| 1 |
/* { dg-do run } */ |
| 2 |
/* { dg-options "-O2 -mavx512vnni -mavx512vl" } */ |
| 3 |
/* { dg-require-effective-target avx512vnni } */ |
| 4 |
/* { dg-require-effective-target avx512vl } */ |
| 5 |
|
| 6 |
static void vnni_test (void); |
| 7 |
#define DO_TEST vnni_test |
| 8 |
#define AVX512VNNI |
| 9 |
#define AVX512VL |
| 10 |
#include "avx512f-check.h" |
| 11 |
#include "vnni-auto-vectorize-1.c" |
| 12 |
|
| 13 |
#define N 256 |
| 14 |
unsigned char a_u8[N]; |
| 15 |
char b_i8[N]; |
| 16 |
short a_i16[N], b_i16[N]; |
| 17 |
int i8_exp, i8_ref, i16_exp, i16_ref; |
| 18 |
|
| 19 |
int __attribute__((noinline, noclone, optimize("no-tree-vectorize"))) |
| 20 |
sdot_prod_hi_scalar (short * restrict a, short * restrict b, |
| 21 |
int c, int n) |
| 22 |
{ |
| 23 |
int i; |
| 24 |
for (i = 0; i < n; i++) |
| 25 |
{ |
| 26 |
c += ((int) a[i] * (int) b[i]); |
| 27 |
} |
| 28 |
return c; |
| 29 |
} |
| 30 |
|
| 31 |
int __attribute__((noinline, noclone, optimize("no-tree-vectorize"))) |
| 32 |
usdot_prod_qi_scalar (unsigned char * restrict a, char *restrict b, |
| 33 |
int c, int n) |
| 34 |
{ |
| 35 |
int i; |
| 36 |
for (i = 0; i < n; i++) |
| 37 |
{ |
| 38 |
c += ((int) a[i] * (int) b[i]); |
| 39 |
} |
| 40 |
return c; |
| 41 |
} |
| 42 |
|
| 43 |
void init() |
| 44 |
{ |
| 45 |
int i; |
| 46 |
|
| 47 |
i8_exp = i8_ref = 127; |
| 48 |
i16_exp = i16_ref = 65535; |
| 49 |
|
| 50 |
for (i = 0; i < N; i++) |
| 51 |
{ |
| 52 |
a_u8[i] = (i + 3) % 256; |
| 53 |
b_i8[i] = (i + 1) % 128; |
| 54 |
a_i16[i] = i * 2; |
| 55 |
b_i16[i] = -i + 2; |
| 56 |
} |
| 57 |
} |
| 58 |
|
| 59 |
static void vnni_test() |
| 60 |
{ |
| 61 |
init (); |
| 62 |
i16_exp = sdot_prod_hi (a_i16, b_i16, i16_exp, N); |
| 63 |
i16_ref = sdot_prod_hi_scalar (a_i16, b_i16, i16_ref, N); |
| 64 |
if (i16_exp != i16_ref) |
| 65 |
abort (); |
| 66 |
|
| 67 |
init (); |
| 68 |
i8_exp = usdot_prod_qi (a_u8, b_i8, i8_exp, N); |
| 69 |
i8_ref = usdot_prod_qi_scalar (a_u8, b_i8, i8_ref, N); |
| 70 |
if (i8_exp != i8_ref) |
| 71 |
abort (); |
| 72 |
} |