |
b_256[i] = (__bf16)float_b[i]; |
| 73 |
|
| 74 |
abs_128 (); |
| 75 |
neg_128 (); |
| 76 |
check_absneg_results (b_128, r_abs_128, r_neg_128, 8); |
| 77 |
|
| 78 |
abs_256 (); |
| 79 |
neg_256 (); |
| 80 |
check_absneg_results (b_256, r_abs_256, r_neg_256, 16); |
| 81 |
} |
| 82 |
|
| 83 |
/* { dg-final { scan-tree-dump-times "loop vectorized using 16 byte vectors" 2 "vect" } } */ |
| 84 |
/* { dg-final { scan-tree-dump-times "loop vectorized using 32 byte vectors" 2 "vect" } } */ |
| 85 |
/* { dg-final { scan-tree-dump-times {(?n)ABS_EXPR <vect} 2 "optimized" { target { ! ia32 } } } } */ |