| 1 |
/* { dg-do assemble } */ |
| 2 |
/* { dg-require-effective-target arm_v8_2a_bf16_neon_ok } */ |
| 3 |
/* { dg-options "-save-temps -O2" } */ |
| 4 |
/* { dg-add-options arm_v8_2a_bf16_neon } */ |
| 5 |
|
| 6 |
#include "arm_neon.h" |
| 7 |
|
| 8 |
/* BF16 DOT without lane. */ |
| 9 |
float32x2_t |
| 10 |
test_vbfdot_f32 (float32x2_t r, bfloat16x4_t a, bfloat16x4_t b) |
| 11 |
{ |
| 12 |
/* vdot.bf16 d, d, d */ |
| 13 |
return vbfdot_f32 (r, a, b); |
| 14 |
} |
| 15 |
|
| 16 |
float32x4_t |
| 17 |
test_vbfdotq_f32 (float32x4_t r, bfloat16x8_t a, bfloat16x8_t b) |
| 18 |
{ |
| 19 |
/* vdot.bf16 q, q, q */ |
| 20 |
return vbfdotq_f32 (r, a, b); |
| 21 |
} |
| 22 |
|
| 23 |
/* 64-bit BF16 DOT with lane. */ |
| 24 |
float32x2_t |
| 25 |
test_vbfdot_lane_f32_0 (float32x2_t r, bfloat16x4_t a, bfloat16x4_t b) |
| 26 |
{ |
| 27 |
/* vdot.bf16 d, d, d[0] */ |
| 28 |
return vbfdot_lane_f32 (r, a, b, 0); |
| 29 |
} |
| 30 |
|
| 31 |
float32x2_t |
| 32 |
test_vbfdot_lane_f32_1 (float32x2_t r, bfloat16x4_t a, bfloat16x4_t b) |
| 33 |
{ |
| 34 |
/* vdot.bf16 d, d, d[1] */ |
| 35 |
return vbfdot_lane_f32 (r, a, b, 1); |
| 36 |
} |
| 37 |
|
| 38 |
float32x2_t |
| 39 |
test_vbfdot_laneq_f32_0 (float32x2_t r, bfloat16x4_t a, bfloat16x8_t b) |
| 40 |
{ |
| 41 |
/* vdot.bf16 d, d, d[0] */ |
| 42 |
return vbfdot_laneq_f32 (r, a, b, 0); |
| 43 |
} |
| 44 |
|
| 45 |
float32x2_t |
| 46 |
test_vbfdot_laneq_f32_1 (float32x2_t r, bfloat16x4_t a, bfloat16x8_t b) |
| 47 |
{ |
| 48 |
/* vdot.bf16 d, d, d[1] */ |
| 49 |
return vbfdot_laneq_f32 (r, a, b, 1); |
| 50 |
} |
| 51 |
|
| 52 |
float32x2_t |
| 53 |
test_vbfdot_laneq_f32_2 (float32x2_t r, bfloat16x4_t a, bfloat16x8_t b) |
| 54 |
{ |
| 55 |
/* vdot.bf16 d, d, d[0] */ |
| 56 |
return vbfdot_laneq_f32 (r, a, b, 2); |
| 57 |
} |
| 58 |
|
| 59 |
float32x2_t |
| 60 |
test_vbfdot_laneq_f32_3 (float32x2_t r, bfloat16x4_t a, bfloat16x8_t b) |
| 61 |
{ |
| 62 |
/* vdot.bf16 d, d, d[1] */ |
| 63 |
return vbfdot_laneq_f32 (r, a, b, 3); |
| 64 |
} |
| 65 |
|
| 66 |
/* 128-bit BF16 DOT with lane. */ |
| 67 |
float32x4_t |
| 68 |
test_vbfdotq_lane_f32_0 (float32x4_t r, bfloat16x8_t a, bfloat16x4_t b) |
| 69 |
{ |
| 70 |
/* vdot.bf16 q, q, d[0] */ |
| 71 |
return vbfdotq_lane_f32 (r, a, b, 0); |
| 72 |
} |
| 73 |
|
| 74 |
float32x4_t |
| 75 |
test_vbfdotq_lane_f32_1 (float32x4_t r, bfloat16x8_t a, bfloat16x4_t b) |
| 76 |
{ |
| 77 |
/* vdot.bf16 q, q, d[1] */ |
| 78 |
return vbfdotq_lane_f32 (r, a, b, 1); |
| 79 |
} |
| 80 |
|
| 81 |
float32x4_t |
| 82 |
test_vbfdotq_laneq_f32_0 (float32x4_t r, bfloat16x8_t a, bfloat16x8_t b) |
| 83 |
{ |
| 84 |
/* vdot.bf16 q, q, d[0] */ |
| 85 |
return vbfdotq_laneq_f32 (r, a, b, 0); |
| 86 |
} |
| 87 |
|
| 88 |
float32x4_t |
| 89 |
test_vbfdotq_laneq_f32_3 (float32x4_t r, bfloat16x8_t a, bfloat16x8_t b) |
| 90 |
{ |
| 91 |
/* vdot.bf16 q, q, d[1] */ |
| 92 |
return vbfdotq_laneq_f32 (r, a, b, 3); |
| 93 |
} |
| 94 |
|
| 95 |
/* { dg-final { scan-assembler-times {\tvdot.bf16\td[0-9]+, d[0-9]+, d[0-9]+\n} 1 } } */ |
| 96 |
/* { dg-final { scan-assembler-times {\tvdot.bf16\tq[0-9]+, q[0-9]+, q[0-9]+\n} 1 } } */ |
| 97 |
/* { dg-final { scan-assembler-times {\tvdot.bf16\td[0-9]+, d[0-9]+, d[0-9]+\[0\]\n} 3 } } */ |
| 98 |
/* { dg-final { scan-assembler-times {\tvdot.bf16\td[0-9]+, d[0-9]+, d[0-9]+\[1\]\n} 3 } } */ |
| 99 |
/* { dg-final { scan-assembler-times {\tvdot.bf16\tq[0-9]+, q[0-9]+, d[0-9]+\[0\]\n} 2 } } */ |
| 100 |
/* { dg-final { scan-assembler-times {\tvdot.bf16\tq[0-9]+, q[0-9]+, d[0-9]+\[1\]\n} 2 } } */ |