| 1 |
/* { dg-do compile } */ |
| 2 |
/* { dg-options "-O1 -march=armv8-a+fp8fma" } */ |
| 3 |
|
| 4 |
#include <arm_neon.h> |
| 5 |
|
| 6 |
void foo(float16_t ap[20000], mfloat8x16_t b, mfloat8x16_t c, int br) |
| 7 |
{ |
| 8 |
float16x8_t a; |
| 9 |
a = vld1q_f16(ap); |
| 10 |
a = vmlalbq_f16_mf8_fpm(a, b, c, 13); |
| 11 |
vst1q_f16(ap, a); |
| 12 |
if (br) |
| 13 |
{ |
| 14 |
a = vld1q_f16(ap + 8); |
| 15 |
a = vmlalbq_f16_mf8_fpm(a, b, c, 13); |
| 16 |
vst1q_f16(ap + 8, a); |
| 17 |
a = vld1q_f16(ap + 16); |
| 18 |
a = vmlalbq_f16_mf8_fpm(a, b, c, 13); |
| 19 |
vst1q_f16(ap + 16, a); |
| 20 |
} |
| 21 |
else |
| 22 |
{ |
| 23 |
a = vld1q_f16(ap + 24); |
| 24 |
a = vmlalbq_f16_mf8_fpm(a, b, c, 13); |
| 25 |
vst1q_f16(ap + 24, a); |
| 26 |
} |
| 27 |
a = vld1q_f16(ap + 32); |
| 28 |
a = vmlalbq_f16_mf8_fpm(a, b, c, 13); |
| 29 |
vst1q_f16(ap + 32, a); |
| 30 |
} |
| 31 |
|
| 32 |
void bar(float16_t ap[20000], mfloat8x16_t b, mfloat8x16_t c, fpm_t mode, int br) |
| 33 |
{ |
| 34 |
float16x8_t a; |
| 35 |
a = vld1q_f16(ap); |
| 36 |
a = vmlalbq_f16_mf8_fpm(a, b, c, mode); |
| 37 |
vst1q_f16(ap, a); |
| 38 |
if (br) |
| 39 |
{ |
| 40 |
a = vld1q_f16(ap + 8); |
| 41 |
a = vmlalbq_f16_mf8_fpm(a, b, c, mode); |
| 42 |
vst1q_f16(ap + 8, a); |
| 43 |
a = vld1q_f16(ap + 16); |
| 44 |
a = vmlalbq_f16_mf8_fpm(a, b, c, mode); |
| 45 |
vst1q_f16(ap + 16, a); |
| 46 |
} |
| 47 |
else |
| 48 |
{ |
| 49 |
a = vld1q_f16(ap + 24); |
| 50 |
a = vmlalbq_f16_mf8_fpm(a, b, c, mode); |
| 51 |
vst1q_f16(ap + 24, a); |
| 52 |
} |
| 53 |
a = vld1q_f16(ap + 32); |
| 54 |
a = vmlalbq_f16_mf8_fpm(a, b, c, mode); |
| 55 |
vst1q_f16(ap + 32, a); |
| 56 |
} |
| 57 |
|
| 58 |
/* { dg-final { scan-assembler-times "msr\tfpmr" 2 } } */ |