| 1 |
/* { dg-do run { target avx512fp16 } } */ |
| 2 |
/* { dg-options "-O2 -mavx512fp16 -mavx512dq" } */ |
| 3 |
|
| 4 |
#define AVX512FP16 |
| 5 |
#include "avx512fp16-helper.h" |
| 6 |
|
| 7 |
void NOINLINE |
| 8 |
emulate_mov2_load_sh(V512 * dest, V512 op1, |
| 9 |
__mmask8 k, int zero_mask) |
| 10 |
{ |
| 11 |
V512 v1, v2, v3, v4, v5, v6, v7, v8; |
| 12 |
int i; |
| 13 |
|
| 14 |
unpack_ph_2twops(op1, &v1, &v2); |
| 15 |
unpack_ph_2twops(*dest, &v7, &v8); |
| 16 |
|
| 17 |
if ((k&1) || !k) |
| 18 |
v5.f32[0] = v1.f32[0]; |
| 19 |
else if (zero_mask) |
| 20 |
v5.f32[0] = 0; |
| 21 |
else |
| 22 |
v5.f32[0] = v7.f32[0]; //remains unchanged |
| 23 |
|
| 24 |
for (i = 1; i < 8; i++) |
| 25 |
v5.f32[i] = 0; |
| 26 |
|
| 27 |
*dest = pack_twops_2ph(v5, v6); |
| 28 |
} |
| 29 |
|
| 30 |
void NOINLINE |
| 31 |
emulate_mov3_load_sh(V512 * dest, V512 op1, V512 op2, |
| 32 |
__mmask8 k, int zero_mask) |
| 33 |
{ |
| 34 |
V512 v1, v2, v3, v4, v5, v6, v7, v8; |
| 35 |
int i; |
| 36 |
|
| 37 |
unpack_ph_2twops(op1, &v1, &v2); |
| 38 |
unpack_ph_2twops(op2, &v3, &v4); |
| 39 |
unpack_ph_2twops(*dest, &v7, &v8); |
| 40 |
|
| 41 |
if ((k&1) || !k) |
| 42 |
v5.f32[0] = v3.f32[0]; |
| 43 |
else if (zero_mask) |
| 44 |
v5.f32[0] = 0; |
| 45 |
else |
| 46 |
v5.f32[0] = v7.f32[0]; //remains unchanged |
| 47 |
|
| 48 |
for (i = 1; i < 8; i++) |
| 49 |
v5.f32[i] = v1.f32[i]; |
| 50 |
|
| 51 |
*dest = pack_twops_2ph(v5, v6); |
| 52 |
} |
| 53 |
|
| 54 |
void NOINLINE |
| 55 |
emulate_mov2_store_sh(V512 * dest, V512 op1, __mmask8 k) |
| 56 |
{ |
| 57 |
V512 v1, v2, v3, v4, v5, v6, v7, v8; |
| 58 |
int i; |
| 59 |
|
| 60 |
unpack_ph_2twops(op1, &v1, &v2); |
| 61 |
unpack_ph_2twops(*dest, &v7, &v8); |
| 62 |
|
| 63 |
if ((k&1) || !k) |
| 64 |
v5.f32[0] = v1.f32[0]; |
| 65 |
else |
| 66 |
v5.f32[0] = v7.f32[0]; //remains unchanged |
| 67 |
|
| 68 |
*dest = pack_twops_2ph(v5, v6); |
| 69 |
} |
| 70 |
|
| 71 |
void |
| 72 |
test_512 (void) |
| 73 |
{ |
| 74 |
V512 res; |
| 75 |
V512 exp; |
| 76 |
|
| 77 |
init_src(); |
| 78 |
|
| 79 |
// no mask |
| 80 |
emulate_mov2_load_sh (&exp, src1, 0x0, 0); |
| 81 |
res.xmmh[0] = _mm_load_sh((const void *)&(src1.u16[0])); |
| 82 |
check_results(&res, &exp, 8, "_mm_load_sh"); |
| 83 |
|
| 84 |
// with mask and mask bit is set |
| 85 |
emulate_mov2_load_sh (&exp, src1, 0x1, 0); |
| 86 |
res.xmmh[0] = _mm_mask_load_sh(res.xmmh[0], 0x1, (const void *)&(src1.u16[0])); |
| 87 |
check_results(&res, &exp, 8, "_mm__mask_load_sh"); |
| 88 |
|
| 89 |
// with zero-mask |
| 90 |
emulate_mov2_load_sh (&exp, src1, 0x0, 1); |
| 91 |
res.xmmh[0] = _mm_maskz_load_sh(0x1, (const void *)&(src1.u16[0])); |
| 92 |
check_results(&res, &exp, 8, "_mm_maskz_load_sh"); |
| 93 |
|
| 94 |
emulate_mov3_load_sh (&exp, src1, src2, 0x1, 0); |
| 95 |
res.xmmh[0] = _mm_mask_move_sh(res.xmmh[0], 0x1, src1.xmmh[0], src2.xmmh[0]); |
| 96 |
check_results(&res, &exp, 8, "_mm_mask_move_sh"); |
| 97 |
|
| 98 |
emulate_mov3_load_sh (&exp, src1, src2, 0x1, 1); |
| 99 |
res.xmmh[0] = _mm_maskz_move_sh(0x1, src1.xmmh[0], src2.xmmh[0]); |
| 100 |
check_results(&res, &exp, 8, "_mm_maskz_move_sh"); |
| 101 |
|
| 102 |
// no mask |
| 103 |
emulate_mov2_store_sh (&exp, src1, 0x0); |
| 104 |
_mm_store_sh((void *)&(res.u16[0]), src1.xmmh[0]); |
| 105 |
check_results(&exp, &res, 1, "_mm_store_sh"); |
| 106 |
|
| 107 |
// with mask |
| 108 |
emulate_mov2_store_sh (&exp, src1, 0x1); |
| 109 |
_mm_mask_store_sh((void *)&(res.u16[0]), 0x1, src1.xmmh[0]); |
| 110 |
check_results(&exp, &res, 1, "_mm_mask_store_sh"); |
| 111 |
|
| 112 |
if (n_errs != 0) { |
| 113 |
abort (); |
| 114 |
} |
| 115 |
} |